Continual learning for cuffless blood pressure estimation

Despite extensive studies on cuffless continuous blood pressure (BP) estimation through machine learning models, those models are typically constrained by a one-off training strategy resulting in fixed model parameters and inadequate adaptation in response to new patterns of data. BP is a dynamic vital sign with a concept drift characteristic. With static models trained with fixed BP datasets and traditional learning (TL) technique, the estimation performance would degrade when the to-be predicted BP distributions deviate from trained ones. In this paper, we propose a novel continual learning (CL) framework for continuous BP estimation. Such framework enables deep learning models to dynamically and sequentially learn continuous BP signals even in the presence of concept drift. We validated the proposed framework through two types of CL models on the data from 3,850 samples (403.67 h) in the University of California Irvine (UCI) database, and compared with the TL model under several controlled experiments. The results showed that the CL model learns well even when different level of concept drift exists in continuous BP. Further, the BP estimation performance of the CL model equipped with 1D-CNN improved by 20.86% on average compared to that of the TL model. These findings suggest that the CL model has a great advantage over the TL model for dynamic continuous BP estimation.