Closed-Loop Neural Prostheses With On-Chip Intelligence: A Review and a Low-Latency Machine Learning Model for Brain...

Area of Science:

• Systems neuroscience
• Neurotechnology
• Biomedical engineering

Background:

• Closed-loop neuromodulation systems are crucial for understanding the brain and developing therapies.
• Current devices lack sufficient on-chip processing and intelligence for advanced applications.
• Neural prostheses require multi-channel recording, on-site processing, and rapid detection for effective closed-loop stimulation.

Purpose of the Study:

• To review existing closed-loop neuromodulation devices and neural prostheses with on-chip machine learning.
• To propose a new energy-area efficiency metric for comparing on-chip classifiers.
• To present techniques for optimizing tree-based classifiers for improved performance and efficiency.

Main Methods:

• Review of commercial and investigational closed-loop neuromodulation devices.
• Analysis of application-specific integrated circuits (ASICs) for neural prostheses.
• Development of a novel Depth-Variant Tree Ensemble (DVTE) and cost-aware learning approach.

Main Results:

• Proposed an energy-area (E-A) efficiency figure of merit for on-chip classifiers.
• Introduced DVTE to reduce processing latency by 2.5x in seizure detection.
• Demonstrated algorithm-hardware co-design for energy- and memory-optimized tree-based models with high accuracy and low latency.

Conclusions:

• Algorithm-hardware co-design is key to developing efficient and accurate on-chip classifiers for neural prostheses.
• Novel tree-based models offer reduced latency and interpretable decision-making for safety-critical applications.
• Advancements in on-chip processing are essential for next-generation closed-loop neuromodulation therapies.