Curiosity-0826-4B extends generative medical event modeling to children: zero-shot asthma and readmission prediction, with lab-value error down 18–46%
Synopsis
Curiosity-0826-4B extends a generative medical event foundation model trained on Epic Cosmos records from adults to pediatric patients and adds explicit representations of death and right-censoring. The model grows from 1B to 4.3B parameters and from an 8K to a 16K context window, generally matching or exceeding Curiosity-0226-1B on tasks both support, with mean absolute error on laboratory value prediction 18–46% lower. Pediatric AUROCs for incident asthma and asthma exacerbation are comparable to an adult reference cohort, while sickle cell crisis is lower in children.
Interpretation
A generative medical event model now covers pediatric patients and produces zero-shot clinical predictions without training separate task-specific models. The earlier Curiosity-0226 series covered only adults, and pediatric patients were absent from its evaluation. In cohorts aged 2–17 years, AUROC was 0.742–0.785 for incident asthma and 0.829–0.852 for asthma exacerbation, comparable to the adult reference cohort; sickle cell crisis reached 0.820–0.847, below the adult 0.915.
The model represents a patient chart as a sequence of discrete medical event tokens and is trained to predict the next event, so sampling generated sequences simulates possible continuations of the recorded history. New representations include birth events, gestational age, birth weight, vitals, non-numeric labs, medication attributes, and explicit tokens for recorded death and right-censoring. Training runs in two stages: compute-optimal pretraining at an 8,192-token context on 320B tokens, then context extension to 16,384 tokens with 435.6B more tokens, for about 755.6B tokens in total.
With explicit representation of documented death, the model can predict mortality in hospice, inpatient, and ambulatory settings. The earlier model had no death event token and so could not directly address these predictions. Adult AUROCs across settings range from 0.821 to 0.942, including 0.942 for inpatient 30-day and 0.921 for ambulatory 365-day mortality.
Perspective
All evaluations are constructed retrospectively from historical electronic health records, covering adults and six pediatric age cohorts on readmission, length of stay, laboratory values, incident disease, single-encounter event generation, and mortality prediction. The authors call for prospective evaluation and for validation and calibration in local settings and clinically meaningful subgroups.
Deaths not documented in the EHR are under-ascertained, and recorded death dates carry only date precision; tasks such as length of stay discard simulations that fail to complete within the maximum sequence length, which disproportionately affects sequences covering longer calendar time or dense events. Retrospective performance does not by itself establish that simulation-derived estimates are calibrated individual clinical risks, and predicting events after documented treatment changes is not equivalent to estimating the causal effect of changing care.