Augmenting Electronic Health Records for Adverse Event Detection.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Adverse events (AEs) from medical interventions increase patient morbidity, mortality, and healthcare costs.
• Predicting AEs using electronic health records (EHRs) is crucial for timely interventions but hindered by data challenges.
• Classical machine learning methods struggle with imbalanced EHR data, missing labels, and complex interactions.

Purpose of the Study:

• To introduce TASER-AE, a novel data augmentation pipeline for structured EHR data.
• To enhance the prediction of adverse events by addressing class imbalance and improving minority-class representation.
• To improve the robustness and predictive performance of classification models for EHR data.

Main Methods:

• Developed TASER-AE, a data augmentation pipeline inspired by Natural Language Processing (NLP) techniques, adapted for structured EHR data.
• Utilized transformer-based classification models in conjunction with the augmented EHR data.
• Applied the pipeline to sparse and imbalanced clinical datasets to enrich minority adverse event classes.

Main Results:

• TASER-AE achieved minority-class F1 scores up to 0.70, significantly outperforming classical machine learning baselines.
• Demonstrated substantial improvements in adverse event detection performance across two distinct EHR datasets.
• Effectively alleviated class imbalance, enhancing the representation of minority adverse event classes.

Conclusions:

• Structured, NLP-inspired data augmentation methods can overcome data limitations in clinical predictive modeling.
• TASER-AE shows significant potential for improving patient safety outcomes through enhanced AE prediction.
• The TASER-AE pipeline offers a valuable tool for researchers working with imbalanced clinical data.