Transfer Learning for Monitoring Continuous Blood Pressure Changes: Personalizing for SCI Individuals

Le résumé fourni par la source

Continuous monitoring of beat-to-beat blood pressure (BP) is essential in many conditions, utmost important for detecting autonomic dysregulation in individuals with spinal cord injury (SCI), yet existing methods struggle to generalize due to limited data and physiological variability. We propose a multimodal deep learning framework for personalized estimation of systolic BP changes (∆SBP), combining scalogram-based convolutional feature extraction with bidirectional long short-term memory and attention mechanisms. As a time-frequency representations enabling learning from raw photoplethysmography (PPG) and electrocardiography (ECG) signals. Our model was pretrained on the large-scale ambulatory Aurora BP dataset and adapted via transfer learning to a small SCI cohort, enabling subject-agnostic generalization. On the Aurora BP hold-out (n = 93) set, our model achieves mean absolute error (MAE) of 9.05mmHg(9.02). Moreover, ∆SBP prediction yielded consistently lower error than static SBP across all modalities, underscoring its relevance for dynamic cardiovascular monitoring. Transfer learning shows a strong performance on the SCI cohort (n = 14) achieving an MAE of 4.56mmHg(2.54), satisfying ANSI/AAMI SP10 criteria and achieving BHS Grade B. Gradient-weighted class activation mapping visualizations and attention weights reveal physiologically meaningful patterns, such as PPG systolic upstrokes and ECG R-peaks, supporting explainability for clinical relevance. Additionally, predicted uncertainty closely tracked actual subject-level error, indicating that the model reliably quantifies its confidence across individuals. Our work contributes with a personalized, subjectindependent framework for beat-to-beat ∆SBP prediction proven in SCI individuals.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.