Physiological Stratified Models for Personalized Cuffless Blood Pressure Estimation Using Wearable Multimodal Sensors

A one-time IR PPG + pulse-pressure-wave recording assigns each subject to a physiological stratum; per-stratum Transformers are then personalized by fine-tuning 129 parameters on 30 s of data. CAS-BP, 1,272 subjects aged 18–70. Single baseline model: MAE 4.10 / 2.69 mmHg; stratified models: SBP 3.95 / 3.02 / 2.95 and DBP 2.59 / 2.11 / 2.15 by stratum.

• Form factor: other

Cuffless blood pressure (BP) estimation from wearable sensors is challenged by significant intersubject physiological variability. To resolve the technical bottlenecks constraining the clinical translation of cuffless BP monitoring-including sensor modality limitations, dataset generalizability issues, and computational complexity-we propose a two-stage framework that integrates established calibration concepts with a physiologically stratified pretraining stage. First, a one-time recording with an infrared photoplethysmogram (PPG) and a pulse pressure wave (PPW) is used to assign subjects to a physiological stratum. Subsequently, cuffless BP estimation is performed using the signals with efficient stratified models. By training distinct models for each stratum, the BP estimation accuracy is significantly enhanced compared to one model trained on the entire cohort. The analysis was conducted on the CAS-BP dataset, which contains 1272 subjects’ recordings (age: 18–70). Transformer-based models were trained to perform estimation and personalized for test subjects using a fine-tuning process on 30-s data. This process required updating 129 parameters. Our method trains stratified models for each stratum; a baseline model was trained on all subjects’ data for comparison. The baseline model yielded a mean absolute error (MAE) of 4.10 mmHg (mean error (ME) ± standard deviation (SD): - 0.15~± ~5.77 ) for systolic BP (SBP) and 2.69 mmHg (ME ± SD: - 0.05~± ~3.93 ) for diastolic BP (DBP). For the stratified models, the SBP MAE was reduced to 3.95, 3.02, and 2.95 mmHg, and the DBP MAE was reduced to 2.59, 2.11, and 2.15 mmHg. We improved estimation accuracy over a generalized baseline model, demonstrating that physiological stratification provides an effective initialization space for computationally lightweight model personalization on wearable devices, requiring minimal calibration data.

Verbatim from Semantic Scholar (opens in a new tab) · Abstract via Semantic Scholar; © the authors/publisher

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

Wavelengths

IR