A Multidimensional Data-Driven Hybrid Transformer Framework for Non-invasive Continuous Blood Pressure Prediction

This study investigates whether a cuffless continuous blood pressure (BP) estimation system can achieve high accuracy by modeling multi-source physiological feature sequences rather than raw waveform samples. Specifically, the authors ask if combining diverse machine learning paradigms—such as Transformers, Kolmogorov-Arnold Networks, and XGBoost—can better capture the complex temporal and nonlinear dynamics of cardiovascular regulation.

The study utilizes synchronized ECG, PPG, and arterial blood pressure (ABP) signals from the MIMIC-III database, supplemented by demographic covariates like age and body weight. Instead of feeding raw waveform segments directly into a neural network, the framework extracts 10-step sequences of eight physiological and demographic descriptors, including pulse transit time, heart rate, PPG foot-to-peak interval, and maximum PPG slope. A Multi-Source Temporal Encoder Module processes these feature sequences using parallel branches, while a Dynamic Conditional Fusion-Decoder applies differential multi-head attention and gated residual correction to generate final continuous DBP and SBP estimates.

Evaluated on 2,431 segment-level held-out test windows, the framework achieved a mean error and standard deviation of 0.41 plus-minus 3.74 mmHg for diastolic blood pressure and -1.60 plus-minus 5.95 mmHg for systolic blood pressure. The 95% limits of agreement were [-6.93, 7.74] mmHg for DBP and [-13.25, 10.06] mmHg for SBP. Furthermore, the proportions of predictions within 10 mmHg of the reference values reached 98.48% for DBP and 94.36% for SBP, outperforming locally retrained baseline models.

Continuous non-invasive blood pressure monitoring is crucial for early hypertension detection and long-term cardiovascular risk management. By demonstrating that feature-sequence fusion can surpass traditional single-modality or purely deep learning baselines, this work advances the development of cuffless monitoring technologies that avoid the discomfort and intermittent nature of traditional arm cuffs.