Respiratory Status Detection with Video Transformers

The embedding-based comparison strategy demonstrates merit, achieving F1=0.75 compared to 0.58–0.69 for two-tower cross-attention approaches, while requiring substantially less training time. The addition of LieRE and MGM yields incremental improvements (F1 rising from 0.75 to 0.81), consistent with the claimed benefits of these methods for spatiotemporal modeling. The temporal-ordering formulation is a clever solution to the lack of fine-grained labels.

The test set size (7 videos, 3 participants) is critically small for drawing statistical conclusions. The 30% exclusion rate (23 of 75 participants excluded due to insufficient resolution, poor framing, or subjective breathlessness assessment) introduces selection bias. The use of healthy volunteers recovering from exercise—rather than actual patients with pathological respiratory distress—creates a substantial domain gap; the authors acknowledge perspiration as a confounding signal that may correlate with exercise recovery but not clinical dyspnea.

The claim of fully automated processing is undermined by participant self-positioning in controlled recordings. Critically, the paper lacks any ablation comparing the video transformer approach to simpler baselines (e.g., optical flow statistics, 2D CNNs, or even frame-level ResNet features), making it impossible to assess whether the architectural complexity is warranted. The performance drops to 55–65% accuracy for small temporal differences (Section 4), suggesting limited sensitivity for subtle but clinically relevant changes.

The comparison to Nawaz et al. is fair in noting the limitations of manual cropping, but the authors fail to compare their method against the actual clinical standard—human clinician assessment—or against simpler video architectures like TimeSformer or even frame-level models. The F1=0.81 metric lacks context: without knowing the class balance in the 7-video test set, this could be misleading. The relationship between temporal ordering performance and actual respiratory distress detection in clinical populations remains theoretical; the authors correctly note this as future work but the title and framing suggest clinical applicability that is not demonstrated.

Reproducibility is severely limited. No code, dataset, or trained model weights are made available. Hyperparameters are sparsely documented: training capped at 5 epochs and <100 hours, but learning rate, batch size, optimizer, and warmup schedule are omitted. The LieRE implementation details (number of learned matrices, initialization) are unspecified. The participant exclusion criteria rely on subjective judgment from a single author ('author TS could not reliably determine the presence of shortness of breath'), which is neither replicable nor objective. The weak supervision approach using first/last third split introduces sensitivity to clip duration thresholds that is not characterized.