A feasibility study on AI-controlled closed-loop electrical stimulation implants.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Miniaturized electrical stimulation (ES) implants offer therapeutic potential but face challenges in real-time control due to computationally intensive biophysical models.
• Developing efficient algorithms is crucial for practical application and advanced control of ES devices.

Purpose of the Study:

• To investigate the feasibility of using computationally efficient machine learning (ML) methods for real-time control of ES implants.
• To assess the accuracy of ML models in predicting muscle response (normalized twitch force) under ES.

Main Methods:

• Utilized a random forest regressor model for predicting normalized twitch force in the extensor digitorum longus muscle of Wistar rats.
• Performed intra-subject and cross-subject calibration with 11 rats, involving 2000 training stimulations.
• Evaluated model performance based on mean absolute error (MAE).

Main Results:

• Achieved a mean absolute error of 0.03 in an intra-subject setting, indicating high prediction accuracy.
• Obtained a mean absolute error of 0.2 in a cross-subject setting, highlighting variability and the need for further research.
• This research represents the first experimental demonstration of AI simulating complex ES mechanistic models.

Conclusions:

• Machine learning, specifically random forest regression, shows significant promise for controlling ES implants with improved computational efficiency.
• While intra-subject control is highly accurate, cross-subject control requires further investigation and development of error reduction techniques.
• These findings open new avenues for AI-driven advancements in personalized and adaptive electrical stimulation therapies.