Design and Evaluation of a Hybrid Intelligence Framework for Explainable Diabetes Risk Prediction and Clinical Decision Support

Abstract

In this study, we propose a hybrid intelligence system for diabetes-risk prediction and clinical decision support, which remains favorable from the perspective of explainability.In this study, a hybrid intelligence approach for explaining diabetes risk prediction and clinical decision support is presented, which remains favorably by the perspective of explainability. The framework tackles the challenges of the traditional healthcare predictions systems that might give risk estimations without providing a clear explanation, evidence-based knowledge, and interactivity decision support. The architecture proposed suggests that the system includes components for data preprocessing, predictive machine learning, software agents, retrieval augmented generation (RAG), large language model (LLM) for medical knowledge retrieval and explainability using SHAP.The proposed architecture includes components such as data preprocessing, predictive machine learning, multiple software agents, retrieval augmented generation (RAG), large language model (LLM) for medical knowledge retrieval, and explainability using SHAP, along with an interactive Streamlit dashboard. The experiments were performed on a high number of diabetes records (100,000+) where the variables comprised a set of demographic, lifestyle, and clinical variables, and the goal was that diabetes should be the prediction target. All the models were trained and evaluated using accuracy, precision, recall, F1 score, ROC-AUC, PR-AUC, Matthews correlation coefficient,Brier score and confusion-matrix analysis. A hybrid ensemble of the predictive probabilities of the seven models was created next based on the validation process. With a validation chosen decision threshold of 0.79, the proposed hybrid model achieved 97.22% accuracy, 81.34% F1 score, 97.82% ROC-AUC and 88.58% PR-AUC on an independent test set having 15,000 records. The proposed architecture for decision support may include the integration of SHAP-based explanations for their transparent global and patient-level interpretation of model outputs [25] and responsible (and trustworthy) principles in the context of healthcare AI [23, 34]. The framework shows how these four techniques together (predictive machine learning, explainability, knowledge-grounded generation, and agent-based interaction) can be combined to provide transparent support for the decision on using diabetes-risk. But it is still in development, so it needs to be validated outside of clinical settings, calibrated, evaluated as a proto-type, validated as a fair device, and monitored in clinical usage before being deployed in real life. [35] [36] [38]