Weighted Temporal Decay Loss for Learning Wearable PPG Data with Sparse Clinical Labels
This work addresses the challenge of label sparsity for health monitoring using wearable PPG data, offering an incremental improvement in prediction accuracy and interpretability for biomarker analysis.
The paper tackles the problem of sparse clinical labels in wearable PPG data by introducing a training strategy that weights samples based on time gaps to labels, improving biomarker prediction accuracy. It achieves an average AUPRC of 0.715, outperforming baselines like a fine-tuned self-supervised method (0.674) and Random Forest (0.626).
Advances in wearable computing and AI have increased interest in leveraging PPG for health monitoring over the past decade. One of the biggest challenges in developing health algorithms based on such biosignals is the sparsity of clinical labels, which makes biosignals temporally distant from lab draws less reliable for supervision. To address this problem, we introduce a simple training strategy that learns a biomarker-specific decay of sample weight over the time gap between a segment and its ground truth label and uses this weight in the loss with a regularizer to prevent trivial solutions. On smartwatch PPG from 450 participants across 10 biomarkers, the approach improves over baselines. In the subject-wise setting, the proposed approach averages 0.715 AUPRC, compared to 0.674 for a fine-tuned self-supervised baseline and 0.626 for a feature-based Random Forest. A comparison of four decay families shows that a simple linear decay function is most robust on average. Beyond accuracy, the learned decay rates summarize how quickly each biomarker's PPG evidence becomes stale, providing an interpretable view of temporal sensitivity.