Development of Machine Learning Models for Prediction of Psychosocial Dysfunction in Children and Adolescents using the Pediatric Symptom Checklist

Abstract

Mental health issues, particularly early-onset psychopathology, are a neglected domain of public health. In low middle-income countries like Pakistan, the management of these problems is hindered by stigma, limited mental healthcare resources, and protracted nature of consultations. The access and convenience of mental healthcare can be improved by leveraging Artificial Intelligence (AI) for the development of datadriven decision support systems. This study aims to develop Machine Learning (ML) based predictive models for psychosocial dysfunction in Pakistani children and adolescents. The Pediatric Symptom Checklist (PSC) is used to collect data of 2,372 individuals. To the best of our knowledge, this study is the first to use data of Pakistani pediatric populations collected using the PSC. Feature selection methods reveal that items related to attention issues, hopelessness, sadness, and irritability are the most significant predictors of psychosocial dysfunction for both age groups. Exclusively for children, externalizing problems like disobedience are also significant. Six ML algorithms viz. Decision Tree, Random Forest, XGBoost, Support Vector Machine, Multilayer Perceptron, and Logistic Regression are selected to develop the predictive models. Among these, Logistic Regression provides the best results with accuracies of 0.98 and 0.99 for children and adolescents, respectively. For children, it is observed that exclusion of item 23 of the PSC (Wants to be with you more than before) improves model performance. Our encouraging findings suggest that the proposed models have the potential to be deployed in healthcare facilities and educational institutes for early and accurate detection of pediatric psychosocial dysfunction.