Search papers, labs, and topics across Lattice.
MIT CSAILA*STARAmerican Heritage SchoolHarvardHopewell Valley Central High SchoolInstitute of Computing ScienceJHUNeuqua Valley High SchoolNorth Hollywood High SchoolSiriraj Informatics and Data Innovation CenterSNUUNIGEUniversity of North FloridaAug 2, 2026arXiv:2608.01462This paper introduces a model-agnostic framework for analyzing modality failures in multimodal clinical AI, focusing on identifying which modality contributes to errors and whether these failures are detectable ("loud") or not ("silent"). By employing a mask-aware probing approach, the framework generates a per-example failure taxonomy and a complementarity matrix that attributes errors to specific modalities, validated against known ground truths. In practical application, the framework reveals that dropping echocardiogram data nearly doubles error rates in a cardiac model, highlighting significant modality dependencies in clinical settings.
Dropping echocardiogram data nearly doubles error rates in cardiac AI models, revealing critical modality dependencies that could impact clinical decisions.
Multimodal clinical models are usually judged on accuracy with every modality present, but deployment removes modalities; an echocardiogram is often unavailable where an ECG is routine. Two questions then matter beyond the size of the accuracy loss: which modality was responsible, and whether the model fails loudly or silently once that modality is dropped. The distinction is per-example and modality-level, and is separate from post-hoc feature attribution (e.g. SHAP). Models are replaced often; the evaluation that answers these questions is reused. We present a model-agnostic modality-failure framework: given N modality embeddings, any mask-aware probe, and labels, it returns a per-example failure taxonomy, a per-modality complementarity matrix that attributes error to modalities, and a loud-vs-silent dropout profile separating monitorable failures from those that pass unflagged far from the decision boundary, using only deployment-observable signals. We release it as a small, unit-tested harness and validate it against planted ground truth. Across seeds it recovers that planted modality dominance and complementary subset, reports per-modality loud-vs-silent rates, and scales to a three-modality complementarity matrix; because the planted structure is known by construction, this validates recovery of per-example attribution rather than clinical performance. We then instantiate the framework on frozen EchoJEPA and HuBERT-ECG embeddings for LVEF and the EF<= 40% HFrEF gate over a paired MIMIC-IV cohort, where on the held-out test split (n = 245) dropping echo nearly doubles error. The narrow echo-to-ECG overlap that bounds cohort size is itself a deployment finding for cardiac foundation models. All of our work can be found at https://github.com/criticaldata/PRIMED-AI.
