Hybrid dataset of real-and-simulated pulse waveforms for deep learning-driven blood pressure estimation using MXene-based iontronic sensors

초록

Deep learning-based non-invasive blood pressure (BP) estimation is emerging for cardiovascular disease management, yet it remains constrained by the scarcity of biomedical data and the difficulty of building physiologically faithful training datasets from limited cohorts. In this study, a hybrid framework integrating physiologically grounded PDA-based data augmentation, CNNLSTM architecture, and MXene-based pseudocapacitive iontronic pressure sensor is presented for BP estimation from radial artery pulse waveforms. The sensor, incorporating MXene electrodes and an ionic liquid/polymer dielectric layer, exhibited high sensitivity (101 kPa- 1), fast response time (43 ms), and mechanical durability exceeding 10,000 load cycles. With a high areal capacitance of 161.9 mF/cm2, the sensor enabled stable acquisition of radial artery waveforms. To overcome the data limitations of biomedical research, the proposed simulator was utilized to expand the limited reference cohort within its physiological distribution, enabling the hybrid architecture to jointly learn time-series pulse characteristics and BP features. The model achieved R2 values of 0.907 and 0.904 for systolic and diastolic BP, with mean absolute errors of -2.16 and 2.99 mmHg, respectively. The proposed framework thus offers a physiologically grounded approach to BP estimation that addresses biomedical data limitations, with the potential to be applied to other limited cohorts in clinical and biomedical research.

키워드