An NLP-Inspired Data Augmentation Method for Adverse Event Prediction Using an Imbalanced Healthcare Dataset

This paper proposes a data augmentation method for imbalanced healthcare datasets. This method was inspired by a data augmentation method in natural language processing (NLP) that generates synthetic sentences for training by replacing some words with similar words. The proposed method generates synthetic patient records by replacing patient backgrounds with similar backgrounds. In this paper, the cosine similarity of the distributed representations was used as the similarity metric between patient backgrounds. The distributed representations of the patient backgrounds were generated by the skip-gram model. To confirm the performance improvement with the proposed data augmentation method, the prediction performance of adverse events (AEs) caused by drug administration was experimentally evaluated on a real-world medical dataset with 1,510,137 records. The combination of the proposed data augmentation method and a conventional undersampling method resulted in an 80.0% improvement in accuracy and a 40.0% improvement in the precision and F1-score. The multifaceted evaluation demonstrated that the proposed method is effective, especially for predicting AEs with positive ratios ranging from 1.0% to 2.1%, which are difficult to predict with conventional machine learning methods but should be predictable in the medical field.