Hidden-uncertainty Assessment via Non-verbalized Signatures for Medical QA

Large Language Models (LLMs) have shown growing potential in medical applications, yet reliable uncertainty estimation remains a key safety bottleneck in medical QA. Reasoning-capable models can still produce incorrect answers with high confidence, both when confidence is self-reported and when inferred from answer-probability proxies (answer log-probabilities / option-softmax). While self-consistency can improve reliability, it requires multiple forward passes and is costly at inference time. We propose a lightweight Trust-Abstain supervised classifier that predicts error risk from a single reasoning trace, shifting attention from answer tokens to the reasoning process by combining trace-level log-probability dynamics, textual markers, and medical-domain signals. We evaluate five open-weight reasoning LLMs, including GPT-oss-120B and DeepSeek-R1-Distill-32B, across ten medical benchmarks (N = 21,324) under a Leave-One-Dataset-Out protocol. Our approach improves ROC-AUC by 2.4-6.0 points over strong single-pass baselines, including self-verbalized confidence and answer-probability proxies. In selective answering regimes (10-40% coverage), it reduces absolute error rate by up to 9.7 percentage points and consistently increases Accuracy-Coverage AUC. A simple logistic regression matches non-linear alternatives. Finally, we identify 12 generalizable features that remain stable across model families. Freezing the feature space to these predictors incurs a small degradation (< 1%) relative to model-specific selection, supporting largely model-agnostic uncertainty signals. Code, trained classifiers, enriched datasets, and an interactive demo are released to support reproducibility and further research.