ASD-Bench: New Benchmark Evaluates AI Models for Autism Screening Across Age Groups

Researchers have created a new benchmark called ASD-Bench to evaluate different machine learning and deep learning models aimed at automating autism spectrum disorder (ASD) screenings. Unlike existing tools, which often target only one architecture and adult demographics, this benchmark assesses models across three age ranges: kids (1-11 years), teens (12-16 years), and adults (17-64 years). It employs four criteria for evaluation: predictive performance, calibration, interpretability, and adversarial robustness. Using a tailored dataset of 4,068 AQ-10 records, ASD-Bench includes various models like XGBoost, Random Forest, and neural networks. It also introduces a new metric, the Heuristic Aggregate Penalty (HAP), to better address false negatives and improve early diagnosis based on age-specific patterns.