Attention-Based MIL for Medical Malpractice Prediction

Abstract

Medical malpractice prediction is challenging due to the weakly labeled, heterogeneous, and multi-instance structure of claims data. We introduce Deep Attention MIL (DAMIL), an attention-based Multiple Instance Learning model that learns to identify the most informative instances within each claim. By optimizing attention weights end-to-end, DAMIL improves both prediction and interpretability.

We evaluate DAMIL on two datasets: (1) a synthetic benchmark with controlled risk patterns, and (2) a real-world dataset from the Col·legi de Metges de Barcelona. DAMIL outperforms traditional MIL and a Bag-of-Words baseline, reaching AUCs of 0.715 (synthetic) and 0.714 (real). Instance-level attention provides interpretable insights into risk-relevant claim components.