Enhancing the Reliability of Medical AI through Expert-guided Uncertainty Modeling

Abstract: AI systems accelerate medical workflows and improve diagnostic accuracy in healthcare, serving as second-opinion systems. However, the unpredictability of AI errors poses a significant challenge, particularly in healthcare contexts, where mistakes can have severe consequences. A widely adopted safeguard is to pair predictions with uncertainty estimation, enabling human experts to focus on high-risk cases while streamlining routine verification. Current uncertainty estimation methods, however, remain limited, particularly in quantifying aleatoric uncertainty, which arises from data ambiguity and noise. To address this, we propose a novel approach that leverages disagreement in expert responses to generate targets for training machine learning models. These targets are used in conjunction with standard data labels to estimate two components of uncertainty separately, as given by the law of total variance, via a two-ensemble approach, as well as its lightweight variant. We validate our method on binary image classification, binary and multi-class image segmentation, and multiple-choice question answering. Our experiments demonstrate that incorporating expert knowledge can enhance uncertainty estimation quality by $9\%$ to $50\%$ depending on the task, making this source of information invaluable for the construction of risk-aware AI systems in healthcare applications.

• The paper introduces an expert-guided uncertainty framework that improves reliability in medical AI tasks, achieving high accuracy and robust calibration.
• It leverages multi-expert soft labeling, probabilistic aggregation, and ensemble methods to quantify both epistemic and aleatoric uncertainties.
• The approach enhances risk stratification and human-AI interplay, supporting more accountable and transparent clinical decision-making.

Expert-guided Uncertainty Modeling for Enhanced Reliability in Medical AI

Methodological Overview

This paper systematically addresses the integration of expert-guided uncertainty modeling into medical AI systems to increase their reliability across diverse image and NLP tasks. The core methodology leverages datasets comprising multiple expert annotations to generate "soft" labels for confidence-aware model training. Recognizing the nuanced uncertainty embedded in expert judgments, the approach incorporates probabilistic aggregation rather than relying solely on traditional hard labels.

Four distinct datasets—BloodyWell, RIGA, LIDC-IDRI, and PubMedQA—form the experimental backbone. Each is characterized by rich expert annotation, facilitating the quantification and modeling of epistemic and aleatoric uncertainties. Ensembles of models (size 10) are constructed using cross-validation and random initialization to ensure diversity. For segmentation tasks, soft labels result from expert prediction averaging; for classification, additional contextual variables (e.g., reagent type for BloodyWell) are encoded alongside images.

The technical arsenal spans state-of-the-art architectures such as MobileNet-V3-small for classification and UNet with batch normalization for segmentation, together with test-time augmentation, MC Dropout, MCMC (via SGLD), Hybrid Uncertainty Quantification (HUQ), and deep ensembles. ModernBERT serves as the backbone for PubMedQA. Calibration and uncertainty quantification methods are carefully matched to the task structure, with advanced probabilistic modeling adopted for multi-class aggregation.

Experimental Results

The models trained under the expert-guided uncertainty framework exhibit superior reliability metrics. Notable outcomes include:

• Image classification accuracy: $96.71\% \pm 0.58\%$
• Multiple-choice QA accuracy: $71.18\% \pm 0.85\%$
• Dice score (binary segmentation): $0.6348 \pm 0.0229$
• Dice score (multiclass segmentation): $0.8638 \pm 0.0081$

Critical to these results is the observation that ensemble-based and soft-label models achieve not only high accuracy but also improved robustness in uncertainty estimation compared to state-of-the-art baselines. The proposed uncertainty decomposition—epistemic from model variance, aleatoric from expert disagreement and inherent data ambiguity—is validated both theoretically and empirically.

A particularly strong claim is the natural emergence of the proposed uncertainty aggregation from the categorical probabilistic structure, justifying the summation of uncertainties across classes as theoretically sound. No gains were found from spatially varying label smoothing compared to simple averaging for segmentation uncertainty, emphasizing the practical efficacy of the method.

Theoretical Implications

The probabilistic formalization provided in the supplementary section clarifies how aggregate uncertainty in multi-class settings decomposes via trace operations into class-wise epistemic and aleatoric components. This approach aligns with contemporary research on uncertainty quantification, offering a principled pathway for scalar aggregation of variance matrices in categorical output spaces. The formalism ensures that total uncertainty metrics (TU, EU, AU) can be meaningfully interpreted and utilized in risk-aware clinical decision-making.

The use of expert-generated soft labels is substantiated by recent literature as improving both calibration and uncertainty estimation [de Vries & Thierens, 2025]. The method demonstrates compatibility with Bayesian and non-Bayesian techniques (e.g., ensembles, dropout, SGLD), thereby accommodating diverse practical constraints in medical AI pipelines.

Practical Significance and Future Directions

The paper emphasizes operational reliability by integrating uncertainty quantification as a first-class concern in medical AI, directly leveraging expert knowledge. This expert-in-the-loop structure has implications for:

• Risk stratification: Improved uncertainty estimates enable dynamic rejection options and selective prediction, mitigating clinical risks.
• Model calibration and auditability: Enhanced calibration empowers deployment in settings requiring high accountability.
• Human-AI interplay: The framework supports more informed interaction between medical practitioners and AI systems, especially in ambiguous cases.

Looking forward, the methodology is extensible to additional domains where uncertainty and expert consensus are prominent, such as radiology, pathology, and interactive diagnostics. The theoretical foundation facilitates integration with explainable uncertainty methods [molchanova2025explainability], human-centered annotation, and adaptive calibration strategies.

Conclusion

The expert-guided uncertainty modeling paradigm described in this paper introduces a robust mechanism for enhancing the reliability of medical AI. By harnessing multi-expert soft labeling and rigorous probabilistic uncertainty quantification, the approach achieves strong numerical performance, improves risk-awareness, and is theoretically justified. The implications for clinical deployment, auditability, and human-AI collaboration are significant. This paper sets the stage for further advances in trustworthy medical machine learning, particularly through principled integration of expert knowledge and uncertainty analytics (2604.01898).