A clinical reasoning pipeline predicts pediatric diabetes outcomes
Agentic LLM reasoning in pediatric diabetes outcome prediction from multimodal medical records · arXiv:2610.40379
In plain words
The paper introduces Cedar, a computer system that predicts outcomes for young people with diabetes from medical records, notes and, for type 1 diabetes, glucose-monitor readings. It gives a probability and lists the record details behind it, which can be checked against the chart.
At one hospital, Cedar ranked patients’ outcomes better on average than trained statistical models across the tested outcomes. The comparison was exploratory, and the study does not show that the system works as well elsewhere.
The research
Can a model predict pediatric diabetes outcomes from sparse records that combine structured medical data, social-worker notes and, for type 1 diabetes, glucose-monitor readings, while showing evidence clinicians can check?
The study tests whether a model can make those predictions and provide a rationale grounded in the chart.
Cedar combines guideline context, chart-checked citations and repeated reasoning to predict pediatric diabetes outcomes from patient records.
Across pediatric diabetes, can a model use sparse, mixed records to predict outcomes while showing evidence clinicians can verify? The authors tested Cedar against trained machine-learning models and separately checked whether its citations matched patient charts and influenced its predictions.
On shared held-out patients, Cedar’s pooled ROC-AUC was higher across the tested type 2 and type 1 outcomes than the trained model groups. ROC-AUC measures how well scores rank patients with and without an outcome across thresholds.
Chart audits found few mismatched values, and masking Cedar’s leading citations changed its probability more than masking random features. These results support stronger average ranking and traceable, relevant evidence in this study.
This was a retrospective study at one hospital, and pooled model comparisons are exploratory because outcomes and patient samples overlap. The results do not establish performance at other hospitals or show that Cedar’s full pipeline causes its accuracy advantage over trained models.
What the evidence shows
- The prediction problem
Cedar, the Clinical Evidence-Driven Agent for Risk-prediction, is a single-model pipeline that predicts pediatric diabetes outcomes from structured medical records, social-worker notes and, for type 1 diabetes, glucose-monitoring data.
The study asks whether this kind of reasoning can predict outcomes better than trained models and whether clinicians can check the evidence behind its predictions. The authors evaluated predictions from the first 18 months of records against outcomes at about 24 months.
- A checkable prediction
For each patient and outcome, Cedar adds fixed guideline statements, produces a probability and a table of cited chart values, checks the citations against the record, then averages the verified probability with three resampled-prompt probabilities.
The comparison used the same held-out patients for Cedar and trained classical and temporal models. Cedar also received outcome-specific guideline information that those statistical models did not, so the comparison covers both model family and prediction procedure.
- Comparing model scores
ROC-AUC measures how well scores rank patients with and without an outcome across possible decision thresholds. Across seven type 2 diabetes outcomes, Cedar’s pooled mean ROC-AUC was 0.815, compared with 0.700 for classical models and 0.656 for temporal models.
Across four evaluable type 1 diabetes outcomes, the means were 0.756, 0.631 and 0.603. These are pooled comparisons across outcomes, not a claim that Cedar led on every outcome.
- What the evidence supports
The authors checked Cedar’s cited values against patient charts and tested whether its most important cited features affected its probability more than randomly selected features did. They also found that two specialists often reconstructed Cedar’s prediction from its rationale.
These checks support the claim that the evidence was traceable and relevant to Cedar’s output, but they do not show that its predictions were clinically correct. A single-call comparison also detected no pooled score difference from the full pipeline.
Deep
- Original paper authors
- Daniel Palacios, Guantong Qi, Anna Kadi, Natali Borrego, Sukru Aras
- Submitted to arXiv
- License
- CC BY 4.0