ACFA: A Hybrid Deep Learning Framework for Cuffless Continuous Blood Pressure Estimation Using Time–Frequency Adaptive PPG Features

Non-invasive blood pressure (BP) monitoring plays a critical role in the diagnosis and prevention of hypertension and related disorders. In recent years, deep learning–based methods leveraging photoplethysmography (PPG) signals have shown promising progress in BP estimation. However, traditional approaches still face several limitations: (1) long-term dependencies within PPG sequences are difficult to capture effectively, (2) frequency-domain feature extraction often lacks sufficient depth to reveal multi-scale patterns, and (3) High-dimensional redundancy increases model complexity and weakens generalization. To overcome these challenges, this paper proposes a hybrid deep learning framework, termed the Adaptive Cross-domain Fusion Architecture (ACFA), for non-invasive BP prediction. ACFA integrates four specialized modules: DyCASNet, for adaptive time–frequency dual-domain feature extraction; xLSTM, for enhanced modeling of long-term dependencies; Transformer, for capturing global contextual information; and FKAN, for characterizing complex nonlinear relationships. Experiments conducted on the public MIMIC-III and PPG-BP databases demonstrate that ACFA substantially reduces the mean absolute error (MAE) in predicting systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean arterial pressure (MAP). The prediction results not only outperform mainstream approaches but also satisfy the Class A standards established by the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS). Overall, this study presents an end-to-end solution for continuous, non-invasive blood pressure monitoring that offers high predictive accuracy and strong generalization capabilities, underscoring its potential for integration into wearable health devices and for broad clinical deployment.