We analyzed data from 145 participants, the majority of which were diagnosed with BD I (92 (63.4%) and 53 (36.6%) with BD II. The participants have been enrolled in the study for a duration of 449±224 days until June 23, 2023. Participants were provided with an Oura smart sensor (Oura Health Oy, Generation 2, Oulu, Finland), a wearable ring that continuously measures activity (e.g., number of steps), sleep (e.g., total sleep duration), and cardiorespiratory variables (e.g., heart rate). The participants were mailed a sizing kit for individualized sensor size selection to ensure optimal skin-sensor contact and data quality. Additionally, participants received a secure e-mailed link asking them to complete a weekly Patient Health Questionnaire (PHQ-9) through a secure email link. Participants must complete all items on both self-rating scales to submit their ratings. For the analysis of activity data, we selected the number of daily steps as a representative variable. As for sleep data, we analyzed the minutes of total sleep per night. We performed multivariate oscillatory mode decomposition using the CEEMD-AN algorithm, followed by Hilbert-based instantaneous frequency computation, and data-driven spectral derivative spike detection. The total PHQ-9 self-rating scale at the vicinity of each detected spike in activity or sleep variability rate was used for labeling episodes of illness on a weekly basis. Using the daily step variable to represent sample activity data, the implementation of the MR-TF-SD2 algorithm on our database showed decreasing levels of episode onset detection sensitivity with decreasing time resolution. Similarly, using the daily total sleep variable as sample sleep data, episode onset detection sensitivity decreased from day-to-day patterns to monthly total sleep patterns

We recruited 53 BD participants at two Canadian academic psychiatric hospitals (the Centre for Addiction and Mental Health, Toronto; the Royal Ottawa Hospital, Ottawa) between April 2016 and December 2019. Participants were provided with a BioHarness™ 3.0 wearable physiological electronic (e-) monitoring device, which they wore continuously for 24 hours. Posture data were recorded in units of degrees from vertical, sampled every sec (1 Hz) with a sensitivity range of 1° to 8°, and a dynamic range of ±180°. The sensors were configured such that a posture value of -90° indicates a supine posture (i.e., lying face up), and a 90° posture indicates a prone posture (i.e., lying face down). Posture was represented as a 1 Hz-discretized single channel of angular positions of a participant’s chest over the course of 24 hours. We extracted a set of 9 time-domain features to characterize postural dynamics in terms of amplitude, energy, variability, and transitions for 3 different periods: day (from 7:00 AM to 2:59 PM), evening (from 3:00 PM to 10:59 PM), and night (from 11:00 PM to 6:59 AM). To assess posture amplitude, we computed the mean posture (angle in degrees) and its range; to assess posture dynamics’ energy content, we computed the root mean squared (RMS) value; to assess posture variability, we computed the coefficient of variation (CV), interquartile range (IQR), and median absolute deviation (MAD). Kurtosis and skewness were computed to assess the postural statistical distribution in terms of distribution sharpness and symmetry. Lastly, the number of postural transitions was computed using Bayesian Online Changepoint Detection (BOCD) which identifies the abrupt changes in sequential data generative parameters, such that each changepoint is indicative of a postural transition (e.g., being upright to bending over, or lying down to sitting upright). We used the Kruskal-Wallis test to assess the level of inter-cluster statistical significance for each posture feature, and corrected the p-values using the Benjamini-Hochberg (BH) method. Then, in each posture-specific cluster, we assessed the median and IQR of cluster-specific illness burden variables. We computed pairwise Spearman correlation coefficients to assess the strength and direction of the association between the postural dynamics descriptors (e.g., mean, IQR) and illness burden continuous variables (e.g., lifetime number of depressive episodes). We used a Chi Square test to assess the association between posture and categorical illness burden variables (e.g., history of suicide attempts, family history of suicide). We controlled for age, baseline functional capacity, and body mass index (BMI) by setting them as control variables in a multiple linear regression model. The p-values were corrected using the BH method. To identify cluster members (i.e., participants who shared similar postural dynamics), we performed hierarchical clustering of BD participants using posture features as model input.