Introduction: Widespread adoption of AI for medical decision-making is still hindered due to ethical and safety-related concerns. For AI-based decision support systems in healthcare settings, it is paramount to be reliable and trustworthy. Common deep learning approaches, however, have the tendency towards overconfidence when faced with unfamiliar or changing conditions. Inappropriate extrapolation beyond well-supported scenarios may have dire consequences highlighting the importance of the reliable estimation of local knowledge uncertainty and its communication to the end user. Materials and Methods: While neural network ensembles (ENNs) have been heralded as a potential solution to these issues for many years, deep learning methods, specifically modeling the amount of knowledge, promise more principled and reliable behavior. This study compares their reliability in clinical applications. We centered our analysis on experiments with low-dimensional toy datasets and the exemplary case study of mortality prediction for intensive care unit hospitalizations using Electronic Health Records (EHRs) from the MIMIC3 study. For predictions on the EHR time series, Encoder-Only Transformer models were employed. Knowledge uncertainty estimation is achieved with both ensemble and Spectral Normalized Neural Gaussian Process (SNGP) variants of the common Transformer model. We designed two datasets to test their reliability in detecting token level and more subtle discrepancies both for toy datasets and an EHR dataset. Results: While both SNGP and ENN model variants achieve similar prediction performance (AUROC: $\approx 0.85$, AUPRC: $\approx 0.52$ for in-hospital mortality prediction from a selected MIMIC3 benchmark), the former demonstrates improved capabilities to quantify knowledge uncertainty for individual samples/patients. Discussion/Conclusions: Methods including a knowledge model, such as SNGP, offer superior uncertainty estimation compared to traditional stochastic deep learning, leading to more trustworthy and safe clinical decision support.

Deep learning models offer transformative potential for personalized medicine by providing automated, data-driven support for complex clinical decision-making. However, their reliability degrades on out-of-distribution inputs, and traditional point-estimate predictors can give overconfident outputs even in regions where the model has little evidence. This shortcoming highlights the need for decision-support systems that quantify and communicate per-query epistemic (knowledge) uncertainty. Approximate Bayesian deep learning methods address this need by introducing principled uncertainty estimates over the model’s function. In this work, we compare three such methods on the task of predicting prostate cancer–specific mortality for treatment planning, using data from the PLCO cancer screening trial. All approaches achieve strong discriminative performance (AUROC = 0.86) and produce well-calibrated probabilities in-distribution, yet they differ markedly in the fidelity of their epistemic uncertainty estimates. We show that implicit functional-prior methods-specifically neural network ensembles and factorized weight prior variational Bayesian neural networks—exhibit reduced fidelity when approximating the posterior distribution and yield systematically biased estimates of epistemic uncertainty. By contrast, models employing explicitly defined, distance-aware priors—such as spectral-normalized neural Gaussian processes (SNGP)—provide more accurate posterior approximations and more reliable uncertainty quantification. These properties make explicitly distance-aware architectures particularly promising for building trustworthy clinical decision-support tools.

Background/Objectives: Valve replacement (VR) patients are at high risk of postoperative complications, but reproducible Electronic Health Record (EHR) benchmarks for evaluating sequential AI models in this setting are lacking. We develop a reproducible pipeline that extracts two EHR datasets from MIMIC-IV (a general-purpose and a predictive benchmark dataset) capturing perioperative histories, high-resolution time-series, and clinically motivated outcome labels. Methods: The cohort comprises 3890 VR patients with clinician-guided feature selection across diagnoses, procedures, laboratory measurements, medications, and physiological monitoring. As an exemplary use case, we define ICU readmission at first ICU discharge as a surrogate for postoperative risk and derive a predictive benchmark under strict label-leakage control. We then compare a Transformer model trained on tokenized longitudinal EHR sequences with Transformer and XGBoost baselines trained on aggregated feature statistics, and assess performance differences using paired statistical tests across validation splits. Results: ICU readmission stratified in-hospital and 100-day outcomes, including mortality, complications, and rehospitalization, confirming the clinical relevance of the prediction target. The sequential Transformer achieved 0.87 AUROC and 0.69 AUPRC. Corrected resampled t-tests confirm improved performance over the non-sequential Transformer, while the comparison with XGBoost indicates a favorable trend without conclusive evidence. Conclusions: Our findings suggest that leveraging longitudinal EHR sequences yields higher predictive performance than static feature summaries for postoperative risk prediction. The publicly released preprocessing pipeline and cohort-construction code enable researchers with MIMIC-IV access to reproduce the datasets and provide a robust benchmark for developing and comparing time-series models in post-valve replacement care.