Seizure detection in neonates: Improved classification through supervised adaptation.

Area of Science:

• Medical technology
• Neurology
• Artificial intelligence

Background:

• Neonatal seizures require timely detection for effective intervention.
• Current patient-independent seizure detection systems lack personalized accuracy.
• Developing adaptable systems is crucial for improving neonatal intensive care unit (NICU) monitoring.

Purpose of the Study:

• To investigate the efficacy of supervised adaptation for patient-specific neonatal seizure detection.
• To enhance the performance of a patient-independent seizure detection classifier using individual patient data.
• To improve the accuracy and reliability of seizure alerts in the NICU.

Main Methods:

• Utilized supervised adaptation techniques, including gradient descent and least mean squares.
• Adapted a pre-existing patient-independent seizure detection classifier.
• Evaluated classifier performance using receiver operating characteristic (ROC) area and accuracy metrics.

Main Results:

• Achieved a 3% increase in mean ROC area with the adapted classifier.
• Observed a 7.7% increase in mean accuracy compared to the patient-independent model.
• Demonstrated the effectiveness of patient-specific data in improving seizure detection performance.

Conclusions:

• Supervised adaptation of patient-independent classifiers shows significant potential for neonatal seizure detection.
• Personalized adaptation using patient-specific data leads to improved accuracy in identifying neonatal seizures.
• This approach can enhance the reliability of seizure alerts in the NICU, aiding clinical decision-making.