Machine Learning for Arterial Blood Pressure Prediction

Jessica Zheng (MIT), Hanrui Wang* (MIT), Anand Chandrasekhar (MIT), Aaron Aguirre (Massachusetts General Hospital and Harvard Medical School), Song Han (MIT), Hae-Seung Lee (MIT), Charles G. Sodini (MIT)

Abstract: High blood pressure is a major risk factor for cardiovascular disease, necessitating accurate blood pressure (BP) measurement. Clinicians measure BP with an invasive arterial catheter or via a non-invasive arm or finger cuff. However, the former can cause discomfort to the patient and is unsuitable outside Intensive Care Unit (ICU). While cuff-based devices, despite being non-invasive, fails to provide continuous measurement, and they measure from peripheral blood vessels whose BP waveforms differ significantly from those proximal to the heart. Hence, there is an urgent need to develop a measurement protocol for converting easily measured non-invasive data into accurate BP values. Addressing this gap, we propose a non-invasive approach to predict BP from arterial area and blood flow velocity signals measured from a Philips ultrasound transducer (XL-143) applied to large arteries close to heart. We developed the protocol and collected data from 72 subjects. The shape of BP (relative BP) can be theoretically calculated from these waveforms, however there is no established theory to obtain absolute BP values. To tackle this challenge, we further employ data-driven machine learning models to predict the Mean Arterial Blood Pressure (MAP), from which the absolute BP can be derived. Our study investigates various machine learning algorithms to optimize the prediction accuracy. We find that LSTM, Transformer, and 1D-CNN algorithms using the blood pressure shape and blood flow velocity waveforms as inputs can achieve 8.6, 8.7, and 8.8 mmHg average standard deviation of the prediction error respectively without anthropometric data such as age, sex, heart rate, height, weight. Furthermore, the 1D-CNN model can achieve 7.9mmHg when anthropometric data is added as inputs, improving upon an anthropometric-only model of 9.5mmHg. This machine learning-based approach, capable of converting ultrasound data into MAP values, presents a promising software tool for physicians in clinical decision-making regarding blood pressure management.