Abstract 19524: Time Series Feature Extraction and Machine Learning for Prediction of In-Hospital Cardiac Arrest
Introduction: Electrocardiogram (ECG) predictors prior to in-hospital cardiac arrest (I-HCA) have not been fully investigated. ECG from telemetry may provide information about diagnostic clues prior to I-HCA. Machine learning algorithms can be used for integration of information from several ECG features in order to build predictive models for clinical decision-support.
Hypothesis: Machine learning (ML) can be used to develop a model to predict I-HCA.
Methods: In a multi-center study, full-disclosure ECG papers (25 mm/s paper speed) from telemetry were obtained for two minutes prior to I-HCA and three other same-duration segments within one hour of I-HCA. Signal segments were also obtained for two minutes within each of the seven hours prior to I-HCA. The full-disclosure ECG was digitized using the connected-component method for lead II and R detection was performed by HAT algorithm. Multiple computational features were extracted including heart rate, QRS duration and morphology, and ST-segment changes. A binary prediction target (variable) was defined by categorizing segments into two classes: within one hour of I-HCA and within seven hours of I-HCA excluding the last hour. Feature selection was performed using Support Vector Machines (SVM) in a nested cross-validation setup. An ML model was inducted and validated on test sets.
Results: There were 199 signal segments from 35 subjects with initial rhythms of pulseless electrical activity (PEA) and asystole (37% female, 69±13 years old, ejection fraction of 51±16%) which comprised the dataset. ECGs showed decreased heart rate (≥15 bpm; 49%), QRS prolongation (≥20 ms, 34%), QRS morphology changes (11%), and ST segment changes (29%) within one of hour of I-HCA. Classification, of instances from the test sets, with the inducted SVM model yielded 78.9% accuracy, 77.6% sensitivity and 80.4% specificity. The ROC area under the curve for the model was 85.9%.
Conclusions: The ML model shows promise through multiple performance metrics including a high ROC area and can prove to be viable for clinical decision-support. With a larger dataset, as the feature space gets more densely populated, we anticipate model performance to improve and the strength of predictors to vary.
Author Disclosures: S. Shandilya: None. P. Sabouriazad: None. M. Attin: None. K. Najarian: None.
- © 2014 by American Heart Association, Inc.