A transfer learning-based calibration-free model for blood pressure prediction using smart monitors

Abstract

Blood pressure monitoring is critical because it enables effective management of hypertension, empowering individuals to take control of their cardiovascular health and prevent serious health complications. In recent years, smart blood pressure monitors have been gradually replacing traditional ones due to their convenience. Collaborating with a company manufacturing smart blood pressure monitors, we develop a calibration-free blood pressure prediction model using electrocardiogram (ECG) and photoplethysmogram (PPG) signals, thereby eliminating the need for initial cuff-based measurement in smart blood pressure monitors. Initially, a pre-trained blood pressure prediction model is established using the publicly available Medical Information Mart for Intensive Care (MIMIC-III) dataset. The pre-trained model, which employs a ResNet deep learning model, achieves a mean absolute error (MAE) of 3.60 for systolic blood pressure (SBP) and 2.97 for diastolic blood pressure (DBP). Subsequently, to ensure the capability of the model in predicting blood pressure based on limited signal data from a smart electronic blood pressure monitor, a novel transfer learning approach known as TL-SQEBPP (Transfer Learning-based Signal Quality Enhanced Blood Pressure Prediction model) is adopted. This framework utilizes the ResNet deep learning model for blood pressure prediction while also incorporating a signal quality model based on the autoencoder as well as an adaptation layer which minimizes the gap between the source domain (MIMIC-III) and the target domain data. Target domain data includes both company-provided and experimental data gathered from subjects using the smart blood pressure monitor. Transfer learning using the target domain data is applied to test and validate the TL-SQEBPP model. The results demonstrate that our proposed method performs well, with TL-SQEBPP achieving an MAE of 4.9 for SBP and 4.19 for DBP with transfer learning applied based on the experimental data. In addition, when transfer learning was applied using the company-provided data, TL-SQEBPP was shown to yield MAEs for SBP and DBP substantially lower compared to alternative versions of the architecture in which the signal quality model and/or the adaptation layer were not included.