Applying Masking Techniques in LSTM Models for ASD Prediction

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by atypical patterns of brain connectivity. Functional Magnetic Resonance Imaging (fMRI) provides a non-invasive way to capture these patterns as time series, which can be modeled with sequence-based neural networks such as Long Short-Term Memory (LSTM) networks. This thesis investigates whether masking strategies can improve the performance and interpretability of LSTM models for ASD classification using data from the Autism Brain Imaging Data Exchange (ABIDE). Two types of masking are explored: input masking, which restricts the model to subsets of brain regions of interest (ROIs) based on L1-SCCA method, and weight masking, based on threshold, which prunes functional connectivity edges to enforce sparser representations. In addition, we reimplemented the L1-Sparse Canonical Correlation Analysis (L1-SCCA) pipeline to evaluate its utility for ROI selection. The results show that the baseline LSTM using all 200 regions gives the most reliable accuracy and F1 scores. L1-SCCA did not yield stable region subsets, input masking only matched the baseline in isolated cases, and weight masking consistently reduced performance. Overall, the full connectivity input remains most effective, while masking strategies did not provide clear benefits in this setting.