A recruitment target of 30 was set for this study, commensurate with usual practice for feasibility studies [33]. A sample size of 30 allows estimation of a compliance rate of 80%, with 95% CIs of ±12.8%. This sample size also provides 80% power to detect correlations of r=0.5.

A total of 556 adults underwent an initial screening for eligibility to participate in the study through a patient recruitment company with links to primary care providers and depression patient groups, to identify individuals with depression potentially suitable for the study. In total, 72 individuals were contacted for more detailed medical history information and to complete the Patient Health Questionnaire-9 (PHQ-9) [34] to obtain an index of depression severity. Participant eligibility was determined according to the following inclusion and exclusion criteria before study entry:

Inclusion criteria were primary psychiatric diagnosis of MDD; treated with antidepressant monotherapy; mild-to-moderate depression, defined by PHQ-9 scores between 5 and 15; aged 18 to 65 years; able to read and understand English; and owning their own iPhone.

Exclusion criteria were personal history of other psychiatric disorder (except nonprimary concurrent anxiety); manic or hypomanic episode; mental retardation, organic mental disorders, or mental disorders owing to a general medical condition as defined in the Diagnostic and Statistical Manual of Mental Disorders 5th Edition; neurological or neurodegenerative disorder; alcohol or other substance abuse or dependence (excluding nicotine or caffeine); responding only to combination or augmentation therapy in the current episode; hospitalization for MDD in 3 months or suicide attempt in 6 months before screening (or the participant was considered to be at significant risk of suicide or hospitalization); having received any investigational compound within 30 days before screening or 5 half-lives before screening, whichever is longer; concurrent participation in other clinical studies; or participation in 2 or more interventional studies in the year before screening.

In total, 30 of the 72 screened individuals were recruited into the study. Of the remaining screened individuals, 7 were eligible but not recruited. Others were excluded because of lack of an iPhone (n=4), insufficient time on medication (n=18), lack of antidepressant medication treatment history (n=2), polypharmacy (n=3), other psychiatric diagnosis or neurological condition (n=5), PHQ-9 higher than 15 (n=1), or insufficient information obtained in screening (n=2).

The study began with a visit to the study site and a short semistructured interview to explore each participant’s expectations and motivations for taking part. Researchers provided study hardware (an Apple Watch Series 2, paired with the participant’s own iPhone), presented the tasks, and gave participants the opportunity to practice using the tasks and device and ask questions. Participants were given contact details for the study center, where they could get in touch by email or phone if they experienced technical issues or had questions or concerns regarding their participation. Testing was completed in the subsequent 6 weeks (42 days), corresponding to the time when antidepressant pharmacotherapy shows efficacy in treating the mood symptoms of MDD. Participants were encouraged to respond to cognitive assessment wherever possible but not to worry when individual assessments were missed.

Data collected on the Apple Watch and iPhone were transferred automatically through Wi-Fi or data roaming via the participant’s iPhone to a secure data center held on Amazon’s Web service. This service provided identity and access control mechanisms to ensure participants (and only participants) had write access, and study managers only had read access. Where data for individual participants were not uploaded for 4 days, the research team made contact to ensure that the study equipment was working and to gain a better understanding of why assessments were not completed.

Full-length cognitive and validated self-report assessments were completed via a Web-based testing interface. Familiarization with the tests was completed during in-person assessments on the first day of participation. Full assessments were completed on 3 occasions: week 1 (between days 1 and 2), week 3 (days 18-24), and week 6 (days 40-46). Participants were sent a unique link to a secure Web page that delivered the test. On completion of assessment, and when the device established an internet connection, data were transferred to a secure Health Insurance Portability and Accountability Act of 1996 ̶ compliant data center in the United States.

The study was completed with a 90-min, semistructured qualitative interview during week 6 at participants’ homes. Interviews explored participants’ experiences of assessment with the wearable technology, changes in motivation and adherence, and contextual factors that might have contributed to those changes. Study hardware was returned at this time.

The study was reviewed and approved by the Proportionate Review Sub-Committee of the Wales Research Ethics Committee 6 at Swansea University (REC reference: 17/WA/0042) and performed in accordance with the current version of the Declaration of Helsinki. All participants provided written informed consent before enrollment.

Of the 37 eligible participants, 30 were enrolled (19 women and 11 men). Participants were aged between 19 and 63 years (mean age 37.2 years; SD 10.4) and had been on their current medication for an average of 9.9 months (range 0.4−94.3 months; SD 9.5). Current medications included serotonin antagonist and reuptake inhibitor (n=1), serotonin and norepinephrine reuptake inhibitors (n=5), selective serotonin reuptake inhibitors (n=20), and tricyclic antidepressants (n=4). Mean depression symptom severity, measured by the PHQ-9, was 9.1 (range 5-15; SD 3.1).

Descriptive statistics for adherence across the duration of the study and by study week are shown in Table 1. Full adherence (100%, 42/42 days) was seen in 21 of 30 participants for cognitive assessment, 15 of 30 participants for mood testing, and 13 of 30 participants for activity assessment. Periods of low adherence tended to cluster temporally (Multimedia Appendix 2). Because of a technical issue on the final study day, the evening session was not administered, resulting in lower adherence on day 42. However, logistic mixed modeling showed no deterioration in adherence (ie, responding at least once daily) over time for assessments of mood, cognition, or activity. Logistic regression confirmed that self-reported depressive symptoms, assessed by the PHQ-9, PDQ-D, and UCLA-LS, were not associated with level of adherence in mood or cognitive assessments. Adherence was not significantly correlated with any CANTAB measures at week 1 (maximum rho=0.15; P=.44).

Participants completed a mean of 86.8% of all possible n-back assessments (range 50%-99%, 63-125 of 126 assessments). Rate of responding in the morning (84%) was lower than the afternoon (87%) and evening (89%; χ²₂=12.9). Furthermore, although adherence (responding at least daily) remained high throughout the duration of the study, logistic regression confirmed modest reductions in individual assessments (morning, afternoon, and evening) over the study duration (morning: fixed-effects estimate=−0.03, P=.02; afternoon: fixed-effects estimate=−0.02, P<.001; evening: fixed-effects estimate=−0.08, P<.001).

Descriptive data for n-back assessments are presented in Table 2. Multilevel analysis of dprime score by study day confirmed a better fit for a cubic term rather than quadratic or linear models (Bayesian information criterion=1298.05; likelihood ratio=10.36; P=.001), indicating an initial rapid improvement in performance followed by a plateau. Model fits for each study participant are shown in Figure 2. Dprime slope showed no significant relationship with the number of n-back assessments completed (rho=−0.02, 95% CIs −0.37 to 0.34; P=.91).

Correlations between task performance metrics from the n-back and overall means from CANTAB cognitive assessments and self-report questionnaires were explored (Table 3). Participants with better performance on CANTAB showed higher intercept and better mean performance on the n-back. Depressive symptoms assessed with the PHQ-9 correlated with mean dprime, and correlations with dprime intercept approached but did not reach statistical significance (P=.06). No significant correlations were seen with PDQ-D or UCLA-LS, or dprime slope.

The 3 mood items showed overall good reliability indices, supporting the combined use of the 3 question items. Between-person reliabilities were high (R=0.97 averaged over time and with time nested within individuals), and within-person generalizability was moderate to high (R=0.75 for within-person variation with time nested within individuals).

Descriptive data for total mood are presented in Table 2. Multilevel analysis of total mood by study day confirmed the best fit for a linear model (Bayesian information criterion=73.38; likelihood ratio=6.14; P=.01). This model showed a modest overall linear improvement in mood over the course of the study (estimate of fixed effect of study day on mood=−0.0026, P=.01). However, there was a great deal of heterogeneity on mood trajectories over the study duration, as shown in model fits for each study participant in Figure 2.

Mean overall scores from daily mood assessments were correlated with full-length self-report questionnaires, showing moderate correlations (Table 4). Self-reported depression (PHQ-9) and cognitive symptoms (PDQ-D) correlated more highly with daily mood assessments than self-reported loneliness as measured by the UCLA-LS.

Significant correlations between dprime mean and intercept were seen for total mood scores, for question items assessing lack of interest, and for low mood (Table 4). Correlations between n-back performance and daily reported cognitive symptoms and all correlations with dprime slope were nonsignificant (P=.12-.79). Examining the relationship between daily mood assessment and CANTAB measures, SWM between errors and strategy showed moderate correlations with daily reported mood, whereas correlations with RVP outcome measures were nonsignificant.

Descriptive statistics for step counts and heart rate are presented in Table 2. A moderate correlation was seen between step counts registered on the 2 devices (rho=0.61; 95% CI 0.57-0.65; P<.001), but there were also instances of marked discrepancy (Figure 3). Overall, the Apple Watch provided a higher step count estimate than the iPhone. Measurement issues were noted for heart rate using the Apple Watch, with individual heart rates registered including a minimum of 22 beats per minute, which was not biologically plausible.

This study demonstrated the feasibility of daily assessments of cognition and mood in mild-to-moderate MDD. The study spanned 6 weeks, corresponding to the time during which response to antidepressant pharmacotherapy efficacy would expect to be demonstrated, indicating that high levels of adherence can be achieved and retained over this time frame.

Exploratory analyses examined the relationship between high-frequency mood and cognitive assessment and validated full-length cognitive assessments and questionnaires. These analyses aimed to establish the degree to which brief frequent assessments capture similar information to validated cognitive assessments and rating scales. Daily mood assessments showed moderate to strong correlations with validated self-report questionnaires of depression, cognitive problems, and loneliness. Correlations were highest for the PHQ-9, a scale designed as both a diagnostic instrument and a severity measure [34], which also showed the highest item overlap with high-frequency assessments. Daily n-back performance correlated moderately with performance on standardized tests of working memory and sustained attention. Findings support the concurrent validity of the measures examined during daily assessments.

Adherence, defined as engaging with cognitive and mood assessments at least once daily, was very high (95%-96%), did not deteriorate over time, and was not predicted by depressive symptoms or cognitive function at study onset. These adherence rates, as well as the overall rate of responding to high-frequency assessments in the current study (≈87% for all possible cognitive assessments), are in keeping with previous compliance rates reported in high-frequency assessments in psychopharmacology, around 50% to 90% [45]. However, it is notable that although this study was significantly longer in duration than most previous high-frequency assessment studies, spanning 6 weeks rather than the typical 1- to 2-week duration, the daily frequency of assessment was lower, with most other studies typically sampling 5 to 10 times per day [45]. Previous studies in patients with mood disorders have shown good overall feasibility and acceptability of high-frequency assessments, although there is likely to be an interaction between protocol burden and burden of illness [46]. The brevity of the current protocol in conjunction with the proximity to wearable assessments may have helped to support the high levels of compliance seen here.

Participants reported that completing assessments was easier when study sessions fit into their daily routines, and that periods of high and low mood affected their motivation to complete assessments. Adherence was also affected by technical problems for some participants, and by forgetting to wear the Apple Watch because of low mood or bereavement. Study center support and reminders during nonadherent periods provided a framework to enable participants to maintain a high level of engagement with the study.

Participants’ performance on the n-back improved over time. Overall, mood symptoms showed a modest concurrent improvement, albeit with great heterogeneity in the trajectories observed over the assessment period. Participants were stabilized on monotherapy at the time of assessment, and many had started their current treatment many months before study participation (9.9 months on average). Improvements on the n-back, therefore, likely reflect the influence of practice effects and task specialization. Participants reported continued improvement in task performance as a motivator for engagement. This finding is supported by studies exploring gamification of tasks, where the use of game design elements (eg, points and scoreboards) can improve motivation [47,48].

Importantly, very few participants reached and maintained ceiling levels of performance on the n-back. The symbols presented were designed to be hard to name, and each testing occasion drew 9 items from a stimulus pool of 227 items. Almost all participants felt that the task was challenging yet achievable. Attainability encouraged them to set personal goals to improve or maintain their scores, indicating that striking a balance between difficulty and attainability can promote engagement [49].

Individual learning rates for each participant were reflected in their n-back slope, which did not correlate significantly with either CANTAB cognitive test measures or self-reported mood. This suggests that the capacity to improve performance is not directly affected by either depressive symptoms or cognitive impairment, consistent with research in a previous study showing that practice effects in cognitive tasks were not moderated by depressive symptomatology [50].

The n-back paradigm is commonly used alongside functional neuroimaging, where it activates a network of frontoparietal areas [51]. Research suggests that n-back is not simply a measure of working memory capacity but depends on functions such as updating, inhibition, and attention [52]. Consistent with this suggestion, n-back mean and intercept correlated with full-length CANTAB cognitive tests of attention and working memory, supporting the use of n-back performance as a sensitive but nonspecific marker of cognitive function.

The trajectory of moods reported by patients during the course of the study was highly heterogeneous, showing no clear relationship with change in cognitive performance (Figure 2). However, we observed a significant association between aggregate daily mood measures with cognitive measures from CANTAB and n-back task performance (mean and intercept).

The relationship between self-report questionnaires and high-frequency assessments of symptoms has been examined in a number of clinical conditions. Although in some cases the correlations are good [53], there can be a mismatch, with questionnaires relying on retrospective recall tending to overstate the severity and frequency of symptoms [54]. Retrospective recall shows distortion in favor of more salient or unique events at the expense of the more mundane [54], and depression is associated with negative biases in recollection during periods of low mood [55]. High-frequency assessment may be particularly useful in patients with MDD for ensuring accurate recording of the course of their illness and treatment response.

In this study, correlations between daily measures and validated self-report questionnaires were moderate to high. Discrepancies between objectively and subjectively assessed cognitive function have been reported before, with the latter being affected by depressed mood [15,24,25]. Our results confirm this association. PDQ-D scores were correlated with daily mood assessments but not with cognitive performance, indicating that self-reported cognitive function cannot substitute for objective assessments.

As our study focused on patients with mild-to-moderate MDD who volunteered for participation, it is unclear whether results would generalize to patients with different severity or to those who are less motivated. In addition, assessment using a small touch screen may not be feasible for patients with visual impairments or those requiring a larger typeface.

Step counts collected via the Apple Watch and the iPhone were discrepant, which could be accounted for by differences in wearing patterns but undermines the reliability of activity data from either device. Measurement issues with heart rate data may reflect that the equipment was not of medical grade, or occasions when the Apple Watch was not fitted sufficiently tightly for reliable measures to be obtained. Although the wearable nature and ease of use of the technology allow for data to be collected over longer periods of time, our findings indicate that caution is required when this equipment is used to examine heart rate in scientific research. Variable accuracy for wrist-worn heart rate monitors, including the Apple Watch, compared with electrocardiogram measurement has also been noted previously in brief comparisons of bouts of exercise [56].

This study supports the feasibility and validity of high-frequency assessment on wearable devices to assess cognitive function and mood in patients with MDD. The study spanned 6 weeks, indicating that high levels of adherence can be achieved and retained over this time frame. Our study suggests that these methods can be used to monitor cognitive function and mood symptoms after the initiation of treatment for depression.