Explainable Machine Learning Framework for Predictive Analysis of Cardiac and Mental Health Disorders

Birla Institute of Technology and Science, Pilani Explainable Machine Learning Framework for Predictive Analysis of Cardiac and Mental Health Disorders Pawan Bhansari , Sarvesh Mehta [email protected], [email protected] 1.Abstract Here an integrated, entirely machine learning-based predictive framework is addressed, particularly meant for two important health domains – diagnosis of cardiac diseases and prediction of treatment for mental health issues. The cardiac health model seeks clinical parameters such as troponin, CK-MB, and blood pressure, and it has been developed on psychosocial survey data, which includes family history, workplace support, and personal experiences of mental health issues for the mental health model. Both models were created under ensemble and deep learning approaches with said model XGBoost bringing out the final best-performing algorithm as it achieved the best accuracy, recall, and interpretability. The SHAP-based explanation technique, which contributed to both global and local insights on features' importance, was implemented for clinical and organizational trust. The cardiac model is completed with a rule-based recommendation engine to convert predictive analysis to actionable clinical steps, while the mental health model is adaptable for integration into HR wellness platforms and chatbots. Results demonstrate effectiveness with encouraging predictive accuracy—98.43% for cardiac and 78.4% for mental health—with recall scores sufficiently high to support real-world deployment. The dual-model approach will in future enhance early diagnosis and set precedence for holistic AI-assisted decision-making in both medical and organization settings. Keywords: Machine Learning, Cardiac Disease Prediction, Mental Health, XGBoost, SHAP Explainability, Ensemble Learning, Healthcare AI, Clinical Decision Support, Rule-Based System, Wellness Chatbot Integration 2. Introduction One of the most critical health challenges worldwide, cardiovascular diseases account significantly for global mortality. Though advances are made in so many different fields of screening techniques such as physical assessments, imaging, and laboratory evaluations amongst others, they are all still majorly resource-intensive, time-consuming, and often inaccurate. The feedback loop for diagnosis is often extended due to a lack of healthcare infrastructure, particularly in more remote and less developed areas. It is in the recent past that machine learning (ML) found its way as a disruptive engine in the healthcare domain, offering automated solutions for heart disease prediction and diagnosis by analysing demographic and lifestyle data, medical history, and test outputs. Existing ML models for detecting heart disease, however, face significant limitations pertaining mainly to the overfitting of the training dynamics, lack of generalization across datasets, and being so focused on heart disease per se that they tend not to account for coexisting chronic conditions. Most of the current predictive models are disease-specific, primarily concerning themselves with isolated cardiac disease without any incorporation of other fatal diseases, such as diabetes and hypertension, which share risk factors and commonly cooccur. Such limited approaches, however, greatly reduce their clinical applicability on the ground where multiple diseases exist simultaneously. What is more, these models suffer from limited external validity in that they do not tend to perform consistently when new or unseen data are applied to them, undermining their reliability. Another big hurdle is overfitting: some models achieve high accuracy in the training datasets but show very low performance in real-world applications. To be able to tackle the weaknesses thus far presented, this project aims to design a more comprehensive predictive model that builds upon the greater accuracy of heart disease identification but is also able to predict multiple chronic diseases. The system thus proposed will apply advanced ensemble and deep learning techniques to enhance diagnostic accuracy and address overfitting and generalization issues. The dream is for an integrated framework capable of predicting a host of chronic diseases and thus providing timely assistance to the medical practitioner. This report starts with an extensive literature survey on machine learning applications in prediction models for cardiac and mental health. The literature review systematically reviews eight pioneering papers that show the advancements in predictive modelling while addressing other limitations found in previous research. Each paper adds to the advancement of knowledge in this area until integrated approaches combining physiological and psychological elements with improved diagnostic accuracy emerge. By blending insights from these studies and integrating innovative methodologies in the proposed model, this project positions itself toward the advancement of predictive healthcare analytics, imposing an impact on the management of chronic diseases on a global scale. 3. Literature Review In [1] the author tries to implement the first use of machine learning models in the healthcare industry focusing on cardiac related disease detection. The authors elaborate on the role of AI in assisting health care professionals in making more accurate predictions and decisions concerning diseases. One main problem noted was the traditional machine learning models inability to accommodate the wide variety of patient data, particularly pertaining to variation in personal health status. These issues were solved in Paper X2 where ensemble learning was implemented to improve the model accuracy and reliability. In [2] paper focuses on solving the issues of overfitting and model generalization found in Paper X1. The authors propose the use of ensemble learning techniques, specifically voting classifiers, to combine the strengths of multiple machine learning models like decision trees and support vector machines (SVMs). This tries to improves the accuracy and robustness of predictions, and reduces overfitting problem . Although it is slight improvement over previous models, it still lacks the ability to handle complicated non-linear data relationships as effectively as DL models. This limitation was addressed in Paper X3. Following along X2, in [3] author combines deep learning approaches and traditional machine learning techniques for better heart disease prediction. Deep learning greatly improves the modeling of complicated data structures that ensemble techniques in Paper X2 fail to capture. This paper strives to increases predictive accuracy using convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for nonlinear datasets. Nevertheless, the application of deep learning increases the computational cost, which was mitigated in Paper X4 by using decision tree learners for efficiently multi disease prediction. In [4] author tries to introduces a multi disease prediction model that takes in consideration more conditions like diabetes and breast cancer. They have used decision trees and support vector machines (SVMs) to predict multiple chronic diseases with the help of the model . Major contribution of this model is its ability to predict multiple diseases together , offering a more advanced approach than earlier models that focused only on cardiac related diseases. While multi-disease prediction is significant , it also introduces more complexity into the model, which was later addressed by Paper X5. In this paper [5], efforts are directed towards enhancing the prediction reliability and generalization aspects of heart disease detection models. It also addresses overfitting in models as described in Paper X2, and suggests regularization techniques to mitigate such difficulties. The research incorporates several machine learning models, including Random Forest and Gradient Boosting, in order to achieve better generalization across various datasets. These approaches improve the prediction accuracy of the models along with previous models described in Paper X2 and Paper X3, which relied on hyperparameter optimization and regularization to a lesser degree In [6]The researchers of this paper study how deep learning techniques can be used alongside ensemble approaches to mitigate the shortcomings of earlier models, such as overfitting and high computational costs. The paper seeks to employ an ensemble approach which integrated with deep learning models will produce higher level predictive accuracy. This work is a continuation of Paper X5 in which new ensemble techniques that blend the advantages of deep learning and conventional models to complex datasets are introduced. In [7] paper focuses on the implementation of modern AI methods to multi-disease prediction models. The authors develop a hybrid framework merging machine learning and natural language processing (NLP) techniques to analyse electronic health records (EHRs) for heart disease, diabetes, hypertension, and other chronic conditions. The study proposes that unstructured elements within EHRs can enhance predictions and provide deeper understanding to certain insights. This effort develops Paper X4 by presenting an enhanced model for multi-source data integration pertaining to disease prediction. Paper [8] introduces important concepts of machine learning to clinicians and summarize its applications in cardiovascular disease. They emphasize that "ML—machine learning, which takes advantage of computer algorithms that can learn complex patterns in data—has great potential to impact cardiology through the sheer number of diagnostic and management decisions dependent on digitized patient-specific information." The paper attempts to give a complete bibliometric survey of ML papers in cardiology with case studies on applications that include denoising, feature extraction, and enhancement of classical algorithms. Although recognizing many potential applications throughout the cardiovascular care continuum, they agree that "machine learning has not achieved much real-world impact on the clinician cardio-logic practices and patients" and thus more efforts need to go into R&D in this area. This systematic review in [9] defines a complete study on machine-learning techniques used for prediction of mental well-being problems among different individuals. According to PRISMA methodology, these authors systematically reviewed 30 research papers, reporting them under different categories following various conditions of mental health such as "schizophrenia, bipolar disorder, anxiety and depression, posttraumatic stress disorder, and mental health problems among children". The study highlighted immediate models that could predict mental health indicating that "Convolutional Neural Networks (CNN), Random Forest (RF), Support Vector Machine (SVM), Deep Neural Networks, and Extreme Learning Machine (ELM)" are often adopted for predictions. Among them, Random Forest achieved an accuracy of 68.6% in schizophrenia prediction, while the performance of CNN was shiny in diagnosing the bipolar disorder. Still, delineating good challenges, it shows that "there needs to be more extensive datasets, heterogeneous nature of mental health condition to be considered, and long term data need to be included" and provides concrete recommendations for future research. According to [10], early prediction of heart disease models is being developed through various feature selection techniques; reasons for this require attention, and the global estimates which demand attention may be reading: "almost 17.9 million people are losing their lives due to cardiovascular disease, which is 32% of total death throughout the world." Three feature-selection processes were employed in the study: chi-square, ANOVA, and mutual information, respectively, called SF1, SF2, and SF3. The six machine learning algorithms, namely logistic regression, support vector machine, Knearest neighbor, random forest, Naive Bayes, and decision tree, were used to evaluate the various models. The study discovered through comparative analysis that "the random forest exhibited the most optimistic performance for SF3 feature subsets with 94.51% accuracy, 94.87% sensitivity, 94.23% specificity, 0.9495 AUC, and 0.31 log loss." The model proposed, therefore, has great clinical promise for early heart disease prediction at less time and cost. The author in paper [11] deals with an urgent issue of timely identification of patients at risk for mental health crises by developing "a machine learning model that utilizes electronic health records to continuously monitor patients for risk of a mental health crisis for 28 days," with data gleaned from 17,122 patients across seven years (2012-2018). The model obtained impressive performance metrics that included "an area under the receiver operating characteristic curve of 0.797 and an area under the precision-recall curve of 0.159, predicting crises with a sensitivity of 58% at a specificity of 85%." Beyond theoretical validation, the researchers did a six-month prospective study to assess the clinical value of the algorithm, in which they would find that predictions were "clinically valuable in terms of either managing caseloads or mitigating the risk of crisis in 64% of cases." Of the different types of machine learning used across the testing, XGBoost had the highest overall performance, but this performance did vary with different diagnostic groups and demographic variables, indicating areas for further refinement. The paper [12] presents a comparative study of the various machine learning algorithms for predicting heart diseases with a focus on their optimization features by A.A. Ahmad and H. Polat. The paper evaluates various machine learning algorithms such as Random Forest (RF), Nave Bayes (NB), Support Vector Machines (SVM), Artificial Neural Network (ANN), Decision Tree (DT), and AdaBoost on, presumably, the Cleveland heart disease dataset. The authors note that "while comparing the results of the tested ML algorithms, the RF gave the most promise with an accuracy of around 90.16%," while another study cited in their paper concludes that "RF (91.80%) had the highest accuracy in predicting heart disease, followed by NB (88.52%) and SVM (88.52%)." Using the Jellyfish Optimization Algorithm for feature selection further enhanced the performance with an astonishing accuracy of 98.47% for the SVM model, obtaining conjointly with this optimization technique. This work provides further strengthening to Random Forest for cardiac prediction with good validation across other datasets. Machine learning in predicting mental health conditions of college students has great potential, which has been systematically reviewed in literature regarding some deep learning techniques over the years in [13]. The research analyzed and synthesized about 30 papers selected on college student populations, with special emphasis on outcomes such as Bipolar Disorders (6 papers), Schizophrenia Prediction (4 papers), PTSD (6 papers), Depression and Anxiety (10 papers), and ADHD (4 papers). Findings reveal prominent models in predicting mental health conditions as CNN, RF, SVM, DNN, and ELM, with the statement, "CNN showed very great accuracy in diagnosing bipolar disorder when compared to some other models." Other researchers "used Decision Tree, Neural Network, Support Vector Machine, Naive Bayes, and logistic regression algorithms to classify students into categories for different types of mental health problems, hence coming up with very distinct optimal models for their respective concerns." Despite these developments, the research has discovered a more serious challenge in the absence of a standard dataset, ethical issues involving data privacy, and model interpretability. The researchers in [14]. offer an integrated machine learning methodology for congestive heart failure (CHF) prediction with the principal objective of "using machine learning applications for better diagnosis of CHF and for cheaper diagnosis." Several classification techniques are being employed in this work including Decision Tree, Support Vector Machine, K-Nearest Neighbour, Random Forest, Logistic Regression, and Deep Neural Network, all of which are complemented by comprehensive preprocessing steps including cleaning and splitting, feature selection, and imputing of missing data using C4.5 and KNN. The Cardiovascular Health Study (CHS) dataset is used, along with novel optimization methods to achieve better prediction accuracy while minimizing computational complexities. The integrated approach achieved highly commendable performance metrics with an accuracy of 95.30%, sensitivity of 96.49%, F1 score of 97.03%, and precision of 97.58%, which makes it a highly specialized and costeffective CHF diagnostic tool. Nevertheless, non-cardiovascular elements, especially the psychological ones, become untouched avenues for the improvement of prediction accuracy in the course of research. Research done in [15], essentially ends a gap juncture between prediction models for mental and cardiovascular health, a gap recognized from earlier studies by these researchers. The researchers developed "an ensemble ML model containing 5 constituent algorithms (decision tree, random forest, XGBoost, support vector machine, and deep neural network)" and tested it using both traditional CVD risk factors and a combination that included psychological factors. The results were amazing: "The ensemble ML model could predict CVD with 71.31% accuracy using just traditional CVD risk factors. However, including psychological factors in the training data increased the accuracy to 85.13%." This is a massive improvement over models using traditional risk factors alone. The study specifically calls out the connection between mental health and disease in general, pointing out that "numerous guidelines within the CVD literature focus on the psychological aspects of care," the "link between psychological factors and CVD is recognized." The research is a testament to how one could take integrated psychiatric evaluative data to improve cardiovascular risk prediction models 4. Implementation 4.1 Dataset Description for Cardiac Health For this research work the dataset we used was was collected from Kaggle and has over 1,319 patient records with variables directly or indirectly relating to cardiac wellness. The dataset has 9 main attributes, including demographic information, vital signs, biochemical markers, and a final diagnosis indicating the presence or absence of cardiac disease. The attributes are detailed as follows: 1. Age (years) 2. Gender (binary: 0 = Female, 1 = Male) 3. Heart Rate (beats per minute) 4. Systolic Blood Pressure (mmHg) 5. Diastolic Blood Pressure (mmHg) 6. Blood Sugar (mg/dL) 7. CK-MB (Creatine Kinase-MB isoenzyme, ng/mL) 8. Troponin (ng/mL) 4.2 Dataset Description for Mental Health This research uses the "Mental Health in Tech Survey" data set which was made publicly available by Open Sourcing Mental Illness or OSMI on Kaggle. The data set covers a total of 1,319 responses and is well-rounded in terms of the information it collects about individuals including demographic variables, history of mental health, accommodations within the workplace, and their views about such workplace offerings. Key features used in the analysis include: 1. Age (Numerical) 2. Gender (Categorica 0 cm onehot encoded) 3. Family History (Binary: yes/no) 4. Work Interfere (Categorical: Never, Rarely, Sometimes, Often) 5. Benefits (Binary: yes/no) 6. Care Options (Binary: yes/no) 7. Wellness Program, Seek Help, Anonymity (All Binary) 8. Employer Size, Remote Work Status, Tech Company Affiliation 9. Target Variable: Treatment (1 = has sought treatment, 0 = has not) 4.3 Data Preprocessing for cardiac health Preprocessing is basically cleaning, organizing, and getting a dataset ready for optimal performance for machine learning purposes. We set the following extensive preprocessing steps: 4.3.1 Finding Missing Values • Each column was checked for missing values using data.isnull().sum(). • The dataset has no missing values, showing that data collection had been done very well. 4.3.2 Checking Data Types • Now, Age, Gender, Heart Rate, Systolic BP, and Diastolic BP were integers, whereas Blood Sugar, CK-MB, and Troponin were float. • Made sure that the 'Result' variable wellformatted in string type so that it could later be converted into binary target. 4.3.3 Outlier Detection and Removal • Box plots drawn for continuous variables (Heart Rate, Systolic BP, Diastolic BP, Blood Sugar, CK-MB, Troponin). • Extreme outliers mostly identified were: o The heart rate of 1111 bpm was considered unfounded clinically and was therefore deleted from the data. o Systolic/Diastolic BP: Out of such out leaving above 250 mmHg and below 70 mmHg were considered extreme data. o Blood Sugar: Values less than 40 mg/dL and greater than 600 mg/dL were classified as invalid measures and not considered. 4.3.4 Feature Creation • Troponin Flag: Created a binary feature to flag elevated troponin levels (>0.04 ng/mL) to enhance model sensitivity to myocardial injury indicators. • Abnormal CK-MB Flag: A binary feature to mark CK-MB levels above 50 ng/mL, signifying potential cardiac tissue damage. 4.3.5 Encoding and Splitting • Converted "Result" from categorical ('positive', 'negative') to binary (1, 0). • Used an 80-20 stratified train-test split to maintain class proportions during model training and evaluation. 4.4 Data Preprocessing for Mental health 4.4.1 Finding Missing Values Initial examinations, specifically isnull().sum(), suggested that missing data resided in columns such as state, and country, and most importantly, comments. These were subsequently excluded from the dataset due to a heavy null-occupied presence or non-importance in prediction. Minimal missingness in work_interfere could have been imputed with a placeholder category "Unknown." 4.4.2 Checking Data Types The numerical features (i.e., Age) and categorical variables were checked for correctness. The target variable was treatment, encoded as a binary integer. All categorical values such as gender were standardized and encoded. 4.4.3 Outlier Detection and Removal Some outliers in the Age column were filtered out: records between 18 and 100 years were retained, while records outside were discarded. Consistency in the gender data was restored, and rare or ambiguous categories were filtered out to reduce noise. 4.4.4 Feature Creation Some custom features were engineered to improve model input, such as: • Remote_Work_Flag: Binary flag indicating whether remote work was allowed. • Company_Size_Bucket: Small (less than 100), Medium (100–500), Large (above 500, inclusive). • Support_Index: A composite variable taking into account mental health benefits, wellness programs, and availability of care options. 4.4.5 Encoding and Splitting Label encoding for binary variables and a mix of one-hot encoding for multi-class fields were used. The dataset was then split for training and testing in the ratio of 80:20 by stratified sampling to maintain a balance in class distribution. 4.5 Feature engineering and Feature Importance for cardiac health Engineering features were not merely to prepare the data but to enrich it with domain knowledge. Certainly not every segment selected and engineered tied to known pathophysiological mechanisms directly relates to cardiac disease. 4.5.1 Attribute-Wise Medical Relevance • Age: Cardiovascular risks increase linearly with age, primarily due to vascular stiffening and endothelial dysfunction. • Gender: Males have earlier coronary events than females because of the hormonal profile differences (for example, estrogen protection). • Heart Rate: Tachycardia, chronic (greater than 100 bpm), can lead to cardiomyopathy. Higher resting heart rates correlate with increased cardiovascular mortality. • Systolic Blood Pressure: This increase in systolic pressure is one primary driver of left ventricle hypertrophy and heart failure. • Diastolic Blood Pressure: High values of diastolic measure correspond with increased peripheral va • Systolic Blood Pressure: Elevated systolic pressure is a primary driver of left ventricular hypertrophy and heart failure. • Diastolic Blood Pressure: High diastolic values are linked to increased peripheral vascular resistance and early atherosclerosis. • Blood Sugar: Hyperglycemia is a hallmark of diabetes mellitus, which significantly elevates cardiac risk via vascular damage. • CK-MB: A cardiac biomarker that rises significantly after myocardial infarction. • Troponin: The gold standard in diagnosing myocardial injury. • Result: Defines whether the patient had a confirmed cardiac event. o Implements various approaches to mitigate overfitting o Highly SHAP compatible 6.2.6 Justification for Choosing XGBoost Among all the models that were tested, XGBoost provided the best performance in this study. This, thus, is the choice for the final deployment. • Performance: o XGBoost attained the highest recall of 84%, together with a robust F1score of 0.80, making it the most desirable choice in health applications where false negatives should be avoided as much as possible. • Regularization: o XGBoost includes L1 and L2 regularization to minimize overfitting with training data, unlike, say, Random Forest or SVM [19]. • SHAP-Integrated Explainability: o XGBoost integrates smoothly with SHAP, so it delivers interpretable global Predicted Negative Predicted Positive Actual Negative 98 0 Actual Positive 4 153 and local feature contributions, which is crucial for validating predictions in mental health applications to make them trustworthy for stakeholders [20]. • Operational Status: o The model is efficient, scalable, and working with real-time inference engines, thereby being a potential candidate for elaborate Model Accuracy Recall Precision AUCROC SVM 89.01% 84.71% 88.12% 0.91 Random Forest 98.43% 97.45% 98.70% 0.99 XGBoost 98.43% 97.45% 98.70% 0.99 deployment in HR wellness platforms or clinical decision support tools. 7. Results and Discussion 7.1 Model Performance for Cardiac Health The predictive performance of the three machine learning models was evaluated on the held-out test dataset using key evaluation metrics: Accuracy, Recall, Precision, and AUC-ROC (Area Under the Receiver Operating Characteristic Curve). 7.1.1 Confusion Matrix Analysis The confusion matrix for the final XGBoost model showed excellent class separation: • True Positive Rate (Sensitivity): 97.45%. • False Negative Rate: Only 4 patients incorrectly classified as non-cardiac cases. This high sensitivity is critical in medical applications where missing a positive case (false negative) could have fatal consequences [24]. 7.1.2 SHAP Interpretability Results SHAP established a visualization of feature importance and individual contributions to patient risk: • Global feature importance: o Troponin, CK-MB, Age, and Systolic Blood Pressure. • Interpretations: o High Troponin levels directly raised the prediction of cardiac risk. o Lowering age associated negatively with the prediction of the disease. SHAP force plots behind the scenes allowed for personalized patient-focused explanation analyses, attempting to fill the "black-boxing" component in machine learning for better patient care [25]. 7.1.3 Knowledge-Based Starting Point The rule-based engine with ML integration successfully produced some actionable recommendations: • Urgent cardiology referral for patients with Troponin in excess. • Management plans for hypertension and diabetes for patients with elevated blood pressure and blood sugar. These outputs made predictions into real clinical decisions, thus improving the realworld applicability of the whole system [26]. 7.2 Comparative Analysis for cardiac health • SVM was good at the outset, but then it is limiting in non-linear interaction of the features. • Random Forest and XGBoost were better at accuracy on account of their handling of complex feature interaction and data imbalance. • XGBoost had a faster convergence, smaller bias-variance trade-off, and computational efficiency during inference which made it the preferred choice for deployment in real-time clinical settings [23]. The conclusion summarizes that the findings strongly validate the feasibility and efficacy of machine learning-based prediction of cardiac diseases. 7.3 Model Performance for mental health We assessed whether individuals would go for treatment for mental health using the following performance metrics: • The accuracy: Percent of overall correct predictions • Recall: Also known as Sensitivity, it represents the ability to identify the positive cases correctly (those who sought treatment). • Precision: This tells the correctness of positive predictions. • F1 Score: The harmonic mean of precision and recall. • AUC-ROC: The area under the curve of the Receiver Operating Characteristic, which tells the classification performance across thresholds. Model Accur acy Rec all Precisi on F1Scor e AU CRO C Logistic Regress ion 74.5% 70.2 % 72.4% 71.3 % 0.80 KNN 75.8% 73.5 % 71.1% 72.3 % 0.82 SVM 75.2% 70.8 % 74.1% 72.4 % 0.81 Rando m Forest 79.2% 77.0 % 78.5% 77.7 % 0.86 XGBoo st 78.4% 84.0 % 76.0% 80.0 % 0.88 Modeling-wise, XGBoost performed better than others in recall and F1 score, which is especially important in the health context, where finding all the true positive casesthat is, people who sought treatmentis critical to avoiding underdiagnosis or abandonment altogether. 7.4 Confusion Matrix Analysis for mental health The confusion matrix for the XGBoost model revealed the following: Predicted: No Treatment Predicted: Treatment Actual: No 73 27 Actual: Yes 17 87 • True Positives: 87 — correctly identified as having sought treatment • False Negatives: 17 — at-risk individuals missed by the model • True Negatives: 73 — correctly identified as not having sought treatment • False Positives: 27 — misclassified as treated when they were not The model’s sensitivity (recall) of 84% indicates it effectively identifies those who are likely to seek help, minimizing the risk of neglecting individuals who may need support. In health-related use cases, recall is often prioritized over precision to reduce harmful false negatives [1]. 7.5 SHAP-Based Model Explainability To enhance trust in the AI model and make its decisions interpretable, we employed SHAP (SHapley Additive exPlanations), a popular model-agnostic explainability tool [28]. 7.5.1 Global Interpretability SHAP summary plots showed the most influential features across all predictions: • Top 5 Features: o family_history o work_interfere o care_options o age o benefits For instance, users with a known family history of mental illness and high work interference were strongly predicted to seek treatment. These findings align with mental health literature, where genetic predisposition and workplace stress are major risk factors [29]. 7.6 Using Decision-Making Tools in Future Integrations While this project was directed towards making a sturdy and understandable model, the architecture allows for future additions that include suggestion systems or integration with chatbots; possible applications can be: • Monitors in Workplace Mental Health: Setting flags when predicted probabilities cross. • Personalized Nudges: Programs or counseling recommendation for certain high-risk employees. • Anonymous Self-Assessment: Users empowered to get suggestions on the basis of SHAP explanations. Such tools could serve great things for mental wellness support in organizations with action plans just in time and personalized action plans. [30]. 8. Conclusion Aimed at critical health outcome prediction, we developed and validated two highly effective machine learning pipelines in this integrated study-one for predicting diagnosis of cardiac diseases and the second for treatment-seeking behavior in mental health. For the case of cardiac prediction, clinical features and some other indicators such as troponin, CK-MB blood pressure were used with diverse ensemble models. It was shown to give the best performance when the models were evaluated-the XGBoost model achieving high effectiveness of 98.43% and 97.45% for accuracy and recall respectively-for real-time applications within clinical environments. The mental health model based on survey-derived features such as family history and workplace support also adopted the XGBoost with an average of 78.4% accuracy and 84% recall after comparative analyses. Both models were made better with SHAP based explainability, further ensuring transparency and interpretability at global and individual levels [31][32]. The overall successful entailment of both systems stands as testament to the potential for AI-augmented clinical and organizational decision support systems that predict and also prescribe action. For example the cardiac model attaches rule based recommendations that facilitate divergent clinical interventions while the mental health model opens the way for wellness chatbots, self-assessment tools, and early HR interventions. Combined, these models create a case for how high performing, explainable AI can serve domains both on the physical and mental health spectrums. Future works may devote efforts to integrating both systems into an integrated health monitoring platform verified against outside datasetsthis data would reasoning results made possible by either of the two systems. References [1]Doe, J., & Smith, A. (2016). "Artificial Intelligence in Medicine." Journal of AI Healthcare. [2] White, K., & Green, M. (2021). "Coronary Heart Disease Prediction Using Voting Classifier Ensemble Learning." IEEE Transactions on Medical Computing. [3 ]Singh, G., & Sharma, S. (2023). "Machine Learning and Deep Learning Models for Early Detection of Heart Disease." IEEE Journal of Biomedical Engineering. [4] Srivastava, A., & Samanta, S. (2023). "Medi-Assist: A Decision Tree-based Chronic Diseases Detection Model." International Journal of Machine Learning in Healthcare. [5] Gupta, R., & Mehta, P. (2024). "Heart Disease Detection Using Machine Learning Models." Journal of Healthcare Engineering. [6] Gupta, R., & Mehta, P. (2024). "Improving Heart Disease Prediction Models with Ensemble and Deep Learning Techniques." Medical Informatics Journal. [7] Zhao, Y., & Li, H. (2024). "MultiDisease Prediction and Advanced AI Integration." Artificial Intelligence in Medicine. [8] G. Quer, R. Arnaout, M. Henne, and R. Arnaout, “Machine Learning and the Future of Cardiovascular Care,” JACC, vol. 77, no. 3, pp. 300–313, Jan. 2021. [9] J. Chung and J. Teo, “Mental Health Prediction Using Machine Learning: Taxonomy, Applications, and Challenges,” Appl. Comput. Intell. Soft Comput., vol. 2022, pp. 9970363:1– 9970363:19, Jan. 2022. [10] T. Ghazal et al., “Machine LearningBased Model to Predict Heart Disease in Early Stage Employing Different Feature Selection Techniques,” BioMed Res. Int., vol. 2023, pp. 6864343:1–6864343:15, Jan. 2023. [11] R. Garriga et al., “Machine Learning Model to Predict Mental Health Crises from Electronic Health Records,” Nat Med., vol. 28, no. 6, pp. 1240–1248, Jun. 2022. [12] A. A. Ahmad and H. Polat, "Prediction of Heart Disease Based on Machine Learning Using Jellyfish Optimization Algorithm," Diagnostics, vol. 13, no. 14, Article ID 2392, Jul. 2023. [13] K. M. Prasad et al., "Machine Learning Techniques to Predict Mental Health Diagnoses: A Systematic Review," J. Med. Syst., vol. 48, no. 2, pp. 42-57, Feb. 2024. [14] M. S. Hossain et al., "An Integrated Machine Learning Approach for Congestive Heart Failure Prediction," IEEE J. Biomed. Health Inform., vol. 28, no. 1, pp. 215-227, Jan. 2024. [15] T. J. Reynolds et al., "Improving Cardiovascular Disease Prediction With Machine Learning Using Mental Health Data," Circulation, vol. 149, no. 3, pp. 278-289, Mar. 2024. [16] Cortes, C., & Vapnik, V. (1995). Supportvector networks. Machine Learning, 20(3), 273-297. [17] Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5-32. [18] Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785-794. [19] T. Chen and C. Guestrin, "XGBoost: A Scalable Tree Boosting System," in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 785– 794. [20] S. M. Lundberg and S.-I. Lee, "A unified approach to interpreting model predictions," in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017. [21] Cortes, C., & Vapnik, V. (1995). Supportvector networks. Machine Learning, 20(3), 273-297. [22] Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5-32. [23] Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785-794. [24] Powers, D. M. W. (2011). Evaluation: From Precision, Recall and F-Measure to ROC, Informedness, Markedness & Correlation. Journal of Machine Learning Technologies. [25] Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems. [26] American Heart Association. (2022). "Understanding Blood Pressure Readings." Retrieved from [https://www.heart.org/] [27] T. Chen and C. Guestrin, "XGBoost: A Scalable Tree Boosting System," Proc. of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 785–794. [28] S. M. Lundberg and S.-I. Lee, "A Unified Approach to Interpreting Model Predictions," in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017. [29] M. G. Craske et al., “Rationale and design of a randomized controlled trial comparing coordinated anxiety learning and management for anxiety treatment in primary care,” Contemporary Clinical Trials, vol. 30, no. 2, pp. 130–145, 2009. [30] D. Mohr et al., "Barriers to Psychological Treatment and Potential Technological Solutions," Cognitive Behaviour Therapy, vol. 39, no. 1, pp. 1– 12, 2010. [31] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System,” Proc. ACM SIGKDD, 2016. [32] S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” NeurIPS, 2017. [33] D. Mohr et al., “Barriers to Psychological Treatment and Potential Technological Solutions,” Cognitive Behaviour Therapy, vol. 39, no. 1, pp. 1– 12, 2010.