NeuroCypher ASD: Machine Learning with Knowledge Graph-Based Representation for ASD Screening in Toddlers

Autism Spectrum Disorder (ASD) screening often relies on structured questionnaires. Yet these data are not always easy to interpret or use in practice. There is a need for screening frameworks that support risk estimation, case comparison, and pattern exploration in a clear and accessible way. In this work, we present NeuroCypher ASD, a framework that combines tabular machine learning with a knowledge graph for ASD screening, interpretation, and data exploration. The tabular model uses questionnaire responses and demographic features to estimate ASD risk. The knowledge graph organizes the data and supports case comparison, pattern exploration, and natural language querying. The system is evaluated under a leakage-aware protocol using tabular, graph-based, and hybrid models. The results show strong performance for the tabular model and clarify the role of graph-based representations within the overall workflow. In this setting, the knowledge graph supports structured analysis of screening data by helping users examine relationships between cases, inspect shared features, and interpret predictions in context. To strengthen interpretability, the framework includes SHAP-based analysis and approximate mappings between embedding dimensions and original questionnaire features. It also includes an anomaly detection module that highlights atypical cases for closer review. Overall, NeuroCypher ASD provides a practical framework for combining prediction, interpretation, and data exploration in ASD screening. The results show the value of integrating machine learning, graph-based context, and natural language interaction in more transparent and accessible screening workflows, particularly in school-based environments.