Skip to main content
QUICK REVIEW

[论文解读] Weighted Temporal Decay Loss for Learning Wearable PPG Data with Sparse Clinical Labels

Yunsung Chung, Keum San Chun|arXiv (Cornell University)|Feb 2, 2026
Non-Invasive Vital Sign Monitoring被引用 0
一句话总结

论文提出一种加权时间衰减损失,基于时间间隙在 PPG 片段与真值标签之间学习 Biomarker 特异的样本权重衰减,从而在时序上相隔的标签也能学习,并提升智能手表 PPG 的多 Biomarker 预测性能。

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.

研究动机与目标

  • 解决基于 PPG 的 biomarker 预测中稀疏临床标签的挑战。
  • 引入 biomarker 特异的可学习时间衰减率,以按时间间隔对训练样本加权。
  • 在来自智能手表数据的 10 个 biomarker 上展示相对于基线的性能提升。
  • 通过在训练中利用时间上远离的信号来展示数据效率。
  • 通过分析学习到的每个 biomarker 的时间敏感性来提供可解释性。

提出的方法

  • 定义一个加权损失,其中样本权重 w_i = g(alpha_b_hat * Delta_t_i),其中 Delta_t_i 为离最近健康记录的天数,alpha_b_hat = softplus(alpha_b) >= 0。
  • 使用带权二分类交叉熵并加一个平均权重奖金来防止权重退化(lambda 固定为 0.5)。
  • 探索四种衰减族(指数、线性、余弦退火、反比),并选择线性在平均意义上的鲁棒性。
  • 使用 450 名参与者的 10 秒 PPG 片段,经质量筛选,且与最近的临床记录在 30 天内对齐,按被试者进行 5 折交叉验证进行评估。
  • 对比基于手工特征的随机森林,以及在所提出损失下微调的 PAPAGEI。
Fig. 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 r
Fig. 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 r

实验结果

研究问题

  • RQ1当临床标签稀疏时,是否存在一种时序感知加权方案能改进 PPG 的 biomarker 预测?
  • RQ2哪种衰减族在多种 biomarker 上提供最鲁棒的性能?
  • RQ3学习一个 biomarker 特异的衰减率是否在固定速率或无时间加权的基础上带来额外收益?
  • RQ4在两端极端标签(顶四分位与底四分位)设定下,该方法的表现如何?
  • RQ5潜在的局限性及影响性能的 biomarker 特异性动态是什么?

主要发现

  • 在所有 biomarker 的平均 AUROC 为 0.712、AUPRC 为 0.715,优于 PAPAGEI(AUROC 0.660、AUPRC 0.674)和 RF(AUROC 0.599、AUPRC 0.626)。
  • 线性衰减在平均性能上表现最佳(AUROC 0.712、AUPRC 0.715)。
  • 学习 biomarker 特异的衰减率相较于固定速率提供了额外的增益。
  • 时序感知加权减少对远距离片段的依赖,并改善对每个被试的片段预测聚合。
  • 两端极端协议下,快速变化的 biomarker(如 WBC 和 Potassium)尤其取得显著增益。

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。