Classifying Adverse Drug Reactions from Electronic Medical Records using Machine Learning and Natural Language Processing

Abstract

Background and Objectives: Patient drug safety is a priority in the World Health Organization (WHO)'s action plan. In Thailand, psychiatric medications are among the top five drug classes linked to adverse events. However, analyzing adverse drug reactions (ADRs) can be a time-consuming process, and variations may arise due to differences in the expertise of medical personnel. Classifying ADRs in psychiatric patients who use bilingual (Thai-English) electronic medical records is complicated by linguistic structures, transliterations, and non-standard spellings, which can hinder the accuracy of existing classification models. This study aims to develop a model for classifying ADRs associated with psychiatric medications in bilingual text. It will compare machine learning and deep learning techniques to enhance medication safety surveillance within Thailand's healthcare context.

Methodology: The process began with collecting ADE data from psychiatric patients between October 2019 and March 2023. These data were then processed to develop a model for classifying ADRs using natural language processing and five traditional machine learning techniques: Gaussian Naive Bayes, K-Nearest Neighbors, Linear Support Vector Machine, Logistic Regression, and Random Forest. Two deep learning techniques were utilized: Convolutional Neural Networks and Bidirectional Long Short-Term Memory Neural Networks. Finally, the model's performance was evaluated using 70% of the data for training and 30% for testing.

Main Results: The findings indicate that the model developed using the Random Forest technique, combined with feature extraction methods incorporating both 1-gram and 2-gram approaches, achieved the highest level of classification performance. When tested with a dataset of 384 samples, this model achieved an accuracy, precision, recall, and F1-score of 84.6%, 82.6%, 84.6%, and 83.2%, respectively. The developed model demonstrated high effectiveness in accurately classifying adverse drug reactions, showcasing its robustness as a machine learning method in the medical domain. Additionally, regarding computational efficiency, the Linear Support Vector Machine and Logistic Regression models exhibited the fastest processing speeds, making them noteworthy options for scenarios requiring rapid decision-making.

Discussions: The findings of this study align with several previous studies, although variations may arise due to the characteristics of the data and the complexity of bilingual (Thai and English) content in electronic medical records. Deep learning methods are resource-intensive and time-consuming; thus, selecting the appropriate technique requires balancing accuracy and speed. Real-time systems may favor Linear Support Vector Machines, even though they are slightly less accurate than Random Forests. Future research should incorporate data balancing techniques, advanced feature extraction, expanding data sources, and experimenting with language models specifically developed for the Thai language and medical contexts.

Conclusions: This study demonstrates the potential of utilizing machine learning and deep learning techniques to detect adverse drug reactions, highlighting important implications for developing an efficient surveillance system. Such a system can enhance the speed and accuracy of identifying events that might otherwise be missed, while reducing the workload for healthcare personnel, allowing them to concentrate more on patient care. The proposed approach aligns with the WHO’s guidance and Thailand’s digital health policies.