Addressing Zero-Inflated Scores in Self-Reported Psychiatric Questionnaires: Utilizing Synthetic Minority Over-Sampling Technique (SMOTE) to Classify Mood and Energy Based on Extreme Activity Levels for Adolescents with Bipolar and ADHD

Abstract

Accurate classification of bipolar disorder (BD) and attention deficit hyperactivity disorder (ADHD) in adolescents is challenging due to symptom overlap and the prevalence of zero-inflated bias in self-report questionnaires. This study tests machine learning algorithms with and without Synthetic Minority Over-sampling Technique (SMOTE) to solve classification problems. We analyzed data from 354 adolescent psychiatric inpatients between 2014 and 2023 diagnosed with BD, ADHD, or other psychiatric conditions. Self-reported mood scores and energy scores were categorized into four severity classes and machine learning models, with and without SMOTE, were tested to address zero inflation and correct class imbalances in the scores. Actigraphy was mapped onto these categories. Applying SMOTE improved the cross-validated overall classification performance across all labels, with accuracy increasing from 58% to 98%, and substantial enhancements in F1 score, precision, and recall, particularly for minority classes. This approach not only boosts the models' sensitivity to minority classes but also enhances the accuracy and reliability of mood and energy assessments in clinical settings for adolescents with BD and ADHD.