Abstract
Generative models trained using self-supervision of tokenized electronic health record (EHR) timelines show promise for clinical outcome prediction. This is typically done using Monte Carlo simulation for future patient trajectories. However, existing approaches suffer from three key limitations: sparse estimate distributions that poorly differentiate patient risk levels, extreme computational costs, and high sampling variance. We propose two new estimators: the Sum of Conditional Outcome Probability Estimator (SCOPE) and Risk Estimation from Anticipated Conditional Hazards (REACH), that leverage next-token probability distributions discarded by standard Monte Carlo. We prove both estimators are unbiased and that REACH guarantees variance reduction over Monte Carlo sampling for any model and outcome. Empirically, on hospital mortality prediction in MIMIC-IV, SCOPE and REACH match 100-sample Monte Carlo performance using only 10-11 samples, representing a ~10x reduction in inference cost without degrading calibration.
Introduction: The Limitations of Monte Carlo in EHR Modeling
Generative Transformers are increasingly applied to Electronic Health Records (EHRs) to simulate potential patient futures and predict clinical risks. The standard approach, Monte Carlo (MC) simulation, involves generating numerous future trajectories and calculating risk as the frequency of an outcome token (e.g., "death").
However, MC simulation presents significant challenges in clinical deployment:
- Computational Intensity: Generating hundreds of trajectories per patient is resource-prohibitive for real-time monitoring.
- High Sampling Variance: Stochasticity in token generation leads to noisy risk estimates.
- Sparse Estimates ("Stair-stepping"): With a finite number of samples (e.g., 100), risk scores are restricted to discrete increments (e.g., 0.01), failing to distinguish between varying levels of low-probability, high-impact risks.
Methodology: Leveraging Next-Token Distributions
To address these flaws, we propose two estimators that utilize the full "next-token" probability distribution produced by the model at each step, rather than discarding it after sampling.
- SCOPE (Sum of Conditional Outcome Probability Estimator): This estimator sums the probability of the outcome token occurring at every step across simulated paths.
- REACH (Risk Estimation from Anticipated Conditional Hazards): A survival-analysis-inspired estimator that generates "outcome-free" timelines. It treats the model's predicted probability of the outcome at each step as a "hazard" rate.
We theoretically prove that both estimators are unbiased and that REACH provides a guaranteed variance reduction over standard MC sampling for any model or outcome.
Experimental Results
We evaluated these estimators using the ETHOS-ARES framework on the MIMIC-IV dataset, focusing on hospital mortality and ICU admission predictions.
10x Efficiency Gains
For hospital mortality prediction, REACH and SCOPE achieved performance equivalent to a 100-sample Monte Carlo run using only 10 to 11 samples. This represents a tenfold reduction in inference cost without compromising calibration or predictive accuracy.
Variance Reduction
As shown in the figure below, REACH and SCOPE maintain significantly lower variance than Monte Carlo, particularly when the number of sampled timelines is small. REACH, in particular, maintains near-zero variance even with minimal sampling.
Discussion: The Role of "Spontaneity"
The efficiency gains were found to be task-dependent. While mortality prediction saw a 10x speedup, ICU admission prediction saw a more modest ~1.2x gain. We attribute this to a property we call "spontaneity": the likelihood of an event occurring abruptly at any given token. Mortality is highly spontaneous in critical care contexts, whereas ICU admission is often preceded by specific clinical indicators (e.g., transfer orders), making it less "surprising" to the model and reducing the relative benefit of distribution-based estimation.
Conclusion and Future Work
SCOPE and REACH significantly improve the feasibility of deploying generative EHR models by reducing computational overhead and providing more granular risk estimates. Future work will focus on extending these methods to multi-token outcomes and developing real-time survival curves to assist clinical decision-making.
Summary updated