Key result
Deep learning models predicting age from 12-lead ECGs demonstrated that mortality risk increased when the predicted age was higher than the patient's actual age (HR 2.49).
Why the study?
Applying AI to infer health-related information from 12-lead ECGs has gained interest, but the performance of deep learning using reduced-lead configurations to predict sex, age, and mortality is less clear.
Can a deep learning approach accurately predict age and sex from reduced-lead ECGs compared to 12-lead ECGs, and do these predictions correlate with mortality risk?
Observational
Can a deep learning approach accurately predict age and sex from reduced-lead ECGs compared to 12-lead ECGs, and do these predictions correlate with mortality risk?
Hazard Ratio: 2.49
Deep learning models can accurately predict age and sex from reduced-lead ECGs, and discrepancies between predicted and actual age or sex are associated with increased mortality risk.
Reduced-lead ECGs support feasible AI age/sex prediction; leaves open whether this improves clinical risk stratification or outcomes.
The resting 12-lead electrocardiogram (ECG) is a widely used diagnostic tool in modern medicine, providing crucial insights into various heart conditions. Recently, the application of Artificial Intelligence (AI) to infer health-related information from the 12-lead ECG has gained significant interest. In this study, we propose a deep-learning approach to predict sex and age from both 12-lead and reduced-lead ECGs (12L, 6L, 4L, 3L, 2L, and 1L), and analyze their implications for predicting patient mortality. We employed a ResNeXt-based architecture and trained our model using the CODE15 dataset. Our best sex prediction model achieved an F1-score of 0.800 ± 0.007, Sensitivity (Se) of 0.807 ± 0.016, Positive Predictive Value (PPV) of 0.793 ± 0.022, and an Area Under the Curve (AUC) of 0.910 ± 0.007 when using the 12-lead ECG configuration. Similarly, our best age estimation model achieved a mean absolute error (MAE) of 8.961 ± 0.180, a Pearson Correlation (ρ) of 0.810 ± 0.004, and a coefficient of determination (R2) of 0.637 ± 0.014 when using the 4-lead ECG configuration. Through the evaluation of different lead-set configurations, we demonstrated that even with a reduced number of leads, our models achieved comparable performance to those ob- tained using the conventional 12-lead ECG setup. Moreover, we found that the mortality risk, assessed by the hazard ratio (HR), increased when our age model predicted an age higher than the actual age by a certain threshold for all lead sets (12L: 2.49, 6L: 2.17, 4L: 2.53, 3L: 2.54, 2L: 2.76, 1L: 2.65). Likewise, when our model misclassified the patient’s actual sex, the mortality risk also increased (12L: 1.36, 6L: 1.26, 4L: 1.38, 3L: 1.33, 2L: 1.20, 1L: 1.12). Additionally, we observed a decrease in mortality risk when our method predicted an age lower than the actual age by a certain threshold (12L: 0.71, 6L: 0.71, 4L: 0.74, 3L: 0.71, 2L: 0.75, 1L: 0.64). Overall, our research shows the efficacy of reduced lead ECGs in predicting age, determining sex, and providing valuable insights into patient mortality.
No takes yet. Share an insight, caveat, or question.
Dias et al. (2023) reported an observational. Deep learning ECG-predicted age higher than actual age vs. Actual age was evaluated on Mortality risk (HR 2.49). Deep learning models predicting age from 12-lead ECGs demonstrated that mortality risk increased when the predicted age was higher than the patient's actual age (HR 2.49).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: