New Framework for Per-Modality Failure Analysis in Multimodal Clinical AI

A new framework that is model-agnostic has been developed by researchers to analyze failures in multimodal clinical AI systems when certain modalities are absent during deployment. This framework, outlined in a paper on arXiv (2608.01462), explores two significant issues beyond mere accuracy loss: identifying the modality causing the failure and determining whether the failure is noticeable or subtle when that modality is missing. Unlike post-hoc feature attribution methods such as SHAP, this distinction is made at both the example and modality levels. Designed for reusability across various models, the framework processes N modality embeddings, a mask-aware probe, and labels to generate a failure taxonomy, a complementarity matrix for errors, and a dropout profile distinguishing between detectable and undetectable failures. This research is particularly pertinent in clinical contexts where modalities, such as echocardiograms, might not be accessible, aiming to enhance the reliability and interpretability of AI in healthcare.