STFM: Enhancing Autism Spectrum Disorder Classification Through Ensemble Learning-Based Fusion of Temporal and Spatial fMRI Patterns

Ziyu Zhou, Yiming Huang, Yining Wang, Yin Liang

Abstract: This paper presents the novel spatial and temporal fusion model (STFM), an effective approach for Autism Spectrum Disorder (ASD) detection and classification tasks using foundational machine learning models. Utilizing ensemble learning principles, STFM improves the classification performance by integrating weak classifiers. The process begins with the sliding window method applied to fMRI data, constructing brain networks through Pearson correlation calculation between brain regions. This infuses the network with both temporal and spatial patterns. Then, bidirectional LSTM (Bi-LSTM) and 2DCNN are applied for temporal and spatial feature extraction respectively. The model further ensures smoother data variations between patterns through interpolation, and utilizes a basic cross attention mechanism for fusion of patterns. The fused patterns are then classified by a simple SVM classifier. The presented STFM model demonstrates a remarkable classification accuracy of 70.42%, surpassing most fundamental machine learning models in ASD detection.