Automated validated tool for epileptic seizure detection using deep learning

This paper explores an innovative approach for the automatic detection of epileptic seizures from audio recordings and Heart Rate Variability (HRV) using Convolutional Neural Networks (CNNs). In medical settings, accurately labeling seizure events is critical for patient monitoring. However, manual annotation by experts is not only time-intensive but also highly repetitive. To address this challenge, we developed a structured questionnaire for patients and eyewitnesses, concentrating on observable characteristics during typical seizure events. This questionnaire was used to prospectively study 198 consecutive adult patients with either Psychogenic Non-Epileptic Seizures (PNES) or Epileptic Seizures (ES). For each question, specific signs, symptoms, and risk factors were extracted as variables. The results showed a sensitivity of 95.10% and a specificity of 97.06%, confirming the reliability of the questionnaire. Also, the method proposed in the study categorizes all seizure vocalizations into a singular target event class, modeling the detection task as a binary classification problem target (seizure event) vs. non-target (non-seizure event). The CNN is trained to detect seizure events in short time frames. Experimental results indicate that the method achieves over 92.5% detection accuracy. Furthermore, the research leverages the correlation between pre-ictal epileptic states and HRV features. By addressing the noise interference commonly present during seizures, the proposed model can robustly train the CNN to identify pre-ictal states. The model's performance is promising, yielding an accuracy of over 91.5% for both positive and negative predictions. The proposed system underwent a human evaluation by a group of physicians at Mansoura University Hospital. The results were highly satisfactory, with the doctors expressing strong approval of the system's performance.