Maximum and minimum activity in inpatient adolescents with Bipolar Disorders: Daily-Variability classification of actigraphy pattern with artificial intelligence

Abstract

Measures of daily activity may be objective markers to help differentiate adolescent bipolar disorder (BD). We used chart reviewed actigraphy data collected from 2014 to 2023, and AI methods to classify well-characterized inpatient adolescents diagnosed with BD-without-attention deficit/hyperactivity disorder (ADHD), BD-with-ADHD, ADHD-without-BD, and other diagnoses (OD). 389 inpatient adolescents (232 female, mean age 15.07), wore an actigraphy monitor for the duration of their inpatient stay (mean number of unique days = 13.04 days). Activity was characterized into four 60-min maximum and minimum daily activity bins, automatically identified using a novel Python script. Feature engineering further described time-series data. 5193 days of data were split into training and testing sets. Random Forest and XGBoost models were trained with cross-validation on the training set and model metrics were compared. The best models were tested on the testing set. XGBoost with feature selection provided the most robust and balanced classification model. The most influential feature was the engineered difference between peak active hours, which along with other activity and age features classified all diagnostic groups with 91.5 % accuracy. Results indicated that daily activity levels, especially the variability between peak activity hours, showed potential for improving diagnostic precision in psychiatric settings. Actigraphy, combined with machine learning, offers a promising approach for classifying diagnostic groups among inpatient adolescent populations and engineered maximum and minimum hourly activity features may provide objective markers to improve diagnostic accuracy. Future studies should aim to test and validate these findings and assess their clinical implications in larger, diverse cohorts in the natural environment.