Enhancing Automated Patient Detection and Tracking in Epilepsy Monitoring Unit Videos.

Area of Science:

• Medical technology
• Artificial intelligence in healthcare
• Neurology

Background:

• Automated seizure detection via video monitoring can be biased by incidental events.
• Existing algorithms lack patient-specific identification, impacting accuracy.
• Patient tracking can isolate individuals, reducing bias and improving specificity.

Purpose of the Study:

• To develop and evaluate a patient tracking system integrated into seizure detection algorithms.
• To mitigate biases in automated video-based epilepsy monitoring.
• To enhance the accuracy and reliability of seizure detection.

Main Methods:

• Fine-tuning a YOLOv8 object detection model with 1,590 labeled frames from 131 epilepsy monitoring unit patients.
• Implementing hyperparameter optimization via grid-search with cross-validation.
• Developing a novel post-processing pipeline for video analysis.

Main Results:

• The fine-tuned YOLOv8 model achieved 97.6% recall and 96.8% precision in patient detection.
• The post-processing pipeline corrected 577 out of 580 errors in new seizure videos.
• The combined approach demonstrated high performance in diverse patient conditions and video quality.

Conclusions:

• The proposed patient tracking and video analysis pipeline significantly enhances video-based seizure detection.
• This bias-free approach improves the reliability of epilepsy monitoring.
• The system enables faster, more accurate interventions, ultimately improving patient safety.