TRACE: Expert-Aligned ECG Representation Learning with Rigorous Benchmarking and Real-World Validation in Acute Cardiac Care

1 Department of Computer Science, Emory University, GA, USA

• The affiliation names the Institute for Medical Engineering and Science at MIT in Massachusetts.The passage begins with the affiliation number 3.Source: §2 · p. 1

6 School of Medicine, Emory University, GA, USA

• The listed keywords are ECG representation learning, multimodal learning, multi-task learning, acute coronary occlusion, and arrhythmia classification.Source: §6 · p. 2

1. Introduction

• Existing multimodal methods can conflate ECG patterns with irrelevant report content, while contrastive alignment misses fine-grained signal structure and reconstruction lacks semantic grounding.The introduction also notes that combining complementary objectives can be difficult because of scale mismatch, allowing one objective to dominate optimization.Source: §1 · p. 3
• TRACE uses an LLM-based pipeline to extract ECG-specific findings from cardiologist reports and filter reports lacking interpretable cardiac content.Source: §1 · p. 4
• A shared encoder jointly optimizes masked autoencoder reconstruction and contrastive multimodal alignment, balanced through learned task uncertainties.Source: §1 · p. 4
• TRACE is evaluated on public benchmarks for structural abnormality detection and arrhythmia classification, and on a large private ACO dataset.The introduction frames TRACE as a specialized representation model for structured predictions in high-stakes, time-sensitive triage rather than open-ended narrative reasoning.Source: §1 · p. 4
• TRACE maintains state-of-the-art performance on public benchmarks and significantly outperforms real-world clinical practice in ACO detection.Source: §1 · pp. 4–5
• The paper presents its contribution as the curation, integration, and clinical validation of established techniques for ECG interpretation, rather than isolated algorithmic invention.Source: §1 · p. 5

2.1.1 MIMIC-IV ECG Dataset

• 609,272 ECG-report pairs from 105,293 patients remain after linking MIMIC-IV-ECG waveforms with cardiologist reports and excluding records with missing data.The source dataset contains 800,035 waveforms and 623,566 reports.Source: §2.1.1 · p. 5

2.1.2 Report Curation Pipeline

• The three-stage Qwen3-32B curation pipeline filters non-informative reports, extracts ECG-attributable findings, and standardizes terminology to address noisy, extraneous report content.Extraction retains findings such as rhythm, axis, and ST-T changes while removing medication, imaging, and historical information.Source: §2.1.2 · pp. 5–6
• 494,252 high-quality ECG-report pairs remain after removing 115,020 samples (19%), providing ECG-attributable supervision such as rhythm, axis, conduction abnormalities, and ST-T morphology.The retained observations are intended to keep contrastive learning from binding ECG embeddings to information not inferable from the tracing.Source: §2.1.2 · p. 6

2.2.1 Overview

• TRACE shares an ECG encoder across masked autoencoder reconstruction of signal patches and contrastive alignment of global ECG representations with matched clinical text.The paired objectives target local signal structure and global clinical semantics.Source: §2.2.1 · p. 7

2.2.2 ECG Encoder

• The ECG encoder embeds non-overlapping temporal patches with lead and position information, processes them through transformer blocks, and uses global and patch representations for alignment and reconstruction.For a 10-second, 500-Hz recording with patch size 50, the passage specifies 1,200 patches.Source: §2.2.2 · pp. 7–8

2.2.3 Masked Autoencoder Reconstruction; 2.3.1 Masked Autoencoder Reconstruction Loss

• Masked autoencoder training encodes visible patches, uses a lightweight decoder to reconstruct masked patches, and computes an L1 reconstruction loss only at masked positions.The asymmetric design is intended to encourage rich encoder representations rather than relying on decoder capacity.Source: §2.2.3 · p. 8, §2.3.1 · p. 9

2.2.4 Text Encoder and Projection; 2.3.2 Contrastive Multimodal Alignment Loss

• A frozen pretrained biomedical text encoder and modality-specific two-layer projectors map ECG and report embeddings into a shared space trained with symmetric InfoNCE at temperature 0.07.The projected representations are unit-normalized and compared by cosine distance.Source: §2.2.4 · pp. 8–9, §2.3.2 · p. 10

2.3.3 Uncertainty-Weighted Multi-Task Loss

• Because reconstruction losses range from 0.02 to 0.05 while contrastive losses range from 4 to 8, TRACE learns task-specific uncertainty weights rather than naively summing or fixing weights.Learned log-variances downweight more uncertain tasks, with logarithmic regularizers preventing unbounded uncertainty.Source: §2.3.3 · p. 10

2.4 Training Parameters

• TRACE combines LLM-based report curation, shared-encoder masked reconstruction and contrastive alignment, and uncertainty-weighted balancing; its final architecture and training differ substantially from MERL.MERL’s released codebase provided the starting point for the contrastive alignment branch.Source: §2.4 · p. 11

2.5.1 Benchmarking TRACE on Public Datasets

• Public benchmarking covers PTB-XL, CPSC2018, and CSN, comparing TRACE with 15 published unimodal and multimodal baselines under official splits.Evaluation uses linear probing and fine-tuning, with AUROC averaged across five random seeds.Source: §2.5.1 · p. 12

2.5.2 Clinical Evaluation on ACO Detection

• TRACE is compared with clinical practice at matched false positive rate and matched sensitivity to quantify sensitivity gain and false positive rate reduction.These complementary operating points reflect rule-in and rule-out trade-offs in practice.Source: §2.5.2 · p. 13
• Qualitative clinical analyses use t-SNE to examine class separability and input-gradient saliency maps on randomly selected true-positive ACO cases.Saliency values are min-max normalized across leads and smoothed with a 10-time-step rolling average.Source: §2.5.2 · p. 14
• The clinical baseline’s sensitivity was 0.781 and false positive rate was 0.503, measured against expert-adjudicated ACO labels and the treating team’s actual STEMI catheterization-lab activation decision at presentation.TRACE results were reported with 10-fold cross-validation, each fold using a 70/10/20 train/validation/test split.Source: §2.5.2 · pp. 14–15

3.1 TRACE Benchmark Performance

• TRACE achieved 96.0% AUROC on PTB-XL rhythm, 95.5% on CSN, 95.0% on CPSC2018, and 93.3% on PTB-XL Superclass under finetuning.PTB-XL Subclass reached 92.6% AUROC under linear probing; MedCPT outperformed a nearly 300× larger Qwen3-32B text encoder.Source: §3.1 · p. 15

3.2 Comparison with Unimodal Self-Supervised Methods

• TRACE outperformed the best unimodal method across all benchmarks and ECGFounder across all five public benchmarks despite ECGFounder’s larger labeled-ECG training set.On PTB-XL Super, adding contrastive alignment to MAE-only pretraining raised AUROC from 0.8623 to 0.9264, a +6.4-point gain.Source: §3.2 · pp. 16–17

3.3 Comparison with Multimodal ECG-Text Methods

• TRACE outperformed K-MERL by 2.81 points and MERL with ResNet by 3.97 points on PTBXL-Super, with gains reported across evaluated benchmarks.The authors attribute the gains to combining MAE-learned local morphology with text-aligned semantics and curated reports.Source: §3.3 · p. 17

3.4.1 TRACE Improves ACO Rule-in and Rule-Out

• At the clinical baseline’s sensitivity of 0.781, TRACE reduced FPR from 0.503 to 0.188; at FPR 0.503, sensitivity rose to 0.929, a 19.0% relative gain.The clinical baseline had sensitivity 0.781 and FPR 0.503; the matched-sensitivity result was a 62.6% reduction in false positives.Source: §3.4.1 · p. 18
• TRACE identified 90 additional ACO patients over six years; initially misclassified as less-urgent NSTEMI, they had mean time-to-catheterization of 20.8 hours versus 0.81 hours for correctly identified ACO, and delay was associated with 26.0% higher MACE incidence.Source: §3.4.1 · p. 18

3.4.2 Physiologic Grounding and Interpretability

• TRACE embeddings showed moderate intrinsic separation between ACO and non-ACO samples without task-specific supervised training.The authors suggest this reflects learned fundamental ischemic features.Source: §3.4.2 · p. 19
• Lead-wise saliency extended beyond ST-segment changes to the QRS complex and T-wave, including morphology associated with conduction slowing and early ischemic changes.The reported salient features included terminal QRS morphology, R-wave progression, and repolarization abnormalities.Source: §3.4.2 · p. 19
• Salient regions in reciprocal and anatomically opposing leads indicate that the model integrates global vectorial and reciprocal electrical patterns characteristic of acute coronary occlusion.These findings reinforce the view of ACO as a complex pathophysiologic entity characterized by subtle, multi-segment ECG abnormalities rather than a discrete ST-elevation phenotype.Source: §3.4.2 · pp. 19–20
• TRACE’s architecture captures nuanced ECG features often overlooked by the conventional ST-elevation paradigm, explaining its superior performance.Source: §3.4.2 · p. 20

4.1 Two paradigms for combining ECG and clinical text

• Text-as-supervision models use reports only during pretraining and take ECGs alone to produce structured predictions at inference, unlike text-as-input systems that generate free-text answers.TRACE belongs to the text-as-supervision paradigm and has no generative decoder.Source: §4.1 · p. 20

4.2 Relation to general-purpose reasoning models

• TRACE targets structured, decision-ready predictions for high-stakes 12-lead ECG triage, whereas general-purpose reasoning models target open-ended explanatory and exploratory use.Without a generative decoder, TRACE cannot produce free-text reasoning or be evaluated on language-generation benchmarks such as ECG-QA.Source: §4.2 · p. 21

4.3 Relation to imaging-paired cross-modal ECG models

• Unlike TRACE’s cardiologist-report pairing, related imaging-paired models use chest X-ray, cardiac magnetic resonance, or echocardiography with ECGs.These imaging modalities are seldom available when an ECG is acquired in prehospital and emergency settings, while paired report text is abundant and requires no co-registered imaging.Source: §4.3 · p. 21

4.4 Concurrent self-supervised ECG models

• A direct numerical comparison with concurrent self-supervised ECG encoders is not faithful because their cohorts, label taxonomies, and data splits differ from TRACE’s evaluation configurations.The authors identify reimplementation and reevaluation under identical protocols as necessary for such a comparison.Source: §4.4 · p. 22

4.5 Limitations

• Clinical validation uses a single county-level EMS cohort and a closed private ACO dataset; external multisite and prospective validation is needed before deployment.The clinical-practice baseline is retrospective rather than from a controlled trial.Source: §4.5 · p. 22
• TRACE is a structured-output representation model, so it does not provide the open-ended explanations or interactive reasoning offered by generative ECG-language models.The authors frame this as a scope boundary rather than a fundamental barrier.Source: §4.5 · pp. 22–23
• TRACE’s text-only multimodal supervision omits complementary imaging, and the model is specialized to standard 12-lead, 10-second ECGs rather than other cardiac data or acquisition settings.The authors identify prospective evaluation and controlled integration of additional modalities as future work.Source: §4.5 · pp. 22–23

5. Conclusion

• TRACE combines LLM-based curation of noisy narratives into expert-grounded supervision with uncertainty-weighted learning of fine-grained morphology and global semantics.The hybrid architecture jointly optimizes the two scales of representation.Source: §5 · p. 23
• TRACE maintains state-of-the-art performance on public benchmarks and significantly outperforms real-world clinical baselines in ACO detection.The conclusion presents this as a result of evaluation across public benchmarks and acute cardiac care.Source: §5 · p. 23
• The authors present TRACE as a scalable pathway for expert-aligned AI in high-stakes acute cardiac triage.The conclusion connects this potential application to improving triage and patient outcomes.Source: §5 · p. 23

A. Downstream Dataset Details

• The downstream datasets include PTB-XL’s 12-lead recordings and three classification tasks, CPSC2018, Chapman-Shaoxing-Ningbo, and a private prehospital ACO cohort.PTB-XL includes 18,885 patients and Superclass, Subclass, and Rhythm tasks; the ACO data come from a county-level EMS system.Source: A. Downstream Dataset Details · p. 33

B. LLM Prompt for Report Curation

• The report-curation pipeline uses a Qwen3-32B prompt to extract ECG-specific findings from raw clinical reports while removing noise and uncertainty language.The prompt specifies a cardiologist persona and explicit decomposition rules.Source: B. LLM Prompt for Report Curation · pp. 33–34

Prompt Design Principles

• Prompt design combines a clinical-expertise role, filtering rules for uncertainty and temporal references, and few-shot examples of transforming narratives into findings.Source: Prompt Design Principles · p. 34

Complete Prompt

• The prompt instructs the model to act as a cardiologist and extract ECG findings into base clinical terms.Source: Complete Prompt · p. 34

4. Extract only ECG-attributable observations: rhythm, rate, axis,

• Extraction is restricted to ECG-attributable observations, including rhythm, rate, axis, intervals, ST-T changes, conduction abnormalities, and morphology.Source: §4 · p. 34

6. Output as comma-separated lowercase terms

• The output is comma-separated lowercase terms, as illustrated by converting a report into terms such as “sinus rhythm” and “first degree av block.”Source: §6 · p. 35

Exclusion Criteria

• Reports yielding the [EXCLUDE] token are removed when they contain no new ECG findings, only administrative or non-ECG information, or illegible or incomplete entries.Source: Exclusion Criteria · p. 36

Deployment and Compute

• The curation pipeline ran on-premise with locally hosted open-source weights and no external or commercial API; it processed 609,272 linked ECG reports in batches of 100.Qwen3-32B ran on four NVIDIA A100 GPUs, with checkpoint-based resumption for fault tolerance.Source: Deployment and Compute · p. 36

C. Clinical Terminology Dictionary

• Terminology normalization expands abbreviated LLM findings such as “RVH” and “RBBB” into full clinical expressions to standardize vocabulary for ECG-text alignment.The clinical dictionary maps SCP codes to clinical term variations.Source: C. Clinical Terminology Dictionary · pp. 36–37

D. Ablation Study

• TRACE’s hybrid CMA+MAE approach outperforms either component alone, with the pattern also reported for ACO detection and Chapman arrhythmia classification.On PTBXL-Super under linear probing, MAE alone achieves 0.8623 AUROC and CMA alone achieves 0.9196 AUROC.Source: Table A.1 · p. 37, D. Ablation Study · p. 37

• At 90% masking, performance partially recovers on ACO, but PTB-XL Super continues to underperform; subsequent experiments use 75% masking.Extreme masking shifts the encoder toward global context and contrastive alignment rather than local reconstruction.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 39
• The MAE loss is 0.02–0.05 while CMA is 4–8, so naive summation lets CMA dominate; fixed weighting offers marginal improvement but requires extensive search.Fixed weighting also lacks adaptability to changing loss dynamics across training epochs.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 39
• Dynamic Weight Averaging is unstable with disparate objective magnitudes, and gradient-based balancing mechanisms fail when loss scales differ by two orders of magnitude.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 39
• TRACE’s uncertainty weighting consistently outperforms alternatives, with gains persisting under fine-tuning and more robust pretrained representations than heuristic balancing schemes.It adapts to changing task difficulties and supports stable joint optimization despite the objectives’ scale mismatch.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 40
• Learned homoscedastic uncertainty parameters model observation noise and use precision terms to give lower optimization weight to objectives with higher intrinsic uncertainty.The probabilistic approach eliminates manual tuning while adapting to changing task difficulties.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 40
• MedCPT outperforms Qwen3-32B by 2.6–4.6 AUROC points (mean 3.7) on every linear-probing task and five of six fine-tuning tasks.Qwen3-32B surpasses MedCPT only on CPSC2018 fine-tuning; the gap is largest under linear probing and narrows under fine-tuning.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 41
• The authors use Qwen3-32B for generative report curation and decomposition, while retaining a retrieval-contrastive encoder for ECG–report alignment.They attribute MedCPT’s advantage to its alignment-compatible embedding geometry, rather than model scale.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 41
• PubMedBERT performed comparably to MedCPT and slightly below it under linear probing, supporting retrieval-pretrained over generic biomedical encoders.Source: LP FT LP FT LP FT LP FT LP FT LP FT · p. 41