Label-efficient cross-population transfer learning for electrocardiographic risk stratification: evaluation across German, Chinese and US cohorts

Figures

• 

  Figure 1. (A) number of 10-second records per cohort, with country of origin annotated. (B) Boxplot of subject age by cohort (MIT-BIH has no age metadata). (C) Sex composition per cohort. (D) Multi-label prevalence of the ten most common unified SNOMED-CT super-classes, expressed as the percentage of records in each cohort with at least one positive annotation for that class. PTB-XL is dominated by MI/STD, Chapman–Ningbo by TINV/LVH and MIT-BIH by PVC.

• 

  Figure 2. Hierarchical clustering (ward linkage) of 22 cohort-by-class prototypes drawn from 13 unified SNOMED-CT super-classes across the three populations.

• 

  Figure 3. (A) representative 10-second lead-II ECGs from each cohort after preprocessing. (B) Mean lead-II power-spectral density (Welch periodogram, n = 40 per cohort). (C) Heart-rate distribution; Chapman–Ningbo is centred at 109 bpm whereas PTB-XL and MIT-BIH are centred near 75–80 bpm. (D) Lead-wise mean RMS amplitude heat-map for the two 12-lead cohorts; Chapman–Ningbo is uniformly higher in amplitude across all leads, with the largest absolute differences in the precordial leads V2–V4.

• 

  Figure 4. (A) PCA projection (PC1 25.2%, PC2 12.6%); MIT-BIH separates along PC1. (B) UMAP 2-D embedding preserving the cohort-specific clusters. (C) Top-15 features ranked by |SMD| between PTB-XL and Chapman–Ningbo. (D) Pairwise MMD2 in the standardised 38-D input space (unbiased estimator, RBF kernel, median-distance bandwidth, 250 observations per cohort, 100 subsamples; error bars are 95 percentile intervals): PTB-XL versus Chapman–Ningbo 0.085, PTB-XL versus MIT-BIH 0.511, Chapman–Ningbo versus MIT-BIH 0.500; all three permutation tests give p = 0.002. (E) The same three statistics recomputed in the learned 64-dimensional embedding across five training seeds (mean ± SD). (F) Embedding-space MMD2 between PTB-XL and Chapman–Ningbo plotted against the Chapman zero-shot four-class macro-AUROC for those five seeds; the two are not associated (Pearson r = + 0.37, p = 0.54), so the alignment measured by MMD does not explain the transfer that is observed.

• 

  Figure 5. (A) schematic of the shared 38-D feature backbone with separate source (PTB-XL, 13 classes) and target (chapman/MIT-BIH) heads. (B) Per-epoch binary cross-entropy for the training and validation partitions on PTB-XL. (C) Validation macro-AUROC (green, left axis) and the cosine learning-rate schedule (purple, right axis); the best validation macro-AUROC of 0.845 is reached at epoch 16, but training always runs to epoch 60 and the final-epoch model is the one evaluated.

• 

  Figure 6. (A) Per-class ROC curves with AUROC. (B) Per-class precision-recall curves with average precision (AP). (C) Per-class AUROC with 95% bootstrap confidence intervals. (D) Row-normalised confusion matrix at the argmax prediction; entries show top-1 recall for each ground-truth class.

• 

  Figure 7. (A) macro-AUROC across training regimes computed on the four evaluable target classes {NORM, STD, LVH, TINV}, with 95% bootstrap intervals; cohort-specific sample sizes are annotated. (B) Per-class transferability gap, that is, the Chapman zero-shot AUROC net of the source AUROC, shown only for classes with at least 10 target positives. (C) Sample-efficiency curve for linear probing, annotated with the zero-shot value and the same-class source-domain comparator (PTB-XL macro-AUROC over the same four classes). (D) ROC curve for NORM on the held-out Chapman–Ningbo test set under zero-shot versus 50%-linear-probing regimes.

• 

  Figure 8. (A) distribution of the composite diagnostic score on the untouched chapman–ningbo test split, stratified by the composite outcome (AUROC = 0.791, 95% CI 0.743–0.836; n = 1,367, 93 positives of which 90 are LVH-positive). (B) Reliability diagram after isotonic recalibration fitted on the dedicated calibration split, with bootstrap uncertainty bands; Chapman–Ningbo Brier 0.055, expected calibration error 0.018, calibration slope 0.78. (C) Decision-curve analysis showing the model against treat-all and treat-none baselines across risk thresholds 0.05–0.30. (D) Subgroup AUROC by age stratum (< 50, 50–64, ≥ 65) and sex with 95% bootstrap intervals and per-stratum sample sizes and positive counts; the age interaction is significant (p = 0.012).

• 

  Figure 9. (A) Top-15 input features by mean |coefficient| across 13 per-class L2-regularised logistic regressions fitted on the standardised 38-D input features of the source training partition, with balanced class weights and at most 2,000 iterations. These are coefficients of an input-feature model and are not attributions of the neural backbone. (B) Class-by-feature coefficient heat-map for the top-11 features of the same input-feature model. (C) Window-level perturbation analysis on a representative PTB-XL record: each of 20 contiguous 0.5 s windows of the raw lead-II signal is replaced by the cohort-median level, all 38 features are recomputed on the modified signal, and the backbone is re-run; bars give the resulting mean absolute change in predicted probability across the 13 source outputs, which is not a gradient saliency map and is uncorrelated with the local amplitude of the window. (D) UMAP embedding of the 64-dimensional learned representation on Chapman–Ningbo, coloured by the predicted composite score.