Abstract:
Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that affects one
throughout life. Early treatment response is strengthened by timely diagnosis of Autism
Spectrum Disorder (ASD) although the current screening algorithms rely on behavioral
assessment, which is both resource consuming and subjective. The use of eye-tracking
needs to provoke characteristic variation in social attention, and in the former studies,
the classical and deep learning models are tested, but comparisons are often limited by
limited participants, confounding tasks, and imprecise measurements. This paper
bridges this gap with a question about whether an integrated and reproducible pipeline
can help make a fair comparison of classical and deep methods to screening ASD and
reduce false negatives in practice. We use our eye-tracking signal analysis as the
standardized workflow data cleaning, data normalization, class balancing and statistical
validation mode (ANOVA and Correlation) and test four baselines (SVM, Random
Forest, MLP, LSTM) using the same splits and measurements and test a stacking
ensemble with a recall-oriented one. Random Forest has the highest total accuracy and
AUC, LSTM models differentiates temporal gazes effectively and the ensemble
maximizes recall, which minimizes the risk of false negative cases. The results obtained
in this research imply whether the combination of classical and deep learning models is
done cautiously, a viable line of constructing supportive screening tools can be offered.
Reproducible implementation, sensitivity of sensitive information, and cost-aware
computation are also in-house priorities that enable us to be responsible adopters of the
technology in the healthcare facilities.
