Epic's Foundation Model: A Vendor-Built EHR "Event Model" for Operational Triage

Generative Model DevelopmentPredictive Process Outcomes

Authors:Shane Waxler; Paul Blazek; Davis White, et al.

Journal:Arxiv

DOI:10.48550/arXiv.2508.12104

Epic has released Curiosity, a foundation model pretrained on 118M patients and 115B events from Epic Cosmos—signaling that major EHR vendors are entering the clinical AI space with vendor-native models. In retrospective testing, the 1B-parameter model outperformed supervised baselines for 30-day readmission (AUC 0.770 vs 0.717) and LOS prediction, but lacks dose/route/frequency tokenization needed for medication-safety CDS.

August 16, 2025

Quick Take

Zero-shot Curiosity‑L (1B parameters) outperformed task‑specific supervised baselines on held‑out operational endpoints in retrospective testing: 30‑day readmission AUCROC 0.770 vs 0.717 and overall hospital length‑of‑stay mean absolute error (MAE) 2.851 vs 3.339 days.

For inpatient pharmacy operations, a pretrained generative event model like Curiosity could support non‑interruptive risk‑sorting (discharge/readmission worklists, long‑stay prioritization) so pharmacist effort focuses on higher‑yield patients rather than universal review.

Why it Matters

Readmissions and prolonged length of stay drive patient risk and hospital throughput strain; pharmacists must balance medication safety with time‑critical discharge and follow‑up tasks.

Turning longitudinal electronic health record (EHR) data into timely, individualized risk signals today typically requires cohort definitions, manual feature engineering, and bespoke analytics that do not scale to routine operational workflows.

A scalable method that converts evolving diagnoses, labs, medication orders, and encounter patterns into timely, calibrated signals matters where stewardship, clinical decision support governance, and staffing constraints limit how much pharmacist attention can be deployed.

What They Did

Pretrained Curiosity on a filtered Epic Cosmos extract (Jan 2012–Apr 2025): 118 million adult patients and 115 billion medical events (151 billion tokens). Inclusion required age 18–120 and ≥2 face‑to‑face encounters; only direct‑care encounters were kept. A random patient‑level 90/10 train/test split was used (no temporal split).

Built three decoder‑only Qwen2‑style transformers (Curiosity‑S 62M, Curiosity‑M 119M, Curiosity‑L 1B parameters) with an 8,192‑token context window. Tokenizer encodes structured events (ICD‑10 diagnoses, ATC medication orders split into 3 tokens, lab test + decile tokens, encounter headers and time tokens) and excludes notes, vitals, allergies, dose/route/frequency, and medication administrations.

Pretrained autoregressively using compute‑optimal schedules derived from scaling‑law fits. At inference the model generates n Monte Carlo patient timelines and aggregates probabilities (n trades off accuracy for linear compute; authors evaluated up to ~80 simulations). Evaluations covered 78 held‑out tasks (diagnosis, prognosis, LOS, readmission, encounter forecasting) against task‑specific baselines (logistic/XGBoost/supervised transformer) using AUCROC, PR‑AUC, and MAE; generations not reaching the target horizon were discarded.

What They Found

Curiosity‑L (1B) improved operational predictions: 30‑day readmission AUCROC 0.770 vs 0.717 for the best supervised baseline (Δ≈+0.05), and reduced LOS MAE from 3.339→2.851 days overall (for admissions <7 days MAE 1.24 vs 2.03).

Generative outputs are plausibly clinical and syntactically robust: invalid multi‑token event rates were vanishingly small (diagnoses 0.011%, medications 0.0087%, labs 0.0002%); prevalence and pairwise co‑occurrence of events matched held‑out data reasonably well; single‑encounter PR‑AUCs reached 0.616 for inpatient medications and 0.899 for labs.

Scaling and inference compute both drive gains: larger model/training scale and lower pretraining loss yielded smoother, often sigmoidal downstream improvements; increasing Monte‑Carlo simulations improved performance with diminishing returns near ~64 generations.

Pharmacy‑relevant ceiling: meaningful gains exist for non‑interruptive operational signals (readmission, LOS, encounter forecasting) that could prioritize pharmacist effort, but the tokenizer omits dose/route/frequency, medication administrations, allergies, vitals, and free text—limiting direct medication‑safety CDS applicability.

Takeaways

One pretrained generative event model can provide multiple operational predictions (LOS, readmission, encounter demand) without task‑specific fine‑tuning in these retrospective tests; phrased more cautiously, results suggest a single 'risk radar' could support triage lists and workload planning.

For pharmacists, the highest‑value near‑term use is non‑interruptive triage (discharge/readmission worklists, long‑stay prioritization) to focus attention on higher‑risk patients rather than universal review.

Do not use Curiosity as a standalone medication‑safety engine: its input tokenization excludes dose/route/frequency, administrations, allergies, vitals, and notes, so it cannot support dose‑specific or administration‑aware CDS without substantial augmentation, integration of additional data types, or fine‑tuning.

Key diligence before operational use: temporal (future‑time) validation, calibration at actionable thresholds (PPV/NPV and patients‑flagged/day), subgroup fairness and calibration checks, and feasibility analysis of inference cost since more Monte Carlo trajectories improve accuracy but increase compute linearly.

Strengths and Limitations

Strengths:

Large multi‑site pretraining and compute‑optimal scaling (118M patients, 115B events, models at 62M/119M/1B) with systematic scaling‑law and loss→downstream analyses across many tasks.; Generative trajectory sampling with low syntactic error rates, encounter‑level calibration analyses, and direct comparisons to task‑specific baselines showing scale‑driven gains across diverse tasks.

Limitations:

Random patient‑level 90/10 split (no temporal split) risks optimistic forward‑time generalization for a deployment setting.; Tokenization excludes critical medication‑safety elements (dose, route, frequency), medication administrations, allergies, vitals, and clinical notes—constraining high‑risk pharmacotherapy CDS.; Discarding simulated generations that fail to reach the target horizon can bias evaluations toward cases that are easier to simulate.; Monte Carlo inference cost scales linearly with the number of trajectories (n), posing challenges for real‑time, bedside interruptive CDS without caching, model distillation, or task‑specific fine‑tuning to reduce sampling variance.

Bottom Line

Treat Curiosity as a strong zero‑shot signal for non‑interruptive operational triage (readmission, LOS, encounter forecasting); require temporal validation, calibrated thresholds, subgroup fairness checks, and a practicable inference strategy before relying on it in pharmacy workflows—it is not ready for dose/administration/allergy‑sensitive medication‑safety CDS without additional data and validation.