VALIDATE ASD: INDEPENDENT MULTIMODAL REPLICATION AND VALIDATION OF AUTISM DATA-SCIENCE MODELS

Abstract This project will develop a robust and transparent framework for the validation and replication of autism data science models through a multi-modal, cross-institutional approach. The initiative will employ comprehensive datasets from various sources, including clinical, genomic, environmental, and physiological data from Texas Children’s Hospital (TCH) and other prominent research repositories. By leveraging advanced AI/ML methods, we will rigorously assess model performance across different pediatric populations, ensuring generalizability and fairness in clinical settings. We will implement a two-pronged validation strategy, including intact model testing and code-blinded replication. This methodology will rigorously evaluate the accuracy, reproducibility, and population-specific calibration needs of predictive models. Key to this effort will be a large-scale integration of structured clinical data from TCH's Research Data Warehouse, combined with genomic data from the SPARK (Simons Powering Autism Research for Knowledge) genetic cohort and metabolomics data from the BaBS (Bacteria and Birth Study). Additionally, environmental exposure models will be validated with placental biomarkers linked to neurodevelopmental outcomes. The project will provide essential insights into the limitations and strengths of autism-related AI models in real-world applications, guiding their future clinical deployment. Through a commitment to transparency, reproducibility, and stakeholder engagement, the project will deliver high-impact validation reports, including community-accessible tools that ensure models are utilized effectively across a broad range of patient populations. The outcomes of this work will enhance autism care and set a new standard for multi- modal, multi-institutional model validation in pediatric health research.