MoWaveQFormer: A Motion-Conditioned Quality-Gated Transformer for Smartphone-Based PPG Heart Rate Estimation

Motivation

• Free-living motion produces distinct spectral corruption patterns, but many methods ignore accelerometer signals or use them without distinguishing motion-induced artifacts.Existing deep models commonly apply one transformation across windows, despite activity-specific spectral signatures.Source: §I · p. 1
• BUT PPG v2.0 labels recordings only as good or bad, which can force HR models to discard borderline-quality segments and their partial physiological information.The paper motivates differentiable internal quality weighting as an alternative to hard accept/discard decisions.Source: §I · p. 1
• The study adds PTT consistency to ECG-derived HR supervision because PTT encodes information about arterial compliance and should remain consistent across short windows.The introduction identifies this as an underused physiological constraint in contemporary PPG HR estimation.Source: §I · p. 1

Contributions

• MoWaveQFormer uses motion-stratified filtering, quality-gated encoding, and a PTT consistency constraint alongside ECG-based HR supervision.Its three stages are a rule-based motion conditioner, a motion-conditioned wavelet filter bank, and a quality-gated Transformer encoder.Source: §I · p. 2
• The proposed framework showed significant improvements over three of four representative motion-agnostic baselines, with no significant difference from a Q-PPG-inspired ResNet1D baseline.The reported comparison uses a Wilcoxon signed-rank test with p < 0.05 for the significant improvements.Source: §I · p. 2

Overview

• In the reported experiments, motion groups come from recorded activity annotations; the accelerometer-derived classifier is evaluated separately as a candidate for annotation-free deployment.R-peak annotations are used only during training to compute the PTT target.Source: A. Overview of MoWaveQFormer Framework · p. 2
• The three-stage pipeline selects a motion-specific filter bank, converts PPG into eight sub-bands, then quality-gates Transformer patch tokens before pooled HR regression.The training objective combines ECG-supervised HR regression with a PTT-consistency term.Source: A. Overview of MoWaveQFormer Framework · p. 2, Figure 1 · p. 2
• Motion information explicitly selects PPG filters before feature encoding, rather than being concatenated with motion features into a shared embedding.This places motion conditioning at the start of representation processing.Source: A. Overview of MoWaveQFormer Framework · p. 2
• Each PPG channel is min-max normalized within its window to [−1, 1], avoiding dataset-level statistics that could leak information across the subject-level split.The normalization also stabilizes the learnable filters and embedding projection.Source: A. Overview of MoWaveQFormer Framework · p. 2

B. Motion Conditioning Module

• The rule-based classifier avoids end-to-end training because coughing and laughing account for only 7.4% of training windows, which the authors judged insufficient for reliable learning.They cite the risk of fitting subject-specific idiosyncrasies rather than motion physics.Source: B. Motion Conditioning Module · p. 3
• The accelerometer classifier uses spectral-energy ratios and fixed thresholds to assign walking, burst, or default subtle/rest motion groups.Walking uses the 1–2.5 Hz stride band; burst uses 4–15 Hz energy relative to the 0.5–4 Hz baseline, with absolute-energy floors for both branches.Source: B. Motion Conditioning Module · p. 3
• Held-out accelerometer classifications agree with annotated activity groups in 56% of cases, reflecting difficulty separating subtle motions using accelerometer statistics alone.The subtle/rest group includes rest, elevated sensor pressure, finger movement, lighting change, and talking.Source: B. Motion Conditioning Module · p. 3

C. Motion-Conditioned Learnable Wavelet Filtering

• The filter bank replaces one globally optimized decomposition with learnable filters indexed by motion group and jointly optimized with the HR regression objective.The design is motivated by activity-dependent artifact spectra and prior observations that one global parameterization can underperform across heterogeneous motion regimes.Source: C. Motion-Conditioned Learnable… · p. 3
• For each of three motion groups, eight learnable 31-sample FIR filters and motion-specific sigmoid gains transform normalized PPG into sub-bands.The selected filter set is applied by zero-padded convolution; its output replaces raw PPG as the Stage 3 encoder input.Source: C. Motion-Conditioned Learnable… · p. 3

D. Quality-Guided Transformer Representation Learning

• The eight sub-bands are split into 30 patches, flattened, projected to 128 dimensions, and combined with learnable positional encodings.Each patch contains 80 values from 8 sub-bands across 10 samples.Source: D. Quality-Guided Transformer… · p. 3
• A learned sigmoid gate reweights each patch embedding from the binary quality label without zeroing low-quality tokens.The gate is applied before the encoder, so low-quality tokens remain available as keys and values in self-attention.Source: D. Quality-Guided Transformer… · p. 3
• A four-layer Transformer processes the gated tokens, whose outputs are globally averaged and passed through a two-layer regression head to predict HR.The encoder uses four attention heads, model dimension 128, feedforward dimension 512, and dropout 0.1.Source: D. Quality-Guided Transformer… · p. 3

E. Physiological Constraint and Training Objective

• Measured PTT is the mean delay from each ECG R-peak to the first subsequent PPG peak, restricted to 50–600 ms.Windows without a valid match are excluded from the PTT term but retain HR supervision.Source: E. Physiological Constraint and… · p. 3
• The training objective combines L1 error against ECG-derived HR with a PTT consistency loss weighted by λ_PTT = 0.1.The PTT loss penalizes the absolute discrepancy between predicted and measured delay for valid targets.Source: E. Physiological Constraint and… · p. 4
• The PTT term adds no trainable parameters and regularizes predictions against the arterial-delay characteristics of the paired ECG–PPG recording.The training procedure applies the combined loss and updates model parameters while keeping the motion index non-differentiable.Source: E. Physiological Constraint and… · p. 4

A. Dataset and Preprocessing

• BUT PPG v2.0 contains 3,888 ten-second recordings from 50 subjects across eight activity conditions, with smartphone PPG and reference ECG for each record.Synchronized tri-axial chest accelerometer data are available for 3,840 records.Source: A. Dataset and Preprocessing · p. 4
• Across the dataset, 78.7% of windows are labeled low quality and 21.3% high quality.The dataset passage identifies this imbalance as motivating the soft quality gate.Source: A. Dataset and Preprocessing · p. 4
• During training, each sample independently has a 0.5 probability of Gaussian noise augmentation and a 0.5 probability of a circular shift of up to ±10 samples.No augmentation is applied during validation or testing.Source: A. Dataset and Preprocessing · p. 4

B. Data Partitioning

• The subject-level split assigns 35 subjects to training, 7 to validation, and 8 to testing; motion groups are strongly imbalanced.The training groups comprise 2,436 subtle/rest windows (88.9%), 102 walking (3.7%), and 204 burst (7.4%); the test groups contain 532, 22, and 44 windows, respectively.Source: B. Data Partitioning · p. 4

C. Evaluation Metrics

• Test performance is measured against ECG-derived HR using MAE, RMSE, and Pearson r, with additional stratification by quality and activity.The evaluation also uses paired Wilcoxon tests and Bland–Altman bias with 95% limits of agreement.Source: C. Evaluation Metrics · p. 4

D. Training Configuration

• The model has 816,445 trainable parameters and is trained for 80 epochs with AdamW; training and evaluation use a cloud-hosted NVIDIA GPU.Inference latency is measured over 200 runs after a 50-iteration warmup on GPU and CPU.Source: D. Training Configuration · p. 4

A. Heart-Rate Estimation Accuracy

• MoWaveQFormer was best among six methods on test-set error and correlation; deep baselines had MAE of 8.2–8.5 bpm, while spectral trackers exceeded 20 bpm.The overall test set contains 598 records.Source: Table II · p. 5, A. Heart-Rate Estimation Accuracy · p. 5
• Improvements over DeepPPG, CNN-BiLSTM, and Harmonic-SP were significant, but the lower mean error versus ResNet1D was not significant (p = 0.319).The Wilcoxon test used paired per-record absolute errors; effect sizes were small versus the two learned CNN baselines.Source: Table III · p. 5, A. Heart-Rate Estimation Accuracy · p. 5

• Representative rest, walking, and coughing windows show visually distinct time- and frequency-domain corruption signatures; at rest, the waveform is clean and periodic, with a dominant spectral peak matching ECG-derived reference rate.Source: B. Motion-Specific Corruption Patterns… · p. 5
• Mean PTT, measured from ECG R-peaks to corresponding PPG peaks, ranges from 0.303 to 0.318 s across eight activity classes, with overlapping standard deviations.Source: B. Motion-Specific Corruption Patterns… · p. 5
• Figure 2 shows representative rest, walking, and coughing windows and spectra, annotates the rest example’s PTT, and plots mean PTT with standard deviations across all eight classes.Source: Figure 2 · p. 5

C. Effect of the Quality Gate

• MoWaveQFormer achieves the lowest MAE on low-quality windows, while ResNet1D is marginally better on high-quality windows; error is substantially higher at SQI=0.Source: Table IV · p. 5, C. Effect of the Quality Gate · p. 5

D. Per-Motion-Class Performance

• MoWaveQFormer has the lowest MAE for walking, coughing, and laughing, but DeepPPG leads on higher-pressure and moving-finger, CNN-BiLSTM on light-change, and ResNet1D on talking.For rest, MoWaveQFormer is comparable to ResNet1D (7.493 versus 7.555).Source: Table V · p. 6, D. Per-Motion-Class Performance · p. 6

E. Ablation Study

• Removing the quality gate or PTT term raises validation MAE by 0.121 or 0.143 bpm, respectively; shared filtering changes it by −0.083 bpm, within run-to-run variation.Source: Table VI · p. 6, E. Ablation Study · p. 6

F. Learned Filter-Bank Analysis

• Walking filters attenuate the stride band slightly more than subtle/rest filters (−13.25 dB versus −12.96 dB), while burst filters attenuate the burst band less (−14.09 dB versus −14.99 dB).Source: Figure 3 · p. 6, F. Learned Filter-Bank Analysis · p. 6

G. Clinical Agreement Analysis

• All eleven errors below −50 bpm occurred on low-quality windows with annotated HR of 120–161 bpm; recomputed QRS-based values were closer to predictions, with CSV-to-recomputed ratios near 2:1.Source: G. Clinical Agreement Analysis · pp. 6–7
• Test-set bias is −5.54 bpm, with 95% limits of agreement of [−29.42, 18.35] bpm; 95.7% of errors fall within those limits.Source: G. Clinical Agreement Analysis · p. 6, Figure 4 · p. 7

H. Computational Efficiency

• MoWaveQFormer has 816,445 trainable parameters and processes a 10-second window in 2.04 ± 0.11 ms on cloud CUDA GPU or 2.63 ± 0.21 ms on CPU.Source: H. Computational Efficiency · p. 7

V. DISCUSSION

• Walking, coughing, and laughing showed the clearest gains, but 88.9% subtle/rest windows dilute motion conditioning’s pooled ablation effect (∆=−0.083 bpm) and remain heterogeneous.Walking artifacts are approximately periodic within 10-second windows, whereas coughing and laughing create brief non-stationary transients that fixed-response filters cannot selectively suppress.Source: §V · p. 7, §VI · p. 8
• With 78.7% low-quality windows, soft gating retains rather than discards data and benefits low-quality windows most; PTT consistency had the largest single-component ablation effect.ResNet1D achieved marginally lower MAE on high-quality windows.Source: §V · p. 7
• The reported results use annotated motion groups; the accelerometer rule classifier agreed with them on only 56% of held-out test data, potentially reducing deployment gains.Misclassification routes windows to a mismatched filter set.Source: §V · p. 7
• Evaluation used one subject-level split of one dataset, and inference latency was measured on cloud infrastructure rather than smartphone hardware.Generalization to other populations, devices, and acquisition protocols remains unevaluated.Source: §V · p. 7

VI. CONCLUSION

• Future work targets finer-grained motion grouping, learned real-time motion classification, cross-dataset validation, and on-device implementation.The proposed taxonomy does not resolve subtle/rest heterogeneity.Source: §V · pp. 7–8
• MoWaveQFormer achieved the best overall performance among six methods, significantly improving over DeepPPG, CNN-BiLSTM, and Harmonic-SP but not ResNet1D.The ResNet1D significance comparison varied across repeated runs, warranting cautious interpretation.Source: §V · p. 7, §VI · p. 8