MERIT: A Clinical Decision Support System Framework for Merging Multi-Perspective Machine Learning Model Insights

Abstract

The data surge and rapid progress in artificial intelligence (AI) performance require fundamental adjustments in analytics, decision-making, and innovation in healthcare technology. One of the major hurdles in modern healthcare is combining these rapidly advancing AI capabilities with clinicians' practical knowledge and experience. This work focuses on eXplainable AI (XAI) to bridge the gap between AI systems' capabilities and end-users' clarity in critical decision-making. Informed by a formative survey of stakeholders to understand the real user demands, a hypothetical machine learning (ML) expert-developer workflow for building model and dataset-agnostic XAI-based Clinical Decision Support Systems (CDSS) is proposed in this work to concentrate on building tools and centralized platforms that remain relevant and usable for a longer time, providing XAI solutions aiding decision-making. An example CDSS, MERIT (Multi-perspective Explanation for Real-time Integrated Treatment), is built following this system, which utilizes the SHapley Additive exPlanations (SHAP), and is based on Django architecture, to cover the full range of user expectations identified from the survey - holistic, contextually grounded, user-adaptable interactive explanations. This web application attempts to explain models' output through global and local model behavior explanation visualizations and additional distribution plots, and it incorporates interactive features for enhanced user exploration. Personalized textual explanations are also enabled through Retrieval Augmented Generation (RAG)-based chat functionality available on the website. The application was experimentally evaluated with the Wisconsin Diagnostic Breast Cancer (WDBC) dataset and a model trained on it. This study aims to support informed clinical decision-making and foster trust in AI-driven care by aligning design decisions with real-world requirements and guiding the best practices in XAI communication.

Ähnliche Arbeiten