Less is more? A hybrid machine learning and psychometric approach to identifying clinically relevant psychopathology in Chinese youth using the Child Behaviour Checklist.
Screening for psychiatric risk in youth at the population level is often constrained by resource limitations and lengthy assessment tools. This study aimed to develop a reduced, psychometrically robust subset of Child Behavior Checklist (CBCL) items to effectively predict transdiagnostic psychiatric morbidity in the youth population.
Data were drawn from a nationally representative sample of 72,109 Chinese youth aged 6-16 years. Initial item screening on the full sample employed unique variance analysis, item-rest correlation tests, and conceptual redundancy checks. A nested case-control subset (approximately 4,500 with diagnoses and 5,000 without) was used for feature selection. Recursive feature elimination with repeated cross-validation was then applied to the nested subset to derive three item sets (n = 35, 69, 98). These were psychometrically evaluated using exploratory graph analysis and confirmatory factor analysis in two age- and gender-stratified samples from the full dataset. Predictive performance was assessed using five machine learning algorithms, trained and tested on a 70/30 split of the nested case-control data.
The 35-item and 60-item subsets achieved high diagnostic accuracy (AUC = 0.88-0.89), with performance comparable to the best-performing larger subsets. Items captured transdiagnostic dimensions including Functional Somatic Symptoms, Neurodevelopmental Dysregulation, Affective-Social Withdrawal, Threat Sensitivity and Cognitive-Perceptual Disturbance, and Disinhibited-Irritable Externalising.
The reduced CBCL sets demonstrated strong diagnostic utility and psychometric soundness. This scalable tool supports transdiagnostic, data-driven screening of youth psychiatric risk at the population level.
Data were drawn from a nationally representative sample of 72,109 Chinese youth aged 6-16 years. Initial item screening on the full sample employed unique variance analysis, item-rest correlation tests, and conceptual redundancy checks. A nested case-control subset (approximately 4,500 with diagnoses and 5,000 without) was used for feature selection. Recursive feature elimination with repeated cross-validation was then applied to the nested subset to derive three item sets (n = 35, 69, 98). These were psychometrically evaluated using exploratory graph analysis and confirmatory factor analysis in two age- and gender-stratified samples from the full dataset. Predictive performance was assessed using five machine learning algorithms, trained and tested on a 70/30 split of the nested case-control data.
The 35-item and 60-item subsets achieved high diagnostic accuracy (AUC = 0.88-0.89), with performance comparable to the best-performing larger subsets. Items captured transdiagnostic dimensions including Functional Somatic Symptoms, Neurodevelopmental Dysregulation, Affective-Social Withdrawal, Threat Sensitivity and Cognitive-Perceptual Disturbance, and Disinhibited-Irritable Externalising.
The reduced CBCL sets demonstrated strong diagnostic utility and psychometric soundness. This scalable tool supports transdiagnostic, data-driven screening of youth psychiatric risk at the population level.
Authors
Luo Luo, Shao Shao, Zavlis Zavlis, Wu Wu, Huang Huang, Xu Xu, Qi Qi, Zheng Zheng, He He
View on Pubmed