While the working population is the driving force of the global economy, common mental disorders (ie, depression and anxiety) alone have cost US $1 trillion each year due to the loss of productivity [1,2]. Pharmacotherapies, psychotherapies, or a combination of both are the primary approaches in managing the symptoms of common mental disorders [3-5]. However, suboptimal service use exists to hinder the influence of the mental health care system [6,7]. The poor mental health service use suggested an immense service gap that might not sufficiently or appropriately support the needs of individuals with mental health difficulties. Against this backdrop, it has become essential to further explore accessible, affordable, and effective alternatives to existing mental health care services for the working population.

Longitudinal studies have demonstrated that work-related stress—such as excessive workload, effort-reward imbalance, and low work social support—is associated with increased risks of depression and anxiety disorders and poor mental well-being [8-11]. Despite a large investment in mental health services and their significant effectiveness, the cost of maintaining these face-to-face interventions is high and often challenging, steering the research direction towards identifying low-cost, evidence-based strategies to optimize intervention cost-efficiency [12-18]. Worse yet, the COVID-19 pandemic has further threatened the mental health landscape of working adults by impacting individuals’ job security, chronic isolation, and uncertainty about the future, as well as adding constraints to traditional face-to-face intervention options [19]. Most of the existing interventions are limited in reaching out to those in need during a pandemic situation, increasing the risk and prevalence of mental ill health [19-22]. This limitation also applies to long-existing underuse issues of mental health services, especially in Asian countries [23,24]. For illustration, previous research in China revealed that only 24% of the participants who were diagnosed with moderate or severe mental health problems sought help from mental health professionals [25]. Reasons included issues such as health service accessibility and costs, as well as associated mental health stigmas. Therefore, mobile health (mHealth) seems a practical approach to improve the low rate of health-seeking behaviors given its advantages of being low cost, easily accessible, and stigma-free. Therefore, attention has increasingly shifted toward digital approaches as a feasible and scalable alternative.

A recent randomized controlled trial (RCT) showed that web-based mindfulness and cognitive behavioral training were efficacious in alleviating psychological distress among college students and young working adults [26]. Several meta-analyses suggested that web-, application-, and/or computer-based interventions were effective in alleviating depression, anxiety, and stress, though with undeniably concerning attrition rates ranging widely from 43% to 99% [27-31]. Taken together, these findings pointed out the effectiveness of internet-based psychological interventions while urging further refinement of their user retention especially for long-term benefits. In addition to internet-based interventions, mHealth platforms that adopt mobile and wireless devices to improve health care also appear promising [32]. Not only has mHealth long been used in physical health care, such as for diabetes management, infectious disease management, and immunization management, increasing application of mHealth to mental health also has emerged, ranging from alcohol and substance use assessment to low-intensity interventions for mental disorders [33-38]. Despite the enlightening trend, less than 5% of the mHealth platforms underwent peer-reviewed research, while the majority of these were conducted in western countries [39,40]. Peer-reviewed, demographically diverse research awaits attention.

Advancements, availability, and affordability of mobile technology have given rise to ecological momentary assessment (EMA), which allows individuals to report experiences and behaviors in everyday lives and in naturalistic settings [41,42]. The rationale for using EMA rests upon 3 major advantages including reducing retrospective recall, improving generalizability and ecological validity, and enabling the exploration of temporal relationships among variables. Similarly, ecological momentary intervention (EMI) refers to the delivery of intervention while individuals are naturally engaged in their everyday lives [43,44]. Emerging evidence suggested the effectiveness of EMI in alleviating anxiety and stress and treating a variety of health behaviors and psychological symptoms [44,45]. However, RCTs of EMA and EMI in mental health promotion remained scarce [45]—and their application in workplace mental health services is even more limited. This is noteworthy given that both approaches could offer unique benefits to working adults, as they deliver mental health care services free from temporal and spatial constraints. Therefore, this gap underscores the need for rigorous investigation of EMA- and EMI-based mHealth tools, particularly in workplace populations and non-Western contexts.

To address this gap, this preregistered RCT was designed to evaluate the effectiveness of a smartphone-based intervention (ie, the Neurum app; Neurum Limited) that leverages both EMA and EMI in reducing workplace stress [46]. Study hypotheses and corresponding analytical schemes were preregistered in September 2020 (registration reference: osf.io/z829u) before the data collection was finished. No statistical analysis was done before the preregistration or the completion of data collection. Specifically, it was hypothesized that participants who received intervention through the Neurum app would display better workplace stress management and mental health improvement as reflected by the reduction in depression, anxiety, and/or stress symptoms. The design, conduct, and reporting of this study followed the CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) guidelines (Checklist 1).

This study was a 2 (between: intervention vs waitlist control groups) × 2 (within: premeasures vs postmeasures) mixed factorial design. A priori power analysis using F test family (repeated measures, between factors) in G*Power was conducted [47]. Assuming a small effect size (ie, f=0.25) and a high correlation between premeasures and postmeasures of the outcome variable (ie, r=.85), with an α level of .05, a total number of 196 participants was needed to achieve a power of 0.80. Assuming an attrition rate of 25%, a sample of 262 participants was determined prior to data collection.

Participant recruitment started in September 2019, Hong Kong SAR. Advertisements were posted on the company websites and populated through themed webinars and offline workshops to reach local working adults. However, due to local social unrest and the COVID-19 pandemic, few signed up within the planned 1-month recruitment period. Therefore, under all authors’ and the funding agency’s approval, the recruitment process was prolonged until the deemed sample size would be achieved. The recruitment process lasted for over a year and was finally terminated by the end of October 2020, when nearly no more participants signed up within 30 days. To maximize the sample size, those who signed up and passed the screening immediately underwent the intervention process. However, no data analysis was conducted before the end of the study. In total, 201 working adults were recruited from a selected list of local companies, who were first directed to an online survey with screening questions, in which they reported their demographic and mental health information. A research assistant who remained unaware of the study design was in charge of the eligibility screening. Specifically, participants were eligible if they (1) were aged 18 years or older at sample recruitment, (2) had access to the internet and a smartphone, and (3) were not diagnosed with major mental disorders. Exclusion criteria included (1) mental illness diagnosis history, (2) current reception of professional mental health services, or (3) current use of psychotropic medication.

Ethical approval was obtained from the Subcommittee on Research Ethics and Safety of the Research Committee of Lingnan University (reference: EC047/2021). All procedures contributing to this work complied with the ethical standards of the relevant national and institutional committees on human study. Informed consent was obtained via electronic signatures indicating the awareness of the study procedure, potential risks and benefits, data confidentiality, and the right to quit the study at any time. Data were encrypted and stored in a financial-sector security level server and deidentified before statistical analysis. No incentives or compensations were provided.

All procedures were delivered via Neurum. Included participants first completed the preintervention assessment, and then they were randomly allocated at a 1:1 ratio to the two groups via a computer algorithm of simple randomization. Specifically, a numerical value between 0 and 1 was assigned to each participant. Those who received a value smaller than 0.5 were assigned to the control group; otherwise, participants were assigned to the intervention group. Double blindness was achieved such that a research assistant who remained unaware of the study design was in charge of the participant assignment, while participants received a unique study ID in a singular format that revealed no group information, and therefore, remained unaware of the grouping. Then, participants in the intervention group received EMA and EMI via Neurum for 4 weeks, while those in the waitlist control group were offered the intervention after the RCT was completed. Finally, all participants completed a postintervention assessment with debrief at the end of the trial (ie, after the designated 4-week intervention [or control] duration).

Neurum is a smartphone app that comprises self-managed mental health modules designed in various combinations of (1) featured techniques, namely cognitive behavioral therapy, mindfulness exercises, and self-regulation skills with (2) various session formats including informational (ie, psychoeducation), evaluative (ie, assessment), and applicable (ie, exercise), aiming at (3) tangible goals in alignment with the biopsychosocial model, namely the improvement and maintenance of biological, psychological, and social health. See Table 1 for a summary of the different components and the rationale.

Neurum is designed with an instant messenger–like user interface where users interact with the virtual app assistant in a conversation format. The palatability of the dialogues was ensured by using layman and local terms in either English or Chinese. The app content was developed with standard app development iterations that involved digitizing of modules based on previously validated mental health education and behavioral changing paradigms, and then user testing with an A/B test method to compare the performance of 2 versions of content. This study describes the foundation of the app, which matches users to an EMI based on information from their EMA via a combination of supervised learning models, but future iterations of Neurum involved personalization of EMI modules based on nuances from EMA, session feedback, and other baseline user data via a more complex deep learning algorithm. As part of the safeguarding protocol, a referral mechanism is established, such that users who score past a certain threshold in their levels of psychological distress are automatically referred to crisis management protocol, which involves the algorithm providing the user with immediate support via distress reduction modules, crisis hotlines, and face-to-face care delivered by a licensed, on-board clinical psychologist. No adverse events were reported during the study.

EMA is delivered such that users are invited to complete a mood journal on Neurum to log their instant emotions, thoughts, and behaviors relevant to workplace stress as well as other psychological distress whenever they feel in need. Users evaluate their general feelings along a negative-to-positive continuum, specify a specific experience they would like to work on, and then identify one or more associated stressors. EMI is delivered such that, upon EMA logs, Neurum would make a recommendation to relevant, evidence-based intervention modules based on an immediate matching of the current demand and available training tools using a self-developed, built-in supervised machine learning algorithm. User feedback is further collected after each training session, which also informs the algorithm. A daily summary of the reported mood experiences and taken exercises was provided for self-monitoring of workplace stress and management performance. See Figure 1 for the overview of Neurum’s EMA or EMI workflow.

To examine the intervention effect, pre– and post–DASS-21 scores of the intervention and control groups were submitted to a 2 (between: intervention vs control groups) × 2 (within: premeasures vs postmeasures) mixed ANOVA. Additional analyses included categorical demographic characteristics (namely sex, education, and marital status; Table 2) as control variables. A supplementary set of mixed ANOVA was conducted to compare intensity level changes for each symptom. Finally, 2 multiple linear regressions with a number of mood journals and exercise sessions completed were conducted to explore any dose-response effect.

Figure 2 shows the participant flow diagram. Demographic characteristics are summarized in Table 2. The majority of study participants were females, had bachelor’s degrees, and were single in terms of marital status. Though the overall attrition rate was low (23/201, 11.4%), there were significantly more participants dropped from the intervention group (n=21) than from the control group (n=2; χ²₁=18.259; P<.001), resulting in 81 participants in the intervention group and 97 in the control group. Considering the intensive EMA or EMI procedures, the differential attrition rates between the 2 groups were not surprising. More importantly, the final sample consisted of 178 working adults and was still sufficient in statistical power.

On average, participants in the intervention group provided 9.74 (SD 18.2; maximum=130) mood journals and completed 6.27 (SD 9.4; maximum=51) exercises, while neither was correlated to preintervention DASS-21 scores (|r|s ≤.156; Ps≥.16), and neither was predicted by sex (t₇₉s≤0.744; Ps≥.46), education level (F_{3, 73}s≤1.947; Ps≥.13), or marital status (F_{2, 77}s≤2.002, Ps≥.14). Daily user engagement on average was 4.95 (SD 6.89; maximum=30) minutes, which was positively correlated with preintervention stress scores (r=0.233; P=.002) and total DASS-21 scores (r=0.175; P=.02), but was not predicted by sex (t₇₈=0.791; P=.22), educational level (F_{3, 73}=1.595; P=.20), or marital status (F_{2, 76}=0.100; P=.91).

Preintervention and postintervention DASS-21 scores, as well as the mean differences, are presented in Table 3. Mixed ANOVAs were applied to compare the pre– and post–DASS-21 total and subscale scores between groups (Table 4). For the total score, equal variance between groups was satisfied (Levene statistics ≤2.127; P≥.15). Results revealed that there was no significant difference between the intervention (mean difference [MD]=21.990) and the control (MD=21.852) groups; but as expected, there was a significant postintervention decrease (MD_{preintervention}=24.721 and MD_{postintervention}=19.121). More importantly, the above effects were quantified by a significant, large-effect size interaction between time and group (Figure 3A) such that the DASS-21 total scores decreased for the intervention group (MD_{preintervention}=29.111 and MD_{postintervention}=14.593) but increased for the control group (MD_{preintervention}=20.330 and MD_{postintervention}=23.649). Similar patterns (Figure 3B-D) were found for the three symptoms such that average scores notably decreased for the intervention group but slightly increased for the control group (F_{1, 176}s≥22.689, Ps<.001), with the strongest effect observed in the stress symptom.

A set of supplementary models found similar patterns controlling for sex, education, and marital status (for condition effects, F_{1, 146}s≤1.601, Ps≥.21, η²≤0.011; for time effects, F_{1, 146}s≥2.933, Ps≤.09, η²≥0.020; for interaction effects F_{1, 146}s≥19.286, Ps≤.001, η²≥0.117), suggesting that the above intervention effects held after controlling for these demographic covariates. Another set of supplementary models (Table 5) found similar patterns using DASS-21 symptom intensity levels as the outcome variables. Change(s) of these intensity levels suggested that the control group ascended while the intervention group descended in symptom severity (χ²₂s≥ 8.605; Ps≤.02). Finally, and surprisingly, a final set of supplementary models (Table 6) using multiple linear regressions revealed that none of the exercise session number, mood journal number, or average daily engagement predicted the postintervention DASS-21 scores.

This RCT examined the effectiveness of a smartphone-based EMA or EMI mental health care services delivered by the Neurum app in mitigating workplace stress and improving mental health for working adults in Hong Kong SAR, China. Consistent with the hypotheses, results demonstrated that the approach was effective in reducing psychological distress after 4 weeks of intervention, suggesting that the EMA or EMI approach via the Neurum app could support working adults’ stress management in real time and in naturalistic settings.

This study demonstrated the effectiveness of integrating EMA or EMI techniques with supervised machine learning algorithms in digital mental health care for a nonclinical population. The findings were broadly compatible with previous studies supporting the effectiveness of mHealth tools in reducing general anxiety, various stress, and depression symptoms [45,50]. Importantly, the effectiveness of Neurum was supported not only by the continuous reduction of psychological symptoms but also by the shifts in symptom intensity levels, suggesting potential use that extends from non–clinical contexts into early-stage clinical applications. One well-known advantage of the ecological momentary approach is its ability to minimize retrospective bias and enhance event trigger monitoring. Previous research has highlighted the promise of such digital phenotyping for routine clinical assessment, particularly for early onset or relapse [51]. Our observed changes in categorical symptom intensity levels further supported this clinical relevance, indicating that symptom improvement was substantial enough to cross the established cutoff thresholds. Because categorical severity shifts could provide a more interpretable indicator of treatment effectiveness than numerical score changes alone, these findings underscore the potential of EMA or EMI-based mHealth tools for early identification, timely intervention, and preventative health care in mental health. Additionally, previous meta-analyses research revealed that multicomponent intervention was tentatively more effective than single-component ones (eg, with cognitive behavioral therapy alone), and that patient preferences for specific intervention components played a pivotal role in treatment outcomes [52,53]. In response to these challenges, the intervention delivered in Neurum integrated multiple modules in various delivery formats for actionable outcomes, enabling participants to decide whether and which exercises to engage in upon each log-in. This inclusive design undoubtedly promoted user retention and engagement and maximized the alignment between individual needs and intervention content, a pattern supported by our user activity data. Taken together, these novel findings highlighted the importance of developing scalable, personalized mHealth tools as accessible, affordable, and flexible alternatives to face-to-face mental health care services.

There were several important issues in the interpretation of results. First of all, more participants dropped out of the intervention group than from the control group. It is noteworthy that the overall dropout did not exceed the predetermined attrition rate for power analysis. It is also noteworthy that the current attrition rate (ie, 11.4%) was smaller than those reported in previous studies (eg, about 26% for cognitive behavioral therapy and about 19% for mindfulness-based traditional interventions, or widely ranging from 43% to 99% for general digital health interventions [27-31,54-56]). Despite this, authors speculated the presence of some exposure effects particularly for the intervention group. Specifically, EMA or EMI required participants to review their mood swings on a daily basis, to exert efforts in managing especially those negative daily experiences, and to review their own performances on the taken exercises. Compared to those assigned to the control groups, these tasks might make the mental health issues more evident than they used to be (namely, before participating in this study) for the intervention group, possibly resulting in a withdrawal from treatment effect frequently observed in traditional psychotherapies [57]. Regrettably, the current design did not include relevant measures to test this speculation. However, to improve its future retention, the intervention could benefit from including methods to cope with the aforementioned exposure effect in the onboarding stage. Meanwhile, a personalized approach that maps personal needs with module recommendations can also improve user retention and performance. In fact, future iterations of Neurum have involved personalization of EMI modules based on nuances from EMA, session feedback, and other baseline user data via a more complex deep learning algorithm. Second, the results somewhat failed to support the association between the number of completed exercises or mood journals with DASS-21 score changes. In other words, the underlying dose-response effect remained unclear. One possible reason was that participants were only encouraged but not forced to report their daily mood journals or to take exercises. Granted that it strategically maximized participants’ acceptance and adherence [58], it undeniably resulted in the limitation of a relatively low daily response rate, and subsequently, the lack of statistical power to reveal the targeted dose-response effect, if any. Therefore, future studies might consider assigning a compulsory number of daily responses to unveil potential underlying mechanisms. Despite this, the current intervention effects could be a result of the improved awareness of one’s own mental health status and consequently improved proactive mental health management effort, or an effect generalized from in-app exercises to realistic life scenarios.

This study also contributes to the relatively nascent field of mHealth research, typically the application of EMA or EMI, in Asian countries [40]. Though with a rapid growth and a wide application, empirical supports of EMA or EMI for mental health were mostly from the United States and European countries (eg, Germany and the Netherlands), with the only exception being a study involving Chinese participants with mild-moderate depression [59-61]. However, western-based mHealth apps were not necessarily applicable to eastern ones [62]. The common mHealth apps in Asian contexts have remained in psychoeducation, counseling services, and self-assessment tools [40,63], while this study expanded them by leveraging the more intense EMA or EMI approach. Furthermore, Neurum takes health equity and cultural adaptivity into consideration by offering multiple language versions (viz, both traditional and simplified Chinese, English, Filipino, and Hindi) to serve local people with different cultural backgrounds. Though ethnicity was not collected in this investigation, the recruitment process was indeed carried out in international enterprises in Hong Kong, SAR. Finally, this study expanded the use of EMA or EMI beyond clinical settings to community populations, which broadened the spectrum of the recipients of such digital health care services. While some digital health tools can be used by patient populations to clinically reduce their depression and/or anxiety, this must be differentiated from digital health tools that are meant to reduce these symptoms in a non–clinical population. The demonstrated effectiveness of Neurum is similar to other well-being and mindfulness apps, such as Calm and Headspace [64-66]. Furthermore, the function of these digital health platforms and the reduction in mental health symptoms are similar to those of other non–clinical interventions like yoga [67]. The wellness technology industry has been growing exponentially with app-based options that are not designed to replace clinical treatment altogether, but to offer complementary early intervention tools, or as alternative options for those that struggle to access in-person care. Relatedly, Neurum essentially represents a shift from disease-centered to health-centered, proactive mental health approach. EMA or EMI, combined with diverse delivery formats (namely, informational, evaluative, and practical), comprehensively supports a proactive mental health management system by addressing foundational mental health literacy, enabling timely monitoring, fostering essential skills, and promoting sustained actions. Such a shift of medical models for mental health corroborates the notion of prevention over treatment, as globally advocated by many governmental and regional health policies [68-71].

This study has several limitations. First, despite the authors’ effort to recruit participants and maintain their engagement during the local social unrest and the COVID-19 pandemic, the final sample size was still slightly smaller than expected. However, the effect size of mixed ANOVA was large, supporting the effectiveness of the intervention. Second, even though Neurum was designed with multiple languages, ethnicity information was not collected, which precluded the possibility of examining cultural differences, if any. Factors such as mental health beliefs, perceptions, and literacy should merit future research attention to improve the generalizability of Neurum in other cultures [72,73]. Third, while the attrition rate of the current study was low, local RCTs on web-based and application-based mental health programs suggested that personalized feedback might mitigate the high attrition rates typically in internet-based programs [26,74]. Indeed, several lines of evidence demonstrated the benefits of personalized intervention including better intervention outcomes, stronger subject adherence, and more persuasive feedback, in both physical and mental health domains [62,75-78]. As such, Neurum could be further advanced by incorporating the personalized intervention approach based on computational algorithms of the EMA or EMI data. Finally, the results of this study did not reveal a dose-response relationship between the reductions in psychological symptoms and the number of EMI sessions completed. A requirement of minimum sessions in the future can help examine the potential dose-response effect and better understand the intervention’s effectiveness.

In sum, this RCT supported that Neurum, a new mHealth intervention that coupled the ecological momentary approach with supervised machine learning algorithms, could effectively reduce workplace stress and psychological distress in working adults. Beyond establishing its efficacy, the findings also highlight the potential of day-to-day, context-sensitive data to deepen the understanding of how EMA or EMI functions within digital psychological interventions. By introducing inclusive design and diverse user groups, this work further underscores the importance of equity and cultural relevance in the next generation of digital mental health care. Together, these insights highlight the promise of adaptive, data-driven mHealth tools in expanding the access to effective psychological support across different populations and in real-time and real-world settings.