Abstract
Objective
Occlusive cuff inflation in ambulatory blood pressure (BP) monitoring disturbs the daily life of the user, and affects efficacy of monitoring. To overcome this limitation, we have developed a novel minimally-occlusive cuff method for stress-free measurement of BP. This study aimed to experimentally evaluate the reliability of this method, and improve the precision of this method by implementing a machine learning algorithm.
Methods
In this method, a thin-plate-type ultrasound probe (Size: 5.6mm-thickness × 28mm × 26mm; weight: 10g) is placed between the cuff and the skin, and used to measure the ultrasonic dimension of the artery (Figure 1). The cuff pressure (Pc), arterial dimension at systole (Ds) and diastole (Dd), systolic BP (SBP) and diastolic BP (DBP) during cuff inflation are theoretically related by the following equations,
SBP-Pc = P0·Exp[α·Ds]
DBP-Pc = P0·Exp[α·Dd]
Where P0 and α are constants, and α indicates arterial stiffness. Since multiple sets of the two equations can be defined over multiple cardiac beats while measuring Pc, Ds and Dd during mild cuff inflation (Pc is controlled less than 50 mmHg, Figure 1), it is possible to estimate SBP (SBPe) and DBP (DBPe) as solutions of the equations. In 6 anesthetized dogs, we attached the cuff and the probe to the right thigh to get SBPe and DBPe, which were one-time calibrated in each animal against reference SBP and DBP measured by using an intra-arterial catheter. We also determined the pulse arrival time (PAT), which is a commonly employed parameter in cuff-less BP monitoring. In all the dogs, BP was changed extensively by infusing noradrenaline or sodium nitroprusside.
Results
DBPe correlated tightly with DBP with a coefficient of determination (R2) of 0.85±0.08, and predicted DBP with error of 3.9±7.9 mmHg after one-time calibration (Figure 2). PAT correlated poorly with DBP (R2=0.49±0.17), and predicted DBP less accurately than this method. SBPe correlated well with SBP (R2=0.78±0.08) (Figure 3). However, even after one-time calibration, difference between SBPe and SBP was 2.6±18.9 mmHg, which was not acceptable. To improve the precision in SBP prediction, we used supervised machine learning approach with use of a support vector algorithm (Python, Scikit-learn), which regressed feature variables (SBPe, DBPe, Ds, Dd heart rate, and PAT) against teacher signal (reference SBP). The support vector algorithm, once trained, predicted SBP with acceptable accuracy with error of 0.7±6.9 mmHg (Figure 3).
Conclusions
This method reliably tracks BP changes without occlusive cuff inflation. Once calibrated, this method measures DBP accurately. With the aid of machine learning, precision in SBP prediction was greatly improved to an acceptable level. This method with machine learning approach has potential for stress-free BP measurement in ambulatory BP monitoring.
Funding Acknowledgement
Type of funding source: Public grant(s) – National budget only. Main funding source(s): Japan Society for the Promotion of Science
Driver Stress State Evaluation by Means of Thermal Imaging: A Supervised Machine Learning Approach Based on ECG Signal
