RhythmFormer's New XAI Framework Reveals Attention Leakage in rPPG Models

Remote photoplethysmography (rPPG) transformers have achieved low heart-rate error on benchmarks, but their decision-making process remains opaque—a growing concern as these models move toward clinical applications. Existing XAI for rPPG relies on qualitative heatmap inspection without quantitative faithfulness metrics or physiology-grounded validation. To bridge this gap, researchers Chen and Nordling systematically analyze RhythmFormer, a transformer with bi-level routing attention and top-k selection. They adapt four attribution methods—raw attention, rollout, flow, and Beyond Intuition—and introduce two novel metrics: a skin coverage metric that quantifies how much attribution mass falls on skin regions, and an adapted SaCo faithfulness coefficient for regression using the MAE between original and perturbed rPPG waveforms.

Applying these tools on the UBFC-rPPG dataset, the team uncovers a multi-hop leakage effect under sparse top-k routing: attention rollout and flow almost completely restore the connections that individual refined-attention layers explicitly set to zero. Beyond Intuition mitigates this leakage via its value-projection-weighted rollout and gradient-supported mask, achieving the highest median refined skin coverage (0.83 vs. 0.57 for vanilla rollout) and faithfulness (F=0.92). A case study on a low-SaCo outlier shows consistent recovery across all four methods once an artefactual region is replaced. While validation across diverse datasets and model variants is still needed, these metrics shift rPPG XAI toward auditable numerical evidence about spatial alignment and perturbation faithfulness—key for trustworthy clinical deployment.