Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records
2026-08-20 • Machine Learning
Machine Learning
AI summaryⓘ
The authors developed BERT-LER, a machine learning model that reads electronic health records (EHRs) like a language, focusing on lab test results by turning them into tokens that keep important value details. They trained this model on data from 75 million patients and tested it on public benchmarks and an asthma study. Their model not only predicts health outcomes well but also shows clear explanations tied to specific medical events, matching known clinical risk factors. This approach improves understanding and prediction using EHR data in various medical areas.
BERTElectronic Health Records (EHR)Laboratory Test ResultsIntegrated GradientsTokenizationPercentile BinningClinical PredictionModel InterpretabilityEHRShot BenchmarkAsthma Severity
Authors
Jun Ni Du, Lukas Adamek, Maxim Kryukov, Flavio Dormont, Ziv Bar-Joseph, Sven Jager, Brandon Rufino
Abstract
Predictive models over structured electronic health records (EHRs) remain central to machine learning for healthcare, but few have jointly emphasized quantitative laboratory information and interpretability with respect to input medical events. We present BERT-LER, a BERT-style model for coded EHR timelines pretrained and fine-tuned from a de-identified EHR dataset of 75 million patients, that encodes laboratory test results as discrete tokens while retaining graded information through percentile-based binning, paired with Integrated Gradients for token-level attributions grounded in the input EHR sequence. We benchmark our approach on the public EHRShot benchmark suite and on an asthma severity progression study based on real-world data. This addresses a methodological gap in EHR foundation-style modeling by unifying laboratory value representation and explainability in a single framework, while assessing whether both predictive performance and explanations generalize beyond standard clinical prediction tasks. Across EHRShot and asthma tasks, BERT-LER achieves predictive performance that is competitive with, and on laboratory-related tasks often exceeds, publicly available benchmark models, and provides attributions that align with clinically known risk factors. Our architecture and explainability approach can be applied to many therapeutic areas and prediction tasks using language models trained on structured EHRs.