Neural shape mediation analysis

Jules Brochard, Etienne Combrisson, Benoît Chatard, Jean-Philippe Lachaux, Andrea Brovelli, Jean DaunizeaubioRxiv 2023 · DOI

Abstract

Neural signal shapes convey significant information about their generating processes. In this study, we introduce a data-driven methodology to identify sensory and behaviourally-relevant traces within neural responses. We present a phenomenological model that characterises temporal variations in intracranial EEG using eight interpretable parameters: peak time, peak intensity, initial and final baselines, accumulation and depletion period, and their respective concavities. This model effectively captures subtle signal variations, especially in sensory decision-making tasks. By decomposing the signals in this manner, we then conduct a comprehensive brain mediation analysis on iEEG data’s shape, pinpointing regions that mediate behavioural processes. Importantly, we can determine which signal dynamics specifically reflect underlying behavioural processes, enhancing the depth of analysis and critique of their role in behaviour. Preliminary applications on a cohort of epileptic patients reveal that our model explains over a third of the signal variance at the trial level across all brain regions. We identified four key regions—encompassing sensory, associative, frontal, and premotor areas—that mediate the impact of task difficulty on reaction time. Notably, in these regions, it was the depletion period, rather than signal amplitude, that correlated with behavioural speed. This study highlights the potential of our approach in providing detailed insights into the neural mechanisms linking stimuli to behaviour.

Introduction

In cognitive neuroscience, the "neural code" serves as the critical bridge between sensory stimuli and behavioral responses. Traditional approaches to identifying this code often rely on simple correlations—for instance, asking whether a stronger signal in the visual cortex correlates with a faster reaction time. However, these methods frequently treat neural activity as a "black box" or depend on averaged data that obscures essential temporal nuances. While mediation analysis can test if a brain region (M) sits between a stimulus (X) and a behavior (Y), standard implementations often lack interpretability. They may identify that a region is involved without specifying which dynamic aspect of the signal—timing, intensity, or duration—drives the behavior.

Methods: Shape Mediation Analysis (SMA)

To address these limitations, we introduce Shape Mediation Analysis (SMA), a data-driven framework that decomposes complex neural signals into eight interpretable "shape" parameters.

The Phenomenological Model

Rather than analyzing raw voltage, SMA fits a model to the high-gamma power envelope of intracranial EEG (iEEG) signals. The model captures:

  • Basics: Peak time, peak intensity, and initial/final baselines.
  • Dynamics: Integration period (ramp-up), depletion period (ramp-down), and their respective concavities (curvature).

Hierarchical Estimation and Temporal Scaling

To manage single-trial noise, a hierarchical "top-down" fitting procedure was employed, starting from grand averages and narrowing down to individual trials. Crucially, signals were aligned using a completion-time scale, where the time axis is rescaled such that $t=0$ represents the stimulus and $t=1$ represents the response. This approach explained significantly more variance (~34%) than standard stimulus-locked (~22%) or response-locked alignments (~22%).

Experimental Paradigm and Datasets

The study utilized a clinical dataset from 62 patients with intractable epilepsy undergoing neurosurgical monitoring. Participants performed a visual search task requiring them to locate a "T" among "L" distractors.

Experimental paradigm: A visual search task where participants identify the location of a target 'T' among distractors.

Trials were categorized as EASY (color pop-out) or HARD (same-color distractors), creating a clear behavioral contrast in reaction times and accuracy.

Key Results

Behavioral Performance

The task design successfully modulated difficulty, as evidenced by significant differences in response times and accuracy between conditions.

Behavioral results showing the percentage of correct answers and response times for Easy vs. Hard trials.

The "Tail" as a Behavioral Mediator

The SMA model explained an average of 33% of signal variance at the single-trial level. The most significant finding was the identification of four key mediating hubs: the right caudal medial visual cortex (R VCcm), right rostral ventral premotor cortex (R PMrv), right ventral inferior parietal cortex (R IPCv), and left ventral orbitofrontal cortex (L OFCv).

In these regions, the depletion window (w2)—the time the signal takes to decay after its peak—emerged as the primary mediator of behavior. While signal amplitude is often the focus of traditional studies, our analysis revealed that the duration of the neural trace was more predictive of reaction time. Specifically, a longer depletion window relative to task difficulty was associated with faster responses, suggesting that neural persistence may function as a "processing buffer" or evidence accumulation phase.

Discussion and Future Directions

While SMA provides a powerful descriptive tool, it remains a phenomenological model. It characterizes the shape of the signal without explaining the underlying biophysical mechanisms, such as neurotransmitter kinetics. Furthermore, the use of a clinical epilepsy population necessitates caution in generalizing to healthy cohorts, despite the exclusion of seizure-prone zones. Future research should aim to bridge this gap by linking SMA parameters to mechanistic circuit models and utilizing more fine-grained anatomical parcellations to further isolate functional sub-regions.

Summary updated

Abstract

Abstract Neural signal shapes convey significant information about their generating processes. In this study, we introduce a data-driven methodology to identify sensory and behaviourally-relevant traces within neural responses. We present a phenomenological model that characterises temporal variations in intracranial EEG using eight interpretable parameters: peak time, peak intensity, initial and final baselines, accumulation and depletion period, and their respective concavities. This model effectively captures subtle signal variations, especially in sensory decision-making tasks. By decomposing the signals in this manner, we then conduct a comprehensive brain mediation analysis on iEEG data’s shape, pinpointing regions that mediate behavioural processes. Importantly, we can determine which signal dynamics specifically reflect underlying behavioural processes, enhancing the depth of analysis and critique of their role in behaviour. Preliminary applications on a cohort of epileptic patients reveal that our model explains over a third of the signal variance at the trial level across all brain regions. We identified four key regions—encompassing sensory, associative, frontal, and premotor areas—that mediate the impact of task difficulty on reaction time. Notably, in these regions, it was the depletion period, rather than signal amplitude, that correlated with behavioural speed. This study highlights the potential of our approach in providing detailed insights into the neural mechanisms linking stimuli to behaviour.
BibTeX
@misc{brochard2023neural,
  author = {Jules Brochard and Etienne Combrisson and Benoît Chatard and Jean-Philippe Lachaux and Andrea Brovelli and Jean Daunizeau},
  title = {{Neural shape mediation analysis}},
  year = {2023},
  doi = {10.1101/2023.09.21.558808},
  howpublished = {bioRxiv},
}
Figure