Development of an artificial intelligence breast cancer diagnostic tool

Universidade do Minho Escola de Engenharia Rute Salomé Pires de Castro Development of an Artificial intelligence Breast Cancer diagnostic tool outubro de 2023 UMinho | 2023 Rute Castro Development of an Artificial intelligence Breast Cancer diagnostic tool University of Minho School of Engineering Rute Salomé Pires de Castro Development of an Artificial intelligence Breast Cancer diagnostic tool Masters Dissertation Master’s in Bioinformatics Dissertation supervised by Óscar Manuel Lima Dias Débora Carina Gonçalves de Abreu Ferreira october 2023 Copyright and Terms of Use for Third Party Work This dissertation reports on academic work that can be used by third parties as long as the internationally accepted standards and good practices are respected concerning copyright and related rights. This work can thereafter be used under the terms established in the license below. Readers needing authorization conditions not provided for in the indicated licensing should contact the author through the RepositoriUM of the University of Minho. CC BY-NC-ND https://creativecommons.org/licenses/by-nc-nd/4.0/ i Acknowledgements Along the sometimes winding path of this investigation, two names stand out as beacons that guided and illuminated my path: Oscar Dias and Débora Ferreira. To both of them, my deepest gratitude. To the supervisor Débora Ferreira, for her wisdom and tireless dedication, for believing in me when I doubted myself, and for being more than an advisor, but a mentor in the truest sense of the word. To the supervisor Oscar Dias, whose insight and enthusiasm for the subject were contagious and inspiring. His ability to simultaneously challenge and encourage was fundamental to my academic and personal growth. Although I express my specific gratitude to these two pillars of my academic journey, I would also like to extend a sincere thank you to everyone who, directly or indirectly, contributed to this journey. Every conversation, every constructive criticism, every word of encouragement, played a role in shaping this work. I would like to thank my family for their unconditional love and for always believing in me, even in the most challenging moments. And to everyone who crossed my path during this period, who in some way touched my life and left their mark, my most sincere thanks. On each page of this thesis, in each line of research, there is a little bit of each of you. And for that, I am eternally grateful. ii Statement of Integrity I hereby declare having conducted this academic work with integrity. I confirm that I have not used plagiarism or any form of undue use of information or falsification of results along the process leading to its elaboration. I further declare that I have fully acknowledged the Code of Ethical Conduct of the University of Minho. University of Minho, Braga, october 2023 Rute Salomé Pires de Castro iii Abstract Breast cancer poses a global healthcare challenge due to its prevalence and multifactorial etiology. Despite advances in early detection and treatment, therapeutic approaches like standardized chemotherapy often fall short, primarily due to tumor heterogeneity and genetic variations. In light of this, personalized treatments are imperative. To address diagnostic and subtype classification challenges, this study utilized Metabric datasets and applied machine learning models after pre-processing, feature selection and hiperparameters otimization. The K-Nearest Neighbors (KNN) model achieved a 94% accuracy rate in cancer detection with a high recall rate, minimizing false negatives. For subtype classification, the Deep Neural Network (DNN) model excelled with a 95% accuracy and an average macro recall value of 0.96. The study in the realm of cancer prediction, only two genes were commonly observed. In the part of predict of the subtype revealed a significant overlap of 32 genes with the PAM50 list, affirming genetic consistency in breast cancer. In conclusion, the integration of KNN and DNN models significantly enhances both diagnostic and therapeutic strategies in breast cancer management. These models not only demonstrated high accuracy but also provided insights into gene-specific roles, potentially revolutionizing personalized, data-driven treatment pathways. Keywords breast cancer, machine learning models, metabric, prediction, cancer subtype, gene expression iv Resumo O cancro da mama representa um desafio global para a saúde devido à sua prevalência e etiologia multifatorial. Apesar dos avanços na detecção e tratamento precoces, abordagens terapêuticas como a quimioterapia padronizada muitas vezes ficam aquém, principalmente devido à heterogeneidade do tumor e às variações genéticas. Diante disso, tratamentos personalizados são imprescindíveis. Para enfrentar os desafios de diagnóstico e classificação de subtipos, este estudo utilizou conjuntos de dados Metabric e aplicou modelos de machine learning após pré-processamento, seleção de recursos e otimização de hiperparâmetros. O modelo K-Nearest Neighbor alcançou uma taxa de precisão de 94% na detecção de cancro com uma alta taxa de recall, minimizando falsos negativos. Para classificação de subtipos, o modelo Deep Neural Network se destacou com uma precisão de 95% e um valor médio de recuperação macro de 0,96. No estudo no domínio da previsão do cancro, apenas dois genes foram comumente observados. Na parte de previsão do subtipo revelou uma sobreposição significativa de 32 genes com a lista PAM50, afirmando consistência genética no cancro de mama. Em conclusão, a integração dos modelos KNN e DNN melhora significativamente as estratégias diagnósticas e terapêuticas no tratamento do cancro da mama. Esses modelos não apenas demonstraram alta precisão, mas também forneceram insights sobre funções específicas de genes, revolucionando potencialmente os caminhos de tratamento personalizados e baseados em dados. Palavras-chave cancro da mama, modelos de aprendizagem de máquina, metabric, previsão, subtipo de cancro, expressão genética v Contents I Introductory material 1 1 Introduction 2 1.1 ContextandMotivation ................................ 2 1.2 Objectives ...................................... 3 1.3 Thesisoverview.................................... 4 2 State of the Art 5 2.1 BreastCancer .................................... 5 2.2 Omicsdata...................................... 6 2.2.1 Omicsdatabases............................... 7 2.2.2 METABRIC.................................. 9 2.3 ArtificialIntelligence.................................. 13 2.3.1 Supervised machine learning models . . . . . . . . . . . . . . . . . . . . . 15 2.3.2 Unsupervised Machine Learning models . . . . . . . . . . . . . . . . . . . 23 2.3.3 DeepLearning................................ 26 2.3.4 Machine learning applied to studying Breast Cancer . . . . . . . . . . . . . . 28 3 Materials and Methods 30 3.1 Methodology for Cancer Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.1.1 Preprocessing ................................ 31 3.1.2 Featureselection............................... 31 3.1.3 Models optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.1.4 Modelsevaluation .............................. 37 3.2 Methodology for Cancer Subtype . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.2.1 Preprocessing ................................ 37 vi Part I Introductory material 1 Chapter 1 Introduction 1.1 Context and Motivation The total collection of DNA instructions present in a cell makes up the genome. The human genome is made up of 23 pairs of chromosomes that are found in the cell’s nucleus and one tiny chromosome that is found in the mitochondria. Everything a person needs to grow and operate is encoded in their DNA. Whether DNA contains millions or billions of letters, its transmission down the generations is the primary method for passing down organismal properties. One of the greatest scientific achievements in history is the Human Genome Project. A multinational team of scientists undertook the project as a biological journey to thoroughly examine the whole DNA (genome) of a particular specie group. The Human Genome Project, which began in October 1990 and ended in April 2003, is best known for its achievement of creating the very first genome sequence for humans. This achievement provided fundamental knowledge about the human blueprint and has since sped up research into human biology and improved medical care [1, 2]. Microarrays and Next Generation Sequencing (NGS) methods, including Whole-Genome Sequencing (WGS), Whole-Exome Sequencing (WES) and Targeted Sequencing are used to generate this sort of data [3]. Although the models cannot be considered completely accurate, they may help doctors and other health science professionals to make the best judgments when treating cancer patients [4]. Our bodies have millions of cells. Cancer starts when cellular changes cause these cells to grow excessively, creating a lump called a primary tumour [5]. Globally, 23% of cancer cases are Breast Cancer (BC), which also kills more women than any other disease [6]. According to the most current data, BC became the most commonly diagnosed cancer in the world in 2020, with 2.3 million new cases and 684,996 fatalities. In South America, Africa, and Asia, BC is more common. Additionally, patients with BC have poorer prognoses and higher rates of recurrence, compared to other cancers, and the current rate of patients living longer than five years is still low [7, 8]. 2 Given these findings, it is critical to learn more about how the subtypes emerge and disappear as well as how genes that govern unchecked cell proliferation are affected by mutations and other forms of functional loss. For instance, deletions and insertions to discover new targeted treatments. As a result, ML provides an approach to fill these gaps and identify differences in gene expression datasets. It is possible to use ML algorithms to understand genomics activity and find relevant genes as indicators for cancer classification and early diagnosis [9]. K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Decision Tree (DT) and Naive Bayes (NB) are some of the examples of Machine Learning (ML) algorithms that have already been tested for this type of cancer to create predictive breast cancer models with datasets containing gene expression data (as well as other types of data) [4]. This work is particularly important in the context of genomics, which is the study of the entire genome and allows for the identification of genetic variations that have a significant role in illness, therapy response, or patient prognosis.[10]. The main mission is not only to offer something intuitive, but above all to help in the projection of an accurate diagnosis for an individual. The motivation behind this approach is based on the perspective of providing health professionals, in particular physician and pathologists, with a tool (a model, the optimal one) that can act as a second opinion, strengthening confidence in the established diagnosis. This additional confirmation can be crucial, accelerating clinical decisions and, consequently, increasing the effectiveness and success of therapeutic interventions. 1.2 Objectives The main objective of this project is to develop a BC diagnostic framework. The framework will be based on training and validating ML models based on gene expression data to predict distinct biological activities relevant to our diagnosis. The work will address the following scientific/technical objectives: −Review the relevant literature on ML models and their applications in cancer classification tasks; −Search for relevant datasets; −Create different ML models for breast cancer detection and classification, evaluating their performance in different case studies; −Develop ML pipelines that encompasses data collection, data preprocessing, feature selection, model training, and validation the model. 3 1.3 Thesis overview Chapter 2 addresses the disease, BC with the aim of providing a biological perspective. It also explores omic data and the databases that store this type of data. The second part of this chapter aims to provide the basics of Artificial Intelligence (AI) and its applications. Also, the basics of ML and its applications and the most used supervised and non-supervised models. Chapter 3 serves to unveil the architecture of the pipelines developed for the project, elucidating each step and the intrinsic details and explain the dataset used. In chapter 4, the results are presented and discussed. Chapter 5 presents the main conclusions and future work. 4 Chapter 2 State of the Art 2.1 Breast Cancer About 10% of BC are hereditary and linked to a family history due to genetic susceptibility. The primary genetic modifications that result in BC include mutations such as minor insertions or deletions, amplifications, and rearrangements [11]. According to its histological and molecular properties, BC is now categorized into five subtypes [12]: −Triple negative: absence of expression of the Human Epidermal Growth Factor (HER2), Progesterone Receptor (PR), Estrogen Receptor (ER), high ki67 index (a measure of tumor cell proliferation and growth), and No Special Type (NST); −Enriched with HER2 (non-luminal): Targeted therapy such as trastuzumab, pertuzumab, lapatinib, neratinib, and trastuzumab emtansine (T-DM1) are effective against HER2 expression, high ki67 index, and NST [13]; −HER2+ luminal type B: expresses ER, PR, and HER2, has a high ki67 index, NST, but pleomorphic cells; responds to targeted therapy; −HER2luminal type B: expresses ER and PR but not HER2. High ki67 index, high-risk gene expression signature, NST, and micropapillary (carcinoma made up of tiny tumor cell clusters inside of gaps resembling dilated vascular canals); −Luminal type A: Comparatively high ER, PR, and HER2 expression to type B, also has a low ki67 index, low risk gene expression signature, NST, tubular cribriform, and classical lobular. It is the most prevalent subtype of BC. Regarding risk factors, it is known that alcohol use, diet, obesity, and physical inactivity are linked to a 5 higher possibility of developing this condition, as well as early menstrual beginning, a delayed menopause, and no breastfeed [14]. Mammography is the first particular diagnostic test that is being offered for BC screening [7]. After a nodule has been confirmed, an ultrasound and a needle biopsy are often conducted [15]. There are a number of treatment options available if a BC malignant tumor is confirmed, including surgery (mastectomy), which is done when the tumor is already big and is followed by adjuvant radiation to prevent recurrences. When there is a localized tumor, breast conserving surgery is done, then adjuvant radiation. Adjuvant chemotherapy and/or neoadjuvant chemotherapy are often used in conjunction with surgery to reduce tumor size. When the BC-hormone is positive, hormone treatment is used. This includes endocrine therapy, targeted therapy, and bone-modifying medications [12]. 2.2 Omics data There are various types of omics data, such as epigenomics (which studies epigenetic alterations), transcriptomics (studying the expression of genes through mRNA), proteomics (studying peptides and proteins), metabolomics (studying the substances and metabolites created by metabolism), microbiomics and genomics (studying the genome through DNA analysis) [3]. These omics enable the generation and analysis of incredibly large volumes of data. Omics data become important as patients may react differently to the same therapy. Precision medicine aims at selecting the patient’s best suitable treatment based on genotype and gene expression data [16]. The best effective treatment with the fewest side effects is what is often meant by the term ”precision medicine”. All fields of pathology can profit from the use of precision medicine, but oncology is the most advantageous. As there are several BC subtypes, numerous genes may be mutated in various subgroups [17]. Knowing the specific genes that have been altered in a tumour might be the difference between administering an ineffective treatment, with substantial side effects, or an effective therapy that has minimal side effects. Precision medicine’s ability to succeed relies on genomic data. This omics provides the information required to build databases that allow for the personalized treatment of each patient. The two primary omics studied in relation to cancer are genomics that is the study of the entire genome in this case with the purpose of identifying genetic variations that have a significant role in illness, therapy response or patient prognosis and transcriptomics [10]. The best course of treatment can be given by knowing a patient’s genotype and gene expression [18]. 6 Omics data supports the development of new drugs, and assist in selecting the appropriate medication, and open the door to the work of diagnosing and categorizing the stage and subtype of cancer. It is feasible to pinpoint the genes and proteins that cause diseases, by examining the genome and how genes are expressed. When opposed to evaluating random chemicals, the process of creating or identifying a medicine is much simpler by knowing the target protein. Understanding the causes of illnesses and how to cure them requires the study of mutations. Presently, a variety of methods are used to identify mutations. Some of the current approaches for finding mutations include DNA sequencing techniques WES and WGS, together with bioinformatics tools [19]. NGS offers a superior molecular and genomic examination of tumors compared to Sanger sequencing. DNA/RNA extraction, DNA/RNA quantification, library preparation and enrichment, NGS sequencing, base calling, alignment, quality control, variant calling, and annotation are some of the phases in the NGS workflow [20]. Its use is mostly applied on cancer research leading to a change in cancer diagnosis. Consequently, this technique helps in the decision of which genes that are relevant for a certain condition [21]. 2.2.1 Omics databases NGS offered less expensive methods to study the human genome. The quantity of data pertaining to the fields of genetics rose significantly with the decrease in sequencing costs. The quantity and quality of genomic data began to get more efficient funding from research institutions and businesses. For the analysis of individual variation and a wide range of disorders, especially cancer, the study of many genomes became crucial. Large volumes of data were and are being created as the cost of getting data decreased thanks to NGS. Because databases act as data repositories and enable data exchange with society, databases are crucial. It makes sense to study the data that already exists rather than invest time and money in gathering new data. Although new data is always beneficial, analyzing current information might lead to new findings. The community has access to a variety of omics datasets like Gene Expression Omnibus (GEO) (more generic) to cancer-specific databases like The Cancer Genome Atlas (TCGA) [22]. 7 Table 1: Available omics databases and the types of data they store Type of Database Name Characteristics Genomics Genomics of Drug Sensitivity in Cancer (GDSC) [23] Identifies biomarkers that have the potential to identify groups that react better to certain anticancer therapy Genomics GenBank [24] Data from DNA of genetic sequences Transcriptomics ArrayExpress [25] Data repository for functional genomics using microarray and sequencing databases Transcriptomics Gene Expression Omnibus (GEO) [26] Global repository for functional genomics data storage and distribution Proteomics Human Protein Atlas (HPA) [27] Mapping of all human proteins in cells, tissues and organs through various omics technologies Proteomics Uniprot [28] Contains protein sequence and annotation data from several European research institutes Compounds PubChem [29] For chemical molecules including information such as IDs and characteristics Compounds ChEMBL [30] Bioactive chemical compounds having druglike characteristics are hand-curated in this database Multiple Omics Cancer Dependency Map (DepMap) [31] Detection of genetic and pharmacological correlations in cancer, with the objective of finding genetic targets and predicting therapeutic susceptibility Multiple Omics The International Cancer Genome Consortium (ICGC) [32] Has data from the 3 main omics: genomics, epigenomics and transcriptomics related to cancer genomes Multiple Omics Catalogue Of Somatic Mutation In Cancer (COSMIC) [33] Manually curated data and GWS data on cancer somatic mutations Multiple Omics GDC Data Portal [34] Contains diverse omics cancer data from various projects 8 2.2.2 METABRIC METABRIC, or the Molecular Taxonomy of Breast Cancer International Consortium, is a groundbreaking project in breast cancer research. Born from a collaboration between UK and Canadian researchers, this initiative aimed to delve deeper into the molecular complexities of breast cancer than any study before it. The goal was to categorize breast cancer on a molecular level to develop more effective, targeted treatments. This dataset is a goldmine of information, containing both clinical and genomic data from nearly 2000 patients. But it’s not just about basic clinical facts like age at diagnosis or hormone status. METABRIC also includes in-depth molecular analysis, essentially serving as a crossroads for four key types of data: genetic expression in cancer patients, their clinical data, copy number alterations, and gene expression data from cancer-free individuals [35]. When it comes to the technical side, high-end genotyping platforms like Illumina microarrays were used to scrutinize the samples. The data gathering process was rigorous, involving complex sequencing, bioinformatics analysis, and validation steps. If you’re interested in using this dataset, be prepared for a stringent access process due to data sensitivity, including paperwork and justification for data use. This is the dataset that was used and access can be requested here: https://ega-archive.org/. Due to the contract signed to have access to it, it is not possible for me to give access to it [36, 37]. The resume of the characteristics of METABRIC datasets are in table 4. Though it might not have an ideal number of samples, METABRIC is unmatched in the depth and richness of its data. Despite its limitations, researchers value it for the unparalleled insights it offers. Table 2: Characteristics of METABRIC datasets Dataset Description Copy Number Alteration (CNA) dataset 22544 features 2174 patients Clinical dataset 24 clinical features 2509 patients Breast cancer genetic expression dataset 20603 features 1981 patients Non-cancerous Genetic Expression dataset 48803 features 144 individuals 9 2.2.2.1. Breast cancer Copy Number Alteration (CNA) dataset Copy Number Alterations Copy Number Alteration (CNA), containing precise information on 22,544 genes for 2,174 patients. While the dataset has some missing values, its significance is paramount. CNAs result in either an increase or a decrease in the number of copies of DNA segments and have a significant role in the development and progression of cancers, including breast cancer [35]. To generate this comprehensive CNA data, METABRIC employed high-resolution genotyping microarrays, specifically the Affymetrix SNP 6.0. This platform offers detailed coverage of the genome. The raw data gathered underwent meticulous bioinformatics processing, including algorithms like Circular Binary Segmentation, to identify areas of the genome where copy numbers were altered. This raw information then passed through additional preprocessing steps, including z-score normalization and quality probe filtering, to prepare it for statistical evaluation. The inclusion of the CNA dataset in cancer predictive models serves multiple purposes. First, it enhances the depth and complexity of the models by capturing molecular details that simpler datasets may overlook. Second, the CNA information allows the identification of genomic patterns related to specific tumor subtypes, thereby improving the accuracy and personalization of predictive models. Finally, the CNA dataset’s role in influencing therapeutic response gives the predictive models a heightened level of clinical relevance [35, 37]. 2.2.2.2. Breast cancer clinical dataset This dataset covers several variables related to the clinical status of patients identified in table 3. It contains 24 features about clinical data (including the one with the patient identifier) and the respective information for all 2509 patients, however, part of the information is missing [35]. This data set is crucial for advancing breast cancer treatment. It helps personalize treatments, make more accurate prognoses, and test new therapies in scenarios that mimic real life. Furthermore, by bringing together clinical data with genomic analyses, we gain a more complete understanding of the disease. This, in turn, allows doctors to adjust treatment plans based on actual results, making care more effective. It is an essential tool for both research and clinical care [38, 37]. 10 Shrinkage and Selection Operator) in LR. Filter methods are a group of feature selection techniques that compare the response of each feature to the response of the target variable. They rank the characteristics according to statistical criteria, such as the correlation between the features and the response or the significance of the features in model construction. These classifications are used to select a subset of relevant attributes for the model. The absence of attention to relationships between characteristics and the potential of picking unnecessary or duplicate features are disadvantages of these techniques. Wrapper approaches are an approach to feature selection that use ML algorithms to assess the efficacy of feature selection. These algorithms evaluate many feature combinations and select the optimal ones based on the specified ML model. They are computationally more costly than filter approaches, but in general, they are more accurate as they account for the interaction between features and how it impacts model performance [67, 68]. As previously mentioned, the last step results in the selection of the optimal model, based on the metrics tables produced and the confusion matrices. Each metrics table contains: −Precision: This metric indicates the proportion of positive hits (class predictions) that were actually correct. It is calculated as: True Positives (TP) True Positives (TP) +False Positives (FP) (2.1) A high precision indicates that the class was predicted well relative to incorrect predictions of the same class. −Recall (or Sensitivity): Shows the proportion of actual positives that were correctly identified. It is calculated as: True Positives (TP) True Positives (TP) +False Negatives (FN) (2.2) A high recall indicates that most of the actual class observations were correctly identified by the model. −F1-score: It is the harmonic average of Precision and Recall and provides a balance between these two metrics. It is particularly useful if classes are unbalanced. It is calculated as: 2× Precision ×Recall Precision +Recall (2.3) A high F1-score indicates a good balance between accuracy and recall. 17 −Support: Refers to the actual number of occurrences of the class in the specified dataset. It indicates how many observations belong to each category in the test set. −Accuracy: It is the total proportion of correct predictions in relation to all predictions made. Gives an overview of the effectiveness of the model. True Positives (TP) +True Negatives (TN) True Positives (TP) +False Positives (FP) +True Negatives (TN) +False Negatives (FN) (2.4) −Macro avg: Calculates the arithmetic mean for each metric (precision, recall, and f1-score) without considering class imbalance. −Weighted avg: Calculates the average of each metric taking into account the class imbalance. Uses the ”support” metric as a weight for each average. 2.3.1.2. Support vector machines An example of a supervised learning method used for classification and regression is the SVM. It operates by identifying a line or hyperplane that best delineates the various dataset classifications. The separating hyperplane that optimizes the margin between the various classes is what the SVM looks for. The margin is the separation between each class’s nearest points, or support vectors, and the hyperplane. The only points that have an impact on where the separating hyperplane is located are these support vectors, which are also the points that are most nearby. Binary classification, which divides data into two classes, and multi-class classification may both be accomplished using SVM (splitting data into more than two classes). Additionally, it may be used for the regression problem. Through the inclusion of a kernel function, which puts the data into a feature space of greater dimension where a linear separation may be found, SVM are also able to handle datasets that are not linearly separable. SVM are extensively utilized in a variety of industries, including bioinformatics, image recognition, and natural language processing. SVM are a sort of supervised learning algorithm that may be used to categorize problems, including the detection of BC [69]. Data from imaging tests, such as mammograms and ultrasounds, as well as information regarding tumour features, such as size, shape, and texture, are often utilized to train an SVM model for the diagnosis of BC. The algorithm is trained using this data to spot patterns that point to BC [70]. The SVM model may be used to categorize new instances of BC as malignant or benign after being trained. As a result, it may be used as a support tool to assist physicians in making treatment choices or in identifying individuals who are more likely to acquire BC. SVM has even been employed by researchers 18 to categorize BC based on gene expression data. To train the algorithm, they chose a number of genes that have previously been linked to BC [71, 72]. 2.3.1.3. k-Nearest Neighbors A supervised learning classification technique called KNN is based on the notion that a data point is categorised based on most of its nearest neighbours. The two primary steps for KNN to function are [73]: −Training: In this phase, a training dataset with each data point labelled with a distinct class is given to the algorithm to load into memory. No further training is done; −Classification: The method locates the k nearest data points (the neighbours) in the training set when a new data point is supplied. A smoother model that is less likely to be overfitted is produced by larger values of k, while a more complicated model that is more likely to be overfitted is produced by lower values of k. The majority of the classes of the k closest neighbours are used to give the class to the new data point. The new point will be categorized as class A if k=3 and two of its neighbours are class A and one neighbour is class B. Then, a test dataset is used to evaluate the KNN model, and accuracy is calculated. The algorithm may be modified to use alternative distance metrics to determine how close data points are to one another or to use weights to give the closest neighbours more weight. KNN is a straightforward algorithm that is simple to comprehend and use, but it has some drawbacks, such as the fact that it needs the storage of the full training dataset and the search for neighbours may be laborious for big data sets [74, 73]. Several studies have employed the KNN to categorize BC, and in some instances, the findings have been encouraging. For instance, in 2017, researchers utilized KNN to categorize BC based on genetic traits. The research revealed that the KNN algorithm has a 97.5% classification accuracy rate. The amount of k and the characteristics used, however, might have an impact on the algorithm’s performance, and it may not always perform as well as more sophisticated algorithms. It is also critical to take into account that the algorithm might be affected by outliers and data noise, and huge datasets can increase its computing cost [75, 76, 77]. 19 2.3.1.4. Linear and logistic regression For the LR a continuous target variable may be predicted using the supervised ML approach of LR using one or more input variables. Finding the best straight line between the data points is the core tenet of LR. The data and an optimization method are used to estimate the coefficients. Finding the coefficient values that reduce the discrepancy between the target variable’s anticipated values and actual values is the goal [78]. The model may be used to generate predictions on fresh data when the coefficients are computed. It is a common practice in several industries, including engineering, finance, and economics, to employ this simple yet effective strategy. Although it is an excellent place to start for regression tasks, LR makes the assumption that independent and dependent variables are always linear, which may not be true for all data sets [79]. Based on certain input characteristics, ML may utilize LR to forecast the probability of developing BC. Risk factors including age, family history, and genetics might be included in these input variables, sometimes referred to as predictors. Predicting BC recurrence is one use of LR in BC research. The association between certain genetic markers and the chance of BC recurrence was modelled by the scientists using LR. To model the link between the predictors and the goal, LR is utilized (recurrence or BC risk). Based on your input variables, the model may be used to forecast the risk of BC in new patients after it has been trained [80, 81]. For the Logistic Regression (LOR) the likelihood of a particular class may be predicted using the supervised ML technique of LOR using a collection of features. It is based on the logistic equation model, which represents the association between the dependent variable (Y) and one or more independent variables (X). Finding the optimal values for the coefficients (b0, b1, b2,...) that maximize the likelihood of correctly predicting the class is the goal. The logistic function also referred to as the sigmoid function, is used by the LOR model to simulate the likelihood of a binary outcome [82, 83]. Numerous applications of LOR include predicting patient recovery in medicine, predicting consumer purchases in marketing, predicting loan default in finance, predicting recidivism in criminology, and many more. It is extremely helpful for issues requiring binary classification, such as the diagnosis of BC, where the class may be either ”cancer” or ”non-cancer”. It is feasible to forecast a patient’s risk of developing cancer using a number of factors and to apply a predetermined cutoff rate to decide whether the patient is diagnosed with cancer or not. Additionally, the possibility that a tumour is benign or malignant may be predicted using algorithms, based on the tumour’s size, shape, density, and levels of gene expression [84, 85, 86]. 20 The fact that LOR is a linear model, which only accounts for linear correlations between input and output variables, should also be noted. Other models, such as DT or NN, may be better suitable for interactions that are more complicated. 2.3.1.5. Decision and Regression Trees Using regression and decision trees for both classification and regression issues, supervised ML algorithms in the form of trees are used. They operate by building a tree-like representation of choices and their outcomes [87]. In a DT, the branches stand in for the choices made in response to the input characteristics, while the leaves represent the anticipated class labels. The procedure begins at the tree’s base and recursively divides the data into classes based on a characteristic that optimizes class separation. The final predictions are performed by moving up the tree from the root to a leaf node [88]. The branches of a regression tree indicate the judgments based on the input characteristics, while the leaves represent the projected continuous value. Similar to a DT, the technique operates by attempting to reduce the variance of the target variable rather than maximising class separation [89, 90]. Both DT and Regression Tree (RT) can handle categorical and numerical data, and they are both simple to comprehend and analyze. But since DT are prone to overfitting, procedures like pruning or ensemble methods are often used to solve this issue. These algorithms are effective for processing huge and complicated datasets since they can handle both numerical and categorical data. In relation to BC, these algorithms have already been used to diagnose BC using mammographic images and also using a patient database [91, 92, 93]. 2.3.1.6. Artificial Neural Networks The structure of the human brain inspires ANN. They consist of layers of linked neurons, each of which performs basic computations and transmits the results to other neurons [94]. Adjusting the weights of connections between neurons based on training examples is the learning process of a neural network. Using optimization methods such as gradient descent, weights are modified to reduce the discrepancy between the expected output and actual data. ANN learns to recognize patterns in the training data and changes the weights to provide correct outputs during training. After training, the ANN may make predictions on fresh data based on the patterns it has learnt. The network receives input data, which the neurons process to generate an output. This procedure is performed several times to fine-tune the weights of the network [95, 96]. ANNs are used in a variety of applications, such as classification, pattern identification, natural lan21 guage processing, and prediction. ANNs may be used to categorize tumours as benign or malignant based on criteria like size, shape, and image intensity in the setting of BC. In addition, they may be used to forecast disease progression and recurrence risk. In the context of BC, ANNs are commonly used for the categorization and detection of lesions, as well as the identification of genetic markers that are crucial for cancer prediction and therapy. ANNs are composed of layers of linked neurons that are trained using vast amounts of data to fulfil certain tasks. In BC gene expression datasets, ANN may be trained to predict the presence or absence of cancer and to identify the disease’s most significant genes. In addition, ANN may be utilized to combine clinical and genetic data for more precise prognosis and therapy of BC [97, 98]. However, it is essential to note that ANN are tools whose efficacy relies on the quality and amount of data available for training. In addition, since ANN are very complicated models, it is essential to review your findings thoroughly to verify that they are accurately understood and not misapplied. 2.3.1.7. Ensembles A ML technique called ensemble learning mixes a number of distinct models to increase the reliability and accuracy of the output. To produce a forecast, many models are trained and combined rather than simply one. There are three primary categories: Random Forests (RF), Boosting and Bagging [99]. By building several independent training models from random samples and replacing the training data, the Bagging (Bootstrapped Aggregation) ensemble approach seeks to increase the stability and accuracy of ML models. By combining the predictions from the several models, it is expected that the final model would be less prone to overfitting and more resistant to data noise. Since each model is trained using a distinct sample of data, it may identify various patterns in the data. The final model may be more accurate than a model trained on all the data by incorporating these patterns. Typically, the various model forecasts are polled or averaged to get the final projection. Bagging is often employed with DT and other overfitting-prone ML models [100]. Boosting is a sort of ensemble ML technique that makes use of weighted combinations of base models to enhance the performance of a poor base model. The approach first fits a base model to the training data, after which it fits additional models to the residual errors that the base model had previously detected. A weighted combination of all base and extra models are created by weighing the additional models according to how well they are able to forecast base model mistakes. Boosting is used to fix base model flaws and increase its capability to generalize to new inputs [100]. RF is an ensemble learning strategy that mixes numerous decision trees to build a final model. It works by separately building multiple DT, and then integrating the findings to get a final forecast. Multiple 22 voting is used in this situation when the predicted class is shared by the majority of trees. The benefit of employing RF is that it is less prone to overfitting and can cope with irrelevant or correlated variables without hurting model accuracy [101]. Ensemble methods, which include classification, regression, and anomaly detection, are ML approaches that aggregate many distinct models to increase the accuracy and stability of findings. Ensemble techniques may be used to incorporate several elements, such as mammographic pictures, clinical data, and gene expression profiles, to increase diagnostic accuracy in the detection of BC [102, 103, 104]. 2.3.2 Unsupervised Machine Learning models Unsupervised learning is a type of ML in which the algorithm is trained on a dataset that has no labels or classifications. Without previous instruction, the data does not have a preset output, hence the goal is to detect patterns or structures in the data. Data compression, anomaly detection, and customer group identification are a few examples of uses [105]. There are several kinds of unsupervised learning algorithms, and each has a unique set of strategies and procedures for identifying patterns in data. Algorithms for unsupervised learning include, among others, dimensionality reduction, and clustering [106]. Since there are no labels to predict, an unsupervised learning model just has one essential component, the training phase. During the training phase, an unlabeled dataset is used to train the algorithm. The model searches for structures or patterns in the data during training. Techniques like DL, dimensionality reduction, and clustering may be used to accomplish this. Whether labels or training responses are present or absent is the primary distinction between supervised and unsupervised learning. When the model has been trained, it may be utilized to explain the input data. This may be achieved by categorizing the incoming data into several groups or by finding important data properties [107, 106]. In brief, when we have labelled data and know the anticipated outcome, we use supervised learning. However, when we do not have labelled data and do not know the expected outcome, we use unsupervised learning. Typically, the stages are as follows [108]: 1. Collecting data: Collecting an unlabeled data set is the initial phase in the process. This data set will be utilized to train the model; 2. Preprocessing: Prior to using the data to train the model, it is important to clean, normalize, and prepare it; 23 3. Model selection: Select the best model and set it up for the dataset and particular job at hand; 4. Model training: Using the dataset as a training set, the model is trained to seek structures or patterns in the data; 5. Description of Data: After training, the model may be used to characterize the input data, clustering it into subsets or highlighting distinctive features; 6. How to utilize the model: The model may be used for tasks like grouping data or dimensionality reduction, as well as identifying patterns or structures in the data. K-means K-means is an algorithm for grouping comparable data into ”k” distinct groups. It works by creating ”k” random starting centroids and then allocating each data point to the centroid that is closest to it. Then, the centroids are re-evaluated based on the average of the given data points, and this procedure is continued until the centroids no longer change or until a stopping condition is satisfied. The outcome is ”k” separate clusters of comparable data [109]. K-means is extensively used in data analysis, particularly in data mining and ML applications. It is a quick and effective method that may be used to identify patterns and trends in massive data sets. In addition, it is often used to execute data collations in several areas, including healthcare, marketing, and finance. K-means is one of ML’s most well-known and widely-used clustering techniques. This technique seeks to identify clusters or groupings of related data within a dataset. K-means can be used to group samples of tumours with comparable genetic traits in the setting of BC. This may be helpful for identifying subgroups of cancers with unique features, which may be significant for the creation of individualized treatment regimens. It is essential to note, however, that K-means is only a data exploration tool, and that the generated findings must be confirmed by additional methods in order to get a more trustworthy conclusion [110, 111]. T-distributed stochastic neighbour embedding T-distributed stochastic neighbour embedding (t-SNE) is a non-linear two-dimensional visualization technique used in ML. It is often used to study and analyze complicated data in a manner that is readily comprehended by people. It is particularly effective for displaying high-dimensional data, such as genetic data, in order to find patterns and correlations among samples. t-SNE relies on retaining closeness 24 between strongly correlated points in high-dimensional space while projecting these points onto a twodimensional representation. This enables t-SNE to show the underlying data structure in a form that is obvious and simple to comprehend [112]. t-SNE is commonly used in ML applications, including BC research. By using t-SNE to the genetic data of BC patients, researchers may find patterns and links between various kinds of tumour cells and their genetic traits, therefore facilitating the identification of novel treatment options and the customization of therapy. However, it is essential to emphasize that t-SNE visualization is a supplementary tool for data analysis and should not be utilized as the only diagnostic method [113, 114]. Principal Component Analysis PCA is a method for reducing the dimensionality of a given dataset. This implies that instead of dealing with each data feature or column, PCA finds and extracts the principal components that represent the majority of the data’s variability. These primary components are derived from the data’s correlation or covariance matrix. PCA has the benefit of reducing the dimensionality of data without sacrificing too much information. This is beneficial for a variety of applications, including data visualization, feature selection, enhancing the performance of ML models, and spotting outliers [115, 116]. PCA is often used with other ML methods to enhance performance. PCA may be used, for instance, to decrease the dimensionality of the data prior to performing a classification or clustering technique. This may be used to examine vast volumes of gene expression data pertaining to BC. These fundamental components may be used for data exploration and visualization, as well as feature selection and categorization. PCA is a crucial tool for identifying complicated patterns in gene expression data and identifying possible BC biomarkers [117, 118]. Hierarchical clustering Hierarchical Clustering is an ML data clustering approach that aims to identify hierarchical patterns within data. The concept is that, as the process progresses, the data is sorted into more dense clusters until all of the points are clustered into a single cluster or a limited number of different clusters. Agglomerative and divisive are the two primary techniques of Hierarchical Clustering. The Agglomerative method begins with each point being a distinct cluster and, over time, the closest clusters are combined into one. The Divisive method, on the other hand, begins with all points in a single cluster, which is then split into smaller clusters over time [119, 120]. Hierarchical Clustering has been used to identify subgroups of BC patients with comparable features 25 based on genomic data, including gene expression. This may be useful for gaining a better understanding of the distinctions between cancers and developing individualized treatment plans. To assure its validity, Hierarchical Clustering, like other clustering algorithms, needs careful parameter modifications and interpretation of findings [121, 122]. 2.3.3 Deep Learning DL is a subfield of ML that dealswith models like DNN. DL employs a sophisticated system of algorithms that makes it possible to handle unstructured data, including documents, photos, and text. DNN consist of numerous hidden layers of nodes, allowing models to learn complex, nonlinear inputoutput interactions. The objective is to discover hierarchical representations of the data, which will enable the models to execute sophisticated ML tasks such as classification, object identification, and natural language processing. These models are trained on vast volumes of data and can effectively identify and forecast from that data without explicit instructions. DL has been utilized extensively in areas such as computer vision, natural language processing, voice recognition, and others, and is increasingly being employed to handle complicated and challenging issues in domains such as medicine. As with ML, research into BC using DL is continually changing and new applications are being created. These procedures must be evaluated and integrated with other methods to guarantee accuracy and safety. Deep Neural Networks DNN are a family of deep ML models based on artificial NN with a large number of hidden layers. They are intended to learn hierarchical representations of complicated data and for the model to automatically learn from them [123]. DNNs are constructed of many layers of neurons, with each layer including multiple interconnected neurons. The input data goes through each layer’s weights and biases, and the layer’s output is transferred to the subsequent layer. During training, weights and biases are modified to reduce model error, enabling the model to acquire knowledge from the data [124, 125]. There are several applications for DNNs, including classification, image recognition, natural language processing, and sentiment analysis. They are particularly beneficial for issues involving complex data, such as pictures and signals, since they enable the model to catch crucial patterns and characteristics that are difficult for shallower models to grasp [126]. DNNs have been used in the context of BC for the categorization of tumor types, identification of tumors in mammography pictures, and survival prediction, among others. They are capable of learning intricate 26 Figure 1: T-distributed Stochastic Neighbor Embedding to visualize the distribution of data in a twodimensional space of class 0 (healthy individuals) and class 1 (individuals with cancer) Then, the feature selection method was applied to the training data. The method chosen was the ANOVA F test, implemented by the ”SelectKBest” class with ”f_classif”. This method evaluated the importance of each characteristic in relation to the objective variable and selected the best 1500 to keep. The training and testing datasets were then transformed to include only these selected features. These transformed versions were saved to files using the pickle format for future use. Finally, the F values and corresponding feature names were retrieved and sorted in descending order of importance. This information was saved in a text file for later analysis. 3.1.3 Models optimization The objective of this step was to identify the best hyperparameters for each model, thus optimizing its performance. This approach offers a systematic and efficient way to perform hyperparameter search, saving time and computational resources. In the presented implementation, a class called ”CustomRandomizedSearchCV” was developed, which 33 acts as an extension of the RandomizedSearchCV functionality provided by the Scikit-learn library. The purpose of this class is to optimize the hyperparameter search process for various MLalgorithms. The class was designed to allow simultaneous and comparative evaluation of different models and their hyperparameter combinations. The ”CustomRandomizedSearchCV” class was initialized with several parameters, including the training and validation data, the ML algorithm to be optimized, the distribution of hyperparameters, the number of iterations allowed, and others. An important point is that this class also allows the use of multiple CPU cores to speed up the training process, defined by the ”n_jobs” parameter. Additional methods were implemented to facilitate the process. The ”calc_total_parameters” method calculates the total number of possible hyperparameter combinations based on the given ranges. The ”random_search_cv” method performs the random search, fitting the model to the data and finding the best hyperparameters. Other methods provide metrics, such as the best set of hyperparameters (”get_best_parameters”), the best score achieved (”get_best_score”), and the average training and testing times (”get_average_train_time” and ”get_average_test_time”). However, choosing the various parameter_distribution for each model and then searching for the best hyperparameters for each model with the ”CustomRandomizedSearchCV” class was done manually, by reading the documentation. (Table: 4) For each ML algorithm including SVM,KNN, LR, DT, RF, AdaBoost, ANN and DNN, an instance of the ”CustomRandomizedSearchCV” class was created. These instances were then used to find the best hyperparameters for each algorithm, using the preprocessed training and validation data. 34 Table 4: Hyperparameters used in CustomRandomizedSearchCV to search for the optimal parameters Model Hyperparameters Support Vector Machines (SVM) Kernel: [linear, poly, rbf] C: [log10(0.1), log10(1000), 1000] Gamma: [log10(0.001), log10(10), 1000] Max_iter: [100, 200, 500, 1000] Num_iters allowed: [1000] K-Nearest Neighbors (KNN) N_neighbors: [1, 30, 30] Algorithm: [ball_tree, kd_tree, brute] Weights: [uniform, distance] Metric: [euclidean, manhattan, minkowski] Num_iters allowed: [1000] Linear Regression (LR) Penalty: [l1, l2, elasticnet, none] Solver: [newton-cg, lbfgs, liblinear, sag, saga] Max_iter: [100, 200, 500, 1000, 2000, 3500, 5000] Multi_class: [auto, ovr, multinomial] C: [0.5, 2, 30] Num_iters allowed: [1000] Decision Tree (DT) Criterion: [gini, entropy] Splitter: [best, random] Max_depth: [None, 2, 4, 8, 16] Min_samples_split: [2, 4, 8, 16, 32] Max_features: [auto, sqrt, log2] Num_iters allowed: [1000] Random Forests (RF) Criterion: [gini, entropy] Max_depth: [None, 2, 4, 8, 16] Min_samples_split: [2, 4, 8, 16, 32] Max_features: [auto, sqrt, log2] Num_iters allowed: [1000] 35 AdaBoost Algorithm: [SAMME, SAMME.R] N_estimators: [10, 50, 100, 200] Learning_rate: [log10(0.01), log10(1), 100] Num_iters allowed: [1000] Artificial Neural Networks (ANN) Hidden_layer_sizes: [128, 256, 512, 1024, 2048] Solver: [lbfgs, sgd, adam] Learning_rate: [constant, invscaling, adaptive] Learning_rate_init: [log10(0.001), log10(1), 1000] Max_iter: [200, 500, 2000] Batch_size: [auto, 64, 512] Num_iters allowed: [1000] Deep Neural Network (DNN) Hidden_layer_sizes: [(512, 256, 128, 128, 128, 128, 128, 128), (256, 128, 128, 128, 128, 128, 128, 128), (512, 256, 128, 128, 128, 128, 128), (512, 256, 128, 128, 128, 128), (256, 128, 128, 128, 128), (512, 128, 128, 128, 128), (512, 128, 128, 128), (256, 128, 128, 128), (256, 128, 128), (512, 128), (256, 128)] Solver: [lbfgs, sgd, adam] Learning_rate: [constant, invscaling, adaptive] Learning_rate_init: [log10(0.001), log10(1), 1000] Max_iter: [200, 500, 2000] Batch_size: [auto, 64, 512] Num_iters allowed: [1000] 36 3.1.4 Models evaluation In this final pipeline the objective was to train multiple ML models, make predictions and evaluate the performance of these models through metrics and confusion matrices. Models with scores below 90% in the previous pipeline were excluded from testing here. First, previously processed datasets were loaded into the previous pipeline in pickle format. This data included training and validation sets (”X_train_validation”, ”y_train_validation”) and test sets (”X_test”, ”y_test”). The best hyperparameters were added to the first model to be trained and predictions were made on the test set. Classification metrics were printed using the ”classification_report” function from the scikitlearn library. Additionally, a confusion matrix was generated and visualized. For each of these models, performance metrics were calculated and confusion matrices were displayed to evaluate the effectiveness of the models in correctly classifying the test data to conclude which model was best. 3.2 Methodology for Cancer Subtype 3.2.1 Preprocessing The first step was to load the patients’ clinical dataset using the Python Pandas library. The Scipy library was also imported to enable the use of statistical normalization. Three options were made for normalization: z-score, min-max and None. Next, the CNA and gene expression datasets were also loaded and their respective gene identifier columns (”Entrez_Gene_Id”) were excluded to standardize the dataset. To deal with any missing values in the datasets, a function was defined to remove them and also print the number of rows removed. This function was applied to all three datasets: clinical, CNA, and gene expression of individuals with cancer. After that, the normalization option “none” was applied to the CNA and gene expression datasets, because the gene expression one already came with z-score normalization and as for the CNA and clinical data did not make sense to normalize given the nature of the data. Then, the gene expression and CNA expression tables were transposed. The genes, which were initially in the rows, were moved to the columns and the patients, which were in the columns, were moved to the rows. Column names were then modified to clearly distinguish between gene expression and CNA data. Later, these transposed datasets were merged based on the indices, which were the patient IDs. The same was done to combine the clinical dataset with this merged mRNA and CNA dataset. Finally, the dimensions of all dataframes were printed for verification and the final dataset was saved to a csv file for use in subsequent pipelines. The option was also made to save the preprocessed dataset in pkl format. 37 3.2.2 Feature selection The purpose of this pipeline was to identify and reduce the data to the most relevant characteristics of the dataset. First, several Python libraries, including pandas, scikit-learn, and other visualization and analysis libraries, were imported. Next, the previously processed dataset was loaded and checked for missing values. Then the clinical characteristics were removed, except the one that was of interest for the prediction, defined at the beginning as “3-Gene classifier subtype”. The “3-Gene Classifier Subtype” column represents breast cancer subtypes based on expression of: ER, PR, and HER2. The dataset was then divided into input variables (x) and target variable (y). The dataset was also divided into training and test sets (80/20) using the train_test_split function from the scikit-learn library. After that, the t-distributed Stochastic Neighbor Embedding visualization technique was applied to visualize the distribution of data in a two-dimensional space (Figure 2). This technique is useful for understanding the structure and relationships between data. Then, the feature selection method was applied to the training data. The method chosen was the ANOVA F test, implemented by the ”SelectKBest” class with ”f_classif”. This method evaluated the importance of each characteristic in relation to the objective variable and selected the best 1400 to keep. This k was verified through the random forest feature importance that 1400 genes or components explained 99.88% of all information, which became the k value used (see the feature_importance.ipynb pipeline). The training and testing datasets were then transformed to include only these selected features. These transformed versions were saved to files using the pickle format for future use. Then, the f-values and corresponding feature names were saved and sorted in descending order of importance. This information was saved in a text file for later analysis. The 1400 features was save in the file ”sorted_f_value.txt”. Subsequently, another feature selection method PCA was tested using the same procedure and the transformed data was also saved in picke format. Finally, the pipeline was executed for the third time, as was the case with f-value and PCA, but here all features were used, the number of features was not reduced as in the previous two times. 38 Figure 2: T-distributed stochastic neighbour embedding to visualize the distribution of data in a twodimensional space of the four classes of “3-Gene Classifier Subtype” 3.2.3 Models optimization The methodology adopted here was exactly the same methodology as the models optimization for Cancer Diagnosis. The only variant here was that the pipeline here was run three times as it ran with data from each feature selection method adopted. The objective was precisely to see which method in general presented the best f1-score results, so that in the future, when the nature of the data was the same and with a view to optimizing time, this method would be used exclusively. 3.2.4 Models evaluation The methodology adopted here was exactly the same methodology as that used in the models evaluation for Cancer Diagnosis. After all the process, a small ”crossgeneslistas” pipeline was created to import the txt files with the list of genes after the feature selection and their respective values to cross with the list of pam50 genes ”pam50.txt”. With this, an ordered list of genes (from highest to lowest value) in common 39 in both lists was obtained. This was done twice, once with the feature selection list for predicting whether you have cancer and once with the feature selection list for predicting the cancer subtype. 40 Chapter 4 Results and Discussion 4.1 Determining the best feature selection method The first results were obtained when testing the three feature selection methods. The scores obtained by each model using the different feature selection methods are illustrated in Figure 3 and 4. In general, the F-value method consistently proves to be the most effective for the majority of models presented, while the PCA and Total methods vary in effectiveness depending on the specific model in question. This consistency suggests that the F-value method is a robust and reliable technique for feature selection in these datasets and contexts. For the SVM model, the best results are obtained when the F-value method is used, reaching a score of 0.934, while the PCA and Total methods reach 0.904 and 0.903, respectively. In the case of KNN, the F-value feature selection technique once again demonstrated superiority with a score of 0.870, followed by PCA with 0.639 and the Total method with 0.642. The LR model exhibits remarkable performance with the F-value method, scoring 0.923, while PCA and Total score 0.885 and 0.912, respectively. The last model in this figure, DT, achieves 0.819 with the F-value method, 0.486 PCA , followed by the , 0.744, and the Total. Moving on to Figure 4, four other models are analyzed: RF, adaboost, ANN and DNN. In the case of the RF model, the best score is achieved with the F-value, 0.919, followed by the PCA, 0.646, and the Total method, 0.918. The adaboost model, in turn, shows solid performance with the F-value method, 0.902, while the PCA and Total methods achieve scores of 0.766 and 0.905, respectively. For the ANN model, the F-value feature selection technique stands out with a score of 0.933, followed by PCA, 0.882, and Total with 0. Lastly, the DNN model shows an impressive result with the F-value method, achieving a score of 0.936, with PCA and Total scoring 0.864 and 0, respectively. 0 means that there was no computational capacity to run the model. 41 SVM KNN LR DT 0 0.2 0.4 0.6 0.8 1 0.934 0.87 0.923 0.819 0.904 0.639 0.885 0.486 0.903 0.642 0.912 0.744 Score F-value PCA Total Figure 3: Comparison between different Machine Learning (ML) models, using three different feature selection methods: F-value, Principal Component Analysis (PCA) and Total(all features) and considering four different models: Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Linear Regression (LR) and Decision Tree (DT). 42 Table 7: Score metrics (precision, recall, f1-score, support, accuracy, macro average, weighted average) of adaboost model Precision Recall F1-score Support Healthy individual (0) 0.92 0.71 0.80 49 Individual with cancer (1) 0.71 0.92 0.80 38 Accuracy 0.80 87 Macro avg 0.82 0.82 0.80 87 Weighted avg 0.83 0.80 0.80 87 The figure 6, the confusion matrix provided for the adaboost model, it is observed that the number of true positives for the Healthy individual (0) is 35, which means that 35 healthy individuals were correctly classified by the model, while 14 Healthy individuals ( 0) were incorrectly classified as Individual with cancer (1). On the other hand, for the Individual with cancer (1), the model was able to correctly identify 35 individuals, while only 3 were incorrectly classified as Healthy individual (0). This result indicates that the model has considerable sensitivity in detecting cancer cases, given the high number of true positives for the Individual with cancer class (1), but shows some limitations in its ability to correctly identify healthy individuals, given the number of false positives for the individual Healthy class (0). Analyzing the metrics table 7 of the model, it aims to predict whether an individual has cancer or not. The model displays an accuracy of 0.92 for the Healthy individual (0), suggesting that when it identifies an individual as healthy, there is a high probability of being correct. However, the recall of 0.71 for this same class indicates that the model correctly identifies only 71% of the truly healthy individuals present in the dataset. On the other hand, for the Individual with cancer (1), the precision is 0.71 and the recall is 0.92, which reveals that, despite some of the predictions being inaccurate, the model is effective in identifying the vast majority of individuals who actually have cancer. The f1-score of 0.80 for both classes suggests that the model maintains a balance between precision and recall. Regarding the averages, the macro average of 0.82 indicates that the model has an almost uniform performance for both classes, while the weighted average of 0.83 demonstrates that, when taking into account the number of samples in each class, the performance remains robust. The overall accuracy of the model is 0.80, indicating that it makes correct predictions in 80% of cases. 49 Figure 7: Confusion Matrix of Artificial Neural Networks (ANN) model. True label represents the real state of the individual and predicted label represents the state predicted by the model for this individual. Top left corner (purple): true negatives (TN). Top right corner (yellow): false positives (FP). Bottom left corner (purple): false negatives (FN). Bottom right (blue): true positives (TP). 50 Table 8: Score metrics (precision, recall, f1-score, support, accuracy, macro average, weighted average) of Artificial Neural Networks (ANN) model Precision Recall F1-score Support Healthy individual (0) 0.53 0.37 0.43 49 Individual with cancer (1) 0.42 0.58 0.48 38 Accuracy 0.46 87 Macro avg 0.47 0.47 0.46 87 Weighted avg 0.48 0.46 0.46 87 Analyzing the confusion matrix in figure 7 provided for the model in question, it is observed that the true Healthy individuals (0), 18 were correctly classified by the model, while 31 were erroneously classified as Individuals with cancer (1). This indicates a tendency for the model to misclassify healthy individuals, potentially leading to a high number of false positives. In contrast, of the true Individuals with cancer (1), 22 were correctly identified by the model, and 16 were misclassified as Healthy individuals (0). Although the model was able to identify a significant portion of individuals with cancer, it still presents difficulties, as it makes mistakes in a considerable portion of these classifications, which can result in false negatives. False negatives, in the context of cancer diagnosis, are particularly worrying as they can lead to delays in treatment. From the metrics table 8 provided for this model, it is clear that its predictive ability to determine whether or not an individual has cancer presents challenges. The accuracy for an individual Healthy (0) is 0.53, which indicates that when the model predicts that an individual is healthy, it is correct just over half the time. The recall of 0.37 for the same class suggests that the model correctly identified only 37% of the Healthy individuals (0) present in the test set. In relation to Individual with cancer (1), the precision is 0.42, while the recall is higher, standing at 0.58. This metric suggests that the model tends to misclassify many individuals as having cancer, but manages to capture 58% of those who actually have the disease. The f1score, which is a harmonic mean between precision and recall, is 0.43 and 0.48 for the Healthy individual (0) and the Individual with cancer (1), respectively. These values point to a moderate performance of the model. As for the general metrics, both the macro average and the weighted average of the metrics are aligned with the global accuracy, which is 0.46, revealing consistency in the model’s predictions, although it is operating below the 50% mark. 51 Figure 8: Confusion Matrix of Deep Neural Network (DNN) model. True label represents the real state of the individual and predicted label represents the state predicted by the model for this individual. Top left corner (yellow): true negatives (TN). Top right corner (purple): false positives (FP). Bottom left corner (purple): false negatives (FN). Bottom right (green): true positives (TP). 52 Table 9: Score metrics (precision, recall, f1-score, support, accuracy, macro average, weighted average) of Deep Neural Network (DNN) model Precision Recall F1-score Support Healthy individual (0) 0.73 0.71 0.72 49 Individual with cancer (1) 0.64 0.66 0.65 38 Accuracy 0.69 87 Macro avg 0.69 0.69 0.69 87 Weighted avg 0.69 0.69 0.69 87 This figure 9 or matrix provides a clear visualization of where the model got it right and where it got it wrong, allowing for a deeper understanding of areas where the model may need improvements or adjustments. The top left corner, represented in yellow, shows the number of true positives for the Healthy individual (0), which is 35, that is, 35 healthy individuals were correctly classified by the model. The upper right corner, in purple, indicates false positives, that is, 14 Healthy individuals (0) were incorrectly classified as Individuals with cancer (1). Analyzing the bottom left corner, also in purple, 13 Individuals with cancer (1) were erroneously classified as Healthy individuals (0), representing false negatives. Finally, the bottom right corner, in green, shows that 25 Individuals with cancer (1) were correctly identified, corresponding to the true positives. By looking at the metrics table 9 for the deep neural network (DNN) model, it is possible to discern several aspects about its performance. Regarding the individual Healthy (0), the model demonstrated an accuracy of 0.73, meaning that 73% of the predictions it made for that class were correct. The recall, or sensitivity, was 0.71, which reveals that the model was able to correctly identify 71% of all true Healthy individuals (0) present in the dataset. The F1-score, which is a harmonic metric between precision and recall, stood at 0.72. As for Individual with cancer (1), the accuracy was 0.64, indicating that 64% of the predictions in this class were correct. Recall was slightly higher, reaching 0.66, which suggests that the model correctly identified 66% of true Individuals with cancer (1). The F1-score for this class was 0.65, remaining consistent with previous metrics. Overall, the model had a weighted average precision of 0.69, and an equally weighted average recall of 0.69. The weighted F1-score followed these values. The avg macro metric also showed values of 0.69 for precision, recall and F1-score. The model’s overall accuracy was 0.69, that is, it made correct predictions 69% of the time. 53 The results in predicting the presence or absence of cancer showed that the KNN model achieved the best results, with an accuracy of 94%. Also, the KNN model demonstrated superior performance, particularly with regard to the recall metric. With a recall of 0.96 for healthy individuals and 0.92 for cancer individuals, KNN exhibited a unique ability to minimize false negatives. Such an ability is critical in a medical environment, where failing to detect a positive case can have devastating consequences. This performance can be attributed to the simplicity and effectiveness of KNN, which makes no specific assumptions about the distribution of the underlying data, making it less prone to overfitting and more capable of generalizing well to new data, especially when the set of data is small or imbalanced [148, 149, 77, 150]. AdaBoost, although it did not perform as effectively as KNN in terms of overall precision, demonstrated a high recall of 0.92 for the ”Individuals with Cancer” class, indicating a remarkable ability to identify positive cases. However, the inherent complexity and sensitivity to outliers can be disadvantageous in medical datasets, where abrupt variabilities are common [151]. The analysis also explored neural network models, where ANN proved to be inadequate for the dataset under study, with a precision of just 46% and low recall values. In contrast, DNN achieved a moderate precision of 69% and had more acceptable recall values [152, 153, 154]. However, even with these limitations, DNN exhibits a greater ability to represent complex relationships in data compared to ANN, which may explain its relatively better performance. The lack of a balanced dataset, with only 144 samples from healthy individuals, may have impacted the robustness of the model. Because of the notable imbalance in the data set (almost 1,900 samples from cancer patients versus just 144 individuals without cancer) a subsampling approach was chosen where both classes were limited to 144, this improved the results. The results were better than those not presented here where all samples were used. Because in the first attempt used all the samples from people with cancer (1980) and all from people without cancer (144) and the results obtained in the metrics tables were very poor, obtaining accuracy equal to zero on some models. The decision not to use the ”oversampling” technique was to preserve the integrity and reliability of the data in a medical context, especially since creating synthetic data in a medical context can introduce variables that deviate the model from biological reality. This problem with the number of samples may also have affected the list of 1500 genes of the feature selection aimed at predicting the presence or absence of cancer, the number of genes in common is notably small, with only 2 genes shared with PAM50 list. The first was ANLN , which is involved in cytokinesis, the final process of cell division [155]. It has been associated with cell proliferation in several types of cancer, 54 making it an interesting target for anticancer therapies. The second was NUF2 , which is an essential component of the centromere-associated complex, playing a fundamental role in the correct segregation of chromosomes during cell division. Its dysfunction can lead to aneuploidy, a common feature in many cancer cells [156]. The fact that all genes, including those in common, have very similar and low trait importance values may suggest that the prediction of this condition is not strongly influenced by a small set of genes, but possibly by a broader grouping of genetic and epigenetic factors [157]. The genes with the highest importance value in predicting the presence or absence of cancer are not on the pam50 list. The first was HIGD1B , which has been implicated in oxidative stress playing a role in cell survival [158]. Secondly, TNNT2 was a gene associated with the heart muscle, however, its relationship with breast cancer is unknown [159]. Third, RASL11A , which is involve in signal transduction processes, is associate with cell proliferation [160] . Next, OTUD6B , which is related to the regulation of deubiquitination, has an indirect relationship with protein stability in cancer cells [161]. Lastly, PMS2P3 is associated with DNA repair and maintenance of genomic integrity [162]. The results presented by ML models, namely in predicting the presence or absence of breast cancer, demonstrated that did not reach an optimal level of precision. This limitation can be, to a large extent, attributed to the biological complexity inherent to the disease. The absence of a specific set of genes that, in itself, is decisive in predicting breast cancer indicates that the genetics of this pathology are multifaceted and are not limited to a limited subset of genes. Biologically speaking, the heterogeneity of breast cancer and the interaction of multiple genetic and epigenetic factors may have contributed to the difficulty of models in making accurate predictions. The PAM50 list, although recognized, may not completely cover all genes relevant to oncogenesis and disease progression. The identification of genes such as HIGD1B , TNNT2 , among others, which are not present in the PAM50 list, highlights this idea. Additionally, imbalance in the dataset, with a significantly lower representation of healthy individuals compared to cancer patients, may have hampered the models’ ability to generalize appropriately. This disparity emphasizes the importance of data balancing in medical contexts, where each sample has critical relevance. In resume, ML models were not able to fully capture the complexity and biological variability of breast cancer, leading to results that, although promising in certain aspects, still have room for optimization. 55 4.3 Breast Cancer (BC) subtype prediction In the context of predicting cancer subtypes, diagnostic accuracy becomes even more critical, as each subtype may require a distinct treatment and specific management approach. Below is a consolidated analysis of the performance of different ML models in predicting cancer subtypes. This analysis is illustrated through the corresponding confusion matrices (Figures 9 to 14) and tables of results (Tables 11 to 16), after applying the best parameters identified in table 10 of the section on hyperparameter optimization. 56 Table 10: Optimal parameters for Machine Learning (ML) models obtained through the CustomRandomizedSearchCV class for prediction of cancer subtype ”three-gene classifier” Model Hyperparameters SVM Max1_iter: [1000] Kernel: [linear] Gamma: [1.7188391428171454] C: [5.220567527846975] LR Solver: [saga] Penalty: [l1] Multi_class: [multinomial] Max_iter: [5000] C: [0.8620689655172413] RF Min_samples_split: [16] Max_features: [sqrt] Max_depth: [16] Criterion: [gini] AdaBoost N_estimators: [200] Learning_rate: [1.0] Algorithm: [SAMME] ANN Solver: [sgd] Max_iter: [500] Learning_rate_init: [0.02596655972934871] Learning _rate: [adaptive] Hidden_layer_sizes: [1024] Batch_size: [auto] DNN Solver: [lbfgs] Max_iter: [500] Learning_rate_init: [0.5831305113526224] Learning _rate: [invscaling] Hidden_layer_sizes: [256, 128, 128] Batch_size: [64] 57 Figure 9: Confusion Matrix of Support Vector Machines (SVM) model. Vertical axis (True label): Represents the actual cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER-/HER2-”, and ”HER2+”. Horizontal axis (Predicted label): Represents the predicted cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER-/HER2-”, and ”HER2+”. Numbers in diagonal cells (yellow and blue) represent correct predictions, while numbers outside this diagonal represent incorrect predictions. 58 Table 13: Score metrics (precision, recall, f1-score, support, accuracy, macro average, weighted average) of Random Forests (RF) model Precision Recall F1-score Support ER+/HER2High Prolif 0.88 0.92 0.90 123 ER+/HER2Low Prolif 0.94 0.92 0.93 125 ER-/HER20.94 0.88 0.91 57 HER2+ 0.88 0.92 0.90 48 Accuracy 0.91 353 Macro avg 0.91 0.91 0.91 353 Weighted avg 0.91 0.91 0.91 353 As refered, the main diagonal of the matrix in figure 11 shows the model hits, that is, the instances in which the predictions correspond to the true labels. Observing this diagonal, it appears that the model was able to correctly classify 114 samples for ”ER+/HER2High Prolif”, 116 for ”ER+/HER2Low Prolif”, 51 for ”ER-/HER2-” and 47 for ”HER2+ ”. These values reinforce the high precision and recall observed in the previously discussed metrics. However, a confusion matrix also highlights where the model made incorrect predictions. For example, 8 truly ”ER+/HER2Low Prolif” samples were classified as ”ER+/HER2High Prolif” and 4 truly ”ER+/HER2High Prolif” samples were classified as ”ER+/HER2Low Prolif”. These confusions between ”High Prolif” and ”Low Prolif” may indicate that there are similar characteristics between these two classes that the model found difficult to distinguish. Additionally, it appears that the ”ER-/HER2-” subtype had 3 samples classified as ”ER+/HER2High Prolif” and 2 as ”ER+/HER2Low Prolif”, suggesting that there may be common characteristics or overlaps between these classes. Finally, the ”HER2+” subtype showed a mixture of classifications, with one sample classified as ”ER-/HER2-” and another as ”ER+/HER2High Prolif”. The RF model, used to predict breast cancer subtypes, shows remarkable performance, evidenced by the metrics provided in table 13. Analyzing the metrics individually by subtype, for ”ER+/HER2High Prolif”, the precision is 0.88, the recall is 0.92 and the f1-score is 0.90. These values indicate that the model has an excellent ability to correctly predict this subtype, being more cautious when assigning a sample to this group, resulting in a recall greater than precision. For the ”ER+/HER2Low Prolif” subtype, the precision, recall and f1-score values are 0.94, 0.92 and 0.93, respectively, denoting an excellent balance between the metrics. The ”ER-/HER2-” subtype has a precision of 0.94, recall of 0.88 and f165 score of 0.91. The ”HER2+” subtype records a precision of 0.88, a recall of 0.92 and an f1-score of 0.90. When evaluating general metrics, such as the macro average and the weighted average, all values center around 0.91, demonstrating the robustness and balance of the model in general terms. This balance is essential to ensure that the model is not over-optimized for a specific subtype, but rather performs consistently across the various subtypes. Accuracy remains in this balance with 0.91. 66 Figure 12: Confusion Matrix of adaboost model. Vertical axis (True label): Represents the actual cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER-/HER2-”, and ”HER2+”. Horizontal axis (Predicted label): Represents the predicted cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER-/HER2-”, and ”HER2+”. Numbers in diagonal cells (yellow and blue) represent correct predictions, while numbers outside this diagonal represent incorrect predictions. 67 Table 14: Score metrics (precision, recall, f1-score, support, accuracy, macro average, weighted average) of adaboost model Precision Recall F1-score Support ER+/HER2High Prolif 0.89 0.91 0.90 123 ER+/HER2Low Prolif 0.90 0.90 0.90 125 ER-/HER20.96 0.88 0.92 57 HER2+ 0.92 0.94 0.93 48 Accuracy 0.91 353 Macro avg 0.92 0.91 0.91 353 Weighted avg 0.91 0.91 0.91 353 The confusion matrix presented in figure 12 represents the model prediction results for the different cancer subtypes. In the ”ER+/HER2High Prolif” class, 116 samples were correctly predicted, while 6 were incorrectly classified as ”ER+/HER2Low Prolif” and 1 as ”HER2+”. Regarding the ”ER+/HER2Low Prolif” class, 116 samples were correctly identified, 7 were incorrectly categorized as ”ER+/HER2High Prolif”, 1 as ”ER-/HER2-” and 1 as ”HER2+” . In the ”ER-/HER2-” subtype, the model made the correct prediction for 52 samples, 2 were classified as ”ER+/HER2High Prolif”, 1 as ”ER+/HER2Low Prolif” and 2 as ”HER2+”. Finally, for the ”HER2+” subtype, the model correctly identified 45 samples and incorrectly classified 3 as ”ER+/HER2High Prolif”. These results reflect the model’s ability to correctly distinguish between different subtypes, as well as identify areas where improvements or adjustments may be needed. The Adaboost model’s metrics table 14 reveals remarkably good performance on several fronts. With a weighted average precision of 0.91 and a weighted average recall and f1-score also at 0.91, the model demonstrates substantial effectiveness in correctly classifying the different subtypes. In particular, the ”ER-/HER2-” class has the highest accuracy of 0.96, indicating that the model has an exceptional ability to identify true positives within this category. However, its recall for this same class is relatively lower (0.88), suggesting a tendency for a greater number of false negatives compared to other classes. The ”ER+/HER2High Prolif” and ”ER+/HER2Low Prolif” classes demonstrate precision, recall and f1-score values close to 0.90, which also indicates good performance of the model in these categories. On the other hand, the ”HER2+” class has a precision and f1-score above 0.90, and a recall of 0.94, thus becoming the class with the highest recall, suggesting a robust ability of the model to capture most true positive cases in this category. 68 Figure 13: Confusion Matrix of Artificial Neural Networks (ANN) model. Vertical axis (True label): Represents the actual cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER- /HER2-”, and ”HER2+”. Horizontal axis (Predicted label): Represents the predicted cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER-/HER2-”, and ”HER2+”. Numbers in diagonal cells (yellow and blue) represent correct predictions, while numbers outside this diagonal represent incorrect predictions. 69 Table 15: Score metrics (precision, recall, f1-score, support, accuracy, macro average, weighted average) of Artificial Neural Networks (ANN) model Precision Recall F1-score Support ER+/HER2High Prolif 0.94 0.92 0.93 123 ER+/HER2Low Prolif 0.94 0.96 0.95 125 ER-/HER20.95 0.96 0.96 57 HER2+ 0.96 0.94 0.95 48 Accuracy 0.94 353 Macro avg 0.95 0.95 0.95 353 Weighted avg 0.94 0.94 0.94 353 The presented confusion matrix in figure 13 provides a visual representation of the single-layer neural network model predictions for classifying breast cancer subtypes. The diagonal values of the matrix indicate the correct predictions of the model, with the ”ER+/HER2High Prolif” class having 114 correct predictions, the ”ER+/HER2Low Prolif” 121, the ”ER-/HER2-” 56 and to ”HER2+” 47. These numbers reflect the high accuracy of the model in correctly identifying cancer subtypes. However, there are also some glaring errors. The ”ER+/HER2High Prolif” class was confused 7 times with ”ER+/HER2Low Prolif” and 2 times with ”ER-/HER2-”. For ”ER+/HER2Low Prolif”, there were 2 cases in which it was confused with ”ER+/HER2High Prolif” and 1 case for each of the other two classes. ”ER-/HER2-” was misclassified once as ”ER+/HER2High Prolif”, and ”HER2+” was also confused once with ”ER+/HER2High Prolif”. Despite these confusions, it is clear that most of the model’s predictions are found on the diagonal of the matrix, indicating that most of the predictions are correct. Analysis of this model’s metrics table 15 reveals high effectiveness in several classification dimensions. The average weighted precision, average weighted recall, and average weighted f1-score are all at 0.94, demonstrating excellent agreement between the model predictions and the true classes. Specifically, the ”ER-/HER2-” class boasts a precision of 0.95 and a recall of 0.96, exhibiting the highest balance between identifying true positives and minimizing false negatives. The ”ER+/HER2Low Prolif” class follows closely with a precision and recall of 0.94 and 0.96, respectively, also representing a very good performance. In contrast, the ”HER2+” class presents a precision of 0.96 but with a slight decrease in recall to 0.94. Although this is a minimal difference, it suggests a slight susceptibility of the model to false negatives in this class. However, it is worth noting that all these metrics remain high, which implies a very robust 70 overall performance. Figure 14: Confusion Matrix of Deep Neural Network (DNN) model. Vertical axis (True label): Represents the actual cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER-/HER2-”, and ”HER2+”. Horizontal axis (Predicted label): Represents the predicted cancer subtypes for patients: ”ER+/HER2High Prolif”, ”ER+/HER2Low Prolif”, ”ER-/HER2-”, and ”HER2+”. Numbers in diagonal cells (yellow and blue) represent correct predictions, while numbers outside this diagonal represent incorrect predictions. 71 Table 16: Score metrics of Deep Neural Network (DNN) model Precision Recall F1-score Support ER+/HER2High Prolif 0.97 0.90 0.93 123 ER+/HER2Low Prolif 0.94 0.96 0.95 125 ER-/HER20.93 0.98 0.96 57 HER2+ 0.94 0.98 0.96 48 Accuracy 0.95 353 Macro avg 0.94 0.96 0.95 353 Weighted avg 0.95 0.95 0.95 353 When observing the confusion matrix provided in figure 14 , it is noted that for the ”ER+/HER2High Prolif” subtype, 109 cases were correctly classified while 9 cases were incorrectly classified as ”ER+/HER2Low Prolif”, 2 cases were incorrectly classified as ”ER-/HER2-”, and 3 cases were misassigned to the ”HER2+” subtype. For the ”ER+/HER2Low Prolif” subtype, 122 cases were correctly categorized, but 1 case was misidentified as ”ER+/HER2High Prolif”, and two other cases, one for ”ER-/HER2-” and another for ”HER2+”, they were also incorrectly classified. Regarding the ”ER-/HER2-” subtype, 55 samples were correctly classified and one case was incorrectly assigned to each of the other three subtypes. Finally, for the ”HER2+” subtype, 45 cases were correctly identified, 1 case was incorrectly classified as ”ER+/HER2High Prolif”, and 2 cases were incorrectly categorized as ”ER-/HER2-”. When evaluating the matrix as a whole, it becomes evident that the majority of samples were correctly classified, When analyzing the metrics table 16 of the multi-layer neural network model for predicting breast cancer subtypes, it is seen that the precision, recall and f1-score values are generally high for all classes. The ”ER+/HER2High Prolif” subtype has a precision of 0.97, although its recall is slightly lower, recording 0.90, which results in an f1-score of 0.93. In contrast, the ”ER+/HER2Low Prolif” subtype exhibits a precision of 0.94 and a recall of 0.96, providing an f1-score of 0.95, slightly higher than the previous one. The ”ER-/HER2-” and ”HER2+” classes boast remarkable performance, with recalls of 0.98 and f1-scores of 0.96 for both. The accuracy for these two classes is 0.93 and 0.94, respectively. The model’s overall precision, or accuracy, reaches an impressive value of 0.95 and the weighted average and macro are around the same value. Overall, the DNN model performed exceptionally high in classifying subtypes, with an accuracy of 95%. DNN stands out as the most appropriate choice, since it was the best at capturing the biological relationships between the features, especially when dealing with complex data such as genetic expressions 72 and clinical information. This model not only achieved high precision but also presented the best balance between the critical metrics of precision and recall, with an average macro recall value of 0.96 [163, 164, 154, 165]. Recall is particularly important in the medical context, as it is more risky not to identify a positive case than to incorrectly identify a negative case as positive. Thus, the model that best balances precision and recall, possibly weighting more heavily toward recall, should be considered the most appropriate. While ensemble methods like RF and AdaBoost did not fare as well, potential reasons include their intrinsic complexity and sensitivity to outliers. On the other hand, models like SVM and LR showcased high accuracy. However, their capabilities might not match the intricate complexities of medical data as efficiently as DNN. On the flip side, ANNs, which typically have a single layer, may not be as suited for such data. Despite their simplicity, ANNs have limitations in capturing complex non-linearities, compared to SVMs with non-linear kernels or DNNs, and may also be more sensitive to the presence of outliers [166]. However, SVM and LR exhibit advantages in specific scenarios due to their inherent simplicity, regularization features, and margin maximizing properties. The prowess of DNN lies in its ability to decipher complex, non-linear relationships in data, crucial for medical applications. It optimizes multiple parameters, making it highly adaptable. Furthermore, its architectural depth allows for the recognition of high-level data features, vital for understanding intricate inter-relationships between variables [163]. The model’s capacity to learn effective internal representations becomes invaluable for high-dimensional datasets like genetic expressions, simplifying data’s complexity and aiding in classification. With noise and measurement error tolerance, coupled with a vast array of regularization techniques, the DNN proves robust against data variations, ensuring efficient performance on both training and unseen data [154, 165]. Given the criticality of recall in medical contexts, and the superior performance of DNN across all metrics, it stands out as the top choice for biomedical applications, underscoring the immense potential of ML in the medical domain. In addition to these assessments, additional analyses were performed to find gene-to-gene correlations by cross-referencing the PAM50 gene list with two lists of 1500 and 1400 genes generated after the f-value feature selection (features that have been most implicated in cancer). The results of this analysis were saved in a two-text file. the resulting lists are named: list_cross1500 and list_cross1400. With the list of 1,500 characteristics crossed with the PAM50 list, only 2 genes were found in common, and have already been discussed in the previous section, despite the importance value of the 1,500 characteristics being very similar and all having low importance values. With the list of 1400 features belonging to the dataset 73 to predict the subtype, 32 genes in common were found, but here the range of feature importance values is from 158.2114 to 0.06139. Furthermore, the significant correlation of 32 genes in common with the genes in the PAM50 panel, which is a reference in the field, strengthens the biological validity of the models. Below the 5 in common with the highest feature importance value. Among these genes, KRT5 stands out, which codes for keratin 5, a protein involved in the structure and stability of epithelial cells. It is commonly used as a marker for basal-like breast cancer, a subtype that is typically more aggressive and has a worse prognosis [167]. Also ESR1 , which is an estrogen receptor and plays a crucial role in modulating gene transcription and regulating the cycle cell phone. Its presence is generally a good indication for hormonal therapy in ER+ cancers [168]. Furthermore, EGFR , which is an epidermal growth factor receptor, is involved in several signaling pathways that control cell proliferation. It is frequently overexpressed in more aggressive breast cancer subtypes, such as HER2+ and basal-like [169]. FOXC1 is associated with invasion and metastasis in breast cancer and is also considered a prognostic biomarker for the basal-like subtype [170]. Lastly, BCL2 regulates apoptosis, a process of programmed cell death. Its imbalance can contribute to the survival of cancer cells [171]. However, there are genes that were not in the PAM50 panel and exhibited a higher feature importance. the lists with the name of all the genes in the feature selection are in the repository with the name: feature1500_list.txt and feature1400_list.txt . Firstly, ROPN1B , that is still the subject of investigation, but initial studies suggest a possible involvement in relevant cellular processes [172]. Second, CELSR2 , which is associated with cell adhesion, may be a factor in the migration of cancer cells [173]. Also CSAD , which is involved in the metabolism of the amino acid serine, has been linked to the growth of cancer cells [174]. MPZL plays a role in cell signaling pathways and may be related to cancer aggressiveness [175]. Finally, MTFR2 plays a role in the formation of mitochondria, which can affect energy production in cancer cells [176]. Overall, it is possible to suggest that, while the models are highly effective in classifying cancer subtypes, corroborated by alignment with the PAM50 panel, they are less consistent in determining the presence or absence of the disease. This difference highlights the importance of balanced datasets and the need for further investigation of identified genes that are not in the panel as the genes were the same for both cases. Furthermore, by incorporating this analysis with existing biological knowledge about the genes involved, the research solidifies the emerging role of ML models in precision medicine, not only as diagnostic tools, but also as platforms for identifying biological pathways, relevant information and the development of targeted therapies. 74 Rosenfeld, Leigh Murphy, David R. Bentley, Ian O. Ellis, Arnie Purushotham, Sarah E. Pinder, AnneLise Børresen-Dale, Helena M. Earl, Paul D. Pharoah, Mark T. Ross, Samuel Aparicio, and Carlos Caldas. The somatic mutation profiles of 2,433 breast cancers refine their genomic and transcriptomic landscapes. Nature Communications , 7(1):11479, May 2016. Number: 1 Publisher: Nature Publishing Group. [36] Mallory Ann Freeberg, Lauren A Fromont, Teresa D’Altri, Anna Foix Romero, JorgeIzquierdo Ciges, Aina Jene, Giselle Kerry, Mauricio Moldes, Roberto Ariosa, Silvia Bahena, Daniel Barrowdale, MarcosCasado Barbero, Dietmar Fernandez-Orth, Carles Garcia-Linares, Emilio Garcia-Rios, Frédéric Haziza, Bela Juhasz, OscarMartinez Llobet, Gemma Milla, Anand Mohan, Manuel Rueda, Aravind Sankar, Dona Shaju, Ashutosh Shimpi, Babita Singh, Coline Thomas, Sabela delaTorre, Umuthan Uyan, Claudia Vasallo, Paul Flicek, Roderic Guigo, Arcadi Navarro, Helen Parkinson, Thomas Keane, and Jordi Rambla. The European Genome-phenome Archive in 2021. Nucleic Acids Research , 50(D1):D980–D987, January 2022. [37] Christina Curtis, Sohrab P. Shah, Suet-Feung Chin, Gulisa Turashvili, Oscar M. Rueda, Mark J. Dunning, Doug Speed, Andy G. Lynch, Shamith Samarajiwa, Yinyin Yuan, Stefan Gräf, Gavin Ha, Gholamreza Haffari, Ali Bashashati, Roslin Russell, Steven McKinney, Anita Langerød, Andrew Green, Elena Provenzano, Gordon Wishart, Sarah Pinder, Peter Watson, Florian Markowetz, Leigh Murphy, Ian Ellis, Arnie Purushotham, Anne-Lise Børresen-Dale, James D. Brenton, Simon Tavaré, Carlos Caldas, and Samuel Aparicio. The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups. Nature , 486(7403):346–352, June 2012. Number: 7403 Publisher: Nature Publishing Group. [38] Ala’a El-Nabawy, Nashwa El-Bendary, and Nahla A. Belal. A feature-fusion framework of clinical, genomics, and histopathological data for METABRIC breast cancer subtype classification. Applied Soft Computing , 91:106238, June 2020. [39] Patrick Henry Winston. Artificial intelligence (3rd ed.), 1992. Archive Location: world. [40] David Gunning, Mark Stefik, Jaesik Choi, Timothy Miller, Simone Stumpf, and Guang-Zhong Yang. XAI—Explainable artificial intelligence. Science Robotics , 4(37):eaay7120, December 2019. Publisher: American Association for the Advancement of Science. [41] Earl B. Hunt. Artificial Intelligence . Academic Press, May 2014. Google-Books-ID: 9y2jBQAAQBAJ. 81 [42] Pavel Hamet and Johanne Tremblay. Artificial intelligence in medicine. Metabolism , 69:S36–S40, April 2017. [43] Batta Mahesh. Machine Learning Algorithms -A Review . January 2019. [44] Issam El Naqa and Martin J. Murphy. What Is Machine Learning? In Issam El Naqa, Ruijiang Li, and Martin J. Murphy, editors, Machine Learning in Radiation Oncology: Theory and Applications , pages 3–11. Springer International Publishing, Cham, 2015. [45] Pramila P. Shinde and Seema Shah. A Review of Machine Learning and Deep Learning Applications. In 2018 Fourth International Conference on Computing Communication Control and Automation (ICCUBEA) , pages 1–6, August 2018. [46] Mohamed Alloghani, Dhiya Al-Jumeily, Jamila Mustafina, Abir Hussain, and Ahmed J. Aljaaf. A Systematic Review on Supervised and Unsupervised Machine Learning Algorithms for Data Science. In Michael W. Berry, Azlinah Mohamed, and Bee Wah Yap, editors, Supervised and Unsupervised Learning for Data Science , Unsupervised and Semi-Supervised Learning, pages 3–21. Springer International Publishing, Cham, 2020. [47] L. P. Kaelbling, M. L. Littman, and A. W. Moore. Reinforcement Learning: A Survey. Journal of Artificial Intelligence Research , 4:237–285, May 1996. [48] Li Yang and Abdallah Shami. On hyperparameter optimization of machine learning algorithms: Theory and practice. Neurocomputing , 415:295–316, November 2020. [49] Nannicini Samulowitz Diaz, Fokoue-Nkoutche. An effective algorithm for hyperparameter optimization of neural networks | IBM Journals & Magazine | IEEE Xplore, 2019. [50] Noemí DeCastro-García, Ángel Luis Muñoz Castañeda, David Escudero García, and Miguel V. Carriegos. Effect of the Sampling of a Dataset in the Hyperparameter Optimization Phase over the Efficiency of a Machine Learning Algorithm. Complexity , 2019:e6278908, February 2019. Publisher: Hindawi. [51] Sheena Angra and Sachin Ahuja. Machine learning and its applications: A review. In 2017 International Conference on Big Data Analytics and Computational Intelligence (ICBDAC) , pages 57–60, March 2017. 82 [52] Mohssen Mohammed, Muhammad Badruddin Khan, and Eihab Bashier Mohammed Bashier. Machine Learning Algorithms and Applications: Algorithms and Applications . CRC Press, Boca Raton, August 2016. [53] Amanpreet Singh, Narina Thakur, and Aakanksha Sharma. A review of supervised machine learning algorithms. In 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom) , pages 1310–1315, March 2016. [54] Vladimir Nasteski. An overview of the supervised machine learning methods. HORIZONS.B , 4:51– 62, December 2017. [55] Tammy Jiang, Jaimie L. Gradus, and Anthony J. Rosellini. Supervised Machine Learning: A Brief Primer. Behavior Therapy , 51(5):675–687, September 2020. [56] Chih-Fong Tsai and Ya-Ting Sung. Ensemble feature selection in high dimension, low sample size datasets: Parallel and serial combination approaches. Knowledge-Based Systems , 203:106097, September 2020. [57] Dejun Zhang, Lu Zou, Xionghui Zhou, and Fazhi He. Integrating Feature Selection and Feature Extraction Methods With Deep Learning to Predict Clinical Outcome of Breast Cancer. IEEE Access , 6:28936–28944, 2018. Conference Name: IEEE Access. [58] Huijuan Lu, Junying Chen, Ke Yan, Qun Jin, Yu Xue, and Zhigang Gao. A hybrid feature selection algorithm for gene expression data classification. Neurocomputing , 256:56–62, September 2017. [59] Computational Methods of Feature Selection - Google Livros. [60] Trevor Hastie, Jerome Friedman, and Robert Tibshirani. The Elements of Statistical Learning . Springer Series in Statistics. Springer, New York, NY, 2001. [61] An Introduction to Variable and Feature Selection. [62] A new feature selection method on classification of medical datasets: Kernel F-score feature selection - ScienceDirect. [63] Sandrine Dudoit, Jane Fridlyand, and Terence P Speed. Comparison of Discrimination Methods for the Classification of Tumors Using Gene Expression Data. Journal of the American Statistical Association , 97(457):77–87, March 2002. Publisher: Taylor & Francis _eprint: https://doi.org/10.1198/016214502753479248. 83 [64] Pradip Dhal and Chandrashekhar Azad. A comprehensive survey on feature selection in the various fields of machine learning. Applied Intelligence , 52(4):4543–4581, March 2022. [65] D Lavanya. ANALYSIS OF FEATURE SELECTION WITH CLASSFICATION: BREAST CANCER DATASETS. 2(5), 2011. [66] Shokoufeh Aalaei, Hadi Shahraki, Alireza Rowhanimanesh, and Saeid Eslami. Feature selection using genetic algorithm for breast cancer diagnosis: experiment on three different datasets. Iranian Journal of Basic Medical Sciences , 19(5):476–482, May 2016. [67] Suhang Wang, Jiliang Tang, and Huan Liu. Feature Selection. pages 1–9. January 2016. [68] Jason Brownlee. How to Choose a Feature Selection Method For Machine Learning. 2020. [69] Christopher J.C. Burges. A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery , 2(2):121–167, June 1998. [70] Rustam Nadira. Classification of cancer data using support vector machines with features selection method based on global artificial bee colony: AIP Conference Proceedings: Vol 2023, No 1, 2018. [71] Ying Liu. Active Learning with Support Vector Machine Applied to Gene Expression Data for Cancer Classification. Journal of Chemical Information and Computer Sciences , 44(6):1936–1941, November 2004. Publisher: American Chemical Society. [72] Mehmet Fatih Akay. Support vector machines combined with feature selection for breast cancer diagnosis. Expert Systems with Applications , 36(2, Part 2):3240–3247, March 2009. [73] Antonio Mucherino, Petraq J. Papajorgji, and Panos M. Pardalos. k-Nearest Neighbor Classification. In Antonio Mucherino, Petraq J. Papajorgji, and Panos M. Pardalos, editors, Data Mining in Agriculture , Springer Optimization and Its Applications, pages 83–106. Springer, New York, NY, 2009. [74] Zhongheng Zhang. Introduction to machine learning: k-nearest neighbors. Annals of Translational Medicine , 4(11):218, June 2016. [75] Md. Milon Islam, Hasib Iqbal, Md. Rezwanul Haque, and Md. Kamrul Hasan. Prediction of breast cancer using support vector machine and K-Nearest neighbors. In 2017 IEEE Region 10 Humanitarian Technology Conference (R10-HTC) , pages 226–229, December 2017. ISSN: 2572-7621. 84 [76] Qingbo Li, Wenjie Li, Jialin Zhang, and Zhi Xu. An improved k-nearest neighbour method to diagnose breast cancer. Analyst , 143(12):2807–2811, June 2018. Publisher: The Royal Society of Chemistry. [77] Shler Farhad Khorshid and Adnan Mohsin Abdulazeez. BREAST CANCER DIAGNOSIS BASED ON K-NEAREST NEIGHBORS: A REVIEW. PalArch’s Journal of Archaeology of Egypt / Egyptology , 18(4):1927–1951, February 2021. Number: 4. [78] Thomas M. H. Hope. Chapter 4 - Linear regression. In Andrea Mechelli and Sandra Vieira, editors, Machine Learning , pages 67–81. Academic Press, January 2020. [79] Dastan Maulud and Adnan M Abdulazeez. A review on linear regression comprehensive in machine learning. Journal of Applied Science and Technology Trends , 1(4):140–147, 2020. [80] S. Murugan, B. Muthu Kumar, and S. Amudha. Classification and Prediction of Breast Cancer using Linear Regression, Decision Tree and Random Forest. In 2017 International Conference on Current Trends in Computer, Electrical, Electronics and Communication (CTCEEC) , pages 763–766, September 2017. [81] Chandrasegar Thirumalai and Rashad Manzoor. Cost optimization using normal linear regression method for breast cancer Type I skin. In 2017 International conference of Electronics, Communication and Aerospace Technology (ICECA) , volume 2, pages 264–268, April 2017. [82] Mohammad Ali Mansournia, Angelika Geroldinger, Sander Greenland, and Georg Heinze. Separation in Logistic Regression: Causes, Consequences, and Control. American Journal of Epidemiology , 187(4):864–870, April 2018. [83] Connelly. Logistic Regression - ProQuest, 2020. [84] Anthony Anene Ude. Classification of Breast Cancer Using Logistic Regression . Thesis, June 2019. Accepted: 2019-08-08T13:40:58Z. [85] Ahmed F. Seddik and Doaa M. Shawky. Logistic regression model for breast cancer automatic diagnosis. In 2015 SAI Intelligent Systems Conference (IntelliSys) , pages 150–154, November 2015. [86] Hsiao-Lin Hwa, Wen-Hong Kuo, Li-Yun Chang, Ming-Yang Wang, Tao-Hsin Tung, King-Jen Chang, and Fon-Jou Hsieh. Prediction of breast cancer and lymph node metastatic status with tumour markers 85 using logistic regression models. Journal of Evaluation in Clinical Practice , 14(2):275–280, 2008. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1365-2753.2007.00849.x. [87] S. B. Kotsiantis. Decision trees: a recent overview. Artificial Intelligence Review , 39(4):261–283, April 2013. [88] Anthony J. Myles, Robert N. Feudale, Yang Liu, Nathaniel A. Woody, and Steven D. Brown. An introduction to decision tree modeling. Journal of Chemometrics , 18(6):275–285, 2004. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/cem.873. [89] Wei-Yin Loh. Classification and regression trees. WIREs Data Mining and Knowledge Discovery , 1(1):14–23, 2011. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/widm.8. [90] Wei-Yin Loh. Fifty Years of Classification and Regression Trees. International Statistical Review , 82(3):329–348, 2014. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/insr.12016. [91] Ronak Sumbaly, N. Vishnusri, and s Jeyalatha. Diagnosis of Breast Cancer using Decision Tree Data Mining Technique. International Journal of Computer Applications , 98:16–24, July 2014. [92] P. Sathiyanarayanan, S Pavithra., M SAI SARANYA., and M Makeswari. Identification of Breast Cancer Using The Decision Tree Algorithm. In 2019 IEEE International Conference on System, Computation, Automation and Networking (ICSCAN) , pages 1–6, March 2019. [93] Gajendra K. Vishwakarma, Pragya Kumari, and Atanu Bhattacharjee. Thresholding of prominent biomarkers of breast cancer on overall survival using classification and regression tree. Cancer Biomarkers , 34(2):319–328, January 2022. Publisher: IOS Press. [94] Mingzhe Chen, Ursula Challita, Walid Saad, Changchuan Yin, and Mérouane Debbah. Artificial Neural Networks-Based Machine Learning for Wireless Networks: A Tutorial. IEEE Communications Surveys & Tutorials , 21(4):3039–3071, 2019. Conference Name: IEEE Communications Surveys & Tutorials. [95] Jinming Zou, Yi Han, and Sung-Sau So. Overview of Artificial Neural Networks. In David J. Livingstone, editor, Artificial Neural Networks: Methods and Applications , Methods in Molecular Biology™, pages 14–22. Humana Press, Totowa, NJ, 2009. [96] Steven Walczak. Artificial Neural Networks, 2019. ISBN: 9781522573685 Pages: 40-53 Publisher: IGI Global. 86 [97] Krishna Mridha. Early Prediction of Breast Cancer by using Artificial Neural Network and Machine Learning Techniques. In 2021 10th IEEE International Conference on Communication Systems and Network Technologies (CSNT) , pages 582–587, June 2021. ISSN: 2329-7182. [98] Filippo Amato, Alberto López, Eladia María Peña-Méndez, Petr Vaňhara, Aleš Hampl, and Josef Havel. Artificial neural networks in medical diagnosis. Journal of Applied Biomedicine , 11(2):47– 58, January 2013. [99] Robi Polikar. Ensemble Learning. In Cha Zhang and Yunqian Ma, editors, Ensemble Machine Learning: Methods and Applications , pages 1–34. Springer US, Boston, MA, 2012. [100] Peter Bühlmann. Bagging, Boosting and Ensemble Methods. In James E. Gentle, Wolfgang Karl Härdle, and Yuichi Mori, editors, Handbook of Computational Statistics: Concepts and Methods , Springer Handbooks of Computational Statistics, pages 985–1022. Springer, Berlin, Heidelberg, 2012. [101] Aakash Parmar, Rakesh Katariya, and Vatsal Patel. A Review on Random Forest: An Ensemble Classifier. In Jude Hemanth, Xavier Fernando, Pavel Lafata, and Zubair Baig, editors, International Conference on Intelligent Data Communication Technologies and Internet of Things (ICICI) 2018 , Lecture Notes on Data Engineering and Communications Technologies, pages 758–763, Cham, 2019. Springer International Publishing. [102] Mohamed Hosni, Ibtissam Abnane, Ali Idri, Juan M. Carrillo de Gea, and José Luis Fernández Alemán. Reviewing ensemble classification methods in breast cancer. Computer Methods and Programs in Biomedicine , 177:89–112, August 2019. [103] Moloud Abdar, Mariam Zomorodi-Moghadam, Xujuan Zhou, Raj Gururajan, Xiaohui Tao, Prabal D Barua, and Rashmi Gururajan. A new nested ensemble technique for automated diagnosis of breast cancer. Pattern Recognition Letters , 132:123–131, April 2020. [104] Mohammad M. Ghiasi and Sohrab Zendehboudi. Application of decision tree-based ensemble learning in the classification of breast cancer. Computers in Biology and Medicine , 128:104089, January 2021. [105] Kamal Kant Hiran, Ritesh Kumar Jain, Dr Kamlesh Lakhwani, and Dr Ruchi Doshi. Machine Learning: Master Supervised and Unsupervised Learning Algorithms with Real Examples (English Edition) . BPB Publications, September 2021. Google-Books-ID: 4VVDEAAAQBAJ. 87 [106] Douglas H. Fisher, Michael J. Pazzani, and Pat Langley. Concept Formation: Knowledge and Experience in Unsupervised Learning . Morgan Kaufmann, May 2014. Google-Books-ID: ATKnCQAAQBAJ. [107] Zoubin Ghahramani. Unsupervised Learning. In Olivier Bousquet, Ulrike von Luxburg, and Gunnar Rätsch, editors, Advanced Lectures on Machine Learning: ML Summer Schools 2003, Canberra, Australia, February 2 - 14, 2003, Tübingen, Germany, August 4 - 16, 2003, Revised Lectures , Lecture Notes in Computer Science, pages 72–112. Springer, Berlin, Heidelberg, 2004. [108] Muhammad Usama, Junaid Qadir, Aunn Raza, Hunain Arif, Kok-lim Alvin Yau, Yehia Elkhatib, Amir Hussain, and Ala Al-Fuqaha. Unsupervised Machine Learning for Networking: Techniques, Applications and Research Challenges. IEEE Access , 7:65579–65615, 2019. Conference Name: IEEE Access. [109] Jason Xu and Kenneth Lange. Power k-Means Clustering. In Proceedings of the 36th International Conference on Machine Learning , pages 6921–6931. PMLR, May 2019. ISSN: 2640-3498. [110] Melissa Zhao, Yushi Tang, Hyunkyung Kim, and Kohei Hasegawa. Machine Learning With K-Means Dimensional Reduction for Predicting Survival Outcomes in Patients With Breast Cancer. Cancer Informatics , 17:1176935118810215, January 2018. Publisher: SAGE Publications Ltd STM. [111] Ade Jamal, Annisa Handayani, Ali Septiandri, Endang Ripmiatin, and Yunus Effendi. Dimensionality Reduction using PCA and K-Means Clustering for Breast Cancer Prediction. Lontar Komputer : Jurnal Ilmiah Teknologi Informasi , page 192, December 2018. [112] Wenbo Zhu, Zachary T. Webb, Kaitian Mao, and José Romagnoli. A Deep Learning Approach for Process Data Visualization Using t-Distributed Stochastic Neighbor Embedding. Industrial & Engineering Chemistry Research , 58(22):9564–9575, June 2019. Publisher: American Chemical Society. [113] Nonita Sharma, Monika Mangla, Sachi Nandan Mohanty, and Suneeta Satpaty. A Stochastic Neighbor Embedding Approach for Cancer Prediction. In 2021 International Conference on Emerging Smart Computing and Informatics (ESCI) , pages 599–603, March 2021. [114] Nonita Sharma, K. P. Sharma, Monika Mangla, and Rajneesh Rani. Breast cancer classification using snapshot ensemble deep learning model and t-distributed stochastic neighbor embedding. Multimedia Tools and Applications , 82(3):4011–4029, January 2023. 88 [115] Ahmed Lasisi and Nii Attoh-Okine. Principal components analysis and track quality index: A machine learning approach. Transportation Research Part C: Emerging Technologies , 91:230–248, June 2018. [116] Tom Howley, Michael G. Madden, Marie-Louise O’Connell, and Alan G. Ryder. The Effect of Principal Component Analysis on Machine Learning Accuracy with High Dimensional Spectral Data. In Ann Macintosh, Richard Ellis, and Tony Allen, editors, Applications and Innovations in Intelligent Systems XIII , pages 209–222, London, 2006. Springer. [117] Hasmarina Hasan and Nooritawati Md Tahir. Feature selection of breast cancer based on Principal Component Analysis. In 2010 6th International Colloquium on Signal Processing & its Applications , pages 1–4, May 2010. [118] Sarthak Sanjay Tilwankar and Bhupendra Singh Kirar. Breast Cancer Detection using Principal Component Analysis and Machine Learning Models. In 2021 First International Conference on Advances in Computing and Future Communication Technologies (ICACFCT) , pages 80–84, December 2021. [119] Moses Charikar, Vaggos Chatziafratis, Rad Niazadeh, and Grigory Yaroslavtsev. Hierarchical Clustering for Euclidean Data. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , pages 2721–2730. PMLR, April 2019. ISSN: 2640-3498. [120] Juan Nunez-Iglesias, Ryan Kennedy, Toufiq Parag, Jianbo Shi, and Dmitri B. Chklovskii. Machine Learning of Hierarchical Clustering to Segment 2D and 3D Images. PLOS ONE , 8(8):e71715, August 2013. Publisher: Public Library of Science. [121] Arjun P. Athreya, Alan J. Gaglio, Junmei Cairns, Krishna R. Kalari, Richard M. Weinshilboum, Liewei Wang, Zbigniew T. Kalbarczyk, and Ravishankar K. Iyer. Machine Learning Helps Identify New Drug Mechanisms in Triple-Negative Breast Cancer. IEEE Transactions on NanoBioscience , 17(3):251– 259, July 2018. Conference Name: IEEE Transactions on NanoBioscience. [122] Alexander J. Titus, Carly A. Bobak, and Brock C. Christensen. A New Dimension of Breast Cancer Epigenetics - Applications of Variational Autoencoders with DNA Methylation:. In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies , pages 140–145, Funchal, Madeira, Portugal, 2018. SCITEPRESS - Science and Technology Publications. [123] Huang Yi, Sun Shiyu, Duan Xiusheng, and Chen Zhigang. A study on Deep Neural Networks frame89 work. In 2016 IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC) , pages 1519–1522, October 2016. [124] Shizhao Sun, Wei Chen, Liwei Wang, Xiaoguang Liu, and Tie-Yan Liu. On the Depth of Deep Neural Networks: A Theoretical View. Proceedings of the AAAI Conference on Artificial Intelligence , 30(1), March 2016. Number: 1. [125] Tao Luo, Zheng Ma, Zhi-Qin John Xu, and Yaoyu Zhang. Theory of the Frequency Principle for General Deep Neural Networks, July 2019. arXiv:1906.09235 [cs, math, stat]. [126] Li Deng, Geoffrey Hinton, and Brian Kingsbury. New types of deep neural network learning for speech recognition and related applications: an overview. In 2013 IEEE International Conference on Acoustics, Speech and Signal Processing , pages 8599–8603, May 2013. ISSN: 2379-190X. [127] Nivaashini Mathappan, R.s. Soundariya, Aravindhraj Natarajan, and Sathish Kumar Gopalan. Biomedical analysis of breast cancer risk detection based on deep neural network. International Journal of Medical Engineering and Informatics , 12(6):529–541, January 2020. Publisher: Inderscience Publishers. [128] K Vijayakumar, Vinod J Kadam, and Sudhir Kumar Sharma. Breast cancer diagnosis using multiple activation deep neural network. Concurrent Engineering , 29(3):275–284, September 2021. Publisher: SAGE Publications Ltd STM. [129] S. Karthik, R. Srinivasa Perumal, and P. V. S. S. R. Chandra Mouli. Breast Cancer Classification Using Deep Neural Networks. In S. Margret Anouncia and Uffe Kock Wiil, editors, Knowledge Computing and Its Applications: Knowledge Manipulation and Processing Techniques: Volume 1 , pages 227–241. Springer, Singapore, 2018. [130] Keiron O’Shea and Ryan Nash. An Introduction to Convolutional Neural Networks, December 2015. arXiv:1511.08458 [cs]. [131] Saad Albawi, Tareq Abed Mohammed, and Saad Al-Zawi. Understanding of a convolutional neural network. In 2017 International Conference on Engineering and Technology (ICET) , pages 1–6, August 2017. [132] Madhusmita Sahu and Rasmita Dash. A Survey on Deep Learning: Convolution Neural Network (CNN). In Debahuti Mishra, Rajkumar Buyya, Prasant Mohapatra, and Srikanta Patnaik, editors, 90