Based on the BCIAUT-P300 dataset, we evaluate the classification intended to diagnose autism using 22 feature extraction methods for EEG signals. The classifiers MDM, SVM, NN, RF, and LDA were examined in relation to PSD, CWT, PCA, and FDC to evaluate onset detection algorithms. Various ensemble learning techniques like AdaBoostM1, Bagging, and GentleBoost were applied to improve the classification performance. Of all the PSD methods, PSD + AdaBoostM1 came out as the best with 96.80% accuracy, 96.88% sensitivity, and 96.00% specificity achieved with this method. The FDA results were also very good, with an accuracy of 96.56% (F1-Score: 98.04%). CWT, PCA, and EEGNet also achieved great performance, but the time-domain method like Shannon Entropy and decomposition methods such as HHT had limited accuracy and F1-scores. The study demonstrates that frequency domain methods and ensemble classifiers can effectively distinguish P300 responses for autism detection. The findings highlight that choosing the right feature extraction methods is important for better classification performance of EEG-based diagnosis systems.