A warning signal with a map
A new preprint describes a medical-AI method that tries to answer two questions at once: how uncertain is a prediction, and which features contribute to that uncertainty. Called egRUE, the method was reported as a top or near-top performer on several uncertainty tests across four medical diagnosis datasets, including tests of how well it identified shifted, unfamiliar images.
The paper frames the contribution as a way to make uncertainty more targeted and interpretable. Its explanations are designed to identify the input features that contribute to a given uncertainty estimate.
Strong results on benchmark tests
Compared with Deep Ensembles, the paper reports relative gains of 151% to 330% in reliability, 4% to 53% in detecting misclassifications, 32% to 77% in selective prediction, 78% to 98% in robustness to false negatives, and 11% to 97% in out-of-distribution detection. These are relative benchmark comparisons, and the supplied analysis reports no confidence intervals or hypothesis-test results for the ranges.
The evaluation covered the Singapore Chinese Health Study lung-cancer and colorectal-cancer datasets, plus OCTMNIST retinal images and BloodMNIST blood-cell images. After preprocessing, the cancer datasets contained 17,464 and 20,689 instances, while OCTMNIST contained 109,309 images and BloodMNIST 17,092. Researchers measured whether uncertainty tracked loss, found wrong predictions, supported selective prediction, signaled false-negative risk, detected shifted data, and yielded informative explanations.
On the cross-dataset comparisons, egRUE was reported to have the highest or second-highest reliability correlation, the highest misclassification-detection AUROC, which tests how well the uncertainty score separates wrong predictions, low AURC on the lung-cancer, OCTMNIST and BloodMNIST tasks, and the lowest or second-lowest σ-risk, the paper's measure for confident incorrect predictions. It also had the highest OOD AUROC on both image datasets.
For the image-shift tests, the reported egRUE OOD AUROCs ranged from 0.946 to 1.000. Against OCTMNIST, the scores were 0.961 for OCTDL-IN, 0.977 for OCTDL-OUT and 1.000 for ChestMNIST, with standard deviations of 0.015, 0.007 and 0.000. Against BloodMNIST, the corresponding scores were 0.997 for Raabin-IN, 0.946 for BoneMarrow-OUT and 0.998 for ChestMNIST, with standard deviations of 0.001, 0.021 and 0.001.
From a score to an explanation
egRUE adapts Expected Gradients to RUE, the underlying reconstruction-based uncertainty method. It uses the RUE encoder and predictor, baseline instances drawn from the training distribution, and the input itself as the target. The resulting feature-importance weights are applied to feature-wise reconstruction errors, producing an uncertainty estimate that gives more weight to features judged more important for that prediction.
In illustrative mispredicted images, the authors report that egRUE focused more specifically than EG or RUE on clinically relevant edema and atypical nucleus features. In the ROAR evaluation, researchers removed uncertain features and retrained the model, then measured performance with area under the curve. The paper says egRUE's ROAR results were generally more stable and higher.
A related ablation compared egRUE with other attribution methods layered onto RUE. The authors state that egRUE outperformed RUE overall; on BloodMNIST, its ROAD AUC was 0.148, compared with 0.144 for igRUE, 0.133 for gradCamRUE and 0.138 for guidedBackpropRUE. The supplied analysis notes a correlation exception in the broader BloodMNIST comparison, so the overall claim should not be read as every metric favoring egRUE.
Experts changed confidence
The human test was narrow: five medical experts answered nine BloodMNIST questions spanning correct predictions with low uncertainty, incorrect predictions with high uncertainty, and incorrect predictions with low uncertainty. When experts saw uncertainty scores alone, they agreed with incorrect predictions in 40% of responses, or 12 of 30.
Adding the explanations was linked to higher confidence in correct predictions and lower confidence in incorrect predictions in the first two categories. The incorrect, low-uncertainty category was mixed, and the supplied analysis says category-level effect sizes were not reported. The study therefore reports a possible shift in trust calibration, rather than evidence about clinical decisions.
The boundary of the evidence
The paper also makes theoretical claims. It states that positive constants derived from the prediction model can bound the total absolute prediction error above and below when uncertainty is measured with RUE or egRUE. It further reports that egRUE explanations satisfy implementation invariance, sensitivity and consistency. The constants are existential rather than reported values, and these results depend on the model assumptions described in the analysis.
On speed, egRUE took 0.00200 seconds per instance, compared with 0.01026 seconds for Deep Ensembles, 0.01249 for Monte Carlo Dropout and 0.01259 for Bayesian Neural Networks. Entropy and Deep Evidential Classification were slightly faster at 0.00132 and 0.00139 seconds. No variability estimate for inference time was reported.
The arXiv document is a version 1 preprint dated 28 Aug 2026. Its evidence combines repeated model comparisons, theoretical analysis and a five-expert survey on one image task. Tabular experiments were repeated 10 times and image experiments five times, with results reported as means and standard deviations. The work was supported by Singapore's Ministry of Health National Medical Research Council and the Centre of AI in Medicine.
The paper evaluates model behavior, explanation quality, computational speed and expert confidence on the reported tasks; it does not report prospective clinical decisions or patient outcomes. The next test is whether the same pattern holds on independent clinical datasets and in larger, more varied expert studies.
Paper data and sources
Original title: Explainable Uncertainty Estimation for Reliable Medical AI
Authors: Li Rong Wang, Jamie Duell, Xinran Xu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text