An Application to automatically detect aggressive events from Dutch narrative electronic health records

From Master Projects
Jump to: navigation, search

Abstract KIM 2

Background: Violence and aggression are a substantial problem for patients and professionals in the care for psychiatric patients. As electronic health records (EHRs) are being used increasingly in Dutch hospitals, the opportunity exists to gather a vast amount of clinical information automatically. Methods for the automated identification of adverse events from the narrative free-text portions of EHRs have been developed but little work has been validated for Dutch-language context. Furthermore, tools that specifically identify aggressive incidents from medical records do not yet exist although this research domain is in need of a more reliable method compared to costly manual chart review inspections of EHRs.

Aim:  The aim of this study was to develop a method based on natural language processing (NLP) and machine learning that automatically identifies aggressive events based on free text notes in Dutch EHRs.

Method: Electronic medical data from the Academic Medical Center of the University of Amsterdam was collected from a total of 120 admissions of patients with a psychotic disorder and were manually reviewed by medical professionals. This resulted in 26829 annotated free text notes. The annotated data was used as the golden standard for the development and evaluation of the automated tool for identifying aggressive events. In our approach we tested a term-feature representation based on tf-idf and a topic feature representation based on Latent Dirichlet Allocation (LDA). Furthermore several supervised machine learning algorithms were compared to see which could best tackle this specific classification problem.

Results: The best machine learning model achieved an f1-score of 0.52 using text-based features. Results showed that using topic feature did not improve model performance and a kernelized support vector machine seems to best tackle the classification problem.

Conclusion: Notwithstanding the relatively limited accuracy prediction, this work offers valuable insights in the exploration of classifying Dutch medical text notes.