An RCT to evaluate LiveWell has been carried out, and data analysis is underway for the following: aim (1) to establish the capacity of LiveWell to reduce relapse and symptom burden and improve QOL in bipolar disorder, aim (2) to investigate the impact of LiveWell on proximal behavioral targets and the relationship between changes in these targets and changes in relapse rate and symptom burden, and aim (3) to identify novel behavioral signatures in individuals with bipolar disorder that predict treatment response and relapse. Aim 1 will test the following hypotheses: primary hypothesis—participants in the intervention group will experience (H1) a longer time to relapse relative to treatment as usual (TAU); secondary hypotheses—participants in the intervention group will experience (H2) a lower percentage of time being symptomatic, (H3) lower symptom severity, and (H4) a better QOL relative to TAU. Aim 2 will test the following additional hypotheses: primary hypothesis—participants in the intervention group will experience (H1) larger improvements in proximal behavioral targets (eg, medication adherence) relative to TAU; secondary hypotheses—variation in the proximal behavioral targets will account for substantial variance in the (H2) primary clinical outcome (time to relapse) and secondary clinical outcomes of (H3) percentage time symptomatic and (H4) symptom severity; proximal behavioral targets will mediate the intervention effect on the (H5) primary clinical outcome (time to relapse) and the secondary clinical outcomes of (H6) percentage time symptomatic and (H7) symptom severity; tertiary hypotheses—participants in the intervention group will experience (H8) an increase in performance determinant scores for each target, and at the final time point, will have (H9) higher performance determinant scores for each target relative to TAU. Aim 3 is exploratory and will develop a database of behavioral sensing (intervention and TAU arms) as well as app use and self-assessment (intervention arm only) data. The relationships between these data and clinical status assessment data will be examined with the long-term goal of better predicting current relapse risk, treatment response, and longitudinal course for individuals with bipolar disorder.

To facilitate recruitment of eligible participants and minimize exclusions while maximizing safety and study power, the following inclusion and exclusion criteria were used: The inclusion criteria were as follows: (1) adults aged 18 to 65 years, (2) individuals with bipolar disorder type I, and (3) a minimum of one acute episode in the last 2 years. Exclusion criteria were as follows: (1) not receiving psychiatric care, (2) current mood episode, (3) current severe suicidal ideation or a recent serious suicide attempt (last 3 months), (4) current substance use disorder (last 3 months), (5) visual impairments limiting mobile phone use, and (6) inability to speak and read English.

The study was reviewed and approved by the Northwestern University Institutional Review Board (STU00202860).

The RCT for LiveWell recruited participants from the Chicago and Minneapolis-Saint Paul areas. At both sites, recruitment letters were sent to eligible individuals (bipolar disorder diagnosis, 18-65 years, and consent to contact for research) whose information was available in site-specific research registries or electronic health record data warehouses. The recruitment letters were followed up with phone calls from the study staff. Study recruitment information was also available for both sites at ClinicalTrials.gov, ResearchMatch.org, and WeSearchTogether.org websites. In the Chicago area, individuals were also recruited via mental health and university clinic presentations, flyers, brochures, e-mails to mental-health providers affiliated with Northwestern Medicine, and advertisements (Facebook, Reddit, Craigslist, Google AdWords, Chicago Transit Authority, digital, and print newspapers).

The study team contacted individuals via telephone (research registries) or individuals contacted the study team on the web, by email, or telephone. Individuals then completed a web based or telephone screening consent and completed a brief web-based or telephone-based eligibility screener (Multimedia Appendix 6). Eligible individuals were scheduled for telephone screening using a modified Mini International Neuropsychiatric Interview [145-147] and the National Institute on Drug Abuse Quick Screen followed by the Alcohol Use Disorders Identification Test and National Institute on Drug Abuse-modified assist for additional substance use disorders screening. A suicide attempt screener was also delivered, and demographic information was obtained. If still eligible, individuals attended a face-to-face clinic visit at which written consent was completed, followed by a structured interview with a trained mental health clinician (psychiatrist or psychologist) using an abbreviated and modified version of the Affective Disorders Evaluation and Clinical Monitoring Form [148-150]. Individuals with a confirmed diagnosis at the clinic visit were scheduled for a baseline telephone assessment and enrolled if no exclusion criteria were present at this assessment. Individuals who exhibited a current mood episode or substance use disorder, severe suicidality, or a recent serious suicide attempt at any step during screening were offered the opportunity to repeat the step at a later date and continue the screening process if the exclusion criteria were resolved (Multimedia Appendix 6).

At their initial face-to-face coaching meeting, intervention arm participants were asked if they wanted to allow any of their mental health providers to access a secure password-protected website that summarizes their self-report data (Daily and Weekly Check-Ins). If a participant consented to provider participation, a letter offering participation was mailed to the provider, and the coach contacted the provider by telephone. Providers interested in participating completed a web-based, verbal, or written consent form before receiving access to the website. Providers could opt to receive alerts via email or telephone when participant self-report data indicated reduced medication adherence, increased or decreased sleep duration, or deterioration in their daily wellness ratings. Providers were not required to opt for alerts to participate in the study. To allow access to the widest range of participants, participants were not required to allow any providers to access the website and providers were not required to consent to access the website or receive alerts for a participant to enroll in the study. However, as part of the written consent to participate in the study, all participants agreed to allow their psychiatrist to be contacted in the event of self-rated crisis situations, including daily wellness ratings of +4 or −4, new onset of depression (PHQ-8 score≥10), new onset of mania (ASRM score≥6), or other indications of emergent clinical problems or mental health deterioration.

A biostatistician, blinded to screening and baseline assessment data, conducted computer-generated randomization on a 2:3 ratio (control:intervention) stratified based on clinical status (low risk—asymptomatic recovery; high risk—continued symptomatic, recovering, prodromal, and symptomatic recovery; Tables 4 and 5). Patients were randomized in permuted blocks of 5 at each site. The unbalanced design increases power to investigate effects within the intervention arm with minimal effect on the investigation of outcomes. Participants were stratified because time to relapse is likely to be significantly shorter for the high-risk participants who have not met the time or symptom number criteria for recovery from the last episode or who are recovered but have subsyndromal or prodromal symptoms [9,94,151,152]. Although many prior face-to-face psychotherapy studies restrict participant inclusion to those in asymptomatic recovery [10-15,18], this study includes additional high-risk individuals to increase access to the intervention.

The primary outcome, time to relapse, will be measured as the number of weeks to episode onset (depression, mania, hypomania, or mixed) based on the Diagnostic and Statistical Manual of Mental Disorders fourth edition (DSM-4) criteria. The DSM-4 episode criteria were chosen to allow comparison with prior RCTs of face-to-face psychotherapy for bipolar disorder, as these studies primarily used the DSM-4 episode criteria [10-18]. To determine the number of weeks to episode onset, weekly clinical status ratings (Tables 4 and 5) were assessed by timeline follow-back rating of bipolar disorder symptom severity (Multimedia Appendix 7) using a modified Longitudinal Interval Follow-Up Evaluation (LIFE) and Clinical Monitoring Form (CMF) [148,149,153]. The secondary outcomes include the percentage of time symptomatic, symptom severity, and QOL. To determine percentage of time symptomatic, weekly psychiatric status ratings (Table 6) were assessed using the LIFE-CMF to determine the percentage of time symptomatic. A week will be scored as symptomatic if the psychiatric status rating is greater than one and a half (Table 6). Symptom severity was assessed using the Quick Inventory of Depressive Symptomatology (QIDS)–Clinician Rating and the Young Mania Rating Scale [154-157]. QOL was assessed using the World Health Organization Quality of Life BREF [158].

The targets selected for assessment were taking medications as prescribed, obtaining adequate sleep duration, maintaining regular routines, and managing signs and symptoms as these targets are readily amenable to monitoring using the Daily Check-In, which is a core feature of LiveWell. Medication adherence was assessed using the Tablet Routine Questionnaire [159-161] and focused on percentage adherence with prescribed psychiatric medications, not including those prescribed as needed. Sleep duration, quality, and excessive sleepiness were assessed using the CMF, QIDS, Pittsburgh Sleep Quality Index, and an additional question to capture excessive sleepiness [162-164]. Sleep duration was primarily considered in this study. Sleep quality and excessive sleepiness will be explored secondarily because interepisode insomnia and hypersomnia are risk factors for relapse [107,163,165,166]. Routine was assessed using a 5-question trait version of the Social Rhythm Metric [167-170] and will be assessed primarily by a frequency score (Multimedia Appendix 7). Finally, the management of residual and prodromal signs and symptoms was assessed using the Symptom Management Scale. This scale was developed from the prodromes coping interview and inventory, which has been used in face-to-face cognitive behavioral therapy for bipolar disorder [99,137,142,171]. The outcome measure will be the total coping score (Multimedia Appendix 7).

Assessors blinded to the study arm conducted outcome and target assessments by telephone at baseline and then every 8 weeks until study exit at week 48 (Figure 6). Windows for complete assessments were up to 4 weeks after the assessment due date, calculated from the baseline assessment date. To minimize assessment burden and because QOL may vary less rapidly than other outcomes [172], the QOL assessment was delivered only at 0, 24, and 48 weeks. Assessments began by focusing on participants’ experiences over the prior 2 weeks. Mood and sleep were assessed using combined mood (CMF, QIDS, and Young Mania Rating Scale) and combined sleep (CMF, QIDS, and Pittsburgh Sleep Quality Index) instruments (Multimedia Appendix 7). The combined instruments streamlined the interviews and reduced the time required to complete the assessments. This was followed by assessment during the prior 2 weeks of routine (Social Rhythm Metric), medication adherence (Tablet Routine Questionnaire), and depending on the week QOL.

Information from the assessment of the prior 2 weeks was then used to anchor the timeline follow-back of weekly depressive and manic symptom severity (LIFE-CMF). In addition to assessing symptom severity, timeline follow-back using the LIFE assessed weekly variations in anxiety, suicidal ideation and attempts, psychotic symptoms, substance use, life events [173], mental health treatment attendance, and psychiatric medication adherence (Multimedia Appendix 7). The Symptom Management Scale was assessed last because it did not focus on experiences during the previous 2 weeks or since the last assessment. Instead, participants were asked to imagine experiencing ongoing or new low-level depressive or manic symptoms and how they might manage them (Multimedia Appendix 7). This approach was used because participant clinical status during the assessment interval may not have required management of prodromal or residual symptoms.

To assess changes in the determinants proposed to govern the enactment of target behaviors, determinant questionnaires were developed for each assessed target (Multimedia Appendix 8). The determinant questionnaires are derived from the question stems of existing medication adherence questionnaires for schizophrenia and bipolar disorder [174,175] and health psychology questionnaires for assessing determinants of physical exercise and other health-related behaviors [76,77,176-183]. The questionnaires measure determinants proposed to underlie the pursuit of LiveWell’s four assessed behavioral targets (taking medications as prescribed, obtaining adequate sleep duration, maintaining regular routines, and managing signs and symptoms). The questionnaires were delivered via a web survey sent by the coaches to the participants. For the intervention arm, they were delivered at weeks 0, 16, and 48 (Figure 6). For participants in the control arm, they were only delivered at week 48 (Figure 6) because the delivery of the determinant questionnaires may serve as an active intervention [76,77,181]. Eventually, the determinant questionnaires may be incorporated into the smartphone app to provide real-time assessment of target change processes to facilitate improved adaptive real-time feedback.

Participants were offered the option of having the LiveWell app (intervention arm only) and an app for sensor data collection and secure transmission of data (Purple Robot, intervention, and control arms) [184] installed on their smartphone. Alternatively, participants had the option of receiving the apps on a study phone with an unlimited national calling and data plan. Individuals who opted to use their own phone were reimbursed for their plan up to the cost of the study phone plan or their own plan, whichever was less. All participants were provided with a wrist-worn actimeter (Pebble). Participants were asked to (1) use the phone with the apps as their only phone, (2) carry the phone with them whenever they left home, and (3) wear the watch 24 hours a day, 7 days a week except while it was charging.

Clinical outcomes and target assessments were performed by assessors with master’s degrees in mental health or counseling. Structured interviews can be validly and reliably administered by such personnel [185-189]. Assessors were trained by mental health professionals (a psychiatrist and psychologist) with instruction and practice on audiotaped ratings, followed by observing, role-playing, and observed assessments (≥3 each). All assessment telephone calls were audiotaped. Assessor fidelity focused on using the CMF for assessing status in the prior 2 weeks. Assessment audiotapes were randomly selected each month (15%) for review and scoring by a trained mental health professional (psychologist) to assess the fidelity of CMF scoring for symptom severity (percent match within 0.5), asymptomatic status (percent match), clinical status (percent match), and psychiatric status rating (percent match).

As the randomization was stratified by risk group, analyses will examine if an interaction between risk strata and the study arm is present. If no evidence of an interaction is evident, the models will be stratified or adjusted for risk strata. For time to relapse, missing data for participants who dropped out or were lost to follow-up will be censored at their last study visit. Percentage time symptomatic will be calculated based only on the observed time in the study. For longitudinal data (symptom severity, QOL, targets), multiple imputation using Markov Chain Monte Carlo methods will be used to address missing data by creating five unique data sets (SAS procedure PROC MI), and the results will be combined (SAS procedure PROC MIANALYZE) to obtain valid statistical inferences and least square means of the outcomes for each time point data will be calculated.

Time-to-relapse curves will be constructed using the Kaplan-Meier method. Log-rank tests will be used to determine if there is a univariate impact of the study arm while adjusting for risk strata. Cox proportional hazard models will then be fit to adjust for risk strata. Proportional hazards assumptions will be assessed, and the estimated hazard ratio and CIs will be presented. To compare percent time symptomatic, linear regression models will be used to determine if there was an interaction between risk strata and the study arm. If no evidence of interaction is seen, the main effect of arm will be reported and least square means will be estimated while adjusting for risk strata. Generalized linear mixed models for longitudinal data will be used for symptom severity and QOL measures. We will examine the interactive effects of the intervention on time. If no interaction is detected, we will test the independent effects of time as well as the treatment group.

Under an initial assumption of 200 participants enrolled and randomized at a 2:3 ratio with a control group relapse rate of 0.60, there would be 80% power to detect a hazard ratio of 0.61 assuming a loss to follow-up rate of 12%. The control relapse rate estimate was based on the reported percent relapse at 12 months for control (63%) seven psychosocial interventions for bipolar disorder [190-198]. This would equate to a reduction in relapse rate to 43% in the LiveWell intervention group. This additionally assumes 30 months of accrual, and 12 months follow-up using a log-rank test at a type I error rate of 5% (PASS 2008) [199]. For secondary outcomes, assuming 12% loss to follow-up, a 2-tailed t test would have 80% power to detect effect sizes of 0.43, for percent time symptomatic, which would equate to differences in mean time of 14% (45% symptomatic vs 69%, assuming an SD of 32%). While power calculations for generalized linear models do exist, they are quite dependent on assumptions that were speculative at the time of study design. Using multiple linear regression, a sample size of 200, adjusting for one covariate that explains 20% of the variability in outcome, we have 80% power to detect an increase in R-squared of 3%.

Generalized linear mixed models will be used to examine the effects of time, risk strata, and study arm on longitudinal target and determinant data. We will use the framework of Muller et al [200] to identify the targets that mediate the effect of the intervention on time to relapse, percentage time symptomatic, and symptom severity. To compare the performance determinant scores for each target at the final time point, t tests with correction for multiple testing between the two groups will be used.

For both arms, behavioral sensor data will be collected including activity (Pebble watch and phone accelerometers), location (via the GPS), timing of incoming and outgoing texts, timing and duration of incoming and outgoing telephone calls, and for some phones, ambient light (lux), and sound (power and frequency). Feature variables will be summarized over 1-week windows, to identify behavioral features strongly correlated with depression and mania symptom severity. The relationship between behavioral features and clinical status will then be explored. In addition, for the intervention arm, self-report assessment and app use data as well as personalized wellness rating anchors and plans, reduce risk plans, sleep duration goals, and routine (bedtime and risetime) goals will be available. These data will be used to explore processes involved in staying well with the aim of improving adaptive delivery of content and tools to decrease relapse risk and symptom burden.