Weighted Temporal Decay Loss for Learning Wearable PPG Data with Sparse Clinical Labels

TL;DR

This paper tackles the problem of sparse clinical labels and temporal misalignment between wearable PPG data and lab biomarker measurements. It introduces a biomarker-specific weighted temporal decay loss, where each sample is weighted by $w_i = g(\hat{\alpha}_b\Delta t_i)$ with $\hat{\alpha}_b = \mathrm{softplus}(\alpha_b)$, and trains a model using $\mathcal{L}_{\text{weighted}} = \frac{1}{N}\sum_i w_i\mathrm{BCE}(\hat{y}_i,y_i) - \lambda\frac{1}{N}\sum_i w_i$; inference uses the base network without weighting. The approach is validated on 450 participants with smartwatch PPG across 10 biomarkers, showing improvements over a strong self-supervised baseline and a Random Forest, with Linear decay providing the most robust performance. The learned decay rates offer interpretable insight into how quickly each biomarker’s PPG evidence becomes stale, advancing data-efficient, non-invasive continuous monitoring in wearable healthcare and highlighting avenues for future cross-domain validation and adaptive time horizons.

Abstract

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.

Figures (1)

• Figure 1: The proposed method uses a weighted decay loss function, parameterized by $\Delta t$ -- the temporal distance between the nearest sparse clinical label and the corresponding sensor data -- to progressively reduce the contribution of the samples as they occur further from the clinical health records.