Most kidney-AI ‘prediction’ studies detect disease already present—what does that mean for patients?
A peer-reviewed review of 169 studies found that only 46 tested a genuinely future outcome, four used an independent external cohort and none measured prospective clinical impact.
The short version
- What happened
- Of 169 primary studies, 122 classified current kidney-disease status and only 46 contained a clearly future-oriented prediction component.
- Who it affects
- Patients should not assume a model labelled predictive has been shown to warn of kidney decline before it happens.
- What remains unproven
- This is a systematic review of published studies, not a prospective trial of a kidney-AI system.
Keep in mind
This is a systematic review of published studies, not a prospective trial of a kidney-AI system.
The search used Scopus only and was restricted to English-language studies, so relevant evidence may have been missed.
Heterogeneous tasks, datasets, outcomes and metrics prevented a pooled performance estimate.
Incomplete reporting in primary studies sometimes limited assessment of timing, preprocessing and validation.
The review found no prospective clinical-impact study and does not establish that every individual model is ineffective.
Source analysis · 1 linked source. Primary analysed source date: 19 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The word “prediction” often hid a different task
Most published artificial-intelligence research described as predicting chronic kidney disease was actually recognising disease at roughly the same time it was measured. That is the central finding of a peer-reviewed systematic review published on 19 October. The authors examined 169 primary studies from 2021 to 2026 and classified a study as genuinely predictive only when it had an identifiable baseline, predictors available by that point, a later renal outcome and a meaningful prediction horizon.
By that rule, 122 studies—72.2%—addressed current-state detection or classification. Forty-six, or 27.2%, included a future-oriented prediction component. This distinction matters for patients and clinicians: identifying patterns associated with kidney disease already present is not the same as estimating who will develop disease, deteriorate, need dialysis or reach kidney failure in time to change care.[1]
Strong-looking scores rarely travelled far
The review found a thin validation trail behind many performance claims. Of 169 studies, 136 lacked a clearly identifiable independent holdout or external validation strategy under the authors' coding rules. Seventeen used geographic or institutional external validation, while only four used an independent external cohort or dataset. None performed a distinct temporal validation using data from a later calendar period.
Calibration—whether predicted risks match observed risks—was assessed in 12 studies. Only one study explicitly evaluated clinical utility using decision-curve or net-benefit analysis. Nine presented a prototype or tool, one showed retrospective operational use and none reported a prospective clinical-impact study. The stronger future-oriented subset did somewhat better on external validation and calibration, but the review still found no direct test of whether an AI warning improves a clinical decision. The literature therefore does not yet show that deploying these systems improves referrals, treatment decisions, kidney outcomes or patient experience.[1]
The distinction that matters
Three different claims that can all be called “prediction”
The review separates recognition of disease now from forecasting an outcome that happens later.
Current-state classification
Use data measured around diagnosis
Detect disease already present
Future-oriented prediction
Use baseline data before a later outcome
Estimate later kidney risk
Clinical impact
Test decisions and patient outcomes prospectively
Not demonstrated in the review
A useful warning, not a verdict on every model
Using PROBAST+AI, the reviewers judged 156 studies at high overall risk of bias during development and 164 during evaluation; no study achieved an overall low-risk judgement. Repeated reuse of public benchmark data also made apparently separate results less independent. Near-perfect classification on a familiar dataset may demonstrate that its labels are separable, but it does not establish that a model will work in a different clinic or forecast a future event.
The review has limits of its own. It searched Scopus only, included English-language studies and could not pool performance because tasks, datasets and outcomes differed. It received no declared financial support, and the authors reported no commercial conflicts. Its counts describe the papers that met the predefined search and coding rules, not every kidney-AI study ever produced. The next evidence that would change the assessment is prospective testing across later time periods and independent populations, with calibration, established clinical comparators and measured patient or service outcomes. Until then, kidney-AI “prediction” should be read as a task label—not proof of useful early warning.[1]