Cascade Windows-Based Multi-Stream Convolutional Neural Networks Framework for Early Detecting In-Sleep Stroke Using Wristbands

Abstract

A stroke, particularly when it occurs during sleep, is likely to have a negative prognosis due to delayed detection. Timely and early detection plays a vital role in ensuring prompt administration of reperfusion therapy and preventing permanent disabilities. To address this, we propose a wearable system comprising two wristbands that monitor asymmetric motion patterns (hemiparesis) during sleep. A novel deep learning framework called Early Detection of In-sleep Stroke (EDIS) serves as the core engine for stroke detection during sleep. The framework employs cascading windows of various sizes for convolutional neural networks (CNNs) to enhance both the detection performance and the detection time. We utilize 1D accelerometer sensor data from both hands to generate 2D matrix images, which serve as input for multiple CNN models. Predictions from these models are combined using blending ensemble learning to make a final decision. Although the EDIS framework requires a larger parameter size and longer inference time due to its network architecture with multiple CNNs, it outperforms five single-CNN models by improving detection performance and reducing detection time. Extensive evaluation results demonstrate that EDIS framework accurately and quickly detects in-sleep stroke within the deadline (3 hours). EDIS-Resnet50 has the best classification performance out of the ten DL model candidates, with an F1-score of 0.955 (0.950, 0.960). We believe that our framework will be a fundamental component of real-time stroke monitoring systems, contributing to a reduction in mortality rates among patients suspected of having a stroke.

Keywords

• Deep learning
• ensemble learning
• stroke detection
• sleep
• wearable computing