Evaluating Multicultural Autism Screening for Toddlers Using Machine Learning on the QCHAT-10 | JoVE Visualize

Area of Science:

Computational neuroscience
Developmental psychology
Artificial intelligence in healthcare

Background:

Early identification and intervention are crucial for improving life outcomes in Autism Spectrum Disorder (ASD).
Traditional diagnostic methods for ASD can be time-consuming, potentially delaying necessary treatment.
The Quantitative Checklist for Autism in Toddlers (QCHAT-10) is a screening tool that requires further validation across diverse populations.

Purpose of the Study:

To evaluate the predictive value of QCHAT-10 features using machine learning (ML) techniques.
To assess the overall accuracy of ML models trained on QCHAT-10 data across different cultural contexts.
To determine the generalizability of ML-based ASD screening tools across international datasets.

Main Methods:

Trained ML models (Decision Tree, Random Forest, XGBoost) on QCHAT-10 datasets from Poland, New Zealand, and Saudi Arabia.
Evaluated model performance using cross-validation and testing on an independent Polish dataset with clinical diagnostic labels.
Employed Recursive Feature Elimination (RFE) to identify the most predictive features.

Main Results:

XGBoost consistently outperformed other models, achieving high Area Under the Receiver Operating Characteristic Curve (AUROC) values (e.g., 0.94 ± 0.06 for the New Zealand model tested on Polish data).
Models trained on international data demonstrated strong predictive capabilities when tested on the Polish dataset, suggesting cross-cultural generalizability.
Feature importance analysis showed some common predictive features across models, although rankings varied by population.

Conclusions:

The QCHAT-10, when analyzed with ML techniques, shows significant potential as a cross-cultural screening tool for Autism Spectrum Disorder.
ML models can effectively predict official autism diagnoses based on QCHAT-10 responses, even when trained on data from different countries.
Further research is warranted to explore cultural variations in feature importance and optimize screening tool performance across diverse populations.