Estimating E/I Balance from Resting-State fMRI Using Biologically Inspired LSTM Networks: Application to ASD Subtyping

This thesis investigates Excitatory/Inhibitory (E/I) balance abnormalities in Autism Spectrum Disorder (ASD) through a novel integration of biologically informed Long Short-Term Memory (LSTM) networks and intra-regional connectivity features. Using resting-state functional Magnetic Resonance Imaging (rs-fMRI) data from the Autism Brain Imaging Data Exchange (ABIDE) dataset, a modified LSTM network was developed to simulate the dynamics of excitatory and inhibitory neuron populations. Instead of modeling interactions between regions, the network was independently fitted to the Blood Oxygen Level Dependent (BOLD) time series of each of the 116 brain regions. From this, biologically grounded intra-area E/I connectivity matrices were extracted, capturing local interaction patterns between excitatory and inhibitory units. Additionally, Ne/Ni were computed for each brain region and statistically evaluated across subjects. These ratios, along with intra-area E/I connectivity features, served as input to a Self- Organizing Map (SOM) clustering pipeline, which revealed ASD subtypes characterized by distinct Ne/Ni profiles and associated functional networks. The findings underscore the importance of regional E/I balance, particularly the Ne/Ni, as a key feature in explaining ASD heterogeneity and offer a biologically interpretable framework for identifying subtype-specific patterns relevant to future diagnostics and interventions.