We propose MedBayes-Lite, a lightweight Bayesian enhancement for transformer-based clinical language models designed to produce reliable, uncertainty-aware predictions. Although transformers show strong potential for clinical decision support, they remain prone to overconfidence, especially in ambiguous medical cases where calibrated uncertainty is critical. MedBayes-Lite embeds uncertainty quantification directly into existing transformer pipelines without any retraining or architectural rewiring, adding no new trainable layers and keeping parameter overhead under 3 percent. The framework integrates three components: (i) Bayesian Embedding Calibration using Monte Carlo dropout for epistemic uncertainty, (ii) Uncertainty-Weighted Attention that marginalizes over token reliability, and (iii) Confidence-Guided Decision Shaping inspired by clinical risk minimization. Across biomedical QA and clinical prediction benchmarks (MedQA, PubMedQA, MIMIC-III), MedBayes-Lite consistently improves calibration and trustworthiness, reducing overconfidence by 32 to 48 percent. In simulated clinical settings, it can prevent up to 41 percent of diagnostic errors by flagging uncertain predictions for human review. These results demonstrate its effectiveness in enabling reliable uncertainty propagation and improving interpretability in medical AI systems.
Transformer-based Large Language Models (LLMs) are rapidly reshaping clinical decision support, offering unprecedented capabilities in medical reasoning, synthesis of biomedical knowledge, and zero-shot diagnostic inference [1][2][3][4][5]. Yet despite their promise, a fundamental obstacle continues to block their safe deployment: they are often most confident when they should be uncertain. In highstakes settings-triage, differential diagnosis, medication dosing-even a single confidently incorrect recommendation can cascade into serious harm [6,7]. Recent studies highlight LLMs asserting definitive answers for rare diseases, misinterpreting contradictory symptoms, and generating unsafe medication guidance [8,9]. These are precisely the situations where clinicians rely more on calibrated uncertainty than on raw prediction accuracy.
Clinical decision-making is inherently probabilistic. Physicians routinely navigate incomplete data, ambiguous presentations, and conflicting evidence, using uncertainty as a signal to slow down, order additional tests, or seek expert consultation [10,11]. This ability to recognize “I might be wrong” is not merely desirable-it is essential for preventing diagnostic errors and mitigating automation bias [12,13]. Current LLMs, however, lack mechanisms for propagating uncertainty through their reasoning processes. As a result, they produce overconfident predictions even when internal token-level evidence is weak or contradictory.
The drive toward developing trustworthy clinical AI has highlighted the need for effective UQ. Accuracy alone is often insufficient in healthcare settings, where models must also signal when their predictions may be unreliable. The landscape of UQ methods for LLMs is broad, spanning foundational Bayesian techniques to clinical risk-based evaluation frameworks. However, the diversity of these approaches makes it challenging to determine which are most suitable for real-world deployment. This section synthesizes these methods to clarify how our work fits within and contributes to this space.
The quest to develop deep learning models that can represent their own uncertainty gained momentum with the work of Gal and Ghahramani [16]. They demonstrated that dropout training can be interpreted as a variational approximation to Bayesian neural networks. This MC dropout technique enables uncertainty estimation without substantial computational cost, and it has since been widely adopted in transformerbased models. However, the method has several limitations: it provides uncertainty only at the output level, lacks contextual or risk-aware calibration, and depends on repeated sampling, which reduces both efficiency and scalability.
Many studies have since applied MC dropout to models such as BERT to measure epistemic uncertainty in Natural Language Processing (NLP) tasks [17,18]. Subsequent work has introduced evidential and beliefbased approaches to improve UQ, yet these methods still tend to treat uncertainty as a post-hoc signal rather than as an integral part of the model’s reasoning process. For example, class-balanced evidential deep learning primarily addresses class imbalance but does not improve calibration or trustworthiness in real-world reasoning systems. Likewise, belief-matching frameworks based on Dirichlet priors can enhance prediction calibration, but they remain limited to output-level uncertainty and rely on fixed priors that cannot adaptively capture both epistemic and aleatoric variance.
Although some researchers have explored how uncertainty may influence attention mechanisms [19], current approaches still fall short of achieving a full Bayesian treatment that propagates uncertainty from embeddings through attention layers and into final decision layers. Developing a comprehensive yet lightweight framework that supports this type of end-to-end uncertainty propagation remains an open challenge for practical and clinical AI systems.
A commonly used strategy for improving confidence reliability is post-hoc calibration. Methods such as Temperature Scaling (TS) [14] and Isotonic Regression (IR) [20] adjust a model’s output probabilities after training to better align predicted confidence with empirical accuracy. These approaches can significantly reduce metrics such as the Expected Calibration Error (ECE) on static validation sets. However, their effectiveness is largely superficial because post-hoc calibration modifies only the final probability layer while leaving the underlying model representations unchanged. As a result, the calibrated probabilities often fail to remain reliable under distribution shift, prompt variation, or changes in input structure. This exposes the inherent limitations of treating calibration as an output-level correction rather than as a modeling-level capability.
Building on these limitations, recent studies including Ye et al. [15] and Yang et al. [21] show that such post-hoc adjustments do not influence the model’s in
This content is AI-processed based on open access ArXiv data.