Background: Previous research has shown that age is highly correlated to ECG findings, and machine learning (ML) can be used to predict the age of a patient based on the findings on a 12-lead ECG. We structured a state-of-the-art deep ML model to correlate with age, which may be used to detect changes following the trends. Methods: We composed a 40-layers self-definite ResNet model. The model was trained with 10-second 12-lead ECG of 25,915 records from 23,625 patients. We then validated the accuracy of age prediction using 6,479 records from a separate dataset. Once the model was validated, we tested the model using randomly selected ECG records, including 20 records from a weight loss center and 20 records from cardiology clinic patients. Results: After the model was developed and validated, we tested it on 40 patients with ages 52.2 ±13.7 years, ranging from 17 to 76 years. The mean absolute error (MAE) was 8.0 years and the root mean square error (RMSE) was 9.7 years. Among these 40 patients, 21 patients (52.5%) had ECG-predicted age within 7 years difference from their chronological age. When we selected the age range of 42-75 years old (29 records), the result of MAE decreased to 7.4 years, with 70% of the prediction results within 7 years and the RMSE decreased to 9.1 years. The average difference between chronological and ECG-predicted ages decreased to 3 years when serial recordings of ECG of the same patients were used. Conclusions: Using ML models, ECGs can be used to correlate with a patient’s age and may be useful to track the changes longitudinally. “ECG-predicted age” may be an important measurement to follow the trend for comprehensive health assessment over time.
Read full abstract