Machine Learning Validates Refined Autism Screening Checklist for Polish Toddlers

A statistical and machine learning evaluation of the Quantitative Checklist for Autism in Toddlers determined that a data-driven Random Forest model predicted autism spectrum disorder diagnoses with 92 percent accuracy using data from young Polish children. The research demonstrated that eliminating eleven checklist questions that lacked consistent statistical support yielded predictive performance comparable to the complete instrument across multiple algorithms, showing that a refined feature set can maintain high diagnostic utility while streamlining the screening process.

Despite the widespread clinical deployment of these checklist variants, critical scientific gaps have persisted regarding their empirical assessment and their formal structural alignment with the revised criteria established under the DSM-5 framework. Evaluating these specific gaps remains essential for determining the overall effectiveness, practical suitability, and diagnostic reliability of the Q-CHAT as an autism screening tool.

To resolve these unanswered questions, researchers Syeda Aneela Zahra Shamsi and Zamir Hussain conducted a systematic statistical investigation to map Q-CHAT items directly onto DSM-5 criteria and evaluate their diagnostic associations. Working with a labeled secondary dataset comprising 252 Polish toddlers, the authors evaluated whether individual behavioral items reliably correlated with the binary target feature denoting the presence or absence of an autism spectrum disorder diagnosis. The study further aimed to identify the most suitable data-driven machine learning architecture for predicting autism classifications using both original and refined feature sets.

The authors conducted a two-step systematic mapping procedure that revealed notable discrepancies between the screening tool and formal psychiatric definitions, specifically showing the underrepresentation of certain DSM-5 subdomains within the Q-CHAT questionnaire. In particular, the analysis showed that items addressing relational communication were underrepresented in the checklist relative to other diagnostic requirements.

To refine the screening tool empirically, the investigators executed an integrated feature-selection strategy that combined classical Chi-square statistical analysis with Recursive Feature Elimination using Random Forest algorithms. This dual selection methodology evaluated every individual checklist item against the target diagnostic classification. The combined analytical process identified eleven specific items within the Q-CHAT that failed to achieve consistent statistical support across both feature-selection criteria, establishing an empirical basis for isolating a refined subset of the most informative screening indicators.

The researchers then conducted comparative benchmarking across four distinct data-driven machine learning algorithms: Decision Tree, Random Forest, Support Vector Machine, and Logistic Regression. Each model was evaluated using both the original full feature set and the refined eleven-item-reduced feature set. The comparative results showed that the refined feature set achieved diagnostic performance comparable to the original feature set across all evaluated models except the Decision Tree, indicating that removing the eleven unsupported questions had minimal overall impact on predictive screening capabilities.

Among the evaluated algorithms, the Random Forest model demonstrated the strongest predictive performance on the Polish toddler dataset, achieving 92 percent accuracy, 91 percent specificity, and 93 percent for precision, sensitivity, and F1-score. The study authors concluded that their two-step systematic mapping and integrated feature-selection methodology provides an adaptable framework that can be applied to evaluate other active autism screening instruments. Extending such analytical processes to other clinical tools is expected to support the empirical standardization of screening items across DSM-5 criteria.

Subject of Research: Statistical alignment and machine learning evaluation of the Q-CHAT autism screening checklist using DSM-5 criteria

Article Title: A systematic statistical analysis of Q-CHAT using DSM-5 criteria and predictive modeling on data from polish toddlers

Article References: Shamsi, S. A. Z., & Hussain, Z. (2026). A systematic statistical analysis of Q-CHAT using DSM-5 criteria and predictive modeling on data from polish toddlers. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03953-9

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03953-9

Keywords: autism spectrum disorder, Q-CHAT, machine learning, Random Forest, DSM-5, predictive modeling, pediatric screening, Polish toddlers, systematic, statistical, analysis, criteria