Debugging Multimodal Clinical AI Failures

Summary & Methodology Analysis

The researchers developed a model-agnostic harness designed to isolate the performance impact of individual input modalities. By taking N modality embeddings, their associated ground truth labels, and a mask-aware prediction function, the framework systemically evaluates model reliability. The harness employs a failure taxonomy that classifies outputs into correct, imprecise, or critical categories based on regression error metrics and predefined clinical gate thresholds, allowing engineers to visualize exactly where performance degrades during deployment scenarios involving missing data. The framework also facilitates a fairness audit by stratifying these failure profiles across demographic attributes. To differentiate between loud and silent failure modes, the system analyzes the margin of predictions relative to the decision boundary (the hyperplane separating different classes in vector space) as modalities are masked, providing an empirical view of how model confidence shifts when sensor data is unavailable. Evaluation against clinical data, specifically using models like EchoJEPA and HuBERT-ECG on the MIMIC-IV dataset, highlights the framework's utility in real-world integration tests. For instance, in a cohort of 245 samples, dropping the echo modality resulted in a substantial increase in LVEF MAE from 10.28 to 18.57, illustrating the high dependency of clinical models on specific input streams. Despite these insights, the study faces limitations regarding external validity, as it relies on data from a single US academic center. Furthermore, the current pipeline processes only one DICOM clip per study, which may constrain the overall absolute accuracy of the models being audited.