Explainable ICD coding via MediThoth: A multimodal clinical decision support system

This study presents MediThoth, a multimodal and explainable framework for automated ICD coding that integrates unstructured clinical notes with structured EHR data from MIMIC-III (46,520 patients; 58,976 admissions) and MIMIC-IV (299,712 patients; 431,231 admissions). Following independent preprocessing of textual and structured modalities, the system evaluates multiple candidate models and finds that GatorTron is the best text encoder, achieving improvements of 11.03% (MIMIC-III) and 3.5% (MIMIC-IV) in Micro-F1 over prior transformer baselines. For structured data, XGBoost outperforms traditional models, achieving gains of 3.6% and 10.2% on MIMIC-III and MIMIC-IV, respectively. MediThoth combines these components through late fusion, yielding a Micro-F1 score of 0.8180 and a Macro-F1 score of 0.7813 on MIMIC-III, representing 8.9–12.2% relative improvement over multimodal baselines, with consistent performance in MIMIC-IV. To enhance interpretability, MediThoth incorporates a dual-layer explanation module that integrates SHAP-based feature attributions with LLM-generated narratives. The results demonstrate that MediThoth delivers both high predictive accuracy and clinically meaningful explanations, establishing a robust and transparent multimodal ICD coding system.