Learning Temporal Basis Vectors for Closed-Loop Neural Stimulation

Area of Science:

• Computational Neuroscience
• Artificial Intelligence in Medicine
• Neurotechnology

Background:

• Accurate forecasting of neural responses is crucial for effective brain stimulation.
• Existing models often lack the efficiency and adaptability required for real-time closed-loop applications.

Purpose of the Study:

• Introduce a novel Temporal Basis Function Model (TBFM) for spatiotemporal neural response prediction.
• Enable model-based control techniques for closed-loop neural stimulation.
• Demonstrate the model's clinical relevance by optimizing efficiency and latency.

Main Methods:

• Developed a TBFM framework learning temporal basis functions.
• Applied TBFMs to micro-electrocorticography (μECog) data from non-human primate optogenetic stimulation experiments.
• Compared TBFM performance against complex non-linear dynamical systems models and linear state space models (LSSMs).

Main Results:

• TBFMs achieved accuracy comparable to complex non-linear models and surpassed LSSMs.
• Required minimal data collection (<20 min) and training time (<5 min).
• Successfully demonstrated shaping neural activity towards desired regimes in simulated closed-loop experiments.

Conclusions:

• TBFMs offer an efficient and accurate approach for modeling neural responses to stimulation.
• The model's optimization in sample efficiency, training time, and latency bridges the gap towards AI-driven closed-loop stimulation therapies.
• This framework holds potential for developing novel treatments for neurological conditions.