Machine Learning Techniques and Chi-square Feature Selection for Diagnostic Classification Model of Autism Spectrum Disorder Using fMRI Data

Background: Autism Spectrum Disorder (ASD) diagnosis relies on subjective observation, hindered by ASD's diverse presentation, symptom overlap, and sex-specific neurobiology, causing misdiagnosis, especially in females. Thus, objective and reliable diagnostic methods are critical. New method: This rs-fMRI study built a sex-dependent ASD diagnostic classification model (DCM) using functional connectivity. After preprocessing, GICA and dual regression were applied. Coherence and mutual information extracted frequency/nonlinear time-domain features. Chi-square feature selection with forward search identified optimal features, evaluated across 4 machine learning (ML) models (Decision Trees, Naïve Bayes, support vector machine (SVM), and K-Nearest Neighbors (K-NN(). Bayesian optimization tuned hyperparameters, and hill-climbing determined feature inclusion. Time-domain features were classified using correlation. Furthermore, the Minimum Redundancy Maximum Relevance (MRMR) algorithm was used to re-evaluate feature selection, assessing the impact of chi-square selected features on classification accuracy. Results: Chi-Squared feature selection with linear full correlation yielded 96.6% accuracy for males, while frequency domain selection using K-NN at specific frequencies achieved 86.7% accuracy for females. Comparison with existing methods: This study achieves comparable accuracy to previous work using fewer features. Prior research using t-test p-values saw male accuracy peak at 96.6% with 11 features and female accuracy at 93.3% with at least six, while this method reaches 86.7% accuracy with a single feature, outperforming single time-domain feature accuracy (83.3%). Conclusion: These results highlight the approach's effectiveness. This study showed that similar features at different frequencies can have varying discriminative power.