Evidence level: B
导言
手腕上的智能手表,正在成为最沉默的心血管预警系统。
光电容积脉搏波(Photoplethysmography,PPG)通过LED光源与光电传感器,捕捉皮肤微血管床的血容量搏动变化,无需穿刺、无需袖带,在每秒数十次的采样中编织出一条连续的生理频率曲线。这条曲线中隐藏着心率变异性、动脉顺应性、外周阻力乃至心律失常的密码。
从房颤检测到高血压风险筛查,从心衰失代偿预警到血糖波动推断,PPG信号与深度学习算法的结合,正将消费级可穿戴设备推向医学诊断的边界。本文系统梳理这一融合技术的科学进展、算法架构与中医频率医学的内在联系。
一、PPG信号的基本原理与频率维度
PPG波形由一个收缩峰(systolic peak)和一个舒张峰(diastolic peak)构成,两峰之间的面积——脉搏指数(Pulse Index,PI)——反映外周血流灌注状态。
从频率视角分析,PPG信号可分解为:
- 心率频率:成人静息约1–2 Hz,对应60–120次/分的生理节律
- 呼吸调制频率:约0.2–0.4 Hz,呼吸周期对心率和每搏输出量的调控
- 交感/副交感频率:通过心率变异性(HRV)分析,0.04–0.15 Hz为交感相关,0.15–0.4 Hz为副交感相关
- 更高频谐波:王唯工团队的脏腑谐波C1–C10理论认为,每个脏腑有其特征共振频率,理论上也应在PPG高频谐波中有所映射
PPG相较心电图(ECG)的局限在于:ECG记录心脏电活动,具有明确的时间基准;而PPG是血液动力学的下游效应,信噪比较低,受运动伪迹、皮肤色素、环境光照干扰显著。但其优势同样明显:连续无感监测,采样频率可达100 Hz以上,覆盖率远超任何临床检测手段。
二、算法演进:从特征工程到端到端深度学习
2.1 传统机器学习阶段
早期PPG疾病检测依赖人工提取特征:
– 峰谷检测 → 计算心率
– 波形面积比值 → 评估动脉僵硬度
– 频域功率谱分析 → 量化HRV频段能量
代表性工作(Bangalore et al., IEEE TBME 2020)使用逻辑回归和随机森林,对房颤的检测灵敏度和特异性分别达到约87%和85%,但误报率高,难以满足临床需求。
2.2 卷积神经网络(CNN)阶段
2018–2021年,CNN开始直接处理原始PPG时序信号,自动学习多尺度特征。典型架构为一维CNN + 短时傅里叶变换(STFT)时频图输入:
- 输入:30秒PPG片段,采样率20–100 Hz
- 卷积核:多尺度(3×、5×、7×),捕捉不同时间尺度的波形形态
- 池化层:压缩特征维度
- 输出:正常/房颤/其他心律失常的多分类
该阶段代表性研究(arxiv 1807.10707)报告了AUC 0.9999的房颤检测性能,但这是受控数据集的结果,真实世界噪声环境下性能显著下降。
2.3 混合架构与端到端学习
BayesBeat(arxiv 2011.00753) 引入了贝叶斯深度学习,使用不确定性量化提升模型在噪声PPG数据上的鲁棒性。其核心思路是:模型不仅输出”是否有房颤”,同时输出预测的可信度,从而在高噪声片段主动降级为”不确定”而非误报。
ResNet + Transformer(2024年高血压筛查研究) 直接从手腕PPG数据预测高血压风险,ResNet残差连接解决了深层CNN的梯度消失问题,5折交叉验证AUC超过0.92,且仅用0.124M参数——模型足够小,可部署在手表端侧芯片上。
2.4 多模态融合
头部可穿戴设备(耳机、眼镜)开始融合加速度计(ACC)+ PPG,实现运动伪迹自动分离;医疗级贴片设备融合ECG + PPG双通道,通过ECG R峰标注PPG的准确心搏时刻,实现监督学习降噪。
三、核心应用场景与验证数据
3.1 房颤(Atrial Fibrillation)
Apple Heart Study(Tamburino et al., NEJM 2021)是一项里程碑式研究:41.2万受试者佩戴Apple Watch,PPG算法筛查出0.52%的房颤高风险人群,其中34.6%在后续ECG贴片确认房颤。这一比例看似不高,但在人口基数下意味着数十万早期发现。
意义:房颤是最常见的心律失常性卒中原因,早期发现抗凝治疗可降低约60%的卒中风险。PPG连续监测填补了传统心电图”看得到的时候抓不住”的盲区。
3.2 高血压风险筛查
2024年基于HBPM纵向数据集(448人,68k+ spot-check实例)的研究显示,深度学习模型从PPG中提取的特征可区分正常血压与高血压前期/高血压,AUC达0.92以上。连续手腕PPG数据相较单次诊室血压测量,更能捕捉”隐匿性高血压”和”白大衣高血压”。
3.3 心衰失代偿预警
心衰患者的体液潴留和外周灌注改变,会引起PPG波形形态的系统性偏移。早期研究表明,PPG衍生的每搏输出量指数和全身血管阻力估计,可提前3–7天预警心衰急性失代偿,为远程医疗干预争取时间窗口。
四、与中医脉诊频率医学的深层联系
PPG连续监测与中医脉诊的关系,是一个极富潜力的交叉领域。
频率同构性:PPG记录的是每一次心搏产生的外周血容量波,脉诊捕捉的是寸口脉的浮沉滑涩——两者在生理本质上共享”血液搏动频率”这一核心信息源。中医脉象28病脉,本质上是不同频率特征的专家分类标签。
互补优势:PPG提供客观量化、可追溯的数字信号;中医脉诊提供整体性、动态性的手感经验判断。两者结合,机器学习算法可以从PPG信号中还原类似”弦脉””滑脉”的特征频率模式,用于客观化中医诊断工具。
预警逻辑的一致性:《黄帝内经》”上工治未病”的治疗哲学,与可穿戴设备”在症状出现之前检测生理偏移”的预警逻辑,高度吻合。PPG频率漂移异常,可能对应中医所说的”气机紊乱”可量化信号。
五、技术瓶颈与未来方向
- 噪声鲁棒性:运动伪迹和皮肤接触不良仍是手表PPG的主要挑战。光学传感器封装改进和算法自适应降噪是主攻方向。
- 验证金标准:多数研究在受控环境或特定人群中验证,跨种族、跨设备、跨年龄段的泛化性研究仍不足。
- 临床落地路径:FDA/NMPA对软件医疗器械(SaMD)的审批路径日趋清晰,但消费级PPG算法的医疗级验证标准仍在建立中。
- 多病种联合建模:从单一病种预警走向心血管风险综合评分(整合房颤、高血压、心衰、动脉硬化等多维度),是下一代算法的目标。
证据与局限
支持证据(Evidence Level B):
- 2项系统综述(PMID 35300400、35127882)确认PPG结合深度学习在心血管疾病检测中的技术可行性,AUC普遍>0.90
- Apple Heart Study等大规模人群研究提供了真实世界连续监测的临床价值初步验证
- 多项研究(BayesBeat、ResNet+Transformer)在公开数据集上验证了算法性能
局限性:
- 大多数高质量研究来自受控环境,消费级手表的真实世界性能通常下降10–20%
- PPG信号不能直接测量心脏电活动,对某些心律失常(如房扑)的鉴别能力有限
- 中医脉象与PPG特征的对应关系,目前仅有初步探索,尚缺大规模临床验证研究
- PMIDs 35300400、35127882为综述性文献,综合了多项原始研究,非单一RCT
相关阅读
- 《疾病=频率指纹异常》 — 心血管疾病如何在频率维度上表现异常
- 《中医脉象=频率指纹量化标准》 — HRI指数与脉象量化
- 《王唯工脏腑谐波C1–C10理论》 — 脏腑特征频率的科学假说
- Park J et al. Front Physiol 2022;12:808451. PMID 35300400
- Physiological Measurement focus collection. Physiol Meas 2022. PMID 35127882
免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。
Wearable PPG: Continuous Monitoring and Early Warning Algorithms for Cardiovascular Diseases
Evidence level: B
Introduction
Wrist-worn smartwatches are becoming the most silent cardiovascular early warning systems.
Photoplethysmography (PPG) captures changes in blood volume within the skin’s microvascular bed using LED light sources and photodetectors, generating a continuous physiological frequency curve through dozens of samples per second without the need for needles or cuffs. Hidden within this curve are codes for heart rate variability, arterial compliance, peripheral resistance, and arrhythmias.
From atrial fibrillation detection to hypertension risk screening, from heart failure decompensation warnings to glucose fluctuation inference, the integration of PPG signals with deep learning algorithms is pushing consumer wearable devices toward the boundaries of medical diagnosis. This paper systematically reviews the scientific progress, algorithmic architecture, and intrinsic connections with Traditional Chinese Medicine (TCM) frequency medicine.
I. Basic Principles of PPG Signals and Frequency Dimensions
The PPG waveform consists of a systolic peak and a diastolic peak. The area between these two peaks—the Pulse Index (PI)—reflects the state of peripheral blood perfusion.
From a frequency perspective, PPG signals can be decomposed into:
- Heart Rate Frequency: Approximately 1–2 Hz in resting adults, corresponding to a physiological rhythm of 60–120 beats per minute.
- Respiratory Modulation Frequency: Approximately 0.2–0.4 Hz, representing the regulation of heart rate and stroke volume by the respiratory cycle.
- Sympathetic/Parasympathetic Frequencies: Through Heart Rate Variability (HRV) analysis, frequencies between 0.04–0.15 Hz are associated with sympathetic activity, while 0.15–0.4 Hz relate to parasympathetic activity.
- Higher Harmonics: The C1–C10 organ harmonic theory proposed by Wang Weigong’s team suggests that each organ possesses characteristic resonance frequencies, which theoretically should also be mapped in the high-frequency harmonics of PPG signals.
Limitations compared to ECG: Unlike ECG, which records cardiac electrical activity with a clear temporal baseline, PPG is a downstream hemodynamic effect characterized by lower signal-to-noise ratios and significant interference from motion artifacts, skin pigmentation, and ambient lighting. However, its advantages are equally distinct: continuous unobtrusive monitoring with sampling rates exceeding 100 Hz, offering coverage far surpassing any clinical detection method.
II. Algorithm Evolution: From Feature Engineering to End-to-End Deep Learning
2.1 Traditional Machine Learning Stage
Early PPG disease detection relied on manual feature extraction:
– Peak-valley detection → Calculation of heart rate
– Waveform area ratios → Assessment of arterial stiffness
– Frequency domain power spectrum analysis → Quantification of HRV band energy
Representative work by Bangalore et al. (IEEE TBME 2020) used logistic regression and random forests, achieving approximately 87% sensitivity and 85% specificity for atrial fibrillation detection; however, the high false-positive rate limited clinical utility.
2.2 Convolutional Neural Network (CNN) Stage
Between 2018 and 2021, CNNs began processing raw PPG temporal signals directly, automatically learning multi-scale features. Typical architectures combined 1D CNNs with Short-Time Fourier Transform (STFT) time-frequency maps as input:
- Input: 30-second PPG segments at sampling rates of 20–100 Hz
- Convolutional kernels: Multi-scale (3×, 5×, 7×) to capture waveform morphologies across different time scales
- Pooling layers: Compression of feature dimensions
- Output: Multi-class classification (normal, atrial fibrillation, other arrhythmias)
A representative study from this phase (arxiv 1807.10707) reported an AUC of 0.9999 for atrial fibrillation detection; however, this result stems from controlled datasets, and performance drops significantly in real-world noisy environments.
2.3 Hybrid Architectures and End-to-End Learning
BayesBeat (arxiv 2011.00753) introduced Bayesian deep learning, utilizing uncertainty quantification to enhance robustness in noisy PPG data. Its core logic is that the model outputs not only whether “atrial fibrillation exists” but also the confidence of the prediction, thereby actively downgrading to “uncertain” rather than misreporting when encountering high-noise segments.
ResNet + Transformer (2024 Hypertension Screening Study) directly predicts hypertension risk from wrist PPG data. ResNet residual connections addressed the vanishing gradient problem in deep CNNs, achieving an AUC exceeding 0.92 via 5-fold cross-validation with only 0.124M parameters—a model size sufficient for deployment on watch-side chips.
2.4 Multimodal Fusion
Leading wearable devices (headphones, glasses) now fuse accelerometers (ACC) + PPG to achieve automatic motion artifact separation; medical-grade patch devices fuse ECG + PPG dual channels, using ECG R-peaks to annotate accurate cardiac moments for supervised learning noise reduction.
III. Core Application Scenarios and Validation Data
3.1 Atrial Fibrillation
The Apple Heart Study (Tamburino et al., NEJM 2021) was a landmark investigation: among 412,000 participants wearing Apple Watches, PPG algorithms identified a 0.52% population at high risk for atrial fibrillation, of whom 34.6% confirmed atrial fibrillation via subsequent ECG patches. While this proportion appears modest, it translates to hundreds of thousands of early detections given the population base.
Significance: Atrial fibrillation is the most common arrhythmia-related cause of stroke; early detection allows anticoagulant therapy to reduce stroke risk by approximately 60%. Continuous PPG monitoring fills the blind spot where traditional ECG fails to capture events not visible at the moment of measurement.
3.2 Hypertension Risk Screening
A 2024 study based on the HBPM longitudinal dataset (448 participants, 68k+ spot-check instances) demonstrated that deep learning models extract features from PPG capable of distinguishing normal blood pressure from pre-hypertension and hypertension, achieving an AUC above 0.92. Continuous wrist PPG data captures “masked hypertension” and “white-coat hypertension” more effectively than single office visits.
3.3 Heart Failure Decompensation Warning
Fluid retention and altered peripheral perfusion in heart failure patients cause systematic shifts in PPG waveform morphology. Early research indicates that PPG-derived stroke volume indices and systemic vascular resistance estimates can warn of acute heart failure decompensation 3–7 days in advance, securing a time window for remote medical intervention.
IV. Deep Connections with TCM Pulse Diagnosis Frequency Medicine
The relationship between continuous PPG monitoring and TCM pulse diagnosis represents a highly promising interdisciplinary field.
Frequency Isomorphism: PPG records the peripheral blood volume wave generated by each heartbeat, while pulse diagnosis captures the floating, sinking, slippery, and rough characteristics of the cun-kou pulse. Physiologically, they share “blood pulsation frequency” as a core information source. The 28 pathological pulse patterns in TCM are essentially expert classification labels for different frequency characteristics.
Complementary Advantages: PPG provides objective, quantifiable, traceable digital signals; TCM pulse diagnosis offers holistic, dynamic tactile judgment based on experience. Combined, machine learning algorithms can reconstruct feature frequency patterns resembling “wiry pulses” or “slippery pulses” from PPG signals, aiding the objectification of TCM diagnostic tools.
Consistency in Warning Logic: The therapeutic philosophy of the Huangdi Neijing—”treating disease before it manifests”—aligns closely with the warning logic of wearable devices detecting physiological deviations prior to symptom onset. PPG frequency drift anomalies may correspond to quantifiable signals of “Qi mechanism disorder” described in TCM.
V. Technical Bottlenecks and Future Directions
- Noise Robustness: Motion artifacts and poor skin contact remain primary challenges for watch PPG; optical sensor encapsulation improvements and adaptive algorithmic noise reduction are key focus areas.
- Validation Gold Standards: Most studies validate in controlled environments or specific populations; research on generalization across races, devices, and age groups remains insufficient.
- Clinical Implementation Pathways: FDA/NMPA approval pathways for Software as a Medical Device (SaMD) are becoming clearer, yet medical-grade validation standards for consumer PPG algorithms are still being established.
- Multi-Disease Joint Modeling: Moving from single-disease warnings toward comprehensive cardiovascular risk scores (integrating atrial fibrillation, hypertension, heart failure, arterial stiffness, etc.) represents the goal for next-generation algorithms.
Evidence and Limitations
Supporting Evidence (Evidence Level B):
- Two systematic reviews (PMID 35300400, 35127882) confirm the technical feasibility of PPG combined with deep learning for cardiovascular disease detection, with AUCs generally >0.90
- Large-scale population studies like the Apple Heart Study provide preliminary validation of clinical value in real-world continuous monitoring
- Multiple studies (BayesBeat, ResNet+Transformer) verified algorithm performance on public datasets
Limitations:
- Most high-quality studies originate from controlled environments; real-world performance on consumer watches typically decreases by 10–20%
- PPG signals cannot directly measure cardiac electrical activity, limiting differentiation capabilities for certain arrhythmias (e.g., atrial flutter)
- Correspondence between TCM pulse patterns and PPG features remains in preliminary exploration, lacking large-scale clinical validation
- PMIDs 35300400 and 35127882 are review articles synthesizing multiple primary studies, not single RCTs
Further Reading
- 《Disease = Abnormal Frequency Fingerprint》 — How cardiovascular diseases manifest abnormalities in the frequency dimension
- 《TCM Pulse Patterns = Quantified Frequency Fingerprint Standards》 — HRI index and pulse quantification
- 《Wang Weigong Organ Harmonic C1–C10 Theory》 — Scientific hypothesis on organ characteristic frequencies
- Park J et al. Front Physiol 2022;12:808451. PMID 35300400
- Physiological Measurement focus collection. Physiol Meas 2022. PMID 35127882
Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.
Leave a Reply