A Machine Learning Framework for Explainable Clinical Diagnosis Using Health Records

The integration of large datasets in healthcare, particularly through Electronic Health Records (EHR), has enabled critical insights into patient care and treatment outcomes. Our research explores the development of a machine learning (ML)-based Clinical Decision Support System (CDSS) for the anamnesis process, focusing on explainable AI (XAI) to enhance diagnostic transparency and practitioner trust. This paper describes a ML model that classifies diagnoses using structured anamnesis data. The development of this ML model is motivated by the availability of over a thousand structured reports, filled by a diverse group of practitioners, which provide a reliable dataset for training a classification model to predict patient diagnoses. By offering consistent diagnostic insights, the primary goal of this ML model is to support reducing consultation times. Secondary goals include standardizing care and minimizing misdiagnoses. We analyzed a subset of the structured reports, covering anamnesis, examination, and treatments, to prepare the input EHR data for the ML pipeline, narrowing 299 variables to 51 and focusing on 20 kinds of diagnoses. Three models were trained: two single-label classifiers and a multi-label classifier (MLC). Practitioners evaluated XAI outputs, which were presented in a way to optimize their understanding without specific data science skills, leading to a final format that highlights key diagnostic variables, irrelevant factors, and alternative options. This work provides a comprehensive review of XAI in healthcare and demonstrates its application in creating an XAI-verified diagnostic model, contributing to the integration of ML in clinical settings.