Leveraging machine learning for diabetes prediction: Ensemble model

*Corresponding author: McDonald Otieno Ogutu Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0. Leveraging machine learning for diabetes prediction: Ensemble model McDonald Otieno Ogutu 1, *, Benson Nzioka Kituku 2 and Simon M. Karume 3 1 Department of Computer Science and Information Technology, Cooperative University of Kenya, Nairobi, Kenya. 2 School of Computer Science and Information Technology, Dedan Kimathi University, Nyeri, Kenya. 3 School of Science, Engineering and Technology, Kabarak University, Nakuru, Kenya. Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 Publication history: Received on 02 August 2025; revised on 04 October 2025; accepted on 07 October 2025 Article DOI: https://doi.org/10.30574/gjeta.2025.25.1.0267 Abstract Diabetes presents great global health challenge, with delayed diagnosis significantly impeding effective management, particularly in resource-constrained regions. This project aimed to enhance timely and accurate diabetes prediction by developing an advanced ensemble machine learning model. A hybrid dataset, compiled from the PIMA Indian (768 instances) and Hospital Frankfurt Germany (2000 instances) datasets, totaling to 2768 datapoints, was utilized to improve generalizability beyond single-source limitations. The methodology involved comprehensive data preprocessing, including the critical imputation of physiologically impossible zero values and feature standardization. F1-score was selected as the primary performance metric due to its ability to provide a vital balance between precision and recall, which is crucial in a medical context where both false positives and false negatives carry significant consequences. Six single classifier models—Logistic Regression, Decision Tree, K-Nearest Neighbors, Support Vector Machine, Random Forest, and XGBoost—were trained on the data and evaluated after hyperparameter tuning. The F1scores of these optimized models were: Logistic Regression (0.6328), Decision Tree (0.9843), K-Nearest Neighbors (0.9869), Support Vector Machine (0.9843), Random Forest (0.9947), and XGBoost (0.9974). Based on these results, XGBoost and Random Forest were selected as base learners for a Stacking Classifier ensemble, which utilized a Logistic Regression meta-learner. The developed ensemble model demonstrated exceptional performance, achieving nearperfect ROC-AUC of 0.9999 and an F1-score of 0.9974. This performance not only surpassed results from recent studies but also highlighted the significant potential of machine learning to predict diabetes accurately. The project recommended further development and integration of the ensemble model into a web application. Keywords: Machine learning; Support vector machine; Gradient boosting; Random Forest; Decision Tree 1. Introduction Diabetes is a disease affecting many people globally, causing serious health problems ((WHO), 2021). Diabetes must be detected early and accurately to be treated effectively. To respond to this, in the recent decade, data science has come up with powerful machine learning tools in the healthcare sector, providing innovative disease prediction and management solutions. This project explores the possibility of leveraging advanced ensemble machine learning classifiers to improve diabetes prediction, with the goal of increasing accuracy, reducing misdiagnosis, and ultimately contributing to better health outcomes. Diabetes ranks among the top prevalent diseases globally. World Health Organization ((WHO), 2021), asserts that this condition’s prevalence among adults of over 18 years is 8.5% and has caused 6.7 million deaths worldwide in 2021. The disease accounts for a substantial portion of premature deaths, alongside cardiovascular conditions, cancer, and respiratory diseases. Despite a decline in diabetes-related deaths from 2000 to 2010, statistics show a resurgence between 2010 and 2016, with mortality rates expected to increase further ((WHO), 2021). The disease also comes along Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 143 with huge health cost implications. In 2021, the disease caused a global health spending of at least US $960 billion. In Sub-Saharan Africa, the burden of diabetes is expected to impact significantly, with its prevalence anticipated to increase 2.5 times between 2021 and 2045. The spending in health related to diabetes is expected to go up from a total of 12.6 billion USD to 46.7 billion USD ((IDF), 2021). From the year 2015, diabetes has been treated with a lot of concern in Kenya. The disease burden in Kenya has been exacerbated by late diagnosis (Manyara, Mwaniki, Gill, & Gray, 2024). A national survey conducted in 2015 revealed that 51% of diabetes cases in urban areas had not diagnosed. Similarly, a cross-sectional study conducted by (Manyara, Mwaniki, Gill, & Gray, 2024) on some 50 patients in Nairobi revealed that 52% were diabetic but had not been diagnosed. This situation is being worsened by the fact that Kenya has a critical shortage of medical professionals, with only 1 doctor available for every 5,263 people according to Kenya National Bureau of Statistics. This is far below the WHO-recommended ratio of 1:1,000. This shortage severely hampers timely diagnosis and management of chronic diseases such as diabetes. As a result, many cases remain undetected until complications arise, contributing to increased morbidity and preventable mortality. The gap in early screening and diagnosis is especially alarming given the rising burden of diabetes in the country. There is a pressing need for innovative, scalable tools—such as machine learning models—that can assist in early prediction and intervention, especially in underserved regions where access to qualified healthcare providers is limited. In healthcare, machine learning techniques can be employed to aid in detecting the disease early enough hence its treatment. The technique analyzes the medical dataset to predict results hence lower costs of identifying complex diseases (Gadekallu, et al., 2020). By incorporating machine learning approaches, researchers and studies have succeeded in building machine learning models that are able to detect diabetes early enough, hence avoiding severe effects of the disease and ensuring early medication. Despite this breakthrough, a good percentage of these models are single classifiers which are prone to overfitting in cases of small datasets and this results into poor generalizability. Another limiting factor with single classifiers is that they are affected by noise and outliers, struggle when it comes to bias-variance tradeoff and they are not robust enough to capture complex patterns. In bid to improve predictive capabilities of single classifier machine learning models in diabetes, recent studies have gone the ensemble way. Compared to single machine learning models, ensemble models do better on accuracy, flexibility and high predictive capabilities. However, majority of the studies and researchers who have developed ensemble models classifiers to predict the risk of the disease have mostly used PIMA Indian dataset. This dataset comes with limitations such as the size; it only has 768 instances. Secondly, it represents a given ethnic group thus devoid of diversity hence not able to consider the different genetic predispositions to diabetes present in other population characteristics like the Europeans and Africans. To add on, using PIMA dataset excludes considerations such as environmental factors and lifestyle which largely influences the risk of diabetes. Lastly, because of the nature of this dataset, it limits generalizability and adoption in healthcare globally. It is on this basis that this project sought to develop from a hybrid of datasets (PIMA (768 datapoints) and Hospital Frankfurt Germany (2000 datapoints)), an accurate ensemble machine learning algorithm from single machine learning models to enhance diabetes prediction. Two best performing models from Random Forest, Support Vector Machines, XGBoost, and k-nearest neighbor were used to develop the ensemble model. 1.1. Study Objectives 1.1.1. General objective This project's primary goal is to develop an ensemble machine-learning model for predicting diabetes. 1.1.2. Specific objectives • To review existing machine learning models used in diabetes prediction. • To evaluate the performance of Logistic regression, XGBoost, Decision trees, SVM, K-NN and Random Forest, in predicting diabetes cases • To determine the best two machine learning models that are highly accurate in predicting diabetes • To develop an ensemble ML model from the two best performing ML algorithms • To evaluate the performance of the resultant ensemble ML model compared to single classifiers 1.2. Significance of the study This study holds an important place in public health, specifically in diabetes prediction. This research study plays part in developing a diagnostic model for predicting diabetes thus helping in early detection of diabetes which in turn allows for proper medication and management of the disease. This goes a long way in reducing the burden and cost associated with the disease. Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 144 A meta-analysis of machine learning models presents the medical field with valuable insights as to which machine learning approaches to adopt after considering the strengths and weaknesses of each. This will give researchers and medical professionals guidance when choosing appropriate machine learning models to use in predicting diabetes. The predictive accuracy of the developed ensemble machine learning model from best performing single classier models is expected to improve the accuracy significantly. This innovation can lead to more robust and flexible prediction systems that can adapt to diverse datasets and populations, enhancing the generalizability of the models. As a country, in the spirit of primary healthcare where the government of Kenya is focusing more on preventive approaches to health as opposed to curative approaches, this innovation will come in handy in helping through screening of members of households at the community level. This will lead to early detection and medication thus preventing the disease from progressing to advanced stages leading to out-of-pocket catastrophic costs for the families. 2. Literature review Lately, following the enormous breakthrough in the machine learning domain, learning techniques have been devised which are quite good at enhancing the detection of cases of diabetes mellitus. In ensemble learning, one has a learning technique, known as the super learner, increasing accuracy by combining outputs from different machine learning algorithms. In their paper, (Dogru, Buyrukoglu, & Ari, 2023) developed the super learner framework based on four classifier models: LR, DT, RF, and gradient boosting; and a meta-learning component using SVM. The study evaluated this model on three datasets to confirm its efficacy. The findings proved that the super learner system performed very well in comparison with each base learner model as a fore-runner of the high-accuracy system for diagnosing diabetes: early-stage risks were forecast with 99.6% accuracy, PIMA data at 92%, and diabetes data from 130 US hospitals at 98%. In their paper, (Dutta, et al., 2022) put forward a new dataset for diabetes from Bangladesh and an automated classification pipeline with a weighted ensemble of machine learning classifiers, namely NB, RF, DT, XGBoost, and LightGBM. It implements Grid Search hyperparameter optimization on K-fold cross-validation, critical hyperparameters, feature selection, and missing value imputation. The results indicated a statistically significant improvement in diabetes prediction performance with the proposed weighted ensemble (Decision Tree + Random Forest + XGBoost + Light GBM) along with preprocessing techniques to attain 0.735 for accuracy and 0.832 for AUC. The study further showed that statistical imputation and RF-based feature selection combined with the identified ensemble technique had given the best results for predicting the risk of diabetes early enough. Employing Bayesian networks and radial basis function, (Mahesh, et al., 2022) came up with a blended ensemble machine learning model to aid in diabetes prediction. They used the developed ensemble model to compare the performance of LR, DT, SVM, K-NN and RF. The study found that the resultant developed ensemble model performed better than the traditional single-learning classifiers by achieving an accuracy score of 97.11%. According to (Atif, Anwer, & Talib, 2022), they opine that single classifier models come with several limitations, the main one being compromising on accuracy because of the models’ lack of generalizability over different datasets. In remedying the situation, they developed a model employing hard voting classifier by combining SVM, LR and DT algorithms. In evaluating their model, they used the PIMA dataset and the Early-Stage Diabetes Risk Prediction Dataset. The results indicated that the ensemble model had superior outcomes with accuracies of 81.17% on PIMA dataset and 94.23% on Early-Stage Diabetes Risk Prediction Dataset. The conclusion from this study was that the ensemble models based on hard voting classifiers especially enhanced prediction in terms of accuracy and reliability different sets of datasets notwithstanding (Atif, Anwer, & Talib, 2022). In another research, (Abnoosian, Farnoosh, & Behzadi, 2023) utilizing imbalanced Iraqi Patient Data, employed a pipeline-based multi-classification approach in prediction of diabetes in three unique groups: the non-diabetic, prediabetes and the diabetic. The study used base classifiers like Gaussian Naive Bayes (GNB), k-NN, RF, SVM, AdaBoost and DT. Since the dataset was imbalanced, they used a weighted ensemble method anchoring on the Area Under the Receiver Operating Characteristic Curve (AUC). To optimize performance, the study employed grid search and Bayesian optimization for hyper-parameter tuning. The resultant ensemble machine learning model compared to single machine learning models, showed superior performance. It gave an accuracy of 0.9887, F1-score of 0.9851, precision of 0.9861, a recall of 0.9792 and an AUC of 0.999. A study by (H. Qi, 2023) developed an ensemble learning model which they named KFPredict. This model incorporated various input models with significant features and different single classifier models for diabetes prediction. The model Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 145 was developed by creating neural network model (KF_NN), that was multi-input in nature. It employed recursive feature in decision trees elimination algorithm together with correlation method to determine significant and non-significant variables. The KF-NN model was then infused with KNN, SVM and RF for the purpose of soft voting and, hence creating a predictive model. The results showed that KFPredict attained accuracy score of 93.5%, sensitivity score of 85% and specificity score of 98%. Compared to single classifier models, this was an improvement of 18.18%. This research underscored the effectiveness of the KFPredict approach in giving robust prediction outcomes on PIMA diabetes data (H. Qi, 2023). An ensemble machine learning algorithm was modelled by (A. Singh, 2021) to predict diabetes employing XGBoost, RF, SVM, Neural Network, and DT. They named it eDiaPredict. Different evaluation metrics like specificity, Gini Index, minimum error rate, area under the curve (AUC), accuracy, area under the convex hull, sensitivity and minimum weighted coefficient were used to evaluate its performance. PIMA Indian dataset was used and eDiaPredict achieved an accuracy score of 95%. According to this study, eDiaPredict enhanced the prediction of diabetes through ensemble modeling. Another voting classifier known as En-RfRsK developed by (Amma, 2024) integrated three single ML classifiers namely, K-Nearest Neighbor, Random Forest and Radial SVM to predict the risk of diabetes. To evaluate this ensemble model, the PIMA data was used. The results showed that En-RfRsK was superior to single classifiers by achieving an accuracy score of 88.89% thus showing its usefulness and effectiveness in enhancing the predictive performance. In predicting blood sugar levels (H. Yang, 2024) used an enhanced stacking ensemble approach that incorporated three enhanced Long Short-Term Memory (LSTM) network models like single classifiers. To ensure adaptive weighting, the study used improved Nearest Neighbor Propagation Clustering Algorithm. To evaluate the model’s performance, the study used the OhioT1DM dataset. Results showed that this model achieved a Root Mean Square Error (RMSE) of 1.425 mg/dL, Mean Absolute Error (MAE) of 0.721 mg/dL, and Matthews Correlation Coefficient (MCC) of 0.982 for a 30minute prediction horizon. The model achieved RMSEs of 3.212 mg/dL and 6.346 mg/dL, MAEs of 1.605 mg/dL and 3.232 mg/dL, and MCCs of 0.950 and 0.930 for 45-minute and 60-minute horizons respectively. The study concluded that LSTM ensemble technique improved RMSE & MAE by 27.92% and 65.32%, in that order compared to non-ensemble model (H. Yang, 2024). 2.1. Gaps identified in previous works A review of recent studies on machine learning models for diabetes prediction reveals that a wide range of algorithms— such as Logistic Regression, Decision Trees, Random Forests, Support Vector Machines, K-Nearest Neighbors, Gradient Boosting, and Naive Bayes have been widely applied. The PIMA Indian Diabetes dataset remains the most frequently used data source across the literature, appearing in studies by (Dogru, Buyrukoglu, & Ari, 2023), (Mahesh, et al., 2022), (Atif, Anwer, & Talib, 2022), (H. Qi, 2023), (A. Singh, 2021), and (Amma, 2024). Despite the variety of algorithms explored, a dominant limitation is the over-reliance on relatively small datasets, particularly the PIMA dataset which only contains 768 instances. This limits the generalizability and robustness of model performance in real-world, diverse populations. In addition to small sample sizes, other studies such as (Dutta, et al., 2022) and (Abnoosian, Farnoosh, & Behzadi, 2023) have used region-specific datasets like the DDC Bangladesh dataset and the Iraqi Patient Dataset. However, these also come with limitations. For instance, the Iraqi dataset is significantly imbalanced, with a disproportionately higher number of non-diabetic cases (91,500) compared to diabetic ones (8,500), posing challenges for fair model training and evaluation. 2.2. Conclusion from the review From the literature review, studies on diabetes prediction predominantly rely on relatively small and often imbalanced datasets most notably the PIMA Indian dataset with only 768 instances. This limitation constrains the development of robust and generalizable machine learning models. Additionally, while a variety of algorithms such as Random Forest, Support Vector Machines, XGBoost, and Decision Trees have been tested, few studies have explored ensemble approaches built from multiple strong-performing base learners, especially using diverse data sources. It is on this basis that this project seeks to bridge these gaps by developing an accurate ensemble machine learning model for diabetes prediction. The model will be built from a hybrid of datasets the PIMA dataset and the Hospital Frankfurt Germany dataset (with over 2,000 datapoints) to improve data diversity and model generalizability. From among the single machine learning models (Random Forest, Support Vector Machines, XGBoost, and K-Nearest Neighbor), the two best-performing algorithms will be selected and combined to form the ensemble model. This Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 146 approach aims to enhance predictive accuracy and support early diagnosis, especially in resource-limited settings where delayed detection contributes to rising diabetes-related morbidity and mortality. 3. Methodology This chapter covers dataset description, the data pre-processing steps, model development, model evaluation metrics, and the selection process for the best models. Figure 1 Approach flow for the ensemble modelling 3.1. Data source and description Aggregated data was sourced from two sources; PIMA Indian dataset (https://www.kaggle.com/datasets/uciml/pimaindians-diabetes-database) and Dataset of diabetes, taken from the hospital Frankfurt, Germany (https://www.kaggle.com/datasets/johndasilva/diabetes) all of which are available in Kaggle. The Indian PIMA dataset has 768 data points while the Hospital Frankfurt dataset has 2000 data points. The variables in the datasets are diabetes status, glucose level, blood pressure, skin thickness, insulin level, BMI (body mass index), pregnancy, diabetes pedigree function and age. Table 1 below shows the attributes. Table 1 Data attributes No Attribute Data type Description 1 Pregnancy Numeric The number of pregnancies 2 Glucose Numeric Glucose plasma levels two hours after consuming glucose 3 Blood pressure Numeric Diastolic blood pressure (mm Hg) 4 Skin thickness Numeric Thickness of the skin fold on the triceps of the upper arm (mm) 5 Insulin Numeric Insulin serum levels in the blood two hours after the glucose test (lh/ml) 6 BMI Numeric Body mass index [weight in kg/(Height in m)] 7 Diabetes Pedigree Function Numeric Numerical value that estimates a person’s genetic risk of developing diabetes based on their family history. 8 Age Numeric Patient age 9 Diabetes status Categorica l Diagnostic results (Diabetic or not diabetic) Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 147 3.2. Data pre-processing steps 3.2.1. Dealing with missing values Measures of central tendency (median) technique of imputation was used to manage missing values. This approach ensures that the dataset remains comprehensive and usable. 3.2.2. Data standardization The project employed Min-Max scaling method to ensure that the range of features are normalized. This will ensure that data falls in a given range, always between 0 and 1. This in turn strengthened the robustness of machine learning models that are sensitive to feature scaling. 3.2.3. Encoding One-hot encoding was employed to assign numerical values to each category of a qualitative variable. This applied to the dependent variable only as it was the only categorical variable in the dataset. The response variable was diabetes status and was dichotomous which is “diabetic” or “not diabetic”. “Diabetic” status was coded as 1 and “not diabetic” status was coded as 0. 3.2.4. Feature selection Feature selection was conducted to identify the most relevant independent variables for model training and to address potential issues of multicollinearity. Correlation heatmaps were employed to assess the relationships between independent variables. During this assessment, no significant cases of multicollinearity were identified among the independent variables as can be observed in figure 2, therefore, all independent variables were retained for subsequent analysis, ensuring the preservation of their individual predictive utility. Figure 2 Correlation matrix of features Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 148 3.2.5. Class balance analysis To gain a comprehensive understanding of the dataset's core, the distribution of the target variable (diabetes status) was analyzed, providing a clear picture of the class balance. This step was vital for identifying potential class imbalance that might necessitate specific handling during model training. The diabetics were 65.6% while the non-diabetics were 34.4%. The analysis result is as shown in figure 3 below; Figure 3 Distribution of diabetes status 3.2.6. Data distribution The individual characteristics of all variables were explored through histograms, which revealed the distribution patterns of each numerical variable. It can be observed that glucose, blood pressure and BMI were moderately normally distributed. The rest of the variables were skewed. The results were as shown in figure 4 below. Figure 4 Distribution of variables Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 149 3.3. Model building and performance analysis 3.3.1. Model building The research dataset was divided into two splits: one was the training set, and the other was the testing set. The training set had 80% of the data while the testing set had 20% of the data. The splitting method allows the model to be trained on a majority of this data while retaining a separate set for evaluating its performance, ensuring that the model performs well at generalization. Six single classifier models—Logistic Regression, Decision Tree, K-Nearest Neighbors, Support Vector Machine, Random Forest, and XGBoost were trained and tested on the data and evaluated after hyperparameter tuning 3.3.2. Performance analysis To evaluate the performance of each model, the project used several evaluation metrics: Accuracy: This was to quantify how much of the total number of samples were correctly predicted. It measures the model reliability on classification of a positive or negative cases. On the other hand, when training datasets are imbalanced, accuracy can be misleading; it quickly goes upwards without paying attention to detection of positive class. 𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 = 𝑻𝑷 +𝑻𝑵 𝑻𝑷 +𝑻𝑵 +𝑭𝑷 +𝑭𝑵 The subject should be classified as either diabetic or not diabetic. Out of the classification, we had subjects truly classified and falsely classified hence the below; • True positive (TP): The subject is predicted as positive (diabetic) and is diabetic. • True negative (TP): The subject is predicted as negative (not diabetic) and is actually not diabetic. • False positive (FP): The subject is predicted as positive (diabetic) and is not diabetic. • False negative (FN): The subject is predicted as negative (not diabetic) but is actually diabetic. • Precision: Indicates the proportion of true positive rightly predicted among all predicted positives. 𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏 = 𝑻𝑷 𝑻𝑷 +𝑭𝑷 Recall: Reflects the ability of the model to identify all relevant instances (true positives). 𝑹𝒆𝒄𝒂𝒍𝒍 = 𝑻𝑷 𝑻𝑷 +𝑭𝑵 F1-Score: A harmonic mean of precision and recall, providing a single metric for model evaluation. Area Under the Receiver Operating Characteristic (ROC) Curve (AUC): Evaluates the model's power to differentiate between positive and negative classes. 3.3.3. Ensemble Machine Learning Development To leverage the complementary strengths of multiple high-performing individual classifiers and enhance predictive accuracy, an ensemble model was constructed using the Stacking Classifier methodology. This approach involved building a two-layer structure: the first layer consisted of base estimators, which were the two best-performing single classifiers identified previously, namely XGBoost and Random Forest, chosen for their superior F1-scores. The second layer comprised a meta-learner, a Logistic Regression model, which was tasked with combining the predictions generated by these base estimators. During the ensemble's training, a crucial step involved employing 5-fold crossvalidation internally, ensuring that the base models generated their predictions on out-of-fold data to prevent data leakage. Furthermore, these base models passed their predicted probabilities to the meta-learner, providing richer information than just hard class labels. The ensemble was configured to utilize all available CPU cores for efficient parallel processing and was set to not pass the original features directly to the final estimator, ensuring the meta-learner solely focused on integrating the insights from the base models' outputs. The complete ensemble model was then trained on the designated training dataset. Global Journal of Engineering and Technology Advances, 2025, 25(01), 142-155 150 4. Results 4.1. Overall performance of single classifier models before and after hyperparameter tuning The six single classifier models were initially evaluated using their default hyperparameters to establish baseline performance. Subsequently, hyperparameter tuning was performed using GridSearchCV to optimize its predictive capabilities on the Diabetes dataset. It can be observed that there were major improvements in K-Nearest Neighbour and Support Vector Machine in terms of the overall performance after hyperparameter tuning. For example, there was a 48% improvement in recall metric in support vector machine after hyperparameter tuning. The results are as in table 2 below. Table 2 Model performance comparison before and after hyperparameter tuning Model Metric Default Tuned Improvement % Logistic Regression Accuracy 0.7726 0.7780 0.70 Logistic Regression Precision 0.7211 0.7361 2.08 Logistic Regression Recall 0.5550 0.5550 0.00 Logistic Regression F1-Score 0.6272 0.6328 0.90 Logistic Regression ROC-AUC 0.8355 0.8355 0.00 Decision Tree Accuracy 0.9892 0.9892 0.00 Decision Tree Precision 0.9843 0.9843 0.00 Decision Tree Recall 0.9843 0.9843 0.00 Decision Tree F1-Score 0.9843 0.9843 0.00 Decision Tree ROC-AUC 0.9880 0.9880 0.00 K-Nearest Neighbor Accuracy 0.8448 0.9810 17.31 K-Nearest Neighbor Precision 0.7807 0.9844 26.08 K-Nearest Neighbor Recall 0.7644 0.9895 29.45 K-Nearest Neighbor F1-Score 0.7725 0.9869 27.76 K-Nearest Neighbor ROC-AUC 0.9391 0.9999 6.47 Support Vector Machine Accuracy 0.8412 0.9892 17.60 Support Vector Machine Precision 0.8411 0.9843 17.03 Support Vector Machine Recall 0.6649 0.9843 48.03 Support Vector Machine F1-Score 0.7427 0.9843 32.53 Support Vector Machine ROC-AUC 0.9002 0.9996 11.04 Random Forest Accuracy 0.9964 0.9964 0.00 Random Forest Precision 1.0000 1.0000 0.00 Random Forest Recall 0.9895 0.9895 0.00 Random Forest F1-Score 0.9947 0.9947 0.00 Random Forest ROC-AUC 0.9999 0.9999 0.00 XGBoost Accuracy 0.9928 0.9982 0.55 XGBoost Precision 0.9845 1.0000 1.58 XGBoost Recall 0.9948 0.9948 0.00