Mumenin Nasirul, Rahman Maisha Mumtaz, Yousuf Mohammad Abu, Noori Farzan M, Uddin Md Zia
Department of Computer Science and Engineering, Bangladesh University of Business and Technology, Dhaka, Bangladesh.
Department of Electrical and Electronics Engineering, Islamic University of Technology, Dhaka, Bangladesh.
Front Artif Intell. 2025 May 30;8:1507922. doi: 10.3389/frai.2025.1507922. eCollection 2025.
Autism Spectrum Disorder (ASD) is a multifaceted neurodevelopmental condition that challenges early diagnosis due to its diverse manifestations across different developmental stages. Timely and accurate detection is essential to enable interventions that significantly enhance developmental outcomes. This study introduces a robust and interpretable machine learning framework to diagnose ASD using questionnaire data. The proposed framework leverages a stacked ensemble model, combining Random Forest (RF), Extra Tree (ET), and CatBoost (CB) as base classifiers, with an Artificial Neural Network (ANN) serving as the meta-classifier. The methodology addresses class imbalance using Safe-Level SMOTE, dimensionality reduction via Principal Component Analysis (PCA), and feature selection using Mutual Information and Pearson correlation. Evaluation on publicly available datasets representing toddlers, children, adolescents, adults, and a merged dataset (Combining children, adolescents, and adults dataset) demonstrates high diagnostic accuracy, achieving 99.86%, 99.68%, 98.17%, 99.89%, and 96.96%, respectively. Comparative analysis with standard machine learning models underscores the superior performance of the proposed framework. SHapley Additive exPlanations (SHAP) were used to interpret feature importance, while Monte Carlo Dropout (MCD) quantified uncertainty in predictions. This framework provides a scalable, interpretable, and reliable solution for ASD screening across diverse populations and developmental stages.
自闭症谱系障碍（ASD）是一种多方面的神经发育状况，因其在不同发育阶段的多样表现而对早期诊断构成挑战。及时且准确的检测对于实施能显著改善发育结果的干预措施至关重要。本研究引入了一个强大且可解释的机器学习框架，用于使用问卷数据诊断ASD。所提出的框架利用了一个堆叠集成模型，将随机森林（RF）、极端随机树（ET）和CatBoost（CB）作为基分类器，以人工神经网络（ANN）作为元分类器。该方法使用安全级别的合成少数过采样技术（Safe-Level SMOTE）解决类别不平衡问题，通过主成分分析（PCA）进行降维，并使用互信息和皮尔逊相关性进行特征选择。对代表幼儿、儿童、青少年、成年人的公开可用数据集以及一个合并数据集（结合儿童、青少年和成年人数据集）的评估显示出高诊断准确率，分别达到99.86%、99.68%、98.17%、99.89%和96.96%。与标准机器学习模型的比较分析突出了所提出框架的卓越性能。使用夏普利值加法解释（SHAP）来解释特征重要性，同时使用蒙特卡洛随机失活（MCD）量化预测中的不确定性。该框架为跨不同人群和发育阶段的ASD筛查提供了一个可扩展、可解释且可靠的解决方案。
