Can an AI agent predict diabetes outcomes in children? A 6,507-patient preprint is promising—not clinical proof
The evidence-citing system ranked future outcomes better than the tested statistical models on average. Its advantage varied by outcome, came from one hospital and has not been prospectively tested in care.
The short version
- What happened
- The retrospective single-centre study used records from 2,489 young people with type 2 diabetes and 4,018 with type 1 diabetes, predicting 24-month outcomes from the first 18 months of follow-up.
- Who it affects
- Children and families should not receive different care because of this preprint. A prediction model could eventually help clinicians notice risk in sparse records, but only if external and prospective testing shows that it improves decisions without widening errors across groups.
- What remains unproven
- Unreviewed preprint
Keep in mind
Unreviewed preprint
Single US hospital
Retrospective records
Outcome-specific 50–250-person test samples
Seven type 2 and four type 1 comparisons
Pooled family tests exploratory
Only dyslipidaemia had clearly separated intervals against the best ML model in both cohorts
Two reviewing clinicians were study authors
Protected patient data not public
No prospective clinical workflow
No treatment or patient outcome
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
The agent ranked future outcomes better on average—but the result was uneven
An evidence-citing AI pipeline called Cedar ranked several 24-month diabetes outcomes more accurately than the statistical models tested by its developers. The retrospective study used records from 2,489 young people with type 2 diabetes and 4,018 with type 1 diabetes at Texas Children's Hospital. Inputs covered the first 18 months after diagnosis; targets included glycaemic control, hypertension, microalbuminuria and dyslipidaemia, plus three treatment outcomes relevant only to type 2 diabetes.
Across the seven type 2 outcome comparisons, Cedar's pooled mean ROC-AUC was 0.815, against 0.700 for classical and 0.656 for temporal machine learning. Across four type 1 comparisons, it was 0.756, against 0.631 and 0.603. Those pooled figures conceal important variation: individual-outcome confidence intervals separated Cedar from the best statistical model only for dyslipidaemia in both cohorts. On two type 1 outcomes, a conventional model was numerically ahead.[1]
Every model saw the same held-out patients, but the test sets changed by outcome
For each outcome, the researchers made one prevalence-stratified train/test split. Trainable models were fitted on the training partition; Cedar was inference-only and read the same held-out patients, represented in six-month windows. The test sample was adapted to the number of assessable patients and outcome prevalence, ranging from 50 to 250 people. Confidence intervals used 2,000 patient-level bootstrap resamples. The family-level significance tests were explicitly exploratory because outcomes overlap and are correlated.
Cedar combined fixed guideline snippets with structured health-record data, social factors extracted from social-worker notes and, for type 1 diabetes, continuous glucose-monitor summaries. The social-factor extraction itself used two language models and iterative expert review, not an independent held-out validation. Most predictive signal came from the outcome's recent history; adding wider context changed discrimination only modestly when that history was complete. The authors estimated Cedar at about $0.10 per patient, compared with less than one millionth of a dollar for a trained model, using vendor list prices rather than measured operating costs.[1]
The distinction that matters
What the study compared—and what it did not
The benchmark tested ranking performance on historical records, not an AI system treating children.
Population
2,489 type 2 + 4,018 type 1 records
One US paediatric hospital
Input
Months 0–18
Structured records, selected social factors and type 1 glucose-monitor summaries
Target
Outcomes at month 24
Seven type 2 and four type 1 comparisons
Measured
Held-out ranking and citation fidelity
No treatment, safety or patient-benefit trial
The citations were faithful to the record; clinical usefulness is still unproven
The authors checked 70,246 values cited by Cedar across 2,490 patient-outcome records; four did not match the chart. That is evidence of transcription fidelity, not correct medical judgement. Two author-clinicians then reviewed 63 deliberately stratified type 2 cases while blinded to the outcome and Cedar's probability. They agreed with one another on 90.5% of forced binary calls, but the sampling design cannot estimate population accuracy and did not compare clinicians with and without AI support.
For families, the immediate implication is caution rather than a new decision tool. The work is an unreviewed, single-centre retrospective preprint; protected data are not public, test sizes varied by outcome, and no prospective workflow measured treatment changes, errors or patient benefit. Confidence would rise with peer review, multi-site external validation, locked prospective testing and trials in which clinicians use the system during real decisions. The authors declare no competing interests and list US public and philanthropic support.[1]