An explainable streaming early identification model for early neurological deterioration based on coordinated fusion of ECG waveforms and vital signs.

Endovascular treatment (EVT) has emerged as a cornerstone in the clinical management of stroke. However, postoperative Early Neurological Deterioration (END) persists as a formidable challenge, frequently correlating with long-term disability and poor prognosis. Consequently, timely and accurate prediction of END is imperative for guiding early clinical intervention and improving patient outcomes. This study introduces DSF-Net, a cost-sensitive, dual-stream multimodal framework engineered for the continuous postoperative monitoring of stroke patients. The architecture leverages a dual-stream design: the first stream employs a 1D convolutional neural network (1D-CNN) to extract latent representations from high-frequency physiological waveforms, while the second stream utilizes a multilayer perceptron (MLP) to encode structured clinical indicators. These fused representations are subsequently processed by a Transformer encoder to capture complex temporal dependencies within granular monitoring windows. To mitigate the challenges posed by extreme class imbalance (i.e., the scarcity of END-positive instances), we integrate cost-sensitive learning with a dynamic threshold optimization strategy prioritized for F1-score maximization. Experimental results demonstrate that DSF-Net achieves superior predictive performance on postoperative monitoring datasets, yielding an Area Under the Curve (AUC) of 0.9996 and an F1-score of 0.9841. Notably, with the optimized threshold, the model enhances the END recall rate to 99%, a significant improvement over the 74% achieved by a conventional LSTM baseline. Furthermore, interpretability analysis using Integrated Gradients (IG) reveals that the model can identify subtle morphological variations in waveforms preceding deterioration, thereby facilitating transparent and actionable early warnings. These findings suggest that DSF-Net provides a robust and interpretable solution for intelligent monitoring in post-stroke care.