A Machine Learning Approach for Enhancing the Prediction of Acute Coronary Syndrome
Recommended Citation
Emakhu JO, Nassereddine H, Monplaisir L, Aguwa C, Arslanturk S, Masoud S, Egbe-Etu E, Tenebe I, Miller J. A Machine Learning Approach for Enhancing the Prediction of Acute Coronary Syndrome. Acad Emerg Med 2023; 30:239.
Document Type
Conference Proceeding
Publication Date
4-20-2023
Publication Title
Acad Emerg Med
Abstract
Background and Objectives: Acute coronary syndrome (ACS) is a leading cause of mortality and morbidity. The incorporation of machine learning models in risk stratification has the potential to support clinical decision support tools, but the inclusion of free text data can be challenging. In this study, we explored machine learning models incorporating free-text clinical documentation to aid ACS evaluation. Methods: This observational cohort study included patients being evaluated for ACS within nine emergency departments (ED). We collected electronic health record data comprised of structured data and free text, unstructured clinical narratives from 1/2017-8/ 2020. We excluded patients that did not have an electrocardiogram and troponin performed in the ED and those with ST-segment elevation myocardial infarction. The primary outcome was non-ST- segment elevation myocardial infarction (NSTEMI) or unstable angina (UA). We performed machine learning models of structured data alone and structured plus unstructured narrative data. Feature selection modeling included BorutaShap and SelectFromModel methods to reduce dataset dimensionality. To address data imbalance, modeling incorporated cost-sensitive classification and a resampling technique (SMOTETomek). We used an 80% training to 20% testing data split to evaluate model performance. Results: There were 31,228 patients with clinical concerns for ACS included in the study, of whom 431 (1.4%) had NSTEMI, 132 (0.4%) had UA, and 30,665 (98.2%) had non-ACS diagnoses. There were 111 total features extracted for model construction, 11 from structured and 100 from unstructured, free text data. The modeling framework that incorporated the unstructured data successfully classified patients as having ACS versus non-ACS diagnoses with accuracy, sensitivity, precision, and F1-score of 95.6%, 96.7%, 96.7%, and 96.7%, respectively. In contrast, without the inclusion of the unstructured data, our proposed model had accuracy, sensitivity, precision, and F1-score of 85.7%, 86.3%, 86.3%, and 86.3%, respectively. Conclusion: Our results suggest that adding features from unstructured clinical notes improves the performance of clinical decision support development for ACS using machine learning techniques.
Volume
30
First Page
239
