A SHAP-BASED MULTI BOOSTING FRAMEWORK FOR INTERPRETABLE DIABETES PREDICTION WITH FEATURE SELECTION

Abstract

The timely and accurate prediction of the disease is essential to avoid the progress of the disease and other severe complications and to enable early interventions. Machine learning (ML) is a powerful approach to support early diagnosis of the disease through maximising the accuracy of the diagnosis. In this study, the authors have proposed a multiple boosting model to aid in the diagnosis of the disease by combining AdaBoost, Gradient Boosting, XGBoost algorithms with SHAP (SHapley Additive exPlanations) feature analysis and feature selection methods. The results of the experiments conducted on the PIMA data set show that XGBoost is the best performing algorithm among the other boosting algorithms with 77.92% accuracy and 83.45% ROC AUC. The features Glucose, BMI, and Age are the top features with the highest contributions to the model output as per the results of the SHAP analysis, with Glucose accounting for 50% of the contribution to the output. Moreover, the subset of features with the top features identified by SHAP still retains 98-99% of the prediction performance with significantly lower computational complexity. Despite the challenges of class imbalance problems, the proposed approach has shown the potential to be applied to real-world problems with decision support systems. Future research will be directed towards improving the minority class detection performance, as well as the inclusion of longitudinal data.