PaperlayerView exact revision on arXiv

Source-linked AI summary

From ECG Signals to Representative-Morphology Heatmaps for Biometric Recognition

Athanasios Angelakis, Marta Gomez-Barrero

arXiv:2609.28183v1cs.CVeess.SP

TL;DR

ECG-image biometric performance depends on how waveforms are rendered, and conventional traces are sparse after rasterization. The paper compares matched traces and representative-morphology heatmaps built from the same ECG samples. Heatmaps improve both FNMR operating points and both identification ranks in all 15 compact model-dataset comparisons, while mean EER improves in 14.

  • Problem

    Image-based ECG recognition depends on rendering, while conventional traces are sparse after rasterization and controlled comparison requires unchanged ECG morphology.

    Source: §1 · p. 2

  • Method

    The method averages representative beats into a matrix and renders that same ECG summary as either a conventional trace or dense cardiac-time-by-lead heatmap.

    Source: §1 · p. 2, §6 · p. 23

  • Results

    Heatmaps improve both FNMR operating points and both identification ranks in all 15 compact model-dataset comparisons, while mean EER improves in 14.

    Source: §6 · p. 23

  • Takeaways & Limitations

    Representative-morphology heatmaps provide a practical bridge between ECG signal processing and computer vision, supporting compact CPU models and larger pretrained architectures.

    Source: §6 · p. 23

    The protocol does not require gallery and probe samples to come from separate recording sessions, so long-term cross-session and cross-device performance remains to be established.

    Source: §5.6 · p. 23

Abstract

from arXiv · show

Electrocardiography (ECG) contains subject-specific morphology that supports biometric recognition, yet image-based performance depends on how the waveform is rendered. We introduce representative-morphology heatmaps, a deterministic ECG-to-image representation adapted from ECGXtractor. Within each block of ten aligned beats, the five beats closest to the block mean are averaged into a 400 by L matrix and rendered either as a conventional trace or as a dense cardiac-time-by-lead heatmap. Since both representations contain identical physiological samples, their comparison isolates the effect of rendering. We evaluate verification and closed-set identification on PTB, ECG-ID, and MIMIC-IV-ECG-DEMO. Five compact models, including ZACH-ViT, are trained from scratch, while six ImageNet-pretrained CNN and transformer backbones assess model scale and visual transfer. Heatmaps improve both FNMR operating points and both identification ranks in all 15 compact model-dataset comparisons, while EER improves in 14. Across the matched experiments, EER decreases by 9.59 percentage points and Rank-1 increases by 24.69 points on average. ConvNeXt-Tiny reaches 2.43% EER on PTB and 5.79% on ECG-ID, whereas DeiT-Base reaches 14.92% on MIMIC-DEMO. ImageNet initialization clearly benefits the two multilead datasets but has a mixed effect on ECG-ID, and performance does not increase monotonically with model size. The best heatmap systems approach the strongest signal-domain EER on PTB and ECG-ID, while DeiT-Base provides the strongest evaluated performance on MIMIC-DEMO. Lead-channel ablation further shows that useful channel combinations depend on the cohort and biometric task. Overall, representative-morphology heatmaps provide an effective image representation for ECG verification and identification.

1 Introduction

The paper tests whether ECG-image rendering affects biometric recognition by comparing traces and dense heatmaps generated from the same representative morphology. Across 15 compact model-dataset comparisons, heatmaps improve both FNMR operating points and both identification ranks.

1 Introduction

  • The method averages the five beats closest to each ten-beat block mean into a 400 by L matrix, rendered as either a trace or a dense time-by-lead heatmap.Both renderings use the same physiological samples, so the comparison focuses on representation.Source: §1 · p. 2
  • The study evaluates verification and closed-set identification on three subject-disjoint datasets using five compact models and six ImageNet-pretrained backbones.Matched scratch and pretrained ConvNeXt-Tiny and DeiT-Base experiments assess initialization.Source: §1 · p. 2
  • Heatmaps improve both FNMR operating points and both identification ranks in all 15 compact model-dataset comparisons.The study also benchmarks six ImageNet-pretrained backbones, signal-domain methods, and lead-channel and image-degradation analyses.Source: §1 · p. 2
  • The evaluation also compares heatmaps with four signal-domain methods and examines lead-channel ablations and image degradation.These analyses complement the controlled trace-versus-heatmap experiment.Source: §1 · p. 2

2 Related Work and Positioning

Prior ECG biometric work spans signal-based methods, ECG-derived images, and compact or pretrained vision models. This study positions its contribution as a representative-morphology heatmap evaluated against a matched trace made from the same numerical source.

2.1 ECG biometric recognition

  • ECG biometric research includes representative morphology templates, while ECGXtractor selects and averages five central beats from ten aligned beats for numerical recognition.Other work motivates direct signal comparisons and analysis of lead utility across cohorts and biometric tasks.Source: §2.1 · p. 3

2.2 ECG-to-image representations

  • Prior ECG images use different signal encodings, whereas this work maps acquired leads against cardiac time and compares a heatmap with a trace from the same matrix.A one-lead ECG-ID heatmap provides a control without inter-lead structure; the table caption identifies Table 1 as a harmonized cohort and protocol overview.Source: §2.2 · p. 3, Table 1 · p. 4

2.3 Compact models, model scale, and visual transfer

  • Compact architectures test the representation with fewer than 0.6M parameters, while larger pretrained CNNs and transformers probe model scale and visual transfer.The study includes ABMIL, TransMIL, Compact Transformers, and ZACH-ViT [21] [22] [23], alongside EfficientNet, ResNet, DeiT, and ConvNeXt.Source: §2.3 · p. 4

3 Materials and Methods

The methods use subject-disjoint cohorts and matched ECG representations, then construct summary segments by averaging morphologically central beats. Preprocessing and sample-key controls prevent identity and sample leakage across partitions and representations.

3.1 Datasets and harmonized cohorts

  • The harmonized cohorts use fixed, subject-disjoint identity partitions, with training on training identities and biometric evaluation on unseen test identities.Validation and test data are not used for checkpoint or hyperparameter selection; representation manifests have one-to-one sample-key correspondence and exclude identity and sample leakage.Source: §3.1 · p. 4

3.2 ECG preprocessing and summary-segment construction

  • ECGs are processed at 500 Hz, R-peaks are detected using one synchronized reference channel, and 400-sample windows are extracted around retained peaks.PTB and MIMIC-DEMO beats are normalized per lead; ECG-ID uses its prepared single-lead pipeline, with full-recording Lead-I normalization before peak detection and extraction.Source: §3.2 · pp. 4–5
  • Each summary segment averages the five beats closest to the mean of a non-overlapping ten-beat block, and the same matrix feeds all compared representations and baselines.Incomplete remainders are discarded, and each beat contributes to at most one segment.Source: §3.2 · p. 5

3.3 Deterministic ECG-to-image transformation

The ECG summary is rendered as either a stacked trace or a time-by-lead heatmap, then encoded by compact or ImageNet-pretrained vision models.

3.3 Deterministic ECG-to-image transformation

  • The same 400 × L summary is rendered as a stacked waveform trace or a heatmap with time, lead identity, and normalized amplitude encoded spatially and by colour.Both renderings are converted to 224 × 224 RGB images; the single-lead ECG-ID heatmap expands its row without adding physiological information.Source: §3.3 · p. 5

3.4 Compact image models

  • Five compact architectures span attention, patch, transformer, convolutional-tokenization, and ZACH-ViT designs, all below 0.6M reference-classifier parameters.ZACH-ViT uses 196 patches without a class token or positional embedding; compact models are trained from scratch for 97 epochs using five seeds.Source: §3.4 · pp. 5–6, Table 2 · p. 7

3.5 ImageNet-pretrained heatmap scale benchmark

  • The pretrained heatmap benchmark spans six CNN and transformer backbones with 4.01M–85.80M extractor parameters, fine-tuned with ImageNet-1K weights.Each model is fully fine-tuned for 23 epochs; its classifier is discarded and the pre-logit representation is used as the biometric embedding.Source: §3.5 · p. 6, Table 3 · p. 7

3.6 Matched initialization ablation

The matched ablation isolates initialization in paired models, while signal-domain comparators apply several scoring and learning approaches to the same ECG summaries.

3.6 Matched initialization ablation

  • Matched ConvNeXt-Tiny and DeiT-Base pretrained–scratch pairs vary only initialization, with data, training protocol, embedding extraction, and evaluation held fixed.Each architecture contributes 15 pairs; comparisons with compact models are not controlled scaling experiments because their training schedules differ.Source: §3.6 · p. 7

3.7 Signal-domain comparators

  • Signal-domain comparators include direct cosine scoring of flattened summaries and the released ECGXtractor verifier and feature extractor.ECGXtractor uses its official 12-lead model on PTB and MIMIC and Lead I on ECG-ID.Source: §3.7 · p. 7
  • Two learned signal comparators are a compact two-layer Conv1D network and a single InceptionTime-style network, trained for 97 epochs with five seeds.Both use identity-supervised cross-entropy and the same Adam learning rate and batch size.Source: §3.7 · p. 8

3.8 Biometric protocols and metrics

The paper fixes verification and identification comparisons, then uses exhaustive channel scoring and controlled image corruptions to probe robustness and scope.

3.8 Biometric protocols and metrics

  • Verification samples up to three mated and 15 non-mated comparisons per test subject, without requiring separate recording sessions.EER and FNMR at 10% and 1% target FMR are reported; closed-set identification uses disjoint gallery and probe samples and reports Rank-1 and Rank-5.Source: §3.8 · p. 8

3.9 Lead-channel scoring ablation

  • Lead-channel scoring evaluates all 4095 non-empty subsets, but removes channels only after full multilead beat selection, so it does not simulate reduced-electrode acquisition.Cross-dataset subset-rank correlations are descriptive because the subsets overlap and are not independent.Source: §3.9 · p. 8

3.10 Heatmap degradation stress test

  • The heatmap stress test evaluates ZACH-ViT and Compact-TransMIL under eight noise, blur, JPEG, and resolution-degradation conditions without retraining.Corruption results are averaged across conditions within each seed, then summarized over five seeds.Source: §3.10 · p. 9

4 Results

Matched heatmaps consistently improve compact-model recognition, while pretrained-model performance, signal-domain comparisons, and useful lead combinations vary by dataset and task. The tested compact models also remain stable under synthetic image degradation.

4.1 Dense heatmaps improve matched trace results

  • EER decreases by 9.59 percentage points and Rank-1 increases by 24.69 points on average across matched trace–heatmap experiments.Heatmaps improve both FNMR operating points and both identification ranks in all 15 comparisons; mean EER improves in 14.Source: §4.1 · p. 9, Figure 2 · p. 10

4.2 Pretrained heatmap scale benchmark

  • ConvNeXt-Tiny has the lowest mean EER on PTB and ECG-ID at 2.43 ± 0.50% and 5.79 ± 0.89%, while DeiT-Base leads MIMIC-DEMO at 14.92 ± 3.08%.Performance is not monotonically related to model size.Source: §4.2 · p. 12, Table 5 · p. 12, Figure 3 · p. 13

4.3 ImageNet initialization helps the multilead cohorts but is mixed for one lead

  • ImageNet initialization improves EER and Rank-1 for ConvNeXt-Tiny and DeiT-Base on PTB and MIMIC-DEMO, but its ECG-ID effects are limited and metric-dependent.The initialization comparisons use a matched 23-epoch protocol.Source: §4.3 · p. 14, Figure 4 · p. 15

4.4 Heatmaps are competitive with direct signal-domain methods

  • Heatmap EER approaches the strongest signal-domain results on PTB and ECG-ID, while DeiT-Base leads four MIMIC-DEMO metrics and ties for Rank-5.Signal methods remain strongest for strict-FMR verification and Rank-1 on ECG-ID.Source: Table 7 · p. 17, Figure 5 · p. 18, §4.4 · p. 18

4.5 Lead-channel scoring is cohort- and task-dependent

  • The best channel combinations differ by dataset and task: PTB benefits from multiple channels, while selected MIMIC-DEMO single channels can match or improve the full-channel result.Verification-optimal combinations differ more across datasets than identification-optimal combinations.Source: §4.5 · pp. 18–19, Figure 6 · p. 19

4.6 Limited average sensitivity to synthetic image degradation

  • Across eight tested image degradations, mean EER changes remain small for Compact-TransMIL and ZACH-ViT, while Rank-1 decreases slightly.The tested corruptions include noise, blur, compression, and resolution changes.Source: §4.6 · p. 19, Figure 7 · p. 21

5 Discussion

The discussion attributes the matched gains to the dense rendering of representative ECG morphology, while describing dataset-dependent pretraining and lead-channel effects. It also identifies limits on cross-session evidence and reduced-sensor conclusions.

5.1 Representative morphology as an image

  • The trace and heatmap share the same representative matrix, so matched comparisons hold ECG samples fixed while testing image encoding.ECG-ID's single-lead result indicates that dense rasterization can help without inter-lead structure.Source: §5.1 · p. 21

5.2 Visual pretraining and model size

  • Pretraining improves both tested architectures on multilead datasets, but its ECG-ID effect is smaller and varies by metric; performance does not rise monotonically with model size.The proposed explanation involving richer joint time-by-lead structure is plausible but not causally established.Source: §5.2 · pp. 21–22

5.3 Comparison with signal-domain methods

  • Heatmap models are competitive with signal baselines, but direct signal methods remain strongest for strict-FMR verification and Rank-1 on ECG-ID.On MIMIC-DEMO, DeiT-Base is best on four metrics and tied for best on Rank-5.Source: §5.3 · p. 22

5.4 Lead-channel contribution

  • Channel utility depends on both dataset and biometric objective, and the strongest subsets are not fully portable across datasets.Verification and identification favour different combinations.Source: §5.4 · p. 22

5.5 Image degradation and deployment

  • Compact-TransMIL and ZACH-ViT remain stable under tested noise, blur, compression, and resolution changes, with Rank-1 slightly more sensitive than EER.Compact models support CPU training and small memory footprints, while larger pretrained models provide stronger absolute performance on several datasets.Source: §5.5 · p. 22

5.6 Limitations

  • The protocol does not separate gallery and probe recordings by session, and post-preprocessing channel removal tests scoring contributions rather than reduced-sensor acquisition.Long-term cross-session and cross-device performance remains to be established.Source: §5.6 · p. 23

6 Conclusion

The conclusion presents representative-morphology heatmaps as a matched image representation for ECG biometrics. Results show broad compact-model gains, while pretrained performance and useful channels depend on dataset and task.

6 Conclusion

  • The method converts a 400×L representative-morphology matrix into a dense time-by-lead image and compares it with a trace from the same physiological samples.This matched design isolates the effect of image representation.Source: §6 · p. 23
  • Heatmaps improve both FNMR operating points and both identification ranks in all 15 compact model-dataset comparisons, while mean EER improves in 14.ZACH-ViT is described as competitive with substantially larger architectures while offering a favourable accuracy–model-size balance.Source: §6 · p. 23
  • ImageNet initialization benefits PTB and MIMIC-DEMO but has mixed effects on ECG-ID; useful lead subsets also depend on dataset and biometric objective.ConvNeXt-Tiny reaches 2.43% EER on PTB and 5.79% on ECG-ID, while DeiT-Base reaches 14.92% EER on MIMIC-DEMO.Source: §6 · p. 23

Data and Reproducibility Statement

The statement identifies dataset access terms and a reproducibility package containing manuscript, code, protocol, and result materials. The appendices describe complete matched and pretrained operating-point tables.

Data and Reproducibility Statement

  • PTB, ECG-ID, and MIMIC-IV-ECG-DEMO are available under their respective access terms, and the manuscript redistributes no source ECG recordings.Source: Data and Reproducibility Statement · p. 23
  • The reproducibility package includes manuscript source, figure code, protocol summaries, seed-level results, and provenance records with source-archive hashes.Code for preprocessing, rendering, training, and evaluation is stated to be released upon acceptance; dataset access remains governed by providers.Source: Data and Reproducibility Statement · p. 23

A Complete matched trace-heatmap operating points; B Complete pretrained heatmap benchmark

  • Appendix A describes complete compact-model trace–heatmap operating points, and Appendix B describes the six-model ImageNet-pretrained heatmap benchmark.Both captions specify five-seed mean and sample SD; the pretrained benchmark uses full fine-tuning for 23 epochs.Source: Table 10 · p. 24, Table 11 · p. 25

C Complete matched initialization ablation

The section reports a matched initialization ablation and exhaustive lead-subset selection, with different selection criteria for verification and identification.

C Complete matched initialization ablation

  • Table 12 compares initialization at matched 23-epoch training, changing only initialization within each model–dataset pair.Values are percentages reported as mean ± sample SD over five training seeds.Source: Table 12 · p. 26

D Best exhaustive subset at each lead count

  • At each lead count, the manuscript-selected subset minimizes EER for verification and maximizes Rank-1 for identification.The table caption specifies selection after exact recomputation.Source: Table 13 · p. 27
  • Table 13 covers the exact selected subset at every lead count.Source: Table 13 · p. 27

E Cross-dataset subset concordance

The section reports Spearman rank concordance between PTB and MIMIC-DEMO for identical lead subsets across subset sizes; nominal p values are qualified by subset dependence.

E Cross-dataset subset concordance

  • Table 14 reports Spearman rank concordance between PTB and MIMIC-DEMO for identical lead subsets at each subset size.Source: Table 14 · p. 28
  • The table treats conventional p values as nominal descriptors because exhaustive, overlapping subsets are not independent random observations.Source: Table 14 · p. 28
  • The concordance comparison spans multiple subset sizes using identical lead subsets across the two datasets.Source: Table 14 · p. 28
Loading 2609.28183v1…