BME PhD Defense: Raj Kasbekar "Applications of Machine and Deep Learning for Continuous, Cuffless Blood Pressure Monitoring" 

PhD Dissertation Defense 

Thursday April 17, 2025 

Gateway Park 1 1002 

9:00am-10:00am 

"Applications of Machine and Deep Learning for Continuous, Cuffless Blood Pressure Monitoring"  

Raj Kasbekar 

Abstract:    Accurate and reliable estimation of beat to beat i.e., continuous, cuffless blood pressure has profound clinical implications, as a powerful prognosticator of cardiovascular, renal and other diseases which are the leading cause of mortality and disease burden globally. The ability to perform reliable and accurate continuous blood pressure (BP)measurements and collect larger Real Word Data (RWD) data sets on diurnal, nocturnal variation and short-term variability can provide fresh clinical insights and new tools for the early detection of diseases like cardiac morbidity, chronic renal failure, malignant and secondary hypertension, pre-eclampsia and autonomic neuropathy. Cuffless, continuous BP estimation is not based on a direct measurement using a pressure sensor, but rather on the principle of a pulse wave percolating through the circulatory system upon ejection of blood from the aortic valve into the arterial system. The transit time of this pulse wave (referred to as the pulse arrival time) has a strong non-linear correlation with the arterial blood pressure. This relationship between BP and transit time is further confounded by the complex interplay of many variables such as blood viscosity, arterial size and thickness, hormones and several others.  

           To address this complex relationship of pressure and pulsatile flow, an effective end-to-end BP monitoring system needs to perform two key tasks; the first task involves extracting a signal with a high signal-to-noise ratio. Among various sensors, Photo-Plethysmogram (PPG) has been the most popular signal (along with ECG) since its optical nature makes it convenient, reliable, repeatable and inexpensive. The second task consists of using this signal to estimate BP using an estimation algorithm. The complex interplay of various factors and the person-to-person variation makes the use of analytical regression or other linear techniques unfeasible, with some meta-analysis raising questions of trusting such a measurement. Researchers have migrated, therefore, towards use of machine learning and deep learning to improve accuracy and lower variability.   

         Our first aim in developing such an end-to-end BP monitoring system consisted of developing a custom prototype sensor using an off-the-shelf optical module (with its associated development board) and researching the factors affecting signal quality in a PPG sensor using a design of experiments approach. We identified a unique combination of these factors that is most optimal in reducing motion artifacts using a full factorial design of experiments methodology and evaluated the effect of these factors on PPG readings with and without motion. Coupled with signal processing, optimal use of these factors gave a high-fidelity PPG signal.  

         Our second aim consisted of developing the theory and demonstrating how multimodal feature datasets, comprising: i) pulse arrival time (PAT); ii) pulse wave morphology (PWM), and iii) demographic data, can be combined with optimized machine learning (ML) algorithms to estimate Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP) and Mean Arterial Pressure (MAP). We used data from healthy subjects as well as hemodynamically compromised subjects in a diseased state. While our bias estimates (-0.65 mmHg-healthy subjects and 1.38 mm Hg- diseased subjects for SBP, -0.08 mm Hg-healthy subjects and 0.45 mm Hg-diseased subjects for DBP , 0.27 mm Hg-healthy subjects and 0.76 mm Hg-diseased subjects for MAP) were within 5 mm Hg of the gold standard intra-arterial BP, as required by the IEC/ANSI 80601-2-30 (2018) standard, our standard deviation (SD) estimates  (15.28 mmHg-healthy subjects and 15.12 mm Hg- diseased subjects for SBP, 7.49 mm Hg-healthy subjects and 7.53 mm Hg-diseased subjects for DBP, 9.51 mm Hg-healthy subjects and 8.84 mm Hg-diseased subjects for MAP) were well over the 8 mm Hg limit for SBP and MAP.   

        Our third aim, therefore, was to apply these algorithms on a large dataset and use personalized calibration (personalization) for estimation of SBP and DBP. It consisted of using a new feature extractor (catch-22) on the photoplethysmogram waveform, training a hemodynamically compromised VitalDB dataset (n=1293), and using personalization triggered by the flow signal, followed by transfer learning (on three machine learning algorithms) to finetune the pretrained model. This approach, when tested on 116 x 3 subjects meeting the AAMI standard criteria, reduced the test error bias to less than 5 mm Hg and the test SD to less than 8 mm for both SBP and DBP—well within the acceptable limits of the AAMI standard.  Using Levene’s test, we found a statistically significant reduction in error variance using our personalization method compared to calibration-free method from a distinct set of training subjects (p<0.05), but no differences between the three machine learning models.  

       Our key contribution, therefore, is based on the premise that the BP, estimated using the pressure flow relationship, is reliable only if the flow or the pulse arrival time is within a certain quantum band. A change in flow or pulse arrival time outside this band would trigger critical personalization or calibration that would ensure that the pressure flow relationship is preserved, and its use would not be that often. This ensures that the accuracy and reliability of the BP estimates is maintained to the AAMI criteria over the long term for hypertensive, healthy or hypotensive subjects and therefore could be suitable for clinical adoption. In practice, personalization would be implemented using a cuff, a calibrated Tono arteriogram pressure sensor or any method for direct pressure calibration.   

       In summary, this dissertation demonstrates that a high-fidelity PPG signal, using catch-22 algorithm for feature extraction, coupled with a ‘flow triggered’ personalization on a large dataset, and use of an optimal machine learning or deep learning algorithm can improve the accuracy and reliability of continuous, cuffless BP estimation, with great promise of clinical adoption. 

For a zoom link, please email kharrison@wpi.edu