Review Papers Claim by Claim

1) Scientific contribution & novelty

The paper's main novelty is systematic evaluation of CWT-derived scalogram and phasogram representations (magnitude + phase) at 12 leads, testing both early and late fusion strategies across modern backbones (ResNet50, Swin, EfficientNet, hybrid CNN–Transformer) and a strong 1D baseline on PTB-XL. This is a meaningful contribution because phase information is under-explored in many ECG DL works and multi-representation early-fusion at full 12 leads is not widely benchmarked on PTB-XL. The incremental performance gain over a strong raw-signal baseline is small (ΔAUC ≈ 0.0014) but consistent across ensemble strategies, which suggests complementary information rather than wholesale replacement of time-domain modeling Chato & Kagozi 2026 (fusion benefit)

2) Methodological rigor

• Positives: Uses standard PTB-XL patient-wise stratified folds (train/val/test 8:1:1), shares code link, documents preprocessing (band-pass 0.05–47 Hz, per-lead z-score, percentile clipping for scalograms), uses modern backbones and ensembles, and reports inference times (ms/sample), which is valuable for deployment considerations Chato & Kagozi 2026 (methods)
• Concerns / gaps: Authors do not provide class-wise AUC/F1/confusion matrices or calibration (reliability) curves for minority super-classes — crucial in clinical multi-label settings where per-class sensitivity matters. Also, while code is shared, exact random seeds, weight initializations, and full hyperparameter sweeps are not exhaustively documented — mildly reducing reproducibility. External validation beyond PTB-XL is absent, so claims about generalizability remain untested (authors acknowledge this). These are standard weaknesses in DL-for-ECG literature and limit clinical translation evidence Chato & Kagozi 2026 (limitations)

3) Signal-processing choices — strengths and risks

The use of Morlet CWT and separate magnitude/phase images is methodologically sound and aligns with prior wavelet-based ECG analyses and delineation work that shows wavelets capture ECG morphology robustly across noise conditions Martinez et al. 2004 (wavelet ECG delineation).

However, compressing 24 channels (early-fusion scalogram+phasogram) to 3 channels via a 1×1 conv adapter for transfer-learning backbones risks information loss; the CWT2DCNN that natively supports 12 or 24 channels is preferable but the paper uses both strategies — the impact of channel compression on performance is not fully quantified beyond net AUCs. That potential fidelity loss is important because lead-specific phase patterns may carry subtle diagnostic cues (authors note mapping but don't show per-lead attribution maps) Chato & Kagozi 2026 (channel compression).

4) Performance claims & interpretation

Observed effect sizes are small: the best ensemble yields AUC 0.9238 vs raw baseline 0.9224. This is statistically small and may be within run-to-run variability; the authors strengthened the claim using focal loss weighting to improve minority labels, but do not provide confidence intervals or repeated-seed experiments to show significance. Therefore, the claim that wavelet fusion materially improves diagnostic accuracy must be interpreted cautiously until cross-run variability, per-class gains, or external validation are demonstrated Chato & Kagozi 2026 (results detail).

5) Reproducibility & openness

Positive: code repository linked (https://github.com/kagozi/MultiModal-EC) and PTB-XL is public (Wagner et al., Sci. Data 2020), which enables replication if the repository contains preprocessing/training scripts and model weights Wagner et al. 2020 (PTB-XL dataset)

Negative: the paper lacks some hyperparameter detail (exact seed(s), optimizer weight decay and scheduler settings, full augmentation/augmentation seed lists) and per-class metrics; the authors state some training hyperparameters (LR ranges, batch sizes) but not full reproducible experiment scripts in the methods section — this reduces reproducibility unless the GitHub repo includes fully runnable pipelines and environment specs.

6) Clinical usefulness and next steps

Practicality: inference times reported (raw model ~1.4 ms/sample; ensemble ~22 ms/sample) are promising for near-real-time use but deployment needs per-class sensitivity, false positive rates, calibration, and prospective clinical evaluation. To move towards clinical use, authors should produce: (i) per-class sensitivity/specificity and confusion matrices by label; (ii) external validation on at least one independent dataset (e.g., CPSC, PTB old dataset, or hospital data); (iii) robustness analyses across leads, noise levels, and demographic strata; and (iv) interpretability maps (saliency/Grad-CAM per lead-frequency band) to show which scalogram/phasogram regions drive decisions Chato & Kagozi 2026 (deployment notes).