AI-Driven Classification of Forensic Autopsy Reports: Balancing Accuracy and Interpretability

Understanding the epidemiology of forensic deaths relies heavily on analyzing autopsy reports, which aim to identify the primary cause of death (CoD). Recent advances in text mining and artificial intelligence now allow for automatic prediction of CoD from these records. This study applied text classification techniques, using both traditional and deep learning algorithms, to predict the mode of death (MoD) from free-text forensic autopsy reports in Tlemcen Hospital - Ouest Algeria region. Our main objective was to analyze and structure the unstructured data in these reports to facilitate medical professionals assessments. The proposed framework includes two key components: first, the collection and preprocessing of 200 autopsy reports using natural language processing (NLP) methods; and second, the classification of these reports using machine learning and deep learning algorithms with TF-IDF as a feature extraction method. Experiment results indicated that the XGB classifier achieved the highest accuracy (95 %). To ensure interpretability of the model predictions, LIME (Local Interpretable Model-agnostic Explanations) was applied, providing transparent insights into feature contributions. This approach balances predictive performance with interpretability, making it a valuable tool for medical and forensic decision-making.