CODE-II: a large-scale dataset for artificial intelligence in ECG analysis

Figures

• 

  Fig. 1. Map of Brazil illustrating the absolute and relative numbers of ECGs for each state served by TNMG in the CODE-II dataset. Basemap polygons (Brazil state boundaries; brazil-states.geojson) were obtained from the Click That 'Hood project (MIT License). The figure was generated using Python/GeoPandas and finalized using Inkscape.

• 

  Fig. 2. Summary of diagnostic class frequencies and group combinations in CODE-II. a Number of ECG exams associated with each diagnostic class in the CODE-II dataset. Diagnostic classes are not mutually exclusive; multiple diagnoses may be assigned to a single exam, except for Normal ECGs, which are exclusive. b Number of exams for each CODE diagnostic group. Exams may belong to more than one group. c Distribution of the number of diagnostic classes assigned per exam. All panels use a logarithmic scale on the y-axis.

• 

  Fig. 3. Global performance of the model on the CODE-II-test dataset. Panels show a micro- and b macro-averaged results for AUROC, AUPRC, F1-score, Recall, Specificity, Precision, and NPV. Threshold-dependent metrics were computed after applying class-specific thresholds selected to maximize the F1-score. Bars represent the mean metric values, with numerical values shown to the right of each bar, and horizontal lines denoting 95% confidence intervals estimated from 1000 bootstrap resamples.

• 

  Fig. 4. Per-class threshold-independent performance of the model on the CODE-II-test. Panels show a AUROC and b AUPRC (average precision) with 95% confidence intervals computed from 1000 bootstrap resamples. Bars indicate the mean metric value for each diagnostic class, with the numerical value shown to the right of each bar. The horizontal lines represent the corresponding confidence intervals. Diagnostic classes are ordered by their prevalence in the test set (shown in parentheses), where prevalence divided by 100 corresponds to the expected AUPRC of a random classifier.

• 

  Fig. 5. Scaling law analysis of model performance. a Macro-AUROC and b macro-AUPRC as a function of dataset size on the fixed CODE-II-test. The curves represent the mean performance across three independent runs, with shaded areas indicating the range between the minimum and maximum values.

• 

  Fig. 6. Generalization performance of supervised and pre-trained models across external ECG benchmarks. Each panel summarizes model performance across training regimes ranging from few-shot to full-data settings: a–c PTB-XL mean macro-AUROC, obtained by averaging each model’s macro-AUROC across the five diagnostic taxonomies (diagnostic, subclass, superclass, format, and rhythm) when trained with 5%, 10%, and 100% of the data; and d CPSC 2018 macro-AUROC. Colors denote distinct model architectures, while bar patterning indicates the training regime: dotted (supervised models trained from scratch), diagonal-hatched (pre-trained models fine-tuned on the target dataset), and solid (same architecture trained from scratch, with random initialization, used to assess the effect of pre-training). Across most architectures, fine-tuned pre-trained models achieve higher performance than their randomly initialized counterparts.

• 

  Fig. 7. Models evaluated on CODE-II-open and CODE-II-test. Panels show a macro-AUROC and b macro-AUPRC for all models trained on the CODE-II-open dataset and evaluated on the CODE-II-test set. Models were either trained from scratch or fine-tuned from pre-trained versions. Colors indicate distinct model architectures, while bar patterning denotes the training regime: dotted for models trained from scratch and diagonal-hatched for pre-trained models fine-tuned on CODE-II-open. For reference, the baseline model trained from scratch on the full CODE-II dataset is also included.

• 

  Fig. 8. Schematic representation of the 10 groups and their corresponding 66 ECG CODE classes.

• 

  Fig. 9. Flowcharts describing dataset curation. a CODE-II (full dataset) and b CODE-II-test. Each panel summarizes eligibility checks, exclusions, and the final number of exams and unique patients.

• 

  Fig. 10. Deep neural network model architecture. Adapted from Fig. 3 in Ribeiro et al.3 (CC BY 4.0).