A CatBoost-Based Framework for Autism Spectrum Disorder Classification from EEG Signals with Blackman–Harris FIR Filtering
TL;DR
These findings support preliminary within-cohort feasibility but not clinical or screening readiness, and support an EEG-based ASD classification framework combining Blackman–Harris FIR filtering, feature extraction and selection, and machine learning.
Abstract
Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder marked by impairments in social interaction, communication, and repetitive behavior; early identification remains difficult given its clinical heterogeneity and the subjectivity of conventional diagnosis. This study proposes an EEG-based ASD classification framework combining Blackman–Harris FIR filtering, feature extraction and selection, and machine learning. EEG data were collected from 16 participants (8 ASD, 8 neurotypical) using a 16-channel system, band-pass filtered at 4–40 Hz, segmented into 4-second epochs with 50% overlap, and represented by 176 temporal and spectral features per epoch (11 per channel), followed by fold-specific Mutual Information feature selection. SVM-RBF, CatBoost, LDA, and Gaussian Naive Bayes were evaluated using nested subject-wise 8-fold cross-validation. Based on 7,145 pooled out-of-fold epoch predictions, CatBoost achieved the highest accuracy (99.58%), followed by LDA (97.47%), SVM-RBF (91.60%), and Gaussian Naive Bayes (57.49%). An exploratory paired t-test on fold-wise accuracy showed a nominally significant CatBoost advantage over SVM-RBF (mean difference 8.00 percentage points, ; Cohen’s ). At the participant level, majority voting correctly classified 16/16 participants with CatBoost and LDA, 15/16 with SVM-RBF, and 9/16 with Gaussian Naive Bayes; the 95% CI for 100% accuracy was 79.41–100.00%. These findings support preliminary within-cohort feasibility but not clinical or screening readiness. ABSTRAK: Gangguan Spektrum Autisme (ASD) adalah gangguan perkembangan neuro yang dicirikan oleh kekurangan interaksi sosial, komunikasi, dan tingkah laku berulang. Pengesanan awal masih sukar disebabkan kepelbagaian klinikal dan sifat subjektif diagnosis konvensional. Kajian ini mencadangkan rangka kerja pengelasan ASD berasaskan EEG yang menggabungkan penapisan Blackman–Harris FIR, pengekstrakan dan pemilihan ciri, serta pembelajaran mesin. Data EEG diperoleh daripada 16 peserta (8 ASD, 8 neurotipikal) menggunakan sistem 16 saluran, ditapis jalur 4–40 Hz, dibahagikan kepada epok 4 saat dengan pertindihan 50%, dan diwakili oleh 176 ciri temporal dan spektral bagi setiap epok (11 setiap saluran), diikuti pemilihan ciri Informasi Bersama khusus bagi setiap lipatan. SVM-RBF, CatBoost, LDA, dan Gaussian Naive Bayes dinilai menggunakan validasi silang 8-lipatan berasaskan subjek. Berdasarkan 7,145 ramalan epok terkumpul luar lipatan, CatBoost mencapai ketepatan tertinggi (99.58%), diikuti LDA (97.47%), SVM-RBF (91.60%), dan Gaussian Naive Bayes (57.49%). Kajian pada ketepatan setiap lipatan bagi ujian-t berpasangan menunjukkan kelebihan nominal yang signifikan bagi CatBoost berbanding SVM-RBF (perbezaan min 8.00 mata peratus, ; Cohen’s ). Pada peringkat peserta, pengelasan undian majoriti adalah betul iaitu 16/16 peserta menggunakan CatBoost dan LDA, 15/16 menggunakan SVM-RBF, dan 9/16 menggunakan Gaussian Naive Bayes; 95% CI bagi ketepatan 100% ialah 79.41–100.00%. Dapatan ini menyokong kebolehlaksanaan awal kohort ini, tetapi bukan kesediaan klinikal atau saringan.