Explainable AI-Based Clinical Decision Support System for Early Prediction of Pathological Findings in Forensic Practice

Abstract

Background: The number of cases, subjectivity of the examiner and difficulty of integrating the different modes of evidence are all increasing challenges in forensic pathology. Whilst identification of the cause of death is strongly provided for, and characterization of trauma is a critical unmet need, neither can be achieved early and accurately. Methods: We propose an Explainable Artificial Intelligence Clinical Decision Support System (XAI-CDSS) combining a hybrid ensemble of Random Forest, XGBoost and Long Short-Term Memory (LSTM) networks, three complementary explainability modules: SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and Gradient-weighted Class Activation Mapping (Grad-CAM). Model training and evaluation were carried out on a subset of a multimodal forensic database 4,872 post-mortem cases collected by five different regional forensic centres which includes clinical data, toxicological data, CT imaging indices and histopathological grades. The class imbalance was addressed by using Bayesian hyperparameter optimisation and Synthetic Minority Over-sampling Technique (SMOTE). Results: The proposed XAI-CDSS outperformed other benchmark models achieved an overall accuracy of 96.2%, AUC-ROC of 0.974, sensitivity of 95.8% and specificity of 96.7% as compared to stand-alone XGBoost (94.1%), Random Forest (93.2%), SVM (90.8%) and Logistic Regression (86.2%). Based on SHAP analysis, the top 3 influential predictors were ICD-10 cause of death, toxicological findings and PM interval. Conclusion: The XAI-CDSS is a decision support system that can be used by forensic practitioners and is clinically interpretable and accurate. This automatic system, which relies on state-of-the-art machine learning and provides human-readable explanations, is a middle ground between algorithmic prediction and evidentiary law requirements..