Neural Decoding and Feature Selection Techniques for Closed-Loop Control of Defensive Behavior.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Psychiatric disorders often involve excessive avoidant or defensive behaviors.
• Predicting these behaviors from neural signals like local field potentials (LFPs) is crucial for developing closed-loop neuromodulation therapies.
• Identifying specific LFP features that encode defensive behaviors presents a significant challenge.

Purpose of the Study:

• To identify and evaluate local field potential (LFP) features for accurately decoding defensive behaviors.
• To investigate the informativeness of various neuro-markers across spectral, temporal, and connectivity domains.
• To assess the performance of machine learning models in predicting freezing, bar-press suppression, and motion.

Main Methods:

• Analyzed LFP signals from rat infralimbic cortex and basolateral amygdala during tone-shock conditioning and extinction.
• Utilized a comprehensive set of neuro-markers and SHapley Additive exPlanations for feature importance.
• Employed Light Gradient-Boosting Machine models to decode three defensive behaviors: freezing, bar-press suppression, and motion.

Main Results:

• Band power and inter-channel band power ratios were optimal features for decoding defensive behaviors.
• High-gamma band power and inter-regional correlations proved more informative than other spectral bands.
• Achieved high accuracy in decoding accelerometry jerk (R²=0.5357, r=0.7579) and bar press rate (R²=0.3476, r=0.6092) with low computational complexity (<110 ms training, <1 ms inference).

Conclusions:

• Accurate, low-latency decoding of defensive behaviors from LFP features is feasible.
• The methodology highlights the potential for real-time decoding in closed-loop psychiatric neuromodulation.
• Identified specific neural circuit features that can serve as physiological targets for therapeutic interventions.