Quantifying LLM Over-Trust in Derived Physiological Measurements

Researchers define Derived-Feature Over-Trust (DFOT) to quantify how LLMs incorrectly assign epistemic status to derived measurements like PPG-based rhythms. Using ECG as a ground truth reference, the study identifies specific failure modes in how models handle conflicting or contextually misleading sensor data.