A bimodal deep learning framework for cerebral perfusion assessment using synchronous ECG and near-field coupled CBF signals

Accurate real-time assessment of cerebral blood flow (CBF) is crucial for optimizing treatment decisions and improving prognostic outcomes in patients with acute ischemic stroke (AIS). In this study, leveraging the advantage of near-field coupling (NFC) technology in capturing blood flow information from intracranial vessels of different sizes, and in combination with synchronous ECG signals reflecting cardiac rhythm fluctuations, we propose a bimodal deep learning framework to characterize cerebral perfusion-related state changes under postural tilt using synchronous ECG and NFC-based CBF (NFC-CBF) signals. The framework consists of a Multi-Scale Feature Extractor (MSFE) and a Depthwise-Separable Cross-Attention Fusion (DSCAF) module, which work collaboratively to achieve hierarchical temporal representation learning and cross-modal fusion of ECG and NFC-CBF signals. Using synchronous ECG and NFC-CBF data collected from young and middle-aged healthy volunteers, we evaluated the proposed framework in binary and four-class classification tasks to assess its feasibility for cerebral perfusion-related state classification. The results show that NFC-CBF signals can capture CBF changes before and after tilt. Synchronous monitoring of ECG and NFC-CBF provided complementary information on cardiac rhythm and cerebral perfusion-related dynamics. The proposed framework outperforms Bimodal Feature Fusion, ResNet-SE-LSTM, and Cross-Transformer Network in terms of accuracy, F1-score, specificity, precision, and recall, achieving accuracies of 95.95% ± 1.54% and 79.29% ± 2.12% in the binary posture classification and four-class age–posture classification tasks, respectively. Ablation experiments indicate that bimodal modeling provided additional discriminative information compared with single-modality inputs, particularly for the four-class task involving both posture and age. The classification performance achieved after incorporating the DSCAF module is superior to that of other conventional feature fusion strategies. These findings support the feasibility of bimodal ECG and NFC-CBF modeling for cerebral perfusion-related state classification under postural perturbation, while further validation in larger, more diverse, and demographically balanced cohorts, including both patient and non-patient populations, is required before its clinical applicability can be established.