The future of cardiac prognostication: supervised machine learning to predict mortality in patients with acute pulmonary embolism

Background Pulmonary embolism (PE) is a significant cause of morbidity and hospitalization worldwide with Right Ventricular (RV) dysfunction identified in association with morality. Two-dimensional speckle tracking echocardiography is an imaging modality used to calculate cardiac strain which can be used to identify early mechanical stress in the RV. Purpose Combining imaging techniques with patient demographics, comorbidities, and clinical parameters, we aim to create a supervised machine learning model to predict mortality within one year after PE and identify important variables. Methods A retrospective cohort of 74 patients who presented with acute PE, confirmed by contrast CT or V/Q scanning, to our hospital system in 2017 who had a transthoracic echocardiogram within 48 hours after the event. STE was used to calculate Endocardial Global Longitudinal Strain (Endo-GLS), End-Diastolic Area (EDA), End-Systolic Area (ESA), and Fractional Area Change (FAC). These parameters were taken along with Size/location of PE, RV size, RV Function, RV Systolic Pressure (RVSP), and TR Velocity. A supervised machine learning model was created using logistic regression and random forest classifier algorithm 100 trees with bootstrapping using Python version 3.7.2. The data was randomly sampled and run through our classifier 50 times which our outcome measurements averaged. Results We identified Endo-GLS as the parameter with the highest weight (30.9%) followed by, FAC (14.0%), and ESA (10.8%). Non-echocardiographic variables that had high importance were patients BMI and Intervention at time of event. We found an Accuracy of 84.3%, Sensitivity of 57.1%, and Specificity of 91.5%. The AUC for our model was 0.87. Conclusion The use of novel imaging methods and supervised machine learning can help identify important variables such as Endo-GLS and help predict mortality in patients with PE. Extrapolated, strain patterns can be sensitively measured in the acute setting of diseases and be used to help predict clinical outcomes. Our next steps are to obtain larger sample size in to create a more robust classifier and to prospectively validate our model. Random Forest Model Funding Acknowledgement Type of funding source: None