Bridging the Gap in ECG-Based Emotion Recognition: A Unified Evaluation of Deep Learning Models
Timothy C Sweeney-Fanelli, Ajan Ahmed, Masudul Imtiaz
eess.SP · cs.LG
Quick verdict
Unified benchmark of five deep models for ECG emotion recognition; accuracies near chance (51.8–57.0%) and model rankings lack statistical testing.
Summary
The paper tackles inconsistent evaluation practices in ECG-based automated emotion recognition (AER) by benchmarking five deep architectures under a single pipeline. The authors introduce ARRC, a Keras-based benchmarking framework, and ARDT, a dataset toolkit, and merge ASCERTAIN, CUADS, and DREAMER into one corpus of 3,216 trials from 119 participants, arguing that inter-dataset training tests generalization where conventional cross-dataset validation tests only adaptability. Three published models (CNN-LSTM, Temporal CNN, PETSFCNN) plus a 1-D CNN and ResNet baseline are trained for binary arousal classification with a combined contrastive/cross-entropy loss, participant-level 80/20 splits, undersampling, and 10-fold cross-validation. The Temporal CNN leads (accuracy 0.570, AUC 0.570) at only 178K parameters and 20M FLOPs; PETSFCNN (0.535) and CNN-LSTM (0.518) underperform; a CUADS-only experiment scores higher (T-CNN 0.598), supporting the claim that intra-dataset validation overestimates generalization. Main limitations: all accuracies sit barely above the 50% chance floor; results are fold means without standard deviations, confidence intervals, or significance tests; hyperparameter 'optimization' reduces to undocumented initial grid searches; the best model and newest dataset are the authors' own prior contributions; and CNN-LSTM is evaluated on ECG input although designed for IBIs, which the authors concede needs re-tuning. Code and pipelines are stated to be open-source.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under a unified pipeline (identical preprocessing, losses, splits, 10-fold CV), the Temporal CNN achieves the best arousal classification accuracy (0.570) and AUC (0.570) with the smallest compute footprint (178K parameters, 20M FLOPs).
- Inter-dataset training across ASCERTAIN+CUADS+DREAMER is a more stringent test of generalization than within-dataset evaluation, which overestimates performance (Table IV: CUADS-only T-CNN 0.598 vs 0.570 inter-dataset).
- Architectural complexity alone does not improve performance: CNN-LSTM (0.518) and PETSFCNN (0.535) underperform smaller/simpler models.
- The open-source ARRC and ARDT frameworks enable reproducible, standardized AER benchmarking and inter-dataset training.
Strengths
Genuinely controlled comparison design
All five architectures share one signal chain (600/200 ms median-filter baseline removal, 12th-order Butterworth 35 Hz low-pass, 256→128 Hz downsample, z-score, 45 s pad/trim; Section III), a common sigmoid head with joint contrastive+BCE loss, and pre-generated participant-level splits reused across models (Section V.A) — differences in Table III are attributable to architecture, not data handling, which is exactly the confound Section II.E criticizes in prior work.
Inter-dataset generalization framing with an empirical contrast
Merging ASCERTAIN+CUADS+DREAMER (3,216 trials, 119 participants, mixed sensors/conditions) is a stricter test than single-dataset CV, and the small CUADS-only experiment (Table IV: T-CNN 0.598 vs 0.570 inter-dataset) gives direct evidence that intra-dataset validation inflates scores — the paper's most transferable finding.
Unusually candid reporting
Efficiency-aware evaluation
Reproducibility infrastructure
Weaknesses
No statistical uncertainty for the central ranking
Near-chance absolute performance limits discriminative value
Hyperparameter 'optimization' overclaimed and undocumented
CNN-LSTM evaluated outside its design regime
Self-preferential benchmark elements
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review — Bridging the Gap in ECG-Based Emotion Recognition: A Unified Evaluation of Deep Learning Models (arXiv 2609.15055; accepted IEEE UEMCON 2026)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy did the ECG emotion recognition models achieve in this benchmark?
On a merged corpus of ASCERTAIN, CUADS, and DREAMER (3,216 trials, 119 participants), 10-fold mean accuracy/AUC were: Temporal CNN 0.570/0.570, ResNet 0.545/0.550, 1-D CNN 0.544/0.540, PETSFCNN 0.535/0.540, and CNN-LSTM 0.518/0.520 — all barely above the 50% chance level of the balanced binary arousal task.
Which datasets were combined and how large is the merged dataset?
The authors merged three public AER datasets — ASCERTAIN, CUADS, and DREAMER — using their ARDT toolkit into a single corpus of 3,216 trials from 119 participants, then balanced the high/low arousal classes by undersampling before training.
Is the code and data for this paper available?
The paper states that all source code (the ARRC benchmarking framework, the ARDT dataset toolkit, and the evaluation scripts) is publicly available on GitHub, though the URLs are withheld in this version for double-blind review; the three datasets are publicly available, with CUADS being the authors' own 2025 release.
Why are the reported accuracies so low (around 50-57%)?
The authors attribute the low scores to using pre-validated stimulus video labels rather than participants' actual felt emotion (which shows only moderate inter-rater agreement), plus the added difficulty of generalizing across different sensors, recording conditions, and demographics from three datasets; their CUADS-only control experiment scored higher (0.559–0.598), supporting this explanation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Table I extraction is garbled (model/dataset/metric values interleaved with body prose), so the original papers' reported scores could not be fully verified; Figures 1-5 (architectures, ROC curves) could not be analyzed from the extracted text.
- ARRC/ARDT source URLs are withheld for double-blind review, so the claimed code availability and the pre-generated splits could not be verified.
- Table III values were reconstructed from flattened extraction text; column alignment was cross-checked against the prose in Sections VI.A-VI.D but pagination-level fidelity is not guaranteed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.