Prehospital Diagnostic Algorithm for Acute Coronary Syndrome Using Machine Learning: A Prospective Observational Study

Background Rapid and precise prehospital recognition of acute coronary syndrome (ACS) is key to improving clinical outcomes. We tested the hypothesis that the machine learning-based prehospital algorithm shows a high predictive power for predicting ACS. Methods We conducted a multicenter observational prospective study that included 10 participating facilities in an urban area of Japan. The data from consecutive adult patients, identified by emergency medical service (EMS) personnel with suspected ACS, were analyzed. The primary outcomes were binary classification models for ACS prediction based on eXtreme Gradient Boosting (XGBoost). Secondary outcomes were classification models for subcategories of ACS, including acute myocardial infarction (AMI) and ST-segment elevation myocardial infarction (STEMI). We evaluated the model performance based on the area under the receiver operating curve (AUC). Results Of the 555 enrolled patients, 388 (70%) were randomly assigned to a training cohort, and 167 (30%) were placed in a test cohort. The XGBoost model for ACS using 43 features performed well (AUC 0.879 [95% CI 0.815–0.935]) in the test cohort. We repeated the analysis with a limited number of selected features, and the performance of the XGBoost model using 17 features remained high (AUC 0.883 [95% CI 0.820–0.939]) in the test cohort. The XGBoost model for ACS using 17 features had a higher AUC than the other four common prediction models: logistic regression, random forest, a linear support vector machine (SVM), and a radial basis function (RBF) kernel SVM. The XGBoost algorithm for prediction of AMI and STEMI using the same 17 features also resulted in high AUC scores of 0.857 [95% CI 0.788–0.919] and 0.871 [95% CI 0.819–0.925] in the test cohort, respectively. Conclusions We found that the machine learning-based prehospital algorithms showed a high predictive power for predicting ACS.