Computationally efficient neural network classifiers for next generation closed loop neuromodulation therapy - a case...

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Medical Devices

Background:

• Responsive neuromodulation systems require efficient real-time biomarker classification.
• Classical filter-based classifiers can be complex to program for patient-specific needs.
• Minimizing computational load is crucial for implantable systems.

Purpose of the Study:

• To explore the utility of compact neural network classifiers for real-time field-potential based biomarker classification.
• To develop a neural network classifier for seizure-state detection in refractory epilepsy suitable for implantable pulse generators.
• To reduce the programming burden on clinicians and decrease detection latency.

Main Methods:

• A compact, feed-forward neural network architecture with minimal units was designed.
• The neural network was trained and tested for seizure-state classification using field-potential data.
• Performance was compared against classical filter-based classifiers using clinician-labeled data.

Main Results:

• The proposed neural network classifier demonstrated comparable accuracy to filter-based classifiers.
• The neural network significantly reduced detection latency compared to classical methods.
• The minimal architecture design accommodates on-board computational constraints of implantable devices.

Conclusions:

• Neural network classifiers offer a viable, efficient alternative for real-time biomarker classification in responsive neuromodulation.
• The developed compact neural network is suitable for on-board processing in implantable pulse generators for epilepsy management.
• This approach promises improved clinical workflow and faster patient response.