信号变乱时，模型到底还能相信多少？How physiological models behave when the signal gets messy.

01

单作者 MPhil 研究 · 可穿戴 PPGSingle-author MPhil work · Wearable PPG

怎样知道一次心率估计是否可信Knowing when a heart-rate estimate is trustworthy

运动会让脉搏、动作和谐波峰同时落进心率范围。CITRUS 不急着压成一个数字，而是保留多个候选，沿时间做因果追踪，并在证据不足时暂不输出。Motion can put pulse, movement, and harmonic peaks in the same heart-rate range. CITRUS keeps several candidates alive, follows them causally, and withholds an output when the evidence is weak.

• 学习排序器从每个窗口约 140 个候选中选择；因果 Viterbi 解码抑制不合理跳变。A learned ranker chooses among roughly 140 candidates per window; causal Viterbi decoding prevents implausible jumps.
• 接受 / 保持 / 拒绝层把覆盖率与误差之间的取舍直接展示出来。An accept / hold / reject stage makes the trade-off between coverage and error visible.

14.39→10.91

运动 MAE · BPMmotion MAE · BPM

6.2 BPM

最佳 50% 覆盖best 50% coverage

156 ms

ESP32-S3 · 每窗口ESP32-S3 · per window

~145 KB

模型大小model size

02

Scientific Data 2026 · HealthRing 数据集Scientific Data 2026 · HealthRing dataset

HealthRing：面向指环健康感知的生理数据集HealthRing: Physiology Dataset for Health Sensing on Rings

HealthRing 同步采集指环 PPG、加速度与临床参考信号，覆盖静息、日常活动和跑步。我作为共同作者参与模型实验、数据检查和 Android 端验证。HealthRing synchronizes ring PPG, accelerometry, and clinical references across rest, daily activity, and running. As a co-author, I contributed model experiments, data checks, and Android-side validation.

• 团队数据集包含 54 名成年参与者和 49.76 小时记录。The team dataset contains 49.76 hours from 54 adults.
• 正式论文把采集协议、传感器布局、生理参考与基线评测放在同一记录中。The published record connects the collection protocol, sensor layouts, physiological references, and baseline evaluation.

Sci Data

期刊journal

2026

已发表published

PPG/ACC

信号signals

HR/RR

参考references

03

单作者预印本 · 非 EEG 睡眠分期Single-author preprint · Non-EEG sleep staging

可穿戴兼容信号究竟能做到什么？How far can wearable-compatible signals go?

这项受控分解研究比较消费级与实验室级心肺信号究竟能恢复多少睡眠结构，并检查系统在什么情况下应该放弃低置信度时段。This controlled decomposition asks what consumer and laboratory cardiorespiratory signals can honestly recover about sleep architecture, and when a system should abstain.

• 所有信号组使用同一个紧凑 Mamba2 模型，避免把换模型带来的收益算到信号上。The same compact Mamba2 model is used across signal groups so gains are not hidden by a model change.
• 置信度报告用于检验：暂不判断最不确定的时段后，保留下来的结果是否真正更可靠。Confidence-aware reporting tests whether withholding uncertain epochs produces a meaningfully better subset.

0.255

消费级 HR / ACC κconsumer HR / ACC κ

0.492

心肺信号 κcardiorespiratory κ

0.796

EEG / EOG κEEG / EOG κ

.452→.512

拒识 20% 后after withholding 20%