Artificial intelligence-guided prediction model for short- and long-term all-cause mortalities using 12-lead electrocardiography

Abstract Introduction Electrocardiography (ECG) can be easily obtained at a low cost and includes voltage and time intervals representing heart conditions. We hypothesized that artificial intelligence (AI) detects a subtle abnormality in 12-lead ECGs and may predict both short- and long-term individual mortality. Methods Among 5,092,736 ECGs of 1,188,665 populations in single our center, 3,076,662 ECGs of 1,013,588 individuals were included after the exclusion of ECGs that had no sinus rhythm. We developed a ResNet-based model to predict 3-day to 5-year all-cause mortalities in 1,735,399 ECGs of 662,885 (training set: age 54 ± 20 years, 52% male) and internally tested the model in 600,107 ECGs of 350,377 (holdout set: age 49 ± 20 years, 51% male). We externally validated the model in 513,628 ECGs of 114,600 individuals of independent data (MIMIC-IV set: age 63 ± 16 years, 50% male). We stratified the 41,949 ECGs from 41,949 patients of clnical external set (UK biobank set: age 64 ± 8 years, 49% male) into four predicted risk groups to compare risk of death among stratified groups. Results The ResNet-based model showed good performance power with the under receiver operating characteristic curve (AUROC) of 0.922, 0.922, 0.916, 0.907, 0.899, 0.861, 0.844 for 3-day, 7-day, 30-day, 60-day, 90-day, 1-year, and 5-year mortality prediction in the holdout set, respectively. Also, in MIMIC-IV set, the model showed good performance power with AUROC of 0.842, 0.836, 0.822, 0.809, and 0.800 for 3-day, 7-day, 30-day, 60-day, and 90-day short-term mortality prediction, respectively, in contrast with long-term mortality prediction (0.698 of 1-year and 0.608 of 5-year mortality). Using the 3-day mortality model, the rate of inhospital death increased as the likelihood of the predicted probability of death within 3 days in MIMIC-IV set (OR 1.41 per 1% of mortality probability, 95% CI 1.41-1.42, p<0.001). Using each mortality prediction model, the group predicted to have the highest mortality probability showed a significantly higher all-cause mortality rate (3-day model [HR 1.42, 95% CI 1.11-1.81], 30-day model [HR 1.60, 95% CI 1.25-2.06], 90-day model [HR 1.56, 95% CI 1.22-2.00]) compared to the group predicted to have the lowest mortality probability in UK biobank set. Conclusions The ECG-based AI prediction model can identify high-risk patients who have short-term future death.AUROC of AI prediction model