Abstract
Hypertension and blood pressure variability (BPV) are major risk factors for cardiovascular disease (CVD). Single-channel photoplethysmography (PPG) has emerged as a promising daily blood pressure (BP) monitoring tool. However, estimating BP trends presents challenges due to complex temporal dependencies and continuous fluctuations. Traditional methods often address BP prediction as isolated tasks and focus solely on temporal dependencies within a limited time window, which may fall short of capturing the intricate BP fluctuation patterns implied in varying time spans, particularly amidst constant BP variations. To address this, we propose a novel deep learning model featuring a two-stage architecture and a new input structure called contextual cycles. This model estimates beat-to-beat systolic blood pressure (SBP) trends as a sequence prediction task, transforming the output from a single SBP value into a sequence. In the first stage, parallel ResU Blocks are utilized to extract fine-grained features from each cycle. The generated feature vectors are then processed by Transformer layers with relative position encoding (RPE) to capture inter-cycle interactions and temporal dependencies in the second stage. Our proposed model demonstrates robust performance in beat-to-beat SBP trend estimation, achieving a mean absolute error (MAE) of 3.186 mmHg, a Pearson correlation coefficient applied to sequences (Rseq) of 0.743, and a variability error (VE) of 1.199 mmHg. It excels in steady and abrupt substantial fluctuation states, outperforming baseline models. The results reveal that our method meets the requirements of the AAMI standard and achieves grade A according to the BHS standard. Overall, our proposed method shows significant potential for reliable daily health monitoring.