Blood pressure estimation from two-channel BCG

Estimated blood pressure from two chair BCG channels using a CNN, BiGRU and temporal attention.

Role
Co-developer of blood-pressure estimation and real-time processing
Contribution
Co-developed blood-pressure elevation experiments, synchronized signal acquisition, the estimation model and real-time analysis.
Period & affiliation
2020.09 ~ 2021.06 BCML, EMMA Healthcare, Team TOG

01 / RESEARCH DETAILTwo-channel BCG model and real-time service

Challenge
Temporal BCG information from the chair back and seat must be connected to real-time blood-pressure display.
Approach
Extracted local features from 10-second windows with a CNN, then used a BiGRU and temporal attention for regression.
Contribution
Co-developed synchronized acquisition, the blood-pressure model and the real-time analysis service.
Model specification in the released final code
Final code specification Implementation
Input 250 Hz, 2 channels, 10 seconds
CNN 10 layers; 64 / 128 / 256 / 512 channels
Regression BiGRU 64 + temporal attention
Evaluation conditions and scopeFinal code: 10 convolutional layers, 64/128/256/512 channels and a 64-unit BiGRU. The presentation figure shows an earlier configuration.

02 / RESEARCH DETAILBlood-pressure tracking and reproduction conditions

Experiment
Recorded resting and post-exercise states for about three hours per participant, with cuff references every two minutes.
Evaluation
Evaluated SBP and DBP errors for six participants using within-person temporal splits of 29/14/57%.
MAE by participant (mmHg) SBP: 1 5.5, 2 3.2, 3 5.25, 4 3.35, 5 4.68, 6 6.2; DBP: 1 4.65, 2 3.12, 3 3.88, 4 4.03, 5 3.8, 6 3.5 0 2 4 6 8 5.5 4.65 1 3.2 3.12 2 5.25 3.88 3 3.35 4.03 4 4.68 3.8 5 6.2 3.5 6 SBP DBP
MAE by participant (mmHg)
Mean blood-pressure error over six participants (mmHg)
Mean across 6 participants (mmHg) SBP DBP
MAE 4.70 3.83
RMSE 6.02 4.90
Evaluation conditions and scopeResults from the 2021 paper: within-person evaluation with overlapping windows and normalization before splitting.