Abstract Background Cardiovascular diseases (CVDs) are among the leading causes of mortality and morbidity globally, with an estimated 1 million hospital admissions for CVD in England in 2019/20, leading to 5.5 million bed days. In addition, patients with CVDs have a notably high rate of hospital readmission, with approximately one in four patients experiencing heart failure (HF) being readmitted within one month of discharge. We explored whether machine learning models trained on electrocardiogram (ECG) and patient electronic health records (EHRs) can support decision-making and risk management and ultimately improve health outcomes for HF patients. Purpose The project aims to develop a machine learning model to predict health outcomes for HF patients. To test the hypothesis, we trained and tested an explainable machine learning model to predict the risk of mortality and hospital readmission rate for HF patients. Methods We used EHR data and tabulated ECG records to train an XGBoost model to predict the risk of 30-day mortality and 30-day hospital readmission for HF patients. EHR data from a cohort of 2,868 patients with their ECG records were used for model development. We extracted a total of 78 features from the ECG data, such as heart rate variability (HRV), P-wave durations, and QT intervals. These ECG features were combined with coded and free-text EHR data (e.g., demographic information, diagnosis, medication history, and lab results) and used as tabular inputs to the XGBoost model. Of the 2,868 patients, an internal set of 574 patients stratified by age, sex, and health outcomes (30-day mortality and 30-day hospital readmission) were held out for testing. Results Among 2,868 HF patients with 10-year follow-up, 1065 patients were reported for mortality or hospital readmission within 30 days of discharge. When the machine learning model was trained on ECG data in addition to EHR data, the model was able to predict patient mortality and hospital readmission rate in the test set with an area under the curve (AUC) of 0.91 and 0.70, respectively. For mortality prediction, the model achieved a sensitivity and specificity of 0.81 and 0.82, respectively, when employing an operating point with minimum difference between the sensitivity and specificity. When the threshold was adjusted to favor sensitivity at 0.91, the model specificity decreased to 0.74. The machine learning model also identified high sensitive troponin levels, QT dispersion, and lactate dehydrogenase levels as features of high importance for predicting the outcomes for HF patients. Conclusions Our developed machine learning model was able to predict mortality and readmission risks in HF patients using multi-modal data. Upon further validation, the machine learning model has the potential to be tested for supporting clinical decision-making and risk management for HF patients.ROC CurveFeature Importance
Read full abstract