Bayesian-Optimized Butterworth-PID Preprocessing for EEG-Based Autism Spectrum Disorder Classification

Electroencephalography (EEG)-based classification of autism spectrum disorder (ASD) remains sensitive to preprocessing quality, yet the role of adaptive filtering in improving downstream classification is still insufficiently established. This study proposes a Bayesian-optimized Butterworth-proportional-integral-derivative (BPID) preprocessing framework and evaluates its contribution within an EEG-based ASD classification pipeline. EEG recordings were collected from 10 participants (5 ASD and 5 neurotypical) using a 16-channel Cyton biosensing board at 250 Hz. Baseline preprocessing used a sixth-order Butterworth band-pass filter with 4–40 Hz cut-off frequencies, while the BPID branch regulated the Butterworth response using Bayesian-optimized PID control. Four preprocessing-classifier schemes were compared: Butterworth-EEGInception, Butterworth-SSM, BPID-EEGInception, and BPID-SSM. PID gains were tuned over 50 iterations, yielding $K_{p}=4.8214\times {10}^{-5}$ , $K_{i}=0.7034$ , and $K_{d}=0.9771$ . Performance was evaluated using subject-wise 5-fold cross-validation, with significance assessed by a one-sided paired Wilcoxon signed-rank test. At the signal level, BPID reduced MSE by 82.69–95.41% in the neurotypical group and 89.97–94.61% in the ASD group compared with Butterworth filtering. At the classification level, BPID-EEGInception achieved $99.42~\pm ~0.37$ % accuracy and $99.42~\pm ~0.36$ % F1-score, while BPID-SSM improved accuracy over Butterworth-SSM ( $98.26~\pm ~0.38$ % vs. $96.13~\pm ~1.26$ %). Wilcoxon testing confirmed significant accuracy improvements for both classifier families. These findings highlight the methodological significance of adaptive preprocessing by linking signal-level fidelity, classifier performance, and statistical validation. Overall, BPID provides an interpretable and practical strategy for limited-cohort EEG-based ASD classification without requiring artifact-free clean targets, although broader validation remains necessary.