JMIR Publications

ABSTRACT

Background:

Distinguishing bipolar disorder (BD) from attention-deficit/hyperactivity disorder (ADHD) and other common psychopathology in adolescents remains a diagnostic challenge due to overlapping symptoms of mood and activity fluctuations

Objective:

To investigate same-day correlations between actigraphy-derived physical activity and self-reported mood/energy ratings, and to evaluate whether these measures can support differential diagnosis of BD, ADHD, and other psychopathologies in adolescents using both traditional statistics and machine learning.

Methods:

We included 209 hospitalized adolescents (2,148 patient-days) across four diagnostic groups: BD without ADHD (n=34), ADHD without BD (n=54), BD+ADHD (n=42), and Other Diagnoses (n=79). Actigraphy data were categorized into quartiles (Max1st–Max4th, Min1st–Min4th), and MET scores (-10 to +10) were classified into severity ranges (OK [<3], Mild [3–4], Moderate [5–6], Severe [>6]). Non-parametric analyses (Kruskal-Wallis, Mann-Whitney U) assessed group differences, and categorical associations were evaluated using chi-square tests with Cramér’s V effect sizes. To account for multiple comparisons, Bonferroni correction was applied (p-corrected < 0.004). Three machine learning models—XGBoost, Random Forest, and Ridge Regression—were developed to predict daily mood (MoodMean) from actigraphy-derived and self-reported energy features.

Results:

This study analyzed 2,148 inpatient days from 209 adolescents across four diagnostic groups (BD without ADHD, BD+ADHD, ADHD without BD, and Other Diagnoses). Actigraphy data were transformed into maximum and minimum quartiles, and daily mood and energy ratings were recorded using the Mood & Energy Thermometer (MET). Non-parametric tests and chi-square analyses with Cramér’s V were used to assess group-level differences and associations. Three machine learning models—XGBoost, Random Forest, and Ridge Regression—were developed to predict daily mood (MoodMean) from actigraphy-derived and self-reported energy features. Feature importance was analyzed within and across diagnostic groups.

Conclusions:

: Our findings supported importance of digital phenotyping where objective actigraphy combined with self-report energy metrics can accurately estimate mood and improve differential diagnosis and may personalize interventions in youth.