From data to diagnosis: Heart disease likelihood prediction using six integrated datasets and rigorous evaluation of some machine learning models for Android app deployment

Heart diseases are one of the primary causes of millions of deaths every year, many of which could be prevented through early detection. Hence, this study aims to develop an Android application capable of identifying high-risk patients using a machine learning model. The model was developed in favor of multiple datasets. Six secondary datasets were combined following some inclusion and exclusion criteria. The merged dataset underwent rigorous preprocessing, including handling missing values, removing duplicates, and excluding scientifically implausible entries. Histograms, box plots, and stacked columns were employed to visualize the data. The cleaned dataset was then divided into 10 folds by stratified k-folds, followed by the SMOTE-NC technique only for the training set, while the test set remained untouched. Thence, six machine learning models — CatBoost, LightGBM, Random Forest, XGBoost, Decision Tree, and AdaBoost — were trained and evaluated. Based on the Wilcoxon Signed-Rank test, LightGBM was initially selected among them. Later, albeit the Recursive Feature Elimination method was employed for about 20% feature elimination, it was not able to perform better than the original 11 features based LightGBM. As a result, the original 11-feature LightGBM model was finally selected. This final selected model was optimized for hyperparameters, followed by diverse performance evaluation under 10 folds — confusion matrices, overall accuracy, F1-score, learning curve, AUC value, and feature importance. It was found that LightGBM had 91.40%, 91.10%, and 96.00% of overall mean accuracy, mean F1 score, and mean AUC value, respectively. learning curve, confusion matrix, and SHAP feature importance plot were also evaluated. Consequently, the LightGBM was integrated into the Android application based on the complete dataset, resulting in an accessible diagnostic tool for early heart disease risk prediction. This application has the potential to support timely interventions and enable more efficient allocation of healthcare resources.