Data Mining Investigation of the Association between a Limb Ballistocardiogram and Blood Pressure
Journal
Physiological Measurement
Date
2018.08.01
Abstract
To investigate the association between a limb ballistocardiogram (BCG) and blood pressure (BP) based on data mining. Approach: During four BP-perturbing interventions, the BCG and reference BP were measured from 15 young healthy volunteers using a custom-manufactured wristband equipped with a MEMS accelerometer and a commercial continuous BP measurement device. Both timing and amplitude features in the wrist BCG waveform were extracted, and significant features predictive of diastolic (DP) and systolic (SP) BP were selected using stepwise linear regression analysis. The selected features were further compressed using principal component analysis to yield a small set of DP and SP predictors. The association between the predictors thus obtained and BP was investigated by multivariate linear regression analysis. Main Results: The predictors exhibited a meaningful association with BP. When three most significant predictors were used for DP and SP, correlation coefficient of r=0.82+/-0.03 (DP) and r=0.77+/-0.05 (SP) as well as root-mean-squared error (RMSE) of 6.0+/-0.6 mmHg (DP) and 9.3+/-0.7 mmHg (SP) were obtained across all interventions (mean+/-SE). The association was consistent in all the individual interventions (r≥0.75 and RMSE≤5.2 mmHg for DP as well as r≥0.69 and RMSE≤8.0 mmHg for SP on the average). The minimum number of requisite predictors for robust yet practically realistic BP monitoring appeared to be 2~3. The association between predictors and BP was maintained even under regularized calibration (r=0.66+/-0.07 and RMSE=8.6+/-0.7 mmHg for DP as well as r=0.63+/-0.05 and RMSE=12.5+/-0.8 mmHg for SP (mean+/-SE)). The requisite predictors for DP and SP were distinct from each other. Significance: The results of this study may provide a viable basis for ultra-convenient BP monitoring based on a limb BCG alone.