Hybrid EEG Feature Fusion Framework for Accurate Autism Spectrum Disorder Diagnosis Using Ensemble Learning

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition with increasing global prevalence and no standardized biological test for early detection. Current diagnosis methods rely heavily on behavioral assessments, which are subjective, time-consuming, and prone to variability. This study proposes a hybrid feature fusion framework for non-invasive ASD diagnosis using electroencephalogram (EEG) signals, specifically event-related potentials (ERPs) such as P300 components obtained from the BCIAUT-P300 dataset. EEG recordings were captured using a g.Nautilus wireless system with eight scalp electrodes, and preprocessed using 0.5-30 Hz bandpass filtering and baseline subtraction to enhance signal quality. Twenty-two EEG features were extracted across time, frequency, and time-frequency domains using methods such as Wavelet Transform, power spectral density, higher-order statistics, and principal component analysis. Five optimal methods, PCA, HOS, PSD, FDA, and CWT, were selected based on their classification potential and fused using both feature-level and decision-level strategies. Ensemble classifiers including SVM, XGBoost, LDA, and Random Forest were trained and evaluated on the fused feature set. The proposed hybrid fusion framework achieved a classification accuracy of 97.7%, sensitivity of 96.8%, and specificity of 98.5%, outperforming traditional single feature or single classifier approaches. The integration of multi-domain feature descriptors with ensemble learning contributes to increased robustness, generalizability, and diagnostic precision. Our work demonstrates the feasibility of combining EEG-based biomarkers with machine learning to support early ASD diagnosis. The framework offers a scalable approach that is aligned with biomedical informatics objectives, with potential for clinical deployment and integration into portable EEG-based screening systems.