Faculty of Engineering & Technology

Permanent URI for this community

https://dspace.nstu.ac.bd/handle/123456789/1

Browse

Browsing Faculty of Engineering & Technology by Author "Argina Akter"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    A Hybrid GA-PSO Optimized Machine Learning Framework with Strategic Data Augmentation for Early Autism Spectrum Disorder Detection and Risk Factor Analysis
    (2023-06-30) Argina Akter
    This research presents a robust machine learning framework for the early detection of Autism Spectrum Disorder and the analysis of its associated risk factors. The study leverages a comprehensive dataset of 288 individuals, encompassing behavioral, developmental, familial, and socio-demographic attributes. Five machine learning classifiers—Support Vector Machine, Decision Tree, Random Forest, Multilayer Perceptron, and Logistic Regression—were implemented and enhanced through a novel two-pronged approach. First, a true hybrid Genetic Algorithm-Particle Swarm Optimization technique was employed for superior hyper-parameter tuning, significantly boosting model performance. Second, strategic data augmentation was performed using Conditional Tabular Generative Adversarial Networks combined with boundary sampling to effectively mitigate data scarcity. The optimized Random Forest model achieved exceptional performance, with an accuracy of 98.85% and an Area Under the Receiver Operating Characteristic Curve of 0.9989. The synergistic combination of GA-PSO optimization and data augmentation proved transformative, enabling multiple models to achieve perfect classification metrics of 100% accuracy on augmented datasets. Model robustness was rigorously validated through stability analysis across multiple data splits (70:30, 75:25, 80:20) and, critically, on a completely unseen test set, where the best model maintained a high accuracy of 94.1%. Beyond predictive performance, a sensitivity analysis was conducted to identify and rank the most influential risk factors. "Limited Speech" emerged as the most critical predictor, followed by parental age at childbirth (i.e., "Mother age at childbirth" and "Father age at childbirth") and key developmental milestones such as "Ability to walk" and "Ability to respond when called by name". This study conclusively demonstrates that the integration of hybrid metaheuristic optimization and strategic data augmentation creates a highly accurate, reliable, and generalizable tool for early ASD screening. The framework provides a practical solution for overcoming data limitations in clinical settings and offers data-driven insights into key risk factors, thereby supporting earlier intervention and informed clinical decision-making. Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by social interaction challenges, restricted interests, and repetitive behaviors. Early detection is critical for timely intervention, yet conventional diagnostic methods are often subjective, time consuming, and inaccessible in low-resource settings. This research develops a robust machine learning-based framework for ASD detection using a multidimensional dataset of 288 individuals aged 3–27 years, encompassing behavioral, developmental, health, familial, and socio demographic attributes. Five classifiers—Support Vector Machine, Decision Tree, Random Forest, Multilayer Perceptron, and Logistic Regression—were implemented and further enhanced through hybrid Genetic Algorithm–Particle Swarm Optimization (GA-PSO) for hyper-parameter tuning. To overcome data scarcity, Conditional Tabular GAN (CTGAN) with boundary sampling was employed for synthetic data augmentation. The optimized and augmented models achieved significant performance improvements, with Random Forest enhanced by GA-PSO yielding near perfect accuracy (>98%, AUROC ≈ 0.999). Rigorous validation, including stability analysis across multiple data splits and testing on a completely unseen dataset, confirmed the model’s reliability and generalizability. Sensitivity analysis identified and ranked key risk factors, with limited speech, parental age, and developmental milestones emerging as dominant predictors. This integrated framework demonstrates the potential of hybrid optimization and augmentation strategies to deliver an accurate, scalable, and clinically meaningful screening tool for ASD, bridging the gap between machine learning innovation and real-world healthcare needs.