An explainable machine learning-based approach for medical diagnosis : interpretability, fairness, and model transparency

Abstract

Machine learning (ML) models have reshaped medical diagnostics, significantly improving accuracy and efficiency. Nonetheless, the intrinsic ”black-box” nature of these models often undermines transparency and trust, constraining their widespread clinical usage . This paper proposes a comprehensive framework to enhance the interpretability, fairness, and causal comprehension of AI-driven diagnosis. We suggest utilising common Explainable AI (XAI) methodologies, notably Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), to elucidate the decision-making processes of machine learning categorisation models. This study presents a Generative AI layer using a locally deployed small language model to translate technical explanations into clinically meaningful insights. This efficient, privacy-preserving system analyses quantitative XAI outputs to generate intuitive, natural-language clinical narratives, offering physicians contextually relevant reasoning for risk assessments. Additionally, we will examine the implementation of causal inference approaches to go beyond mere correlations, aiming to uncover the genuine causal relationships underlying model decisions and enhance the reliability of AI-assisted diagnostics. This framework is designed to increase confidence and accountability of AI for physicians and patients . By combining rigorous statistical interpretability with Language-model-driven narrative generation, this work seeks to promote greater trust and increases the probability of ethical adoption of AI in healthcare.