Machine Learning vs. Rule-Based Methods for Document Classification of Electronic Health Records within Psychiatry - A Systematic Literature Review

Abstract

Throughout the history of artificial intelligence, various algorithm branches have predominantly been used at different times. The last decade has been characterized by a shift from rule-based methods to self-learning methods. However, while the shift towards using ML methods is evident, there is no comparison of both methods for document classification. This systematic literature review focuses on the document classification in healthcare notes from electronic health records within psychiatry. We assess how these methods compare to each other in terms of classification performance and how they have developed throughout time, and we discuss potential directions of the field. We find that rule-based methods have had a higher performance for most of the last decade than machine-learning methods.Yet, the shift in representation techniques and algorithms used in recent years resulted in machine learning methods performing better.Dense document representation techniques, with mostly non-zero cells, outperform sparse representation techniques, with mostly zeros. Also, many neural networks outperform other self-learning- and rule-based methods. We find that state-of-the-art language models are barely employed in the psychiatric domain and expect an increase in the application of federated learning can increase the data availability for model training.