Sistema de Apoio ao Diagnóstico do Risco de Transtorno de Ansiedade em Crianças

Abstract

Childhood anxiety disorders represent a significant mental health concern, particularly due to the challenges associated with early identification, the subjectivity of clinical assessment instruments, and the overlap of behavioral patterns. In this context, computational approaches have been investigated as decision-support tools, especially those based on machine learning techniques applied to psychometric data. However, such approaches still face challenges related to decision robustness, clinical interpretability, and consistent performance in ordinal multiclass scenarios. To address these limitations, this work proposes a computational decision-support system for binary and multilevel classification of anxiety disorder risk in children, grounded in supervised learning models from the perspective of Electrical Engineering applied to mental health. Three established models are comparatively evaluated — Random Forest, Support Vector Machine, and Multilayer Perceptron — considering a binary screening scenario and an ordinal multiclass scenario composed of four risk levels. The methodology employs a public dataset from the Harvard Dataverse, consisting of psychometric, behavioral, and psychosocial information, and is validated through stratified cross-validation and performance metrics aligned with clinical practice, including explicit analysis of ordinal clinical error by distinguishing adjacent and non-adjacent misclassifications. The results indicate that the models achieve high discriminative capacity in the binary scenario (AUC values above 0.94) and consistent performance in the multiclass scenario (overall accuracy between 83% and 85%), with no universally superior classifier. A complementary behavior among the architectures is observed, with Random Forest demonstrating greater global stability and Multilayer Perceptron showing improved discrimination between intermediate risk levels. It is concluded that classical supervised models, when properly validated, constitute promising tools for the ordinal stratification of pediatric anxiety risk. From an Electrical Engineering perspective, the problem is characterized as a latent state inference process based on noisy and partially observable data, contributing methodologically to the development of interpretable computational decision-support systems in child mental health.