NAPIFASD: Design of an Iterative Neuroadaptive Multimodal Framework for Predictive Modeling, Subtype Discovery, and Intervention Simulation in Autism Spectrum Disorders

Abstract: The current increase in the prevalence of Autism Spectrum Disorder (ASD) and the need for early personalized intervention mandate the development of predictive systems that reach beyond static diagnosis. Machine learning models for ASD, as it exists today, are largely inflicted by limitations, including restricted data modalities, low generalization across populations, and limited interpretability — all of which become significant constraints for their clinical relevance and translational value. Furthermore, such systems often do not model the temporal development of neurodevelopmental markers and overlook the crucial interlink between diagnosis and a strategy for actionable intervention. To plug these gaps, we present the NeuroAdaptive Predictive and Interventional Framework for Autism Spectrum Disorder (NAPIF-ASD)—a multimodal, interpretable, and intervention-aware pipeline to enhance ASD prediction and intervention planning. The framework initiates with Temporal Neuro-Behavioral Graph Embedding (TNGE-Net), which jointly models individual developmental trajectories exploiting the power of graph neural networks over longitudinal neuroimaging and behavioral data. Then, Domain-Regularized Adaptive Clustering Ensemble (DRACE) identifies clinically meaningful subtypes of ASD through unsupervised clustering with domain knowledge embedded throughout the process. For interpretability, Neuro-Symbolic Causal Inference Model (NS-CIM) generates subject-specific causal maps that aid with counterfactual reasoning and explanatory insights. For generalization of the model, Federated Meta-Learning for Autism Model Transferability (FML-AMT) preserves data privacy while providing cross-institution robustness. Finally, Intervention-Driven Reinforcement Learning Engine (IDRLE) simulates long-term outcomes of early interventions towards minimizing the harm and maximizing the benefit of therapeutic strategies. By connecting prediction with actionable outcomes, this entire pipeline not only advances predictive accuracy and interpretability but constitutes a paradigm shift in the autism research landscapes. It makes significant contributions to the field of machine learning in healthcare by explicitly showing how longitudinal, causal, and federated approaches can work together towards precision psychiatry processes.