Explainable age prediction using a transformer trained on diverse global electrocardiograms

Figures

• 

  Figure 1. Study overview. Datasets: 12-lead electrocardiograms from 12 datasets across four, from which were 11 public (blue) for model development and the external population-based LEAD cohort (green). Modelling: the four compared models (Transformer and convolutional neural network, each trained on all-public or single-cohort data), with the data each was trained and tested on, coloured by origin. Evaluation: prediction accuracy, subgroup performance, mortality risk, saliency-based explainability, and pathology correlation.

• 

  Figure 2. Pooled subgroup analysis of model performance stratified by chronological age and biological sex. The analysis compares the two transformer models’ (Transformer (Global), left, Transformer (LEAD), right) performance averaged across datasets for male (blue) and female (orange). (Top): Δage, defined as artificial intelligence–electrocardiogram age minus chronological age, across chronological age deciles. (Middle): Distribution of mean absolute error (MAE) across age deciles. (Bottom): Pearson correlation coefficients (r) between predicted and chronological age within each gender across age deciles. Error bars represent 95% confidence intervals and statistical significance indicated as *<0.05, **<0.01, and ***<0.001. Shaded grey background regions correspond to Cohen’s standard effect size thresholds for correlation magnitude (Small: ∣r∣=0.1–0.3, Medium: ∣r∣=0.3–0.5, Large: ∣r∣>0.5).

• 

  Figure 3. Survival analysis and mortality risk stratification based on Δage. (A) Kaplan–Meier overall survival estimates for the CODE and (B) SaMiTrop datasets. Patients are stratified into three distinct groups based on Δage: decelerated ageing (Δage ≤−8 years; green), concordant age (−8 years < Δage < +8 years; grey) and accelerated ageing (Δage ≥+8 years; red). The number of patients at risk over time (in years) is populated below each axis. Global statistical variance among the strata was determined using a multivariate log-rank test. (C) Forest plot depicting the associated relative risk of all-cause mortality derived from Cox proportional hazards regression models utilizing the transformer model predictions across both cohorts. The concordant aging group (−8 years<Δage <+8 years) serves as the baseline reference presenting standard risk. HR, hazard ratio; CI, confidence interval.

• 

  Figure 4. Saliency-based explainability and physiological feature sensitivity of the model. (A) Representative global median electrocardiogram traces for all 12 leads. The colour gradient indicates the median absolute saliency, highlighting regions of highest model focus. (B) Box plots showing the pooled sensitivity of the model to standardized electrocardiogram segments. (C) Heatmaps displaying the mean absolute saliency distributed across all leads and electrocardiogram segments, grouped by age deciles to show developmental changes in model attribution. (D) Global Spearman rank correlations isolating the relationship between model saliency and clinically relevant electrocardiogram features. Features are grouped by electrical timing, wave amplitude, and beat variability. (E) Heatmap of age-stratified Spearman correlations of electrocardiogram features and saliency.

• 

  Figure 5. Impact of clinical comorbidities and lifestyle factors on artificial intelligence–electrocardiogram age predictions and model feature importance. (A) Horizontal bar chart summarizing the patient distribution of lifestyle factors and clinical comorbidities within the study cohort. (B) Forest plot of age- and sex-matched analyses depicting the shift in corrected Δage across clinical subgroups. The markers represent median values with interquartile ranges, indicating accelerated (positive) or decelerated (negative) predicted ageing relative to a baseline healthy reference group. (C) Grouped bar chart displaying Spearman rank correlations between the presence of specific comorbidities or lifestyle factors and continuous age metrics (artificial intelligence–electrocardiogramage and chronological age). (D) Heatmap of Spearman correlations revealing the relationship between physiological electrocardiogram features and model saliency across different disease subgroups, demonstrating how specific underlying pathologies modulate the model’s region-specific focus compared to healthy individuals.