Pretrained Patient Trajectories for Adverse Drug Event Prediction Using Common Data Model-based Electronic Health Records

Abstract

Pretraining electronic health record (EHR) data through language models by treating patient trajectories as natural language sentences has improved various medical tasks. However, EHR pretraining models have never been utilized in adverse drug event (ADE) prediction. Here, we propose a novel pretraining scheme for common data model (CDM) based EHR data, named CDM-BERT. We utilized diagnosis, prescription, measurement, and procedure domains from observational medical outcomes partnership (OMOP)-CDM. We newly adopted domain embedding (DE) to simplify pretraining procedure and to improve comprehension of medical context. ADE prediction was selected as a finetuning task. For drug groups, we included nonsteroidal anti-inflammatory drugs (NSAID), anticoagulants (AC), glucocorticoids (GC), and chemotherapy (Chemo). For corresponding adverse events, we selected peptic ulcer (PU), intracranial hemorrhage (ICH), osteoporosis (OP), and neutropenic fever (NF), respectively. CDM-BERT was validated by internal and external datasets with 510,879 and 419,505 adult inpatients. CDM-BERT outperformed all the other baselines in all cohorts, demonstrating the effectiveness of DE (area under the receiver operating characteristic curve (AUROC) of 0.977, 0.908, 0.980, 0.989 for NSAID-PU, AC-ICH, GC-OP, Chemo-NF cohorts in internal validation, and 0.967, 0.960, 0.972, 0.959 in external validation, respectively). We also identified important features for each cohort, and several prior studies and clinical knowledge suggested the results. CDM-BERT has demonstrated its potential as a foundation model through its prediction performance, interpretability, and compatibility.