Can a high-AUC diabetes model still be unsafe?
New analysis today of a peer-reviewed 2 October audit of 12 machine-learning approaches on two public diabetes datasets. Similarly ranked models can differ materially in calibration, uncertainty and safe deferral, but this is a benchmark study—not a clinical trial, diagnostic approval or evidence of improved patient outcomes.
The short version
- What happened
- The study compared 12 models using a balanced 70,692-record derivative of the 2015 US BRFSS survey and a separate 520-person early-stage symptom dataset.
- Who it affects
- Patients could benefit if uncertainty routes difficult cases to appropriate review, but a poorly calibrated probability could still mislead referrals or reassurance.
- What remains unproven
- Both datasets are retrospective public benchmarks; neither represents a prospective clinical deployment or a completed diagnostic pathway.
Keep in mind
Both datasets are retrospective public benchmarks; neither represents a prospective clinical deployment or a completed diagnostic pathway.
The 70,692-record BRFSS derivative is artificially balanced and based largely on self-reported US survey variables, so its probabilities do not directly represent natural prevalence.
The 520-record symptom dataset is small and highly separable, making near-ceiling results and uncertainty gains vulnerable to sampling variation.
Fairness checks were limited to age and sex, while conformal coverage was marginal and depends on exchangeability.
The composite safety score depends on chosen metrics and weights; it is an analytic device, not a clinical regulatory standard.
Source analysis · 3 linked sources. Primary analysed source date: 2 October 2026. Earliest linked supporting source: 26 August 2016.
Read the full report · 7 min The evidence, methods, implications and limitations
The study asks a broader question than which model scores highest
A diabetes classifier can separate higher-risk from lower-risk records reasonably well and still give probabilities that are too confident, unstable or difficult to act on. The new Scientific Reports paper treats that distinction as a safety problem. Rather than selecting a winner from one accuracy table, the authors compare discrimination, probability calibration, uncertainty, selective prediction and conformal coverage across 12 classical, ensemble and deep tabular models.
That framing matters in health care because a probability is often used downstream: to set a referral threshold, communicate risk or decide which cases need human review. A high area under a curve does not show that a quoted 70% risk occurs in roughly seven of ten comparable people. It also does not show that the model recognises unfamiliar cases. The paper therefore provides an audit framework, not a claim that any tested model is ready to diagnose diabetes.[1][2]
Two public datasets create very different tests
The larger dataset contains 70,692 records and 21 predictors derived from the 2015 US Behavioral Risk Factor Surveillance System. It is balanced between diabetes and non-diabetes classes, unlike natural prevalence, and relies heavily on self-reported survey information. Age, body-mass index, blood pressure, general health and related indicators can be predictive without being a clinical work-up. The balancing makes model comparison convenient but changes the probability environment in which calibration would be used.
The second dataset contains only 520 records with 16 symptom and demographic features. It has much clearer class separation and yields near-ceiling results, including PR-AUC values around 0.99 for several models. That can reveal how uncertainty methods behave in a small, easy benchmark, but it is also a warning: 520 observations leave wide scope for sampling instability, and symptoms collected for a diabetes dataset may not resemble unselected primary-care populations.[1][2][3]
Twelve models were compared inside a leakage-controlled pipeline
The authors evaluated logistic regression, support-vector and probabilistic baselines, tree ensembles including random forest, XGBoost and CatBoost, and deep tabular approaches including multilayer perceptrons, TabNet and FTTransformer. They used a stratified five-fold outer cross-validation design. Preprocessing was fitted inside training data, the held-out fold was reserved for testing, and a separate calibration subset was used for temperature scaling. Those choices reduce a common source of optimistic leakage.
On the BRFSS derivative, FTTransformer had the highest reported mean PR-AUC at 0.804, followed by XGBoost at 0.803 and CatBoost at 0.800. The differences are too small to support a sweeping claim that one model family dominates. The more useful finding is that models with nearly indistinguishable discrimination had different expected calibration errors, uncertainty scores and composite safety rankings. Optimising only the leaderboard metric would conceal those operational differences.[1][2]
Deferral improved error rates, but coverage is a design choice
The researchers estimated epistemic uncertainty with repeated predictions and tested a selective strategy that retained 90% of cases while deferring the most uncertain tenth. On the larger dataset, mean error fell by roughly 7%. On the small early-stage dataset, the largest relative reduction was 43.7% for the multilayer perceptron, although several confidence intervals for that dataset included zero. A large relative gain from a small initial error rate is not the same as a large clinical benefit.
Deferral does not make difficult cases disappear. It transfers them to another process—usually a clinician, laboratory test or follow-up pathway. A real service would need to measure how many patients are deferred, whether uncertainty is concentrated in particular demographic groups, how quickly review occurs and what errors remain among confidently automated cases. The paper reports age- and sex-based checks, but those limited attributes cannot establish equity across ethnicity, disability, income, geography or access to care.[1][2]
Conformal coverage is conditional, not a safety guarantee
Split conformal prediction produced empirical coverage close to nominal 90% and 95% targets. This means the generated prediction sets contained the true class at approximately the promised marginal rate in held-out data under the study's exchangeability assumptions. Some uncertain cases receive both possible labels, making ambiguity visible rather than forcing a single answer. That is useful behaviour for triage systems, where an honest 'both remain plausible' can trigger review.
The guarantee does not survive every distribution shift. If deployment populations differ in prevalence, measurement, access, survey response or disease definition, past residuals may no longer represent future cases. Marginal coverage can also hide undercoverage in smaller subgroups. The datasets were not collected as a prospective clinical pathway, and no threshold was tied to treatment benefits or harms. Conformal validity here is a statistical benchmark result, not evidence that patients would be safer.[1][2]
What would change the assessment
The next step is not another random split of the same tables. Stronger evidence would freeze a model and evaluate it on later patients from multiple health systems, with clinically verified outcomes and prespecified calibration, subgroup, deferral and decision-curve measures. Natural disease prevalence must be retained or corrected explicitly. The evaluation should compare the AI-assisted pathway with usual care and report missed cases, unnecessary tests, time to diagnosis, workload and patient-relevant outcomes.
Until then, the study supports a procurement question rather than a purchasing conclusion: show the calibration plot, uncertainty behaviour and deferred-case workflow, not only AUC. Its most consequential result is that similar headline rankings can mask different safety properties. It does not establish that any of the 12 models should screen, diagnose or prioritise a person outside the two public datasets.[1][2][3]