Evaluating Feature Importance in the Context of Simulation-Based Inference for Cortical Circuit Parameter Estimation

Alessandro Sandron, Alejandro Orozco Valero, Juan Miguel García, Gabriele Mancini, Francisco Pelayo, Christian Morillas, Stefano Panzeri, Pablo Martínez-Cañada

Abstract: Extracellular electrophysiology recordings capturing the activity of neuronal populations, e.g., Local Field Potentials (LFPs), have offered important insights into cortical dynamics. Yet there is still a lack of clarity about how features and characteristics of these extracellular potentials relate to the properties and function of the underlying neural populations. Mechanistic models combined with simulation-based inference (SBI) algorithms have emerged as an effective strategy for developing predictive tools that fit well with available empirical data and can be used to predict key parameters that describe neural activity. Numerous SBI techniques rely on summary statistics or interpretable features to approximate the likelihood or posterior. However, at present, a significant challenge is assessing how each feature impacts the SBI model’s predictions. Here, we developed an approach to determine feature importance in the context of cortical circuit parameter inference. We first created a dataset that includes a million distinct simulations from a spiking cortical microcircuit model of recurrently connected excitatory and inhibitory populations. Biophysics-based causal filters were coupled with spikes to generate realistic LFP data. We then extracted a set of meaningful features from simulated LFP data that were used to train an SBI algorithm. To evaluate feature importance, we employed SHAP values, a prominent tool in machine learning for interpreting the contribution of each feature to the prediction outcomes. Our findings demonstrate the effectiveness of our approach in pinpointing the most critical features, such as dfa, rs_range, low_freq_power or transition_variance, for inferring parameters of a recurrent cortical circuit model based on electrophysiological data.

External IDs:doi:10.1007/978-981-96-3294-7_35