以下内容为一般性专利撰写与策略信息，旨在支持权利要求与说明书草拟，不构成具体法律意见或律师—客户关系。建议在提交前由具备相应执业资格的律师针对目标法域适配并复核。
一、专利介绍（中文）
1) 标题
一种无袖标可穿戴连续血压估计系统及其方法
2) 技术领域
本发明涉及生物医学信号处理、可穿戴设备与人工智能技术领域，具体涉及一种融合光电容积描记信号（PPG）与加速度信号（ACC），通过端侧一维卷积网络与时序Transformer进行特征提取，并结合云端个体化校准，实现连续、无袖标血压估计的可穿戴系统及其方法，涵盖数据质量自清洗、腕带松紧自校正与模型漂移监测报警等功能。
3) 背景技术
无袖标血压估计旨在通过非侵入、连续方式实现血压监测，以提升场景适应性与用户依从性。PPG信号可提供与脉搏波形态相关的血流动力学特征，ACC信号可反映运动状态与佩戴接触状态。然而，现有技术在运动伪迹抑制、个体差异补偿、低功耗端侧推理与长期性能稳定性方面仍存在不足，尤其在动态环境下准确度与一致性欠佳，且受腕带松紧变化与模型漂移影响较大。
4) 发明内容
4.1 要解决的技术问题
- 在端侧功耗受限的条件下，融合PPG与ACC以鲁棒估计连续血压；
- 通过云端个体化校准应对个体差异与时变特性；
- 自动识别并剔除异常搏动与运动伪迹，降低误差与漂移；
- 估计并补偿腕带接触压力变化导致的测量偏差；
- 监测模型性能漂移并提供告警，维持长期可靠性；
- 满足医疗电气安全与数据合规的工程化要求。
4.2 技术方案（概述）
- 设备构成：包含PPG传感器、三轴加速度计、低功耗处理器/SoC、存储器与无线通信模块的可穿戴终端；可选集成光路驱动与AFE前端。
- 信号处理：端侧对PPG与ACC进行同步采集、时频域预处理与多模态融合；采用一维卷积网络提取局部时域形态特征，采用时序Transformer提取长程时序依赖，输出收缩压/舒张压/平均动脉压的连续估计。
- 个体化校准：云端利用受试者标定血压数据（如经袖带式或其他经临床验证设备获得）与端侧特征，训练或更新个体化校准参数/微调模型，并将参数下发至端侧。
- 数据质量自清洗：端侧通过异常搏动检测、伪迹评分与置信度度量，对异常拍点/窗口进行自动识别与剔除/重权重处理。
- 腕带松紧自校正：利用ACC与PPG形态/幅值变化估计接触压力或佩戴松紧度，对输出进行补偿或动态重标定触发。
- 模型漂移告警：端侧/云端基于分布偏移检测、置信度趋势、参考窗口内一致性指标监测性能变化，超过阈值时提示重校准或软件更新。
- 低功耗：通过模型压缩、量化、稀疏注意力与占空比管理，使端侧推理与传感总功耗小于30 mW（以典型工作占空比计）。
- 合规设计：在工程实现上考虑医疗电气安全（如IEC 60601-1、IEC 60601-1-2、IEC 80601-2-30等适用部分）、医疗器械软件生命周期与风险管理（如IEC 62304、ISO 14971、IEC 62366-1）以及数据合规（如GDPR、HIPAA、中华人民共和国个人信息保护法PIPL与网络安全法等适用法规）。
4.3 有益效果
- 在运动状态下提升血压估计稳健性与一致性；
- 降低因个体生理差异、佩戴松紧和环境变化导致的系统性偏差；
- 通过模型漂移监测维持长期可靠性与可追踪性；
- 端侧低功耗实现连续监测，延长续航并提升可用性；
- 便于满足医疗安全与数据合规的产业化落地需求。
5) 可选实施方式要点（非限定）
- 一维卷积网络与时序Transformer的具体拓扑、窗口长度、采样率与特征融合策略；
- 基于心搏周期分割、拍间变异度量、PPG形态畸变指数与ACC能量门限的异常检测；
- 利用PPG直流/脉动分量比值、脉搏上升沿斜率与ACC低频分量估计接触压力；
- 端侧与云端协同的个体化校准流程（如联邦学习、差分隐私可选）；
- 模型漂移检测指标（如分布散度、置信区间宽度、预测残差统计）与告警阈值更新；
- 低功耗策略（模型量化、蒸馏、动态电压频率调节、稀疏注意力、事件驱动采样）。
6) 合规注意事项（工程实现层面）
- 医疗电气与EMC：IEC 60601-1、IEC 60601-1-2；自动无创血压类设备的适用专用标准可参考IEC 80601-2-30；风险管理与可用性工程：ISO 14971、IEC 62366-1；软件生命周期：IEC 62304。
- 临床性能评价：在目标法域内采用相应指南或标准进行临床验证与偏倚/精度评估（例如IEEE Std 1708针对于可穿戴无袖标血压设备的性能评价方法）。
- 数据合规：数据最小化、加密存储与传输、去标识化/匿名化、访问控制与跨境传输合规（如GDPR、HIPAA、PIPL等，依地域适用）。
二、权利要求框架建议（中文）
A. 独立权利要求（装置/系统类，尽量聚焦必要特征）
- 一种可穿戴血压估计装置，包括：
- PPG传感器与ACC传感器，配置为同步采集与佩戴者脉搏相关的PPG信号和加速度信号；
- 处理器与存储器，存储有在处理器上可执行的指令，使所述处理器被配置为：
- 对所述PPG信号与所述加速度信号进行多模态融合预处理；
- 通过一维卷积网络与时序Transformer的组合模型提取时序特征；
- 基于提取的特征生成佩戴者的连续血压估计；
- 应用由远程服务器生成的个体化校准参数对所述血压估计进行校准。
B. 独立权利要求（方法类）
- 一种连续血压估计的方法，包括：
- 同步采集PPG与ACC；
- 进行多模态融合预处理；
- 采用一维卷积与时序Transformer模型提取特征并输出血压估计；
- 利用基于个体标定数据训练得到的校准参数对输出进行个体化校准。
C. 独立权利要求（计算机可读介质/程序）
- 存储有指令的非暂态计算机可读介质，当由处理器执行时实施上述方法。
D. 独立权利要求（系统含云端）
- 包括权利要求A之装置与服务器，服务器被配置为基于佩戴者参考血压与端侧特征训练或更新个体化校准参数，并下发至装置。
E. 从属权利要求（优选特征）
- 数据质量自清洗：异常搏动检测与伪迹剔除策略（基于拍间间期异常、PPG形态畸变指数、ACC能量/频谱门限等）。
- 腕带松紧自校正：基于PPG幅值/直流分量、上升沿斜率与ACC低频分量的接触压力估计与补偿。
- 模型漂移告警：基于分布散度、预测不确定度或残差统计的漂移检测与告警；
- 低功耗约束：端侧推理与传感平均功耗小于30 mW；
- 安全与合规模块：端侧加密、访问控制、数据匿名化/去标识化、传输加密；
- 模型实现：轻量化Transformer（稀疏注意力/线性注意力）、量化/蒸馏、事件驱动或自适应采样；
- OTA更新与回滚；本地缓存与断点续传；
- 训练与校准细节：使用受试者参考血压数据进行监督微调或校准映射学习；支持联邦学习。
三、关键要素清单（中文）
- 传感层：PPG发射/接收结构、采样率与同步；三轴ACC范围与噪声规格。
- 预处理与融合：时域/频域滤波、时钟同步、窗口与步长、多模态特征拼接或注意力融合。
- 模型结构：1D卷积堆叠提取局部形态；时序Transformer捕获长程依赖；输出层映射至SBP/DBP/MAP；不确定度估计（可选）。
- 个体化校准：云端训练/微调个体参数；端侧应用与版本管理；触发策略（定期/事件驱动）。
- 异常自清洗：搏动级质量评分、异常检测、剔除/重加权。
- 松紧自校正：接触压力估计、补偿曲线或动态重标定触发。
- 漂移监测：分布漂移度量、阈值策略、告警与处置（重校准/更新）。
- 功耗与资源：模型压缩、量化位宽、占空比、续航目标；端侧算力与内存约束。
- 安全与合规：硬件隔离、加密、用户同意与数据最小化；适用标准与监管路径预留。
- 可扩展性：方法、装置、系统、介质多类别权利要求；不同佩戴位置/多波段PPG/附加IMU或压力传感器的等同/替换方案。
四、措辞修改建议（中文）
- 将“异常搏动自清洗”表述为“基于质量评价与异常检测的自动伪迹剔除与重加权处理”；
- 将“腕带松紧自校正”表述为“基于接触压力估计的佩戴状态补偿”；
- 将“模型漂移告警”表述为“基于分布偏移与性能指示的漂移检测与告警机制”；
- 避免使用“保证”“完全消除”，改用“用于”“适于”“被配置为”；
- 对“Transformer”可描述为“具有自注意力机制的时序特征提取网络”，并在说明书中将轻量化/稀疏注意力作为等同变型；
- 对功耗限定置于从属权利要求，主权利要求不宜过度限定实现路径；
- 对云端“个体化校准”以“生成个体化校准参数并下发应用”为主述，不限定具体算法，以保留覆盖面；
- 引入术语解释条款，界定“连续”“同步”“校准”“接触压力”等术语在本案中的技术含义；
- 在说明书实施例提供参数范围与多种可选路径，以满足充分公开并支持广度。
五、必要/充分技术特征建议（中文）
- 必要特征（独立权利要求应保留）：
- PPG与ACC的同步采集与多模态融合；
- 端侧一维卷积与时序Transformer组合用于特征提取并输出血压估计；
- 应用云端生成的个体化校准参数进行校准。
- 充分但可下放（从属权利要求表述）：
- 异常搏动检测与伪迹剔除细节；
- 腕带松紧估计与补偿的具体指标与算法；
- 漂移检测的具体统计量与阈值；
- 功耗小于30 mW的量化限定与具体低功耗手段；
- 具体网络结构、核大小、注意力形式与训练细节。
六、可实施性与充分公开提示（中文）
- 提供代表性实施例：采样率（如PPG 100–256 Hz，ACC 50–200 Hz）、窗口长度（如8–30秒）、步长（如1–5秒）、滤波器参数、特征列举（形态学特征、拍间变异、频域功率）。
- 描述模型参数量与端侧推理复杂度范围（如百KB–数MB级，MAdd数量级），并给出量化位宽（如8-bit）与占空比策略，论证<30 mW可达性（以典型SoC与占空比为例说明）。
- 公开个体化校准流程：参考血压来源、最小校准样本量、更新频率、边缘/云协同与安全传输。
- 说明异常检测与松紧补偿的可替代实现（基于规则或学习的质量评分），确保不因单一算法受限。
- 明确数据加密、访问控制与去标识化流程作为实施例，不将具体法规文本写入权利要求，避免地域性限制。
——
I. Patent Introduction (English)
1) Title
Cuffless wearable system and method for continuous blood pressure estimation
2) Technical Field
The invention relates to biomedical signal processing, wearable devices, and artificial intelligence. It particularly concerns a cuffless wearable system that fuses photoplethysmography (PPG) and accelerometer (ACC) signals, employs on-device one-dimensional convolutional networks combined with a temporal Transformer for feature extraction, and performs cloud-based individualized calibration. The system further implements automatic artifact removal, wristband tightness compensation, and model drift alerting, under an on-device power budget below 30 mW, with consideration of medical electrical safety and data protection compliance.
3) Background
Cuffless blood pressure (BP) estimation aims to provide non-invasive, continuous monitoring with improved usability. PPG conveys hemodynamic waveform features, while ACC captures motion and wear state. Existing approaches face limitations in motion robustness, inter-subject variability compensation, low-power on-device inference, and long-term stability, especially under dynamic conditions, strap tightness variation, and model drift.
4) Summary of the Invention
4.1 Problems to be solved
- Robust continuous BP estimation on a power-constrained wearable by fusing PPG and ACC;
- Address subject-specific and time-varying factors via cloud-based individualized calibration;
- Automatically detect and exclude abnormal beats and motion artifacts;
- Estimate and compensate for strap tightness/contact pressure changes;
- Monitor model performance drift and issue alerts;
- Engineer the solution consistent with medical electrical safety and data protection requirements.
4.2 Technical solution (overview)
- Device: a wearable including a PPG sensor, a 3-axis accelerometer, a low-power processor/SoC, memory, and wireless communications; optionally an integrated optical driver and AFE.
- Signal processing: synchronous acquisition and multimodal preprocessing of PPG and ACC; on-device feature extraction using a 1D convolutional network and a temporal Transformer; continuous SBP/DBP/MAP estimation.
- Individualized calibration: a server trains or updates subject-specific calibration parameters/models using reference BP data and device features, and deploys the parameters back to the device.
- Automatic artifact handling: beat-level quality assessment and outlier/artifact detection with exclusion or reweighting.
- Strap tightness compensation: estimation of contact pressure using ACC and PPG morphology/amplitude metrics, followed by compensation or re-calibration triggers.
- Model drift alerting: distribution shift and confidence/residual statistics monitored on-device or in the cloud, with alerts and remedial actions upon threshold exceedance.
- Low power: model compression, quantization, sparse/linear attention, and duty cycling to maintain average power below 30 mW under typical duty factors.
- Compliance: consider medical electrical safety (e.g., IEC 60601-1, IEC 60601-1-2, IEC 80601-2-30 as applicable), software lifecycle and risk management (e.g., IEC 62304, ISO 14971, IEC 62366-1), and data protection (e.g., GDPR, HIPAA, China PIPL and cybersecurity requirements, as applicable).
4.3 Advantages
- Improved robustness and consistency of BP estimates under motion;
- Reduced systematic errors due to inter-subject variability, strap tightness, and environmental changes;
- Sustained reliability via drift monitoring and alerting;
- Low-power continuous monitoring enabling longer battery life and better usability;
- Engineering pathway aligned with safety and data compliance.
5) Optional embodiments (non-limiting)
- Specific topology of the 1D CNN and temporal Transformer, window length, sampling rates, and fusion strategies;
- Artifact detection using beat segmentation, inter-beat variability, PPG morphology distortion indices, and ACC energy thresholds;
- Contact pressure estimation via PPG DC/AC ratio, upstroke slope, and low-frequency ACC;
- Edge–cloud calibration workflows (e.g., federated learning, differential privacy optional);
- Drift indicators (distributional divergence, uncertainty width, residual statistics) and adaptive thresholds;
- Low-power techniques (quantization, distillation, DVFS, sparse attention, event-driven sampling).
6) Compliance considerations (implementation)
- Electrical safety and EMC: IEC 60601-1, IEC 60601-1-2; for automated non-invasive BP devices see IEC 80601-2-30 where applicable; risk management and usability: ISO 14971, IEC 62366-1; software lifecycle: IEC 62304.
- Clinical performance: evaluate per jurisdictional guidance/standards; IEEE Std 1708 provides test methods for wearable cuffless BP devices.
- Data protection: data minimization, encryption at rest/in transit, de-identification/anonymous processing, access control, cross-border transfer compliance (e.g., GDPR, HIPAA, PIPL, as applicable).
II. Claiming Strategy (English)
A. Independent device/system claim (focus on essentials)
- A wearable BP estimation device comprising:
- a PPG sensor and an ACC sensor configured for synchronized acquisition of PPG and acceleration signals from a wearer;
- a processor and memory storing instructions that, when executed, cause the processor to:
- perform multimodal preprocessing of the PPG and acceleration signals;
- extract temporal features using a combined one-dimensional convolutional network and a temporal Transformer;
- generate continuous BP estimates based on the extracted features; and
- apply subject-specific calibration parameters generated by a remote server to calibrate the BP estimates.
B. Independent method claim
- A method of continuous BP estimation comprising:
- synchronously acquiring PPG and ACC signals;
- performing multimodal preprocessing;
- extracting features using a one-dimensional convolutional network and a temporal Transformer to output BP estimates; and
- calibrating the output using subject-specific parameters derived from subject reference BP data.
C. Computer-readable medium claim
- A non-transitory computer-readable medium storing instructions that, when executed by a processor, perform the method above.
D. System including cloud
- A system comprising the wearable device of claim A and a server configured to train or update subject-specific calibration parameters based on subject reference BP and device features, and to deploy the parameters to the device.
E. Dependent claims (preferred features)
- Automatic artifact handling with beat-level quality scoring and outlier exclusion;
- Strap tightness/contact pressure estimation using PPG amplitude/DC components, upstroke slope, and low-frequency ACC, and compensation thereof;
- Drift detection and alerting based on distributional divergence, uncertainty, or residual statistics;
- Power constraint: average on-device sensing and inference below 30 mW;
- Security and compliance modules: on-device encryption, access control, de-identification, encrypted transport;
- Model realizations: lightweight/self-attention variants, quantization/distillation, event-driven or adaptive sampling;
- OTA updates with rollback; local caching and resumable transfer;
- Calibration details: supervised fine-tuning using subject reference BP; federated learning support.
III. Key element checklist (English)
- Sensing: PPG emitter/detector optics, sampling and synchronization; 3-axis ACC range and noise.
- Preprocessing and fusion: time/frequency filtering, clock sync, windowing/stride, feature concatenation or attention-based fusion.
- Model: 1D CNN for local morphology; temporal Transformer for long-range dependencies; outputs for SBP/DBP/MAP; optional uncertainty.
- Individual calibration: cloud training/update of subject parameters; on-device application and versioning; trigger policies.
- Artifact handling: beat quality scoring, detection, exclusion/reweighting.
- Tightness compensation: contact pressure estimation and compensation or re-calibration triggers.
- Drift monitoring: shift metrics, thresholds, alerts and remediation.
- Power/resources: compression, quantization bit-width, duty cycling, battery life; compute/memory limits.
- Safety/compliance: isolation, encryption, consent and minimization; standards and regulatory pathway placeholders.
- Extensibility: multiple claim categories; equivalents for sensor sets, wavelengths, wear locations, and attention variants.
IV. Wording refinements (English)
- Replace “self-cleaning” with “automatic artifact detection, exclusion, and reweighting based on quality assessment.”
- Replace “wristband tightness self-correction” with “contact pressure estimation and compensation for wear state.”
- Replace “model drift alert” with “drift detection and alerting based on distributional and performance indicators.”
- Avoid absolutes such as “guarantee” or “eliminate”; use “configured to,” “adapted to,” “operable to.”
- Describe “Transformer” as “a temporal feature extraction network employing self-attention,” and reserve sparse/linear attention as equivalents.
- Place the <30 mW limit in dependent claims; avoid over-limiting the independent claims.
- Frame individualized calibration as “server-generated subject-specific parameters” without constraining to a single algorithmic form.
- Include a definitions section in the specification to construe “continuous,” “synchronized,” “calibration,” and “contact pressure.”
V. Necessary vs. sufficient technical features (English)
- Necessary for independent claims:
- Synchronized acquisition and multimodal fusion of PPG and ACC;
- On-device feature extraction using a combination of 1D CNN and temporal Transformer to generate BP estimates;
- Application of server-generated subject-specific calibration parameters.
- Sufficient but relegated to dependent claims:
- Specific artifact detection metrics and exclusion policies;
- Specific strap tightness estimation/compensation indicators;
- Particular drift metrics and thresholds;
- Power budget (<30 mW) and concrete low-power techniques;
- Network layer details, kernel sizes, attention forms, and training specifics.
VI. Enablement and sufficiency tips (English)
- Provide representative parameters: sampling (e.g., PPG 100–256 Hz, ACC 50–200 Hz), window length (e.g., 8–30 s), stride (e.g., 1–5 s), filter settings, and enumerated features (morphology, inter-beat variability, spectral power).
- Disclose model size and inference complexity ranges (e.g., hundreds of kB to a few MB; MAdds order), quantization (e.g., 8-bit), and duty cycling to substantiate the <30 mW budget on a typical wearable SoC.
- Detail calibration workflow: sources of reference BP, minimum calibration samples, refresh cadence, edge–cloud cooperation and secure transfer.
- Present alternative implementations for artifact handling and tightness compensation (rule-based or learned quality scores) to avoid undue limitation.
- Describe encryption, access control, and de-identification as implementation embodiments; avoid embedding specific jurisdictional statutes in the claims to preserve breadth.
附注（Notes）
- 标准与法规名称为通行业内公认文本（例如IEC 60601-1、IEC 60601-1-2、IEC 80601-2-30、ISO 14971、IEC 62304、IEC 62366-1、IEEE Std 1708、GDPR、HIPAA、PIPL），各法域适用性与版本以提交时最新有效文本为准。
- 为提高授权与后续执行力，建议在撰写阶段准备对比实验数据、漂移长期跟踪数据与功耗测量记录，用以支撑技术效果与产业可实施性。
