Learning from Bilateral PPG via Cross-Site Attention and Knowledge Distillation

Accurate heart rate estimation from wrist-worn photoplethysmography (PPG) during sports is challenging due to motion artifacts and asymmetric signal quality across wrists. Existing methods typically rely on single-wrist sensing and hand-crafted quality measures, overlooking complementary bilateral information. We propose an end-to-end bilateral PPG framework that jointly models both wrists using dual PapaGei encoders, SQI-conditioned temporal gating, and cross-wrist self-attention to emphasize cleaner signals. Evaluated on a bilateral sports dataset spanning six activities and 30 subjects with ECG ground truth, the model achieves 2.85 bpm MAE and r = 0.96, outperforming unilateral baselines. To improve deployability, we further introduce a knowledge distillation framework that transfers bilateral motion-compensation priors into a single-wrist EfficientNet-1D student. The distilled student achieves 2.79 bpm and 3.72 bpm MAE on the non-dominant and dominant wrist, respectively, surpassing corresponding unilateral baselines while requiring only one sensor at inference. These results establish bilateral PPG as an effective training paradigm for robust single-wrist heart rate estimation under motion.