Overfitting rarely shows up in the number a vendor puts on a slide. It hides one layer down, in how the model was tested and on whom. A model that has memorised its training data — including the noise and the confounders — can post a beautiful internal score and still fail the first time it meets a patient it was not built around. The signs are consistent enough to list. What follows are ten red flags that a reported performance will not generalise, each anchored to a documented, published case, and each paired with the check that would have caught it. As of July 2026.
The ten flags at a glance
| # | Red flag | Documented case | The check it fails |
|---|---|---|---|
| 1 | Internal validation only | Only 6% of imaging studies externally validated 1 | External validation |
| 2 | The score collapses at a new site | Sepsis model: AUROC 0.63, 67% missed 2 | Independent external test |
| 3 | A shortcut, not the disease | Pneumonia CNN: AUC 0.861 by sorting on hospital 3 | Confounder / saliency audit |
| 4 | Performance rides on artifacts | Skin markings raised false positives ~40% 4 | Robustness to incidental features |
| 5 | Data leakage into the test set | Leakage in 294 papers across 17 fields 5 | Leakage audit / model info sheet |
| 6 | Too many parameters, too few events | 87% of ML model studies high risk of bias 6 | Events-per-predictor, overfitting control |
| 7 | The claim outruns the evidence | 61 of 81 claimed parity or superiority 7 | Claim-vs-design match |
| 8 | Discrimination reported, calibration ignored | Miscalibration "misleading and harmful" 8 | Calibration curve |
| 9 | Never tested in a live clinic | Retinopathy model hit real-clinic image limits 9 | Prospective field evaluation |
| 10 | No temporal validation | AKI models over-predicted across nine years 10 | Temporal / drift monitoring |
Work down the list against any claim. A single flag is a question rather than a disqualification; several together are a pattern.
1. Internal validation only
The most common red flag is also the easiest to miss, because it is an absence. Performance measured on held-out data drawn from the same source the model was trained on tells you how well it memorised that source, and routinely overstates how it behaves elsewhere. It is worth knowing how rare the alternative is: across 516 AI diagnostic-imaging studies, "only 6% (31 studies) performed external validation" 1. When a claim reports only internal or cross-validated performance, treat every figure as an upper bound, and ask for a test on data from a different site, scanner, or era. The concept, and why it filters out most of the field, is covered in external validation.
2. The score collapses at a new site
The second flag is what the first one is hiding from you. When a widely implemented proprietary sepsis model was evaluated outside the system that built it, it "predicted the onset of sepsis with an area under the curve of 0.63" and, at its live alerting threshold, "did not identify 1709 patients with sepsis (67%)" 2. The internally reported performance had promised much more. The check is direct: demand the external number, at the operating point that will be used, and weigh it above any internal figure. A large gap between the two is the definition of overfitting made visible.
3. A shortcut, not the disease
A model can score well by learning something correlated with the label that has nothing to do with the pathology. The clearest demonstration: a pneumonia classifier trained across hospital systems learned to detect which hospital produced a scan and exploited the differing disease rates, so that "merely sorting by hospital system achieved an AUC of 0.861" on a joint dataset 3. The network was reading metadata artifacts, not lungs — and its performance fell on outside data where that shortcut no longer held. The check is a confounder audit: saliency maps, tests on sites with different base rates, and a hard look at whether any high-performing feature is clinically implausible.
4. Performance rides on incidental artifacts
A close cousin of the shortcut is the incidental mark. When benign lesions were imaged with ordinary gentian-violet surgical skin markings, a melanoma classifier raised its probability scores and its specificity fell — "increasing the false-positive rate of benign nevi by approximately 40%" 4. The lesions were unchanged; a purple ink dot was enough to move the output. A headline accuracy on clean, curated images would never have surfaced this. The check is robustness testing: perturb the inputs with the artifacts a real clinic produces — markings, rulers, hair, off-centre framing — and see whether the score holds.
5. Data leakage into the test set
Leakage is when information from the test set reaches the model during training, so the evaluation grades memorisation. It is far from rare. A cross-field survey found "17 fields where leakage has been found, collectively affecting 294 papers," and in a re-analysis of one field the correction was decisive: "when the errors are corrected, complex ML models do not perform substantively better than decades-old LR models" 5 — logistic regression matched the deep model once the leak was closed. The check is a leakage audit: confirm that patients — whole patients, and not the images alone — are split cleanly between train and test, that no preprocessing touched the whole dataset before splitting, and that no feature encodes the outcome.
6. Too many parameters, too few events
Flexible models overfit when they are asked to estimate more than the data can support. A systematic review of prediction models built with machine learning found that "148 (87%, 95% confidence interval 81% to 91%) were rated at high risk of bias," that 56% "were developed with an inadequate number of events per candidate predictor," and that 39% "assessed overfitting improperly" 6. The check is arithmetic before it is anything else: how many outcome events, how many candidate predictors, and was overfitting controlled and honestly measured. A large, expressive model trained on a few hundred events is a red flag regardless of the score it reports.
7. The claim outruns the evidence
Sometimes the model is fine and the sentence describing it is the problem. A review of studies pitting AI against clinicians in imaging found that "most non-randomised trials are not prospective, are at high risk of bias, and deviate from existing reporting standards," even as 61 of 81 concluded the algorithm was comparable or superior to clinicians 7. A retrospective, single-site, internally validated model has earned the claim "promising in this setting" — and no more. The check is to hold the conclusion against the design: match the strength of the words to the strength of the test, and discount any claim of parity or superiority that rests on a weak comparator.
8. Discrimination reported, calibration ignored
Most claims report discrimination — usually AUROC — and stop. Discrimination measures whether the model ranks sicker patients above healthier ones; it says nothing about whether a predicted 20% risk means a real 20% event rate. That second property is calibration, and methodologists warn that "poorly calibrated algorithms can be misleading and potentially harmful for clinical decision-making" 8, yet it is the metric most often omitted. A model can rank beautifully and still hand you probabilities that are systematically too high, driving over-treatment. The check is to insist on a calibration curve alongside every discrimination figure; the distinction between the two is set out in sensitivity, specificity, and AUROC.
9. Never tested in a live clinic
A number earned on curated retrospective data is a promising start, and no guarantee about a working clinic. A deep-learning system for diabetic retinopathy, accurate in the lab, was deployed across real clinics and ran into "tensions between the model's thresholds for data quality, and the quality of data that arise from an imperfect, resource-constrained environment" 9: images taken in ordinary lighting failed its quality bar and were rejected, sending patients back around the loop. The statistical accuracy was intact; the deployment was not. The check is prospective, in-workflow evaluation — evidence that the model behaves when it meets the messy inputs and human responses of the setting you run.
10. No temporal validation
The last flag is the slowest. A model frozen at one moment ages against a moving world. Seven acute-kidney-injury models validated over the following nine years kept their ranking ability — "discrimination was maintained for all models" — while "calibration declined as all models increasingly overpredicted risk" 10. Nothing about the models changed; the population and the event rate did. A claim that reports a single point-in-time score, with no plan to re-measure, is a red flag for decay you will not see until it has arrived. The check is temporal validation and ongoing monitoring, the subject of our guide on dataset shift and model decay, with the one-line definition at model drift.
How to read this
Three cautions travel with the list. First, no single flag condemns a model — a retrospective, internally validated study can be excellent groundwork, and the flags are prompts for the next question rather than verdicts. Weigh them together; a claim carrying several is far more worrying than one carrying a single caveat. Second, the flags describe the evidence, not the model: a modest model tested honestly is more trustworthy than a dazzling one tested carelessly. Third, turning these flags into a procurement, deployment, or regulatory decision is a step beyond appraisal — confirm your obligations with your compliance or regulatory reviewers before acting on a claim, because a clean evaluation and a cleared or reimbursable device are not the same thing.
A single thread runs through all ten. Each flag marks a place where the reported number was measured under conditions kinder than deployment. Internal data is kinder than external; a balanced sample is kinder than true prevalence; a clean image is kinder than a marked one; a frozen snapshot is kinder than a drifting year; a curated retrospective set is kinder than a live clinic. Overfitting, in the end, is the distance between the conditions of the test and the conditions of use, and every check on this list is a way to measure that distance — and to close it — before a patient depends on the result.
Sources and method
This guide anchors each red flag to a primary source: an external-validation base rate 1, a deployed-model external-validation collapse 2, two documented shortcut and artifact cases 34, a cross-field leakage survey 5, a risk-of-bias review quantifying overfitting practice 6, a claims-versus-design review 7, the calibration literature 8, a human-centred field evaluation 9, and a nine-year temporal-drift study 10. Every figure is drawn from the primary source cited beside it. We revisit this page on a 180-day cycle and whenever a new documented failure case or reporting standard lands. For the checklist these flags support, see how to read an AI validation study; for the numbers behind the field, see our clinical AI trial results tracker.