This study was approved by the Human Research Protection Office at the University of Pittsburgh Institutional Review Board (ID STUDY21110097). Data were collected as part of clinical care for mood monitoring during inpatient stay and not as part of a funded research study, and thus, participants were not compensated for their data. All data were fully deidentified before analysis, and the Institutional Review Board determined that additional informed consent was not required. Patient privacy and confidentiality were maintained throughout the study.

This retrospective study analyzed data from 459 adolescent inpatients at the Western Psychiatric Hospital, University of Pittsburgh Medical Center, collected from 2014 to 2023. Participants, aged 13 to 18 years, were diagnosed with well-characterized BD without ADHD (mean age 14.97, SD 1.36 years; 82% [28/34] female), BD with ADHD (mean age 15.33, SD 1.91 years; 57% [24/42] female), ADHD without BD (mean age 15.00, SD 1.39 years; 44% [24/54] female), and other diagnoses (mean age 15.23, SD 1.46 years; 75% [59/79] female). The other diagnoses group included patients with a variety of psychiatric disorders (anxiety disorders: generalized anxiety, panic disorder, social anxiety, social phobia, separation anxiety disorder, and other anxiety disorders), disruptive behavior disorders (conduct disorder, oppositional defiant disorder, and intermittent explosive disorder), obsessive-compulsive and related disorders, substance use disorders, and trauma and related disorders (posttraumatic distress disorder, acute stress disorder, and adjustment disorder), each potentially displaying different activity levels. Since our focus was on BD, we made sure that our participants in the other diagnoses group, including those with major depressive disorder, had no history of manic episodes (based on DSM-5 criteria and expert review) to avoid any overlap with BD. The study included all inpatients at the Western Psychiatric Hospital who completed MET daily ratings and wore an actigraphy device during their hospitalization while they were receiving diagnostic assessments and treatments. The study excluded patients with a diagnosis of autism spectrum disorder (levels II and III), intellectual disability, eating disorder, schizophrenia, or severe substance use disorder. After processing and cleaning of actigraphy data according to previously published studies (eg, standardized visual editing procedures and inclusion of additional semiautomated quality assurance procedures on Actiware [Phillips Resipronics], including identification of the main rest interval [defined as the longest rest interval each day] and removal of invalid sleep intervals containing ≥1 hour of off-wrist time or recording errors) [12,13], 209 participants were included in the analysis.

RSD, a licensed, board-certified pediatric psychiatrist specializing in pediatric BD and the director of the Inpatient Child and Adolescent Bipolar Services unit at Western Psychiatric Hospital, directly trained all clinical diagnostic staff and confirmed all psychiatric diagnoses through consensus meetings. The DSM-5 criteria were used for the final identification of BD-I/II, and other specified bipolar was defined as the presence of brief (4 hours/day and 4 in a lifetime) cluster DSM manic symptoms that caused impairment. All instruments had good psychometric properties and included the Kiddie Schedule for Affective Disorders and Schizophrenia for School-Age Children (K-SADS) [14].

The self-reported MET, developed at the Western Psychiatric Hospital, rates mania and increased energy on a scale of 1 to 10 points and depression and tiredness on a scale of −1 to −10 points [1]. Each participant completed the MET scale each morning and evening, and separate scores for mood and energy were averaged across the day. For this study, we included (1) the highest and lowest scores for elated mood (“MoodPosMax” and “MoodMin”) and elated energy (“EnergyPosMax” and “EnergyMin”) on the MET (scale from 0 to +10 points), (2) the highest and lowest scores for depressed mood (“MoodNegMax” and “MoodNegMin”) and depressed energy (“EnergyNegMax” and “EnergyNegMin”) on the MET (scale from 10 to 0 points), and (3) the range for mood (“MoodRange”) and energy (“EnergyRange”) on the MET (scale from −10 to +10 points).

Patients with BD (feeling ok, experiencing mania, or experiencing depression) assessed their mood and energy levels daily using this scale. Our master’s degree clinician (MW) met with the patients daily to assist in identifying and recording mood symptoms.

Continuous activity monitoring was conducted using the Philips Actiwatch 2 (Phillips Respironics), worn on the nondominant wrist, which recorded physical activity in 1-minute epochs throughout each patient’s hospitalization, capturing both daytime activity and sleep patterns. To ensure a robust activity profile aligned with clinical diagnostic criteria for BD, which requires a minimum of 4 days of observation, only patients with at least 4 consecutive days of actigraphy data were included. The initial dataset comprised 459 adolescent inpatients admitted between 2014 and 2023, and it underwent rigorous quality control: 6 cases were excluded due to missing demographic data, 29 cases were removed for insufficient actigraphy data (eg, admission or discharge days or fewer than 4 days of recording), and 35 additional cases were excluded for incomplete mood or energy scores, resulting in 389 cases. Further refinement excluded 180 cases with zero values for all mood or energy reports on over 70% of days in order to focus on clinically informative data, yielding a final cohort of 209 adolescents (2148 patient-days). This study involved a diverse cohort, with the highest participation from the “other diagnoses” group (79 unique IDs and 799 days recorded). This was followed by the “ADHD without BD” group (54 unique IDs and 572 days), “BD with ADHD” group (42 unique IDs and 444 days), and “BD without ADHD” group (34 unique IDs and 333 days).

To quantify activity extremes, a custom Python algorithm (Automatic Max and Min Identifier) identified 4 nonoverlapping 60-minute maximum (Max1st-Max4th) and minimum (Min1st-Min4th) activity periods daily between 07:00 AM and 09:59 PM. Differences between maxima and minima (eg, Max4 – Min4) were calculated to assess activity variability, and timestamps for peak and trough periods were recorded to evaluate diurnal patterns. Continuous actigraphy data were categorized into quartiles (Max1st-Max4th and Min1st-Min4th) using Python’s pandas.qcut() function, ensuring equal observation distribution to mitigate skewness. Mood and energy scores from the MET were classified into severity ranges: OK (<3), mild (3-4), moderate (5-6), and severe (>6), aligning with clinical standards.

We used nonparametric tests to evaluate differences in mood and energy variables across diagnostic groups. Kruskal-Wallis tests assessed overall group differences, and pairwise Mann-Whitney U tests with Bonferroni correction controlled for multiple comparisons (adjusted significance threshold: P<.004). Associations between actigraphy-derived activity quartiles (Max1-Max4 and Min1-Min4) and mood and energy severity categories from the MET were examined using chi-square tests with Cramér V effect sizes. Effect sizes were interpreted using established conventions: negligible (0‐0.1), weak (0.1‐0.2), moderate (0.2‐0.3), and strong (>0.3). All data were anonymized to ensure confidentiality.

Statistical analyses were conducted in three phases: (1) group comparisons of mood and energy variables using Kruskal-Wallis tests and pairwise Mann-Whitney U tests with Bonferroni correction (P_corrected<.004); (2) associations between categorized actigraphy-based activity levels (quartiles) and mood/energy severity ranges using chi-square tests with Cramér V effect sizes; and (3) machine learning classification (logistic regression, HistGradientBoosting, random forest, extra trees, and XGBoost) to predict daily mood severity (SevereDay and SevereTomorrow), based on actigraphy-derived features and self-reported energy scores. Feature importance was extracted to identify the top predictors within each diagnostic group.

Table 1 summarizes the demographic and clinical characteristics. The “BD with ADHD” and “other diagnoses” groups included a higher proportion of females than the “ADHD without BD” group. Participants in these 2 groups were also older than those in the “ADHD without BD” and “BD without ADHD” groups.

Kruskal-Wallis tests, corrected for multiple comparisons (P_corrected<.004), revealed significant differences across diagnostic groups for most mood and energy variables (Multimedia Appendix 1). For mood ratings, the “BD without ADHD” and “ADHD without BD” groups reported the highest maximum positive mood scores (MoodPosMax), whereas the “other diagnoses” group showed the highest maximum and minimum negative mood scores (MoodNegMax and MoodNegMin) and the widest mood range. Minimum positive mood (MoodMin) was the lowest in the “BD with ADHD” and “ADHD without BD” groups.

For energy ratings, the “ADHD without BD” and “BD without ADHD” groups showed the highest maximum positive energy (EnergyPosMax), while the “other diagnoses” group exhibited the most severe negative energy states (EnergyNegMax and EnergyNegMin) and the broadest energy range. Minimum positive energy (EnergyPosMin) was the highest in the “ADHD without BD” group and the lowest in the “BD with ADHD” and “BD without ADHD” groups. Overall, the “BD without ADHD” and “ADHD without BD” groups were characterized by elevated positive mood and energy and lower variability, whereas the “other diagnoses” group—largely including depressive disorders—was marked by greater negative affect and instability.

Pairwise contrasts (Multimedia Appendix 2), adjusted with Bonferroni correction (P_corrected<.004), confirmed distinct symptom profiles across diagnostic groups. The “ADHD without BD” and “BD without ADHD” groups exhibited significantly higher maximum positive mood (MoodPosMax) and positive energy (EnergyPosMax) compared to the “BD with ADHD” and “other diagnoses” groups (P<.001). In contrast, the “other diagnoses” group consistently showed the most severe negative mood and energy scores (MoodNegMax, MoodNegMin, EnergyNegMax, and EnergyNegMin; all P<.001). Within the BD subtypes, “BD without ADHD” reported higher negative mood than “BD with ADHD” (P<.01). Overall, “BD with ADHD” presented the least variability across mood and energy domains.

Box plot visualizations (Multimedia Appendix 1) highlighted that the “other diagnoses” group experienced disproportionately higher negative mood and energy scores, with several outliers indicating extreme cases.

Kernel density estimates for mood and energy variables across diagnostic groups are shown in Multimedia Appendix 3. Most participants across all groups clustered in the “OK” range (<3), consistent with mild symptoms. However, “ADHD without BD” was overrepresented in the upper ranges of positive mood and energy, while “other diagnoses” was overrepresented in severe negative states. The “ADHD without BD” group had the greatest number of severe cases for positive mood and energy, whereas the “other diagnoses” group contributed the largest share of severe negative mood and energy cases. The “BD without ADHD” group generally had the fewest severe cases, consistent with a comparatively more stable profile.

Additional variability analyses (SD or IQR and distributional summaries) are presented in Multimedia Appendix 4, which revealed that energy variables fluctuated more widely than mood variables across all groups. The “other diagnoses” group displayed the greatest instability, with the largest SDs for negative mood (MoodNegMax SD=3.86) and energy (EnergyNegMax SD=3.65). By comparison, the “ADHD without BD” and “BD with ADHD” groups showed similar variability, while the “BD without ADHD” group exhibited narrower energy ranges (EnergyRange: 17 vs 20 in other groups), suggesting a more constrained but distinct fluctuation pattern. IQR analyses reinforced these findings. The “other diagnoses” group had the widest IQRs for negative mood (MoodNegMax IQR=7) and energy (EnergyNegMax IQR=6), whereas the “BD with ADHD” group showed tighter distributions for positive mood (MoodPosMax IQR=3).

The correlation matrix between activity quartiles and mood and energy variables is shown in Multimedia Appendix 5.

This study introduces a novel methodological approach by transforming continuous actigraphy data into quartiles and applying categorical statistical analyses to examine associations among physical activity, mood, and energy in adolescents with BD, ADHD, and other psychiatric diagnoses. Our results revealed significant diagnostic differences in self-reported mood and energy profiles, with distinct patterns of variability and correlation strength across groups. In particular, BD without ADHD was characterized by negative mood-energy fluctuations tied to higher ACs, while ADHD without BD showed a stronger link between activity levels and positive energy states.

The correlation analyses demonstrated strong within-domain associations—positive mood with positive energy and negative mood with negative energy—consistent with prior research showing that mood and energy are closely interdependent in pediatric populations [15]. The ADHD without BD group exhibited a particularly high correlation (r=0.78) between maximum positive mood and maximum positive energy, underscoring the pronounced interplay between these dimensions in the absence of comorbid BD [16,17]. In contrast, BD without ADHD displayed the strongest associations between negative mood, negative energy, and activity fluctuations (Cramér V=0.195‐0.241). The presence of comorbid ADHD in BD significantly impacts mood and energy dynamics, suggesting distinct clinical phenotypes resulting from this co-occurrence [17]. This aligns with prior research showing that mood instability in BD is often accompanied by circadian disruptions and irregular activity levels [5,7,8]. These findings suggest that actigraphy may serve as an objective marker of negative affect dysregulation in BD. Moreover, the moderate association between energy variability and actigraphy indicates that daily energy fluctuations may provide a more reliable marker of dysregulation than self-reported mood alone.

In ADHD without BD, higher activity levels were more strongly associated with positive energy (Cramér V=0.225), suggesting that actigraphy reflects arousal and hyperactivity rather than mood instability. This aligns with prior work differentiating ADHD-driven hyperactivity from BD-related mood changes [17]. Features, such as EnergyMean, EnergyMax/Min, and average ACs, emerged as potential diagnostic markers to distinguish ADHD from BD in clinical settings.

Patients classified under “other diagnoses” showed the greatest variability in negative mood and energy, highlighting instability but weaker overall associations with actigraphy. This may reflect the heterogeneity of this group, which included depressive and disruptive disorders, where actigraphy may be less informative for mood-energy dynamics [18-20]. This aligns with findings that indicate a spectrum of mood dysregulation that extends beyond traditional diagnostic boundaries, necessitating a more comprehensive understanding of the interplay between physical activity and psychological factors [21]. Our findings emphasize the complex interplay among physical activity, mood, and energy levels in pediatric psychiatric conditions [22]. These findings are consistent with previous research [12,23], emphasizing the utility of actigraphy in assessing sleep disorders, evaluating circadian rhythms, and pediatric sleep research. Incorporating mood, energy, and actigraphy assessments can refine clinical phenotyping and enable tailored interventions targeting specific symptom dimensions in each disorder [24-26].

Machine learning results demonstrated that combining actigraphy, sleep, and energy features can meaningfully classify mood severity. Our leak-safe pipeline showed that same-day severe mood could be predicted with good accuracy (ROC-AUC=0.84), while next-day risk was more modest but clinically useful (ROC-AUC=0.78). Energy variability, peak values, and average actigraphy metrics consistently ranked among the most important predictors. Notably, a regularized linear model (logistic regression) performed on par with or better than tree-based ensembles, suggesting that much of the predictive signal is captured by summary features when leakage is controlled. At clinically relevant thresholds, the same-day model achieved severe-class recall of approximately 0.75 with precision of approximately 0.61, supporting its potential for inpatient monitoring. Next-day detection (recall of approximately 0.54; precision of approximately 0.67) may provide actionable early alerts for staffing or clinical check-ins.

Taken together, these findings suggest that wearable-derived actigraphy and sleep measures, when combined with simple daily energy ratings, provide a robust framework for real-time monitoring of mood severity in adolescents with psychiatric conditions [27]. BD without ADHD emerges as a group where actigraphy is particularly informative [28] for identifying negative mood states, while ADHD without BD is better characterized by associations with hyperactivity and positive energy. Clinically, these results support incorporating daily energy tracking into psychiatric assessments, as energy may be a more stable and discriminative marker than mood in certain contexts.

This study has several limitations. First, its cross-sectional design precludes causal inferences. Longitudinal studies are needed to clarify whether activity fluctuations drive mood changes or vice versa. Second, actigraphy alone does not capture contextual influences, such as medication effects, sleep disturbances, and environmental stressors. While our naturalistic inpatient setting increases ecological validity, future research should integrate multimodal assessments, including polysomnography, ecological momentary assessment, and biomarker data to better account for these factors. Third, class imbalance in severe mood states may have reduced statistical power. Although our group-aware, leak-safe modeling approach mitigated this issue and applied the SMOTE only within training folds, rare-event calibration remains a challenge. Moreover, the models did not incorporate within-patient correlation via mixed-effects or adjust for treatment and milieu variables, which are both priorities for future work.

A prior SMOTE-based approach balanced multiclass mood and energy categories (OK, mild, moderate, and severe) but did not enforce patient-wise separation, raising concerns about data leakage. Our revised framework addressed this by restricting analyses to binary severity outcomes, applying the SMOTE only within training folds, and using group-level cross-validation. This design yields more conservative, leakage-safe estimates of real-world performance.

Our findings highlight disorder-specific patterns of mood-energy dysregulation: BD without ADHD was characterized by strong associations between negative mood or energy states and extreme activity fluctuations, while ADHD without BD was characterized by high activity linked to positive energy rather than mood instability. In a leak-safe, patient-wise framework, combining daily energy ratings with actigraphy- and sleep-derived features reliably identified same-day severe mood and provided moderate next-day risk signals (ROC-AUC=0.84 and 0.78, respectively). Energy variability and peaks, alongside average and extreme activity metrics, consistently emerged as the most informative predictors.

Clinically, these results support incorporating daily energy tracking with wearable devices into psychiatric assessments. Actigraphy may be especially valuable for identifying negative affect states in BD, while energy ratings offer a stable and discriminative marker across conditions. The integration of objective activity measures with subjective energy assessments could refine real-time psychiatric monitoring, improve diagnostic precision, and guide personalized interventions for adolescents with BD, ADHD, and related disorders.