Stop Silent-Failure CDS: Demand Brier-Scored Uncertainty

Predictive Model DevelopmentPredictive External Validation

Authors:Mats Tveter; Thomas Tveitstøl; Christoffer Hatlestad-Hall, et al.

Journal:Artificial Intelligence in Medicine

DOI:10.1016/j.artmed.2026.103374

Uncertainty that actually tracks error under dataset shift may be the difference between safe CDS and “silent failure.” In EEG Normal/MCI/Dementia prediction, independently trained multi-run ensembles held up best (AUC ~0.84; Brier ~0.45) while some checkpoint ensembles degraded (AUC ~0.71; Brier ~0.57), pointing pharmacy leaders to demand Brier-based validation and route high-uncertainty cases to pharmacist review.

February 2, 2026

Quick Take

Independently trained ("multi-run") ensembles performed best in this study's EEG Normal/MCI/Dementia classification and uncertainty tracking (best multi-run AUC ≈ 0.84 ± 0.02; Brier ≈ 0.45 ± 0.01). Some checkpoint-based ensembles (e.g., Snapshot) performed worse (AUC ≈ 0.71 ± 0.03; Brier ≈ 0.57 ± 0.03). MCI remained the hardest class (best per-class AUC ≈ 0.71; recall ≤ 0.41).

For clinical AI exposed to dataset shift, favor diverse, independently trained ensembles and include proper scoring rules—particularly the Brier score alongside calibration measures like ECE—to make uncertainty operationally useful (e.g., to route high‑uncertainty cases for pharmacist review after local validation).

Why it Matters

Probability-like model outputs can be misleading: overconfident errors are high-risk in clinical settings, so models should reliably indicate when predictions are likely wrong.

EEG illustrates real-world dataset shift—biologic noise, artifacts, and cross-site differences create out-of-distribution inputs—so uncertainty estimates that track performance are essential for safety.

In inpatient pharmacy and other clinical CDS, uncertainty that meaningfully tracks accuracy is a governance tool to route only the right cases to pharmacist review while preserving stewardship, staffing, and compute constraints; this requires local validation and continuous monitoring.

What They Did

Used the CAUEEG cohort (1,176 subjects; 80/10/10 train/validation/test split) of resting-state EEG preprocessed to 19 channels; four non-overlapping 30 s epochs per subject were selected and averaged to produce a single subject-level prediction; age was included as a normalized input feature (with noise augmentation during training).

Implemented an Inception-style CNN backbone and compared eight uncertainty approaches grouped into multi-run ensembles (N=5 independent models: Deep Ensemble, Bagging, Augmentation, Depth), checkpoint ensembles (Snapshot, Fast Geometric Ensemble extracted from one run), and sampling methods (Monte Carlo dropout, SWAG); applied temperature scaling and tuned hyperparameters with Optuna.

Validated models on the held-out test set, three external out-of-distribution cohorts (Miltiadous, MPI, TDBRAIN), and a battery of ten graded, EEG-specific synthetic shifts (e.g., additive noise, amplitude scaling, interpolation, phase/peak shifts); evaluated discrimination and probabilistic calibration using AUC, Brier score, and Expected Calibration Error (ECE).

What They Found

Multi-run ensembles (including augmentation/depth variants) achieved the best in-distribution performance (AUC ≈ 0.84 ± 0.02; Brier ≈ 0.45), whereas some checkpoint ensembles (e.g., Snapshot) performed substantially worse (AUC ≈ 0.71 ± 0.03; Brier ≈ 0.57). Sampling methods (MCDropout, SWAG) were intermediate or variable across runs.

MCI remained the weakest class across methods (best per-class AUC ≈ 0.71; recall ≤ 0.41), consistent with clinical label ambiguity and heterogeneity.

On external OOD cohorts (note these sets are small and some are single‑class), multi-run models retained high Normal recall (Miltiadous N ≈ 0.82–0.89; MPI/TDBRAIN N ≈ 0.95–1.00) and reasonable dementia recall in Miltiadous; checkpoint and SWAG methods degraded with wide run-to-run variability.

Across the ten synthetic EEG shifts, multi-run ensembles' uncertainty (Brier and often ECE) generally increased in step with accuracy loss and signalled degradation until extreme corruption. In several shift types (e.g., amplitude scaling, Gaussian noise) some checkpoint ensembles showed flat or uninformative uncertainty despite falling AUC—a 'silent-failure' pattern.

Expected Calibration Error (ECE) alone can be misleading: under some shifts checkpoint or underconfident models had favorable ECE despite poor discrimination; the Brier score more consistently captured clinically relevant miscalibration in these experiments.

Takeaways

When evaluating or procuring predictive tools, prioritize methods whose uncertainty increases as performance drops under realistic dataset shift—here, independently trained multi-run ensembles met this criterion while some checkpoint approaches did not; require evaluation across relevant OOD cohorts and synthetic shifts.

Treat uncertainty as a 'smoke alarm,' not an autopilot: use uncertainty to flag cases for human review. Note that under extreme corruption any method can miscalibrate; in this 3-class task the random-guess Brier baseline ≈ 0.667 can serve as a reference for extreme failure.

Do not accept ECE in isolation as proof of safety—ask vendors and developers to demonstrate Brier score behavior across realistic shifts and OOD cohorts so that uncertainty meaningfully tracks error.

Expect weakest performance where clinical labels are fuzzy (MCI) and watch for demographic shortcuts (age was an input and OOD cohorts skewed younger); insist on local validation, continuous monitoring, and pharmacist judgment for edge cases.

Strengths and Limitations

Strengths:

Domain-specific, realistic stress testing: subject-level aggregation, a held-out test set, three external OOD cohorts, and ten EEG-specific synthetic shifts targeted at modality-relevant corruptions.; Rigorous evaluation and transparency: Optuna hyperparameter tuning, temperature scaling, pre-specified AUC/Brier/ECE metrics, run-level confidence intervals, and publicly available code/data to support reproducibility.

Limitations:

Primary training on a single CAUEEG cohort and small, younger OOD samples (e.g., MPI, TDBRAIN) with few or absent MCI cases limits generalizability and raises age-confounding concerns.; Synthetic, isolated perturbations and the observed miscalibration once Brier > 0.667 limit extrapolation to complex, correlated 'organic' operational drift; local validation and continuous monitoring remain necessary for deployment.

Bottom Line

In this evaluation, independently trained multi-run ensembles best preserved discrimination and produced uncertainty estimates that rose with error across held-out, OOD, and synthetically shifted EEG data; checkpoint ensembles and relying on ECE alone risked masking failures. Track proper scoring rules (especially the Brier score) during local validation and route high‑uncertainty cases for human review.