Agentic LLM reasoning in pediatric diabetes outcome prediction from multimodal medical records

What it claims

CEDAR is a single-model LLM pipeline from Baylor College of Medicine and Texas Children's Hospital (arXiv, October 2026). It predicts pediatric diabetes outcomes at 24 months from the first 18 months of records and returns a fixed table of cited chart values with each prediction. It beats trained ML on pooled ROC-AUC in both cohorts. The paper also reports that a naive single-call LLM on the same base model reached comparable discrimination. The authors argue that CEDAR's main addition is a checkable rationale rather than extra accuracy. They also find accuracy is limited by the information in the data, not by model strength.

How it works

Each patient's first three 6-month visit windows (months 0 to 18) are the input. The record combines structured EHR, social determinants of health extracted from social-worker notes and, for type 1 diabetes, continuous glucose monitoring data. The task covers seven T2D outcomes and four T1D outcomes, with prevalence from 58.7% (dyslipidemia, T2D) down to 4.7% (hypertension, T1D). The model reasons over the record and outputs a probability plus a table of cited features, visits and values. Baselines are classical ML over a flattened feature matrix and temporal ML over the visit sequence. A variant gives CEDAR a frozen ML score as an in-context tool.

Evidence

• Per-outcome strength is weaker than the pooled result. The 95% bootstrap intervals of CEDAR and the best ML model separate only on dyslipidemia, in both cohorts. T2D margins range from +0.21 (dyslipidemia) and +0.11 (microalbuminuria) down to +0.03 (optimal glycemic control) and +0.01 (insulin independence). In T1D, random forest leads on optimal glycemic control (0.797 vs 0.785) and logistic regression leads on hypertension (0.688 vs 0.666). The T1D comparison against classical ML alone has p = 0.058.
• Base model. CEDAR's lead over the best ML model stayed positive on all four base models tested, including the open-weight Llama 3.3 70B.
• Carry-forward baseline. CEDAR mostly carries the last observed outcome state forward. It does so about as often when the state changed as when it held (p = 0.39). It beats a pure carry-forward baseline on six of seven T2D outcomes, by +0.035 to +0.083 ROC-AUC.
• Faithfulness. A counterfactual audit covered 70 T2D and 40 T1D records. The cited evidence alone reproduces the prediction (sufficiency 0.95). No single citation is decisive (comprehensiveness 0.08 and 0.05). Masking the top ten citations shifts the prediction by 0.32, against 0.065 for ten random values.
• Clinician review. The review used 63 T2D cases with the probability withheld. Mean usefulness ratings were 4.68 and 3.95 out of 5. On 29 of CEDAR's 31 errors, at least one expert reached the same wrong call. The authors read this as showing that errors cannot be detected from the rationale.
• Error overlap. Errors are nearly independent between CEDAR and ML (phi 0.12 to 0.17 in T2D, 0.03 to 0.06 in T1D). Within the ML family, errors correlate strongly (0.63 and 0.61).
• ML score as a tool. This gave +0.03 ROC-AUC on T1D optimal glycemic control, matching a stacked late-fusion combiner. It gave no gain on T2D dyslipidemia.
• Abstention. Declining the least confident 40% of patients raised T2D ROC-AUC from 0.767 to 0.825.
• Failed improvements. Verifier-gated self-refinement, a diverse-reasoner ensemble, LLM-guided feature engineering and prompting with transition rates did not move discrimination beyond noise.
• Modalities. CGM helped only on T1D optimal glycemic control. SDOH helped on no outcome in any model family.
• Temporal shift. All model families held accuracy when trained on earlier-diagnosed patients and tested on more recently diagnosed ones.
• MIMIC-IV. On this adult critical-care database, CEDAR led on one-year mortality but trailed supervised ML on readmission and in-hospital mortality.

How to use it

Inference costs about $0.10 per patient, against under $10⁻⁶ for a trained model. No training is needed. The page names Llama 3.3 70B as an option an institution could run locally. It gives no code, prompts, weights or licence.

Limits

• Single site, private data. The data comes from Texas Children's Hospital only, and the companion cohort papers are still in preparation.
• Low-prevalence errors. On low-prevalence outcomes, CEDAR makes far more errors than classical ML (0.58 vs 0.15 in T2D, 0.31 vs 0.05 in T1D), because it attempts rare positive calls.
• What the audit covers. The fabrication audit checks transcription, not clinical judgment.
• Calibration. Probabilities need recalibration.
• Truncated page. The fetched text ends partway through the outcome-history ablation, so those results are missing here.