Pretrained Patient Trajectories for Adverse Drug Event Prediction Using Common Data Model-based Electronic Health Records
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Pretraining electronic health record (EHR) data using language models by treating patient trajectories as natural language sentences has enhanced performance across various medical tasks. However, EHR pretraining models have never been utilized in adverse drug event (ADE) prediction.
Methods
A retrospective study was conducted on observational medical outcomes partnership (OMOP)-common data model (CDM) based EHR data from two separate tertiary hospitals. The data included patient information in various domains such as diagnosis, prescription, measurement, and procedure. For pretraining, codes were randomly masked, and the model was trained to infer the masked tokens utilizing preceding and following history. In this process, we introduced domain embedding (DE) to provide information about the domain of the masked token, preventing the model from finding codes from irrelevant domains. For qualitative analysis, we identified important features using the attention matrix from each finetuned model.
Results
510,879 and 419,505 adult inpatients from two separate tertiary hospitals were included in internal and external datasets. EHR pretraining model with DE outperformed all the other baselines in all cohorts. For feature importance analysis, we demonstrated that the results were consistent with priorly reported background clinical knowledge. In addition to cohort-level interpretation, patient-level interpretation was also available.
Conclusions
CDM-based EHR pretraining model with DE is a proper model for various ADE prediction tasks. The results of the qualitative analysis with feature importance were consistent with background clinical knowledge.
Plain language summary
Patient history is like natural language; each medical code corresponds to a word, and the sequence of medical codes corresponds to a sentence. Language models learn from sentences by inferring masked words using remaining unmasked words, and language model-based EHR pretraining models comprehend medical context similarly. As several studies that have utilized EHR pretraining models have achieved great success in various tasks, we applied EHR pretraining models for adverse drug event (ADE) prediction. For better inference of the EHR pretraining model, we introduced domain embedding (DE) to provide a hint of the domain for each masked code. The model pretrained with DE performed the best in various ADE tasks, and regarding background clinical knowledge was well-reflected in our feature importance-based qualitative analysis.