⚡ electrical engineering

Explaining RhythmFormer: A Systematic XAI Analysis of Periodic Sparse Attention for Remote Photoplethysmography

This paper bridges the gap between visual plausibility and auditable evidence in remote photoplethysmography (rPPG) by adapting four attribution methods to RhythmFormer and introducing quantitative metrics for skin coverage and faithfulness, ultimately demonstrating that the "Beyond Intuition" method achieves superior spatial alignment and perturbation faithfulness compared to standard approaches.

Original authors: Louis Chen, Torbjörn E. M. Nordling

Published 2026-06-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Louis Chen, Torbjörn E. M. Nordling

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Black Box" Doctor

Imagine a new type of doctor that can measure your heartbeat just by watching a video of your face. It doesn't need a watch or a chest strap; it just looks at the subtle color changes in your skin caused by blood pumping. This is called Remote Photoplethysmography (rPPG).

The specific "doctor" this paper studies is called RhythmFormer. It's very smart and accurate, but it's also a "black box." We know it gives the right answer, but we don't know why it thinks what it thinks. Is it looking at your forehead? Your cheeks? Or is it secretly looking at your red shirt or the background?

In medicine, you can't just trust a black box. You need to know it's looking at the right place (your skin) to be trustworthy. This paper is like an investigation to open that black box and see what the AI is actually focusing on.

The Problem: The "Top-K" Filter

RhythmFormer has a special trick to be fast. Imagine it has a team of 640 tiny scouts looking at different parts of your face. To save time, the boss tells them: "Only the top 40 most interesting scouts get to report back. The other 600 are ignored."

This is called Sparse Attention (or "Top-K selection").

The Goal: Focus only on the most important skin areas.
The Risk: What if the "boss" ignores a scout who was actually important? Or what if the scouts talk to each other later and bring back information from the 600 ignored scouts anyway?

The Investigation: Four Different Flashlights

The researchers wanted to see where the AI was looking. They used four different "flashlights" (methods) to shine light on the AI's decision-making process. Think of these as four different ways to trace a path through a maze:

Raw Attention: A simple snapshot. It shows who the boss talked to right now.
Attention Rollout: A cumulative map. It traces the path from start to finish, assuming that if Scout A talks to Scout B, and Scout B talks to Scout C, then Scout A is indirectly responsible for Scout C's work.
Attention Flow: A water-pipe analogy. It calculates how much "information water" flows from the input (your face) to the output (the heartbeat number) through the pipes.
Beyond Intuition: The "Super-Flashlight." This is a fancy new method that combines the path-tracing of Rollout with a check on how much the AI actually cares about the information (using gradients). It tries to fix the mistakes of the other flashlights.

The Findings: Leaky Pipes and Hidden Truths

1. The "Leaky Pipe" Surprise (Multi-hop Leakage)

The researchers found a weird glitch with the Rollout and Flow flashlights.

What happened: Even though the boss explicitly told the AI to ignore 600 scouts (setting their importance to zero), the Rollout method showed that those ignored scouts were still "talking" to the final result through a chain of other scouts.
The Analogy: Imagine a teacher tells a student, "Don't look at the back of the room." But then, the student looks at the front, who looks at the middle, who looks at the back. The Rollout method sees this chain and says, "Hey, the student is still looking at the back!"
The Result: This "leakage" meant that the Rollout method showed the AI was paying attention to non-skin areas (like the background) much more than it actually was. It diluted the signal.

2. The Winner: "Beyond Intuition"

The Beyond Intuition flashlight was the best.

Why? It realized that just because a scout is in the chain doesn't mean they are important. It checked if the information actually mattered before counting it.
The Score: It kept the focus tightly on the skin (the forehead and cheeks).
- Skin Coverage: 83% of its attention was on valid skin.
- The others: Only about 57% (Rollout) or less.
The Verdict: Beyond Intuition is the most honest flashlight. It didn't get tricked by the "leaky pipes."

3. Does Looking at Skin Mean Better Heart Rates?

The researchers asked: "If the AI looks more at the skin, does it get a better heart rate?"

The Waveform: Yes. When the AI focused more on the skin, the shape of the heartbeat wave it predicted was cleaner and more accurate.
The Heart Rate Number: Not really. On the specific dataset they used (UBFC-rPPG), the AI was already so good at guessing the heart rate that it didn't matter much if the focus was perfect or slightly off. The "ceiling" was too high to see the difference.
The Takeaway: The paper suggests that in noisier, harder situations (like if you are moving around), focusing on the skin would probably matter much more.

4. The "Hand-Over-Eye" Test

To prove their tools worked, they found one video where the person rubbed their eye (a "glitch" or artifact).

The Problem: The AI got confused, and the "Beyond Intuition" flashlight gave a very weird, negative score, saying, "This doesn't make sense!"
The Fix: They edited the video to remove the hand rubbing the eye.
The Result: Suddenly, all four flashlights agreed that the video was now "normal" and the scores went back to being high and positive.
Why this matters: It proves that their tools can actually detect when something is wrong with the input, not just when the math is right.

Summary

This paper built a toolkit to check if an AI heart-rate monitor is looking at your face or just guessing.

They found that some common ways of explaining AI (Rollout) are "leaky" and show the AI looking at things it's actually ignoring.
They found that a newer method (Beyond Intuition) is much better at spotting the real skin areas.
They proved that when the AI looks at the skin, the heartbeat wave it predicts is cleaner, even if the final number is already pretty good.

Bottom Line: If you want to trust an AI doctor, you need the right flashlight. Beyond Intuition is currently the best one for this specific job, helping us see that the AI is indeed looking at your skin, not your shirt.

Technical Summary: RhythmFormer: A Systematic XAI Analysis of Periodic Sparse Attention for Remote Photoplethysmography

Problem Statement

Remote Photoplethysmography (rPPG) using deep learning, particularly transformer-based architectures like RhythmFormer, has achieved low heart-rate error on benchmarks. However, as these systems move toward clinical deployment, their decision-making processes remain opaque. Existing Explainable AI (XAI) approaches for rPPG are limited by:

Qualitative Dominance: Reliance on visual heatmap inspection without quantitative faithfulness metrics.
Lack of Physiological Grounding: Absence of validation against known physiological signal sources (e.g., facial skin regions).
Methodological Gaps: Raw attention maps are contested as faithful proxies for model focus, as highlighted regions may not actually drive predictions.
Sparse Attention Complexity: RhythmFormer utilizes a "bi-level routing" mechanism with top-k selection, creating sparse attention matrices. Standard cumulative attribution methods (like attention rollout) may fail to account for how these sparsity constraints interact across layers.

Methodology

Model and Architecture

The study focuses on RhythmFormer, a single-stream transformer for rPPG. Its core innovation is the Periodic Sparse Attention mechanism within Hierarchical Temporal Periodic Transformer (TPT) blocks:

Bi-Level Routing: Queries and keys are projected and pooled to generate coarse scores. A top-k selection (k=40) filters regions, retaining only the highest-scoring areas.
Refined Attention: Standard scaled dot-product attention is applied only to the gathered key-value pairs.
Structure: The model consists of three hierarchical TPT levels, each containing two blocks, totaling six attention layers.

XAI Framework

The authors adapt and apply four attribution methods to RhythmFormer's sparse architecture:

Raw Attention: Column-wise aggregation of attention weights.
Attention Rollout: Recursive multiplication of augmented attention matrices to capture transitive token relationships.
Attention Flow: Modeling attention as a flow network to compute maximum flow between input and output tokens.
Beyond Intuition (BI): A hybrid method combining attention rollout with integrated gradients. It decomposes attribution into:
- $P(L)$ : W-corrected attention rollout (re-weighting tokens by their value-projection magnitude to correct for information flow distortion).
- $F^c$ : Integrated gradient mask of the loss with respect to the last-layer attention.
- Final attribution: $T = P(L) \odot F^c$ .

Adaptations for RhythmFormer: The authors reconstruct sparse matrices as dense $N \times N$ matrices (with zeros for excluded regions) to enable matrix operations. They also adapt the baseline for integrated gradients to a per-clip mean-intensity video rather than a black screen to maintain distributional validity for rPPG.

Evaluation Metrics

Two quantitative metrics are introduced to move beyond visual inspection:

Skin Coverage: The ratio of attention mass falling on physiologically valid skin regions (face and neck) versus the total attention mass. This is calculated using a pre-trained BiSeNet face parser.
SaCo Faithfulness: Adapted from classification to rPPG regression. It measures the consistency between the salience ranking of pixel groups and the prediction impact (measured as Mean Absolute Error, MAE, between original and perturbed rPPG waveforms). A higher SaCo score indicates that masking high-salience regions causes a larger prediction error, confirming the explanation's faithfulness.

Key Results

1. Multi-Hop Leakage in Sparse Attention

A critical finding is the "multi-hop leakage" effect in standard attention rollout.

Observation: Although individual refined-attention layers explicitly zero out ~50% of off-diagonal entries via top-k selection, the cumulative rollout product restores these connections.
Quantification: Approximately 48% of the total off-diagonal attribution mass in plain rollout comes from entries that were zeroed out in at least one layer.
Consequence: This leakage causes the "refined" rollout to have lower skin coverage than the "pre-attention" stage, contradicting the intuition that sparsity should improve focus on skin.

2. Performance of Beyond Intuition

Beyond Intuition (BI) emerges as the strongest candidate for this setting:

Mitigation of Leakage: The W-correction in BI's $P(L)$ component attenuates the multi-hop leakage.
Skin Coverage: BI achieves the highest median refined skin coverage (0.83) compared to vanilla rollout (0.57) and attention flow (0.46).
Faithfulness: BI achieves the highest median SaCo faithfulness score (0.92) on the UBFC-rPPG dataset.
Mechanism: The combination of value-projection weighting and the gradient-supported mask effectively restricts attribution to regions the model genuinely relies on, avoiding the "re-filled" zeros of standard rollout.

3. Correlation with Prediction Quality

Skin Coverage vs. Waveform Quality: There is a significant positive correlation between skin coverage and waveform reconstruction fidelity (Pearson $r$ ).
Skin Coverage vs. Heart Rate Error: A weak negative trend exists between skin coverage and Heart Rate (HR) Mean Absolute Error (MAE). The authors attribute the weakness of this correlation to the saturated performance of the UBFC-rPPG dataset (controlled conditions, high SNR), where even imperfect waveforms yield accurate HR estimates, masking the impact of spatial attention.
Faithfulness vs. Accuracy: SaCo faithfulness scores do not correlate significantly with prediction accuracy (MAE or waveform $r$ ). This suggests that a model can be faithful (relying on specific regions) even if those regions do not yield the most accurate prediction, or vice versa.

4. Case Study: Artifact Ablation

In a case study involving a clip with hand occlusion (Subject 46, Clip 11), BI initially produced a strongly negative SaCo score ( $F = -0.53$ ). After replacing the occluded frames with stable neighboring frames:

All four methods showed a recovery in SaCo scores toward $F \approx 0.8$ .
Skin coverage also improved.
This consistency across algorithmically distinct methods suggests SaCo behaves robustly as a metric for detecting attribution misalignment caused by artifacts.

Significance and Claims

The paper claims to bridge the gap between visual plausibility and auditable evidence in rPPG XAI. Its primary contributions are:

Systematic Framework: Providing the first quantitative comparison of multiple attribution methods specifically adapted for sparse, top-k routing transformers in rPPG.
Identification of Leakage: Demonstrating that standard cumulative attribution methods (rollout, flow) suffer from "multi-hop leakage" in sparse architectures, potentially misleading interpretations of model focus.
Method Recommendation: Identifying Beyond Intuition as the most robust method for RhythmFormer, as it explicitly mitigates leakage and achieves superior alignment with physiological priors (skin coverage) and faithfulness metrics.
Metric Adaptation: Successfully adapting the SaCo faithfulness coefficient from classification to rPPG regression, enabling quantitative evaluation of explanation quality.

The authors maintain a modest stance, noting that their findings are specific to the UBFC-rPPG dataset and RhythmFormer architecture. They emphasize that the weak correlation between skin coverage and HR error may be an artifact of the dataset's high quality and that further validation across diverse, noisy datasets is required to generalize these findings.