Machine Learning and Explainable AI for Q-CHAT-10–Based Autism Spectrum Disorder Screening Assistance in Toddlers: Comparative Evaluation and Web Deployment

Abstract

Machine-learning methods applied to questionnaire data may support early developmental screening for Autism Spectrum Disorder (ASD), provided that model outputs are interpreted as screening assistance rather than as evidence of clinical diagnosis. This study develops and evaluates a comparative machine-learning and artificial neural-network framework for Q-CHAT-10–based toddler screening and extends the system with explainable artificial intelligence (XAI) and web-based decision-support functionality. The original dataset contained 1,054 records; after removal of the case identifier and 79 repeated response profiles, 975 unique cases remained, comprising 691 screening-positive and 284 screening-negative observations. Sixteen predictors were retained after exclusion of the pre-computed Qchat-10-Score to reduce direct target leakage. Nine classifiers were evaluated under a common 80/20 stratified hold-out design with training-only SMOTE and StandardScaler transformation: Logistic Regression, Decision Tree, Random Forest, K-Nearest Neighbors, polynomial and radial-basis-function Support Vector Machines, Gaussian Naive Bayes, Quadratic Discriminant Analysis, and an early-stopped Multilayer Perceptron. Logistic Regression achieved 100% held-out accuracy, sensitivity, specificity, F1-score, and ROC-AUC, with a log loss of 0.0272; McNemar’s exact test identified seven discordant cases favoring Logistic Regression over the MLP-ANN (p = 0.0156). A source-data audit showed that the screening label is deterministically derived from the sum of A1-A10 at a threshold of four, indicating that the perfect result reflects reconstruction of the questionnaire-based screening rule rather than independent clinical diagnosis. The Streamlit implementation incorporates SHAP and LIME explanations, with LIME outputs presented in native feature units for improved interpretability, together with valid-domain what-if analysis, exploratory data analysis, batch inference, downloadable screening reports, and all-model confusion-matrix visualization. The framework demonstrates transparent automation of ASD screening assistance while emphasizing the need for external clinician-confirmed validation before clinical decision-support use.