SUPPORT SYSTEM FOR THE DIAGNOSIS OF THE RISK OF ANXIETY DISORDER IN CHILDREN

Resumo

Anxiety disorders in children represent an important public health challenge, considering the subjective nature of the symptoms, the individual variability of clinical manifestations, and the lack of standardized diagnostic support tools. This gap compromises early identification and appropriate preventive interventions. This study investigated computational approaches for the multilevel classification of the risk of anxiety disorders in children, using behavioral and physiological data. The methodology involves the application and comparison of machine learning models, including Random Forest, Support Vector Machine, and Multilayer Perceptron Neural Network, in binary (presence or absence) and multilevel (mild, moderate, severe risk) paradigms. The research used a dataset of 193 children, publicly available on Harvard Dataverse by Carpenter [1], licensed under CC0 1.0 Universal. The evaluation of the models used standardized metrics aligned with the diagnostic criteria of the DSM-5, Diagnostic and Statistical Manual of Mental Disorders - 5th edition, and the ICD-11, International Classification of Diseases - 11th edition. After refinements in the methodology, the results showed significant improvement: Random Forest – accuracy 89.4%, sensitivity 80%, specificity 88.7%; Support Vector Machine – accuracy 88.5%, sensitivity 81.3%, specificity 91.9%; Multilayer Perceptron – accuracy 87.8%, sensitivity 77.7%, specificity 91.1%. Accuracy above 87% indicates excellent overall performance of the models, correctly classifying most cases. Given the topic, high sensitivity is crucial to avoid the omission of relevant cases. Sensitivity values (77% to 81%) demonstrate effective identification of positives, which is clinically important. High specificity (above 88%) shows accurate recognition of negatives, reducing false positives. These results indicate a strong predictive performance, especially for the SVM, which showed balance and robustness, being the most suitable to minimize both false negative and false positives. This highlights the feasibility of machine learning in early detection of anxiety risk in children. Continuous improvements aim to improve accuracy and clinical applicability. This study contributes to the advancement of diagnostic strategies in child mental health by offering a computational approach to support personalized, evidence-based clinical interventions.