Neural decoding and feature selection methods for closed-loop control of avoidance behavior

Jinhan Liu, Rebecca Younk, Lauren M Drahos, Sumedh S Nagrale, Sangeeta Yadav, Alik S. Widge, Mahsa ShoaranJournal of Neural Engineering 2024 · DOI

TL;DR

This paper introduces a novel machine learning framework for real-time decoding of defensive behaviors (like freezing and avoidance) from neural signals (LFPs) in rats. By employing a comprehensive set of neuro-markers and a sophisticated feature selection method (SHAP), they achieve high accuracy with remarkably low computational cost, paving the way for closed-loop psychiatric neuromodulation.

What Problem Does This Paper Address?

Many psychiatric disorders, such as anxiety and PTSD, are characterized by excessive or maladaptive defensive/avoidant behaviors. Developing closed-loop brain-machine interfaces (BMIs) that can predict and potentially modulate these behaviors in real-time is a major goal for psychiatric treatment. However, a significant challenge lies in identifying which neural signals (specifically, local field potentials or LFPs) and which features within those signals accurately encode these complex behaviors with the speed and efficiency required for real-time intervention. Previous work has often focused on simpler features or longer timescales, leaving a gap in understanding how to decode moment-to-moment defensive actions from LFPs for clinical applications.

Methods

The researchers developed a machine learning framework built around the Light Gradient-Boosting Machine (LightGBM), a highly efficient gradient-boosted decision tree algorithm. A key innovation is their comprehensive neuro-marker extraction process, which goes beyond traditional spectral power features to include temporal and connectivity domains. They analyzed 17 types of neuro-markers, resulting in over 1200 features.

Crucially, they employed SHapley Additive explanations (SHAP) for feature importance assessment. This allowed them to identify the most informative features and implement a feature selection strategy. This strategy iteratively selected features based on their SHAP values, aiming to maintain decoding performance while drastically reducing the number of features. This approach balances the need for rich neural information with the computational constraints of real-time applications. They also compared LightGBM against a range of other ML models (including ANNs) and explored the impact of temporal lags in neural data.

Datasets

The study utilized electrophysiological (LFP) and behavioral data collected from 16 adult Long Evans rats. The data was acquired during a tone-shock conditioning and extinction protocol, a standard paradigm for studying defensive behaviors. The dataset was partitioned into training, validation, and testing sets using a sliding-window 5-fold cross-validation approach to respect the temporal nature of the data. No new benchmark datasets were introduced; rather, existing animal models and data collection paradigms were leveraged.

Key Insights and Ideas

  • High-gamma power is king: The study consistently found that high-gamma band activity (80-150 Hz) in both the infralimbic cortex (IL) and basolateral amygdala (BLA) was a critical predictor of defensive behaviors, outperforming other frequency bands. This challenges the traditional focus on theta band activity for fear-related research.
  • Behavioral specificity matters: Different defensive behaviors (accelerometry jerk vs. bar press suppression) are encoded by distinct neural features. Accelerometry jerk is more strongly linked to local spectral power, while bar press suppression relies more on inter-regional connectivity. This suggests that targeted interventions might need to consider these specific neural correlates.
  • Feature selection is crucial for efficiency: By using SHAP to identify and select the most informative neuro-markers, the researchers could achieve high decoding accuracy with a drastically reduced feature set (e.g., 36 features out of 1296), leading to significant reductions in training/inference time and memory usage. This is vital for real-time BMI applications.
  • Freezing is surprisingly hard to decode: Despite being a common measure of defensive behavior, the freezing score was the least decodable, suggesting that its neural correlates might be more complex or less directly captured by LFPs at the fine timescales studied.

Key Results

  • High decoding accuracy: The model achieved an average coefficient of determination (R²) of 0.5357 and a Pearson correlation coefficient (r) of 0.7579 for accelerometry jerk, and R² = 0.3476 and r = 0.6092 for bar press rate.
  • Remarkably low computational cost: Training took <310 ms, inference was <0.051 ms, and memory usage was 16.6 kB on a single CPU core. This is orders of magnitude faster and more memory-efficient than many other ML models.
  • Feature importance: Band power (BP) and band power ratio between channels (BPRC) were identified as optimal features. High-gamma band activity was consistently the most informative across various analyses.
  • Feature selection benefits: Using only the top-ranked features (e.g., 36 out of 1296) maintained decoding performance while reducing training time by over 97% and memory usage by over 90%.

Limitations and Future Work

  • Freezing score decoding: The paper acknowledges the difficulty in accurately decoding the freezing score, suggesting that their focus on second-to-second changes might not fully capture the broader trends that previous studies have found to be more decodable. Future work could explore longer time windows or different neural signals.
  • Offline vs. Online: The current study was conducted offline. The authors plan to deploy their framework within an online paradigm to assess its real-time performance and feasibility for actual interventions.
  • Generalizability: While the study demonstrates promise, validating the model on additional datasets and in different task contexts is necessary to ascertain its broader applicability.
  • Hybrid Models: The authors suggest exploring hybrid models that combine the strengths of manual neuro-marker selection with the automatic feature extraction capabilities of ANNs.
  • Intervention Design: While SHAP identifies important features, translating this into direct, actionable interventions (e.g., "reduce high-gamma power in BLA") requires further research into causal relationships.

Summary updated

Abstract

<i>Objective.</i>Many psychiatric disorders involve excessive avoidant or defensive behavior, such as avoidance in anxiety and trauma disorders or defensive rituals in obsessive-compulsive disorders. Developing algorithms to predict these behaviors from local field potentials (LFPs) could serve as the foundational technology for closed-loop control of such disorders. A significant challenge is identifying the LFP features that encode these defensive behaviors.<i>Approach.</i>We analyzed LFP signals from the infralimbic cortex and basolateral amygdala of rats undergoing tone-shock conditioning and extinction, standard for investigating defensive behaviors. We utilized a comprehensive set of neuro-markers across spectral, temporal, and connectivity domains, employing SHapley Additive exPlanations for feature importance evaluation within Light Gradient-Boosting Machine models. Our goal was to decode three commonly studied avoidance/defensive behaviors: freezing, bar-press suppression, and motion (accelerometry), examining the impact of different features on decoding performance.<i>Main results.</i>Band power and band power ratio between channels emerged as optimal features across sessions. High-gamma (80-150 Hz) power, power ratios, and inter-regional correlations were more informative than other bands that are more classically linked to defensive behaviors. Focusing on highly informative features enhanced performance. Across 4 recording sessions with 16 subjects, we achieved an average coefficient of determination of 0.5357 and 0.3476, and Pearson correlation coefficients of 0.7579 and 0.6092 for accelerometry jerk and bar press rate, respectively. Utilizing only the most informative features revealed differential encoding between accelerometry and bar press rate, with the former primarily through local spectral power and the latter via inter-regional connectivity. Our methodology demonstrated remarkably low training/inference time and memory usage, requiring<310 ms for training,<0.051 ms for inference, and 16.6 kB of memory, using a single core of AMD Ryzen Threadripper PRO 5995WX CPU.<i>Significance.</i>Our results demonstrate the feasibility of accurately decoding defensive behaviors with minimal latency, using LFP features from neural circuits strongly linked to these behaviors. This methodology holds promise for real-time decoding to identify physiological targets in closed-loop psychiatric neuromodulation.
BibTeX
@article{liu2024neural,
  author = {Jinhan Liu and Rebecca Younk and Lauren M Drahos and Sumedh S Nagrale and Sangeeta Yadav and Alik S. Widge and Mahsa Shoaran},
  title = {{Neural decoding and feature selection methods for closed-loop control of avoidance behavior}},
  journal = {Journal of Neural Engineering},
  year = {2024},
  doi = {10.1088/1741-2552/ad8839},
}
Figure