Explainable Artificial Intelligence for Personalized Diabetes Risk Prediction

Abstract

Diabetes keeps affecting millions of people around the world, with the current estimation of 537 million and is expected to rise steadily in the coming decades. So, detecting it at an early stage is very crucial. Recognizing the risk factors at an early stage would help prevent further health complications in the future. Even though machine learning provides algorithms that can predict health risks, offering tools which can analyze complex patterns of medical data effectively than traditional methods. Many of these models lack transparency, so it makes it hard to understand and limits their trust, especially in health care. Such difficulties have led to the development of XAI, which can assist in understanding why those predictive models have arrived at the decision. To clarify these issues the well-known tools SHAP and LIME are outstanding in revealing outcomes. SHAP uses a game theory principle to show how each features assign influences providing both global and local insight, while LIME provides simple easy to follow rules for each individual prediction. Synergistically they offer a surplus of details and simplicity. Due to the balance of performance and simplicity random forest and Logistic Regression predictive models have been commonly applied to diabetes risk classification. With appropriate techniques of data pre-processing such as feature scaling, handling missing values, stratified data sampling both models can pro-vide reliable estimation or predictions using the public dataset widely available medical inputs or metrics such as age, BMI, glucose, blood pressure then display using XAI making them transparent and user friendly. Web based interface designed with usability can serve as effective platform to understand those predictive models with visual aids, educational tool tips for those non-technical users and clearly labelled output enhance comprehension and accessibility. Responses from technical and non-technical users show that combining visual and rule-based explanations facilitates greater understanding and builds confidence on the trust of the system predictions. Ease of interaction along with clear presentation significantly contributes to the user’s trust and willingness to engage with those predictive tools. There has been great emphasis on those predictive models that provide not only optimize accuracy but also offer clarity, ethical disclaimer and meaningful Interactions. In sensitive fields such as health care appliances where lives are affected by decisions, the combination of performance and interpretability is not optional but essential. With continued advancement in model explanation techniques as well as interface design including evaluation by professional medical staff and performing clinical trial will be key to shaping trustworthy AI-driven decision support tools in the future.