MACHINE-LEARNING FOR THE ELECTROCARDIOGRAM-BASED PREDICTION OF SHORT-TERM AND LONG-TERM MORTALITY AT THE TIME OF DISCHARGE IN A POPULATION-LEVEL COHORT OF PATIENTS WITH ACCESS TO UNIVERSAL HEALTHCARE (CCC-230)

Background: The electrocardiogram (ECG) serves as a widely accessible diagnostic instrument, routinely employed in nearly all patient encounters within acute care settings. This tool offers valuable diagnostic and prognostic insights, with discharge ECGs potentially facilitating risk stratification and informing post-discharge management. It is unclear how the use of discharge ECG in addition to the information such as comorbidities that are at clinicians’ disposal at discharge can further improve prognostication in patients.

METHODS AND RESULTS: In this study, we leveraged a large, population-based cohort of 1,308,136 ECGs from 496,303 healthcare episodes of 255,162 patients (2007–2020) in Alberta, Canada under universal health insurance who sought care at emergency departments or hospitals in Alberta, Canada. Our aim was to develop deep learning (DL) models utilizing discharge ECG tracings and extreme gradient boosting (XGB) models based on discharge ECG measurements for predicting short-term and long-term mortality at 30-day, 1-year, and 5 years from the time of discharge. We then compared the performance of these models against an established Charlson Comorbidity Index (CCI) with demonstrable prognostic efficacy.

We divided the overall ECG dataset into a random split of 60% for the model development, and the remaining 40% as the holdout set for validation. The models for 30-day, 1-year, and 5-year mortality were trained on 145,026, 138,384, and 103,784 patients and evaluated on 81,767, 76,368, and 54,852 patients, respectively. In our holdout cohort, 8.7%, 19.7%, and 42.7% patients died by 30-days, 1-year, and 5-years, respectively. DL models with ECG traces, age, sex performed better (84.71%, 79.66%, 81.96%) than the XGB models of ECG measurements, age, sex (78.47%, 74.42%, 78.32%) in predicting outcomes for all three timepoints and the clinical risk model based on CCI, age, sex (81.38%, 80.95%, 82.34%) in predicting 30-day outcome. Moreover, addition of CCI information improved the performance of ECG models at all three time-points by 1.88%, 4.44%, 2.3% in DL and by 3.69%, 6.5%, 3.82% in XGB.


Conclusion: In this study, we developed mixed-input AI-augmented ECG-based prediction models to output a calibrated probability of mortality at 30-day, 1 year, and 5 years from discharge using discharge ECGs. Our observations show that although ECG-based models are more predictive than CCI-based clinical models particularly for short-term mortality, incorporating CCI improves the performance of ECG-based models at all three time- points. Therefore, combining both ECG and CCI information can provide superior prognostic prediction and may be a more effective approach for planning care after a patient's discharge from the hospital.