DeepStroke: An efficient stroke screening framework for emergency rooms with multimodal adversarial deep learning.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neurology

Background:

• Stroke triage in emergency rooms is challenging, often relying on time-consuming CT scans or less accurate clinical tests.
• High misdiagnosis rates persist despite current stroke screening methods.
• The need for rapid, accurate stroke assessment in acute settings is critical.

Purpose of the Study:

• To introduce DeepStroke, a multimodal deep learning framework for computer-aided stroke presence assessment.
• To utilize readily available facial video and audio data for detecting facial paralysis and speech disorders.
• To improve the accuracy and efficiency of stroke screening in emergency settings.

Main Methods:

• Developed DeepStroke, a multimodal deep learning framework integrating facial video and audio data.
• Employed transfer learning to mitigate biases and enhance generalizability.
• Utilized multi-modal lateral fusion and adversarial training for feature extraction and model regularization.

Main Results:

• DeepStroke demonstrated superior performance compared to state-of-the-art models and human expert teams.
• The framework achieved higher sensitivity (10.94%) and accuracy (7.37%) than traditional stroke triage methods.
• Assessments were completed in under six minutes, indicating high efficiency.

Conclusions:

• DeepStroke offers a promising, rapid, and accurate solution for computer-aided stroke assessment in emergency settings.
• The AI framework has significant potential for clinical translation, improving patient outcomes.
• Multimodal deep learning effectively leverages facial and speech analysis for stroke detection.