Autism spectrum disorder screening using DSM-5 fulfillment and machine learning adaptation

• Main Author: Wan Zaini, Wan Azamudin
• Format: Thesis
• Language: English; English
• Published: 2025
• Subjects: Q Science; QA Mathematics
• Online Access: http://eprints.utem.edu.my/id/eprint/29321/

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by persistent challenges in social communication, restricted interests, and repetitive behaviours. The prevalence of ASD has increased globally, prompting the need for more reliable, objective, and scalable screening and diagnostic methods. Traditional diagnostic tools, such as the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), remain widely used in clinical settings. However, these tools are inherently dependent on subjective human judgment and clinician expertise, which can lead to inconsistencies in diagnosis and delayed interventions, particularly in early developmental stages. To address these limitations, this study explores a data-driven approach by integrating DSM-5 diagnostic criteria with advanced machine learning (ML) and deep learning (DL) models to enhance ASD detection and severity classification. Two datasets were employed in this research: the Autism Screening dataset, consisting of 1054 toddler data, 104 adolescence data, and 704 adult data samples, used for binary classification between ASD and non-ASD individuals; and the DSM-5 Diagnostic Dataset from Hospital Canselor Tuanku Muhriz UKM (HCTM), comprising 177 clinical samples after oversampling, used for multi-class classification of ASD severity (mild, moderate, and severe). Given the imbalance in class distribution, particularly in the severity-level dataset, oversampling techniques were implemented to improve model fairness and performance across all severity categories. The machine learning models evaluated in this study include Support Vector Machine (SVM), Decision Tree, and k-Nearest Neighbour (kNN). A Deep Neural Network (DNN) architecture was also designed and trained for comparative analysis. Model performance was assessed using standard classification metrics such as accuracy, precision, recall, and F1-score. The results demonstrate that the DNN model outperformed traditional ML models in both binary and severity-level classification tasks. Notably, the DNN achieved 100% accuracy in detecting ASD among younger children, reinforcing its potential as a tool for early screening. Furthermore, the severity classification results showed improved granularity and consistency compared to outcomes generated by manual assessments alone. This research highlights the value of integrating clinical standards with artificial intelligence to improve the speed, accuracy, and objectivity of ASD screening processes. The findings suggest that such hybrid approaches could support clinicians in making more informed decisions, reduce diagnostic delays, and enable timely interventions. Future research should explore larger and more diverse populations, refine model generalizability, address ethical considerations such as data privacy and bias, and assess real-world clinical deployment feasibility.