Machine Learning-Based Prognostic Modeling for Autism Spectrum Disorder: A Comprehensive Review

Abstract

Autism Spectrum Disorder (ASD) is a heterogeneous neurodevelopmental condition characterised by persistent differences in social communication and interaction, together with restricted or repetitive patterns of behaviour. The considerable variation in symptom presentation, developmental progression and response to intervention makes prognosis particularly challenging. Machine learning (ML) has emerged as a promising computational approach for identifying complex patterns within behavioural, clinical, neuroimaging, physiological and other high-dimensional datasets. This review examines the application of ML-based prognostic modelling in ASD, focusing on early prediction, symptom-severity estimation, developmental trajectory modelling and prediction of intervention outcomes. The review synthesises evidence on conventional ML techniques, including logistic regression, decision trees, random forests and support vector machines, alongside artificial neural networks and deep-learning approaches. Existing studies demonstrate that ML can generate useful predictions from relatively small sets of clinical and developmental variables, while more complex approaches can exploit neuroimaging, EEG, video and multimodal information. However, the evidence base remains constrained by small samples, heterogeneous datasets, class imbalance, overfitting, limited external validation and inadequate representation of diverse populations. Deep-learning studies have reported high diagnostic performance, but substantial heterogeneity across studies limits direct translation into clinical practice (Ding et al., 2024). Evidence specifically addressing prognosis is more limited, although early studies indicate potential for predicting treatment response and developmental outcomes (Thabtah et al., 2022). The review concludes that the future of ML in ASD prognosis depends on longitudinal datasets, independent validation, interpretable models, clinically meaningful outcomes and representative populations. ML should ultimately complement, rather than replace, clinical judgement in individualised ASD assessment and intervention planning.