A Data-Driven Approach for Autism Spectrum Disorder Screening

Autism spectrum disorder (ASD) is a brain development condition influenced by genes and environment, with no known cause or cure. The diagnosis is challenging due to the absence of biological markers, requiring various examinations and specialist input. Delayed diagnosis often results in late intervention, emphasizing the need for early, individualized therapeutic approaches to enhance quality of life. Current ASD research emphasizes early detection, integrating clinical assessments with advanced technology to expedite diagnosis. One alternative for achieving rapid detection is the use of Machine Learning (ML) models, which have shown potential in identifying patterns in datasets, aiding in early ASD diagnosis. For this purpose, a representative dataset needs to be used to train these models for the ASD classification task. However, healthcare providers often lack access to detailed diagnostic data, making ML model applications challenging. This study proposes utilizing non-diagnostic data from healthcare providers, such as requests for consultations and procedures, to develop ML models for ASD classification. The approach involves creating a dataset without diagnostic data, selecting essential variables, and training ML models for classification. The classifiers utilized were Random Forest, XGboost, Catboost and a stacking strategy to combine the base classifiers. Models trained with parameter tuning presented better results, and the stacking strategy achieved the best accuracy (73.62%). The present study demonstrates the predictive power of a new strategy for ASD screening, providing an alternative where traditional diagnostic data is unavailable. The proposed approach is running in a healthcare provider operator for ASD screening and is benefiting many children and adolescents with faster and more accurate diagnoses.