Poor outcomes for patients with serious mental illnesses (SMI) are associated with various factors involving patients, health care providers, and the health care system, in addition to the delivery of inadequate medical services [1-3]. The current clinical ways used to track the progression of SMI may be insufficient or inadequate. Traditionally, health care professionals rely on in-person consultations for assessment and diagnosis [4,5]. However, these clinic-based services come with a variety of logistical difficulties [4], especially for people from impoverished communities [6]. To monitor patients, they must frequently visit a clinical facility, often within restricted operating hours, which can be particularly challenging for individuals dealing with SMI [4]. Additionally, these approaches are demanding in terms of resources because they necessitate one-on-one interactions with a trained health care provider, making widespread adoption difficult. The inherent issues related to accessibility and scalability in these methods create significant obstacles to patient care [7,8].

Remotely collected digital markers may help overcome key challenges in monitoring SMI, such as the reliance on traveling for monitoring and in-person assessments [9,10]. Continuous monitoring of behavioral and physiological data using personal devices has the potential to aid in the identification of personalized indicators for the initiation of illness [4], subsequently enabling the development of tailored treatment strategies [9,11,12]. Smartphones can provide a solution.

Smartphones can be used to collect data to understand mental health status. This technique, called digital phenotyping (DP) [13], is already in use in some behavioral problems and mental disorders, such as posttraumatic stress disorder (PTSD) [14], and SMI such as schizophrenia [15-17] and mood disorders [18-20] to predict relapse, symptom exacerbation, or mood fluctuations [21].

Digital phenotypes are collected either actively or passively [5,22]. Active data collection requires the individual’s active participation through active engagement [23]. This would include self-reported thoughts and symptoms through surveys or scales, whereas passive data collection does not require an individual to perform any specific action outside their regular activities [24]. Passively collected data, such as screen behavior, including screen time, locking or unlocking events, and screen lighting has been shown to reflect mental health states in adults, offering insights into phases of agitation, depression, and variations in sleep quality and patterns [10,23,24]. Accelerometer data from smartphones has also been linked to mental health concerns such as depression and anxiety due to reduced activity, such as mobility and traveling, while audio data, such as read-aloud recordings, help indicate depressive states [23,25]. Within DP, sleep-related features, such as sleep duration, variability, latency, and nighttime phone use, are increasingly recognized as important phenotypes that can reflect underlying mental states. Disturbed sleep, as captured through these DPs, has been linked to elevated stress and SMI [26]. In bipolar disorder, disruptions in sleep patterns can precipitate manic or depressive episodes and persist even during remission, affecting overall functioning [27-29]. Major depressive disorder (MDD) often features insomnia or hypersomnia, with sleep issues potentially preceding depressive episodes and exacerbating their severity [30]. In schizophrenia, sleep abnormalities frequently emerge before psychotic symptoms and are linked to more severe clinical outcomes [30]. Geolocation data offer insights into individuals’ mobility patterns and daily routines [31]. In schizophrenia, reduced GPS-derived mobility, such as spending more time at home and traveling shorter distances, is associated with increased severity of negative symptoms such as social withdrawal and diminished motivation [32]. Similarly, in bipolar disorder, fluctuations in geolocation patterns can reflect mood episodes, with decreased movement often observed during depressive phases [33].

DP has gained significant traction in high-income countries (HICs), where studies have demonstrated its utility in detecting early warning signs (EWS), monitoring symptom fluctuations, and predicting relapse in individuals with SMIs [22,34-37]. For example, in a US-based study, passive sensing data, including GPS location, accelerometer activity, screen use, and call or text logs, was used to identify EWS of psychotic relapse. Machine learning (ML) techniques were applied to detect behavioral anomalies preceding relapse events, demonstrating the potential of DPs for relapse prediction in schizophrenia [22]. A narrative review demonstrated how various digital features, such as geolocation, communication patterns, and activity levels, have been effectively used in schizophrenia research to understand behavioral changes, and studies included were conducted in HICs [16]. In mood disorders, passive data from smartphones and wearable devices, such as GPS location, accelerometer activity, sleep pattern data, and heart rate, were used to estimate depressive and anxiety symptom severity [18,19].

Together, these passively collected DP, such as sleep, mobility, screen use, and activity patterns, offer a continuous, unobtrusive window into individuals’ lived experiences. Their integration into mental health monitoring holds promise for early detection of symptom changes, relapse prediction, and timelier, personalized interventions in the care of individuals with SMI.

Relapse is a significant concern in the management of SMIs. Relapse can lead to adverse outcomes, including poor psychosocial functioning, increased caregiver burden, and higher health care costs [38-40]. Therefore, continuous monitoring and adherence to treatment are crucial in mitigating the risk of relapse and improving long-term outcomes for individuals with SMIs. Most patients go through a period with changes in behavior, which precedes their psychotic relapse, commonly known as EWS [41]. EWS includes changes in sleep patterns, hallucinations, delusions, hostile behavior, cognitive decline, depression, and paranoia [42]. Developing systems that can identify and monitor these EWS could help clinicians intervene early [41]. In a study, anomalies in passive data, such as geolocation, accelerometer readings, and screen state, were 2.12 times more frequent in the month preceding a relapse compared to nonrelapse periods [40,43]. In another study with participants with schizophrenia, the analysis revealed that anomalies in passive data (geolocation, accelerometer data, and screen state) were significantly more frequent before a relapse [43]. Importantly, models incorporating passive data outperformed those relying solely on active survey data in predicting relapse [43]. These findings show the potential of passive smartphone data to serve as early digital markers of relapse, offering a scalable and nonintrusive method for the timely intervention and continuous monitoring in individuals with SMI.

As described above, most of the DP research has been conducted in HICs, raising concerns about generalizability to low- and middle-income countries (LMICs) [44], where digital usage patterns, social determinants of health, and mental health care infrastructure differ substantially [42]. This motivates the need to evaluate whether DP approaches can be adapted and applied effectively in LMIC settings such as Bangladesh, where the burden of untreated SMIs is high and innovative, low-cost monitoring solutions are urgently needed [45,46].

Mental and behavioral disorders account for 12% of the global disease burden, with over 70% of this impact affecting LMICs [47]. In Bangladesh, the treatment gap (ie, the gap between people who need care and the people who obtain care) is over 92% [47], which means that less than 1 in 10 people get the mental health care they need, and the gap is more evident in slums [6,45,47]. People living in slum communities have high rates of SMIs, limited access to mental health services, and conditions of chronic hardship [48]. Therefore, a low-cost, easy-to-use solution to support the identification and monitoring of mental disorders is needed, and DP may provide a solution. DP has paved its way in mental health research [9-11], but most of the research conducted has been in high-income or developed countries; these do not account for findings in LMIC settings. Yet, 80% of people with mental disorders live in LMICs [42]. Social factors such as sudden changes in life, urbanization, and poverty often lead to a high burden of mental illness in several LMICs [42,49-51]. Additionally, there is a significant lack of awareness regarding mental health conditions among LMICs [49]. Slum dwellers here have a high burden of mental health conditions, which remains unnoticed due to a lack of help-seeking or lack of resources [48,49]. The addition of DP will enhance the detection of mental disorders among these communities with little to no cost, as the use of smart devices such as smartphones is common among slum dwellers in Dhaka.

The primary outcome for prediction modeling will be the occurrence of a relapse event within a 6-month follow-up period defined using prespecified, quantitative, diagnosis-specific criteria based on validated clinical scales, supplemented by structured clinician adjudication. For participants with psychotic disorders, relapse will be defined as either (1) a ≥25% increase from baseline on the Positive and Negative Syndrome Scale (PANSS), sustained for at least 7 days, or (2) deterioration in global functioning as indicated by a decline of ≥10 points on the Global Assessment of Functioning (GAF), confirmed by clinician rating. These thresholds align with relapse definitions used in prior studies in psychosis, where relapse was operationalized using scale-based symptom worsening and functional decline [17,22,43]. For participants with nonpsychotic disorders, including MDD and PTSD, relapse will be defined as clinically significant symptom worsening on the Brief Symptom Inventory (BSI), operationalized as a ≥30% increase from baseline on the relevant symptom subscale and/or meeting caseness on the BSI Global Severity Index (GSI), defined as a total GSI T-score ≥63 together with an increase of at least 7 points from the mean of the 2 preceding assessment scores, indicating a move into or further into the clinically significant distress range [52]. In repeated-measurement and longitudinal studies involving patients with mood, anxiety, and trauma-related disorders, changes in BSI symptom subscales and the GSI have been widely used to capture clinically meaningful worsening over time [52-55]. In these populations, sustained increases from an individual’s own baseline, rather than isolated score fluctuations, are commonly interpreted as reflecting true symptom deterioration. To reduce the risk of misclassifying temporary or situational symptom changes as relapse, clinically meaningful worsening is typically operationalized using relative change thresholds or standardized deviations from baseline sustained across multiple assessments.

Across all diagnostic groups, any psychiatric hospital admission or emergency psychiatric presentation during follow-up will be classified as a relapse event. All algorithm-identified relapse events will undergo blind clinician adjudication by an independent clinician using only symptom scale trajectories and clinical information; the clinician will be blinded to all DP features. Clinician judgment may confirm or refute events meeting quantitative thresholds but will not introduce relapse events in the absence of predefined criteria, thereby preventing post hoc outcome redefinition. Each relapse will be timestamped at the earliest point at which the criteria are met, enabling prediction within predefined risk horizons (eg, relapse within 30 d).

This study is a 6-month prospective cohort study designed to explore the feasibility and reliability of identifying DP to predict relapse of SMIs among slum residents in Dhaka, Bangladesh. This study is part of the broader NIHR (National Institute for Health and Care Research) initiative, Transforming Access to Care for Serious Mental Disorders in Slums (TRANSFORM project) [47], which provides the overarching framework. The TRANSFORM project is dedicated to improving access to mental health care for individuals with SMIs residing in urban slum environments [47]. The TRANSFORM study aims to bridge the treatment gap by developing, implementing, and assessing innovative community-based interventions [47].

Participants will be recruited from the Korail slum, one of the largest and most densely populated slums in Dhaka [47]. It is situated between 2 affluent areas and is home to an estimated 200,000 people living in densely populated conditions with limited access to basic services [47,51,56,57]. The community consists of a mix of long-term residents and recent migrants from rural areas, and many inhabitants face precarious employment, housing instability, and health inequities [56,57]. Mental health needs are high, yet access to formal mental health services remains scarce [47].

Participants will be selected using purposive sampling. Participants will be identified through the TRANSFORM project [47], where community engagement representatives are well known and respected in the area. A community-based approach will be used to identify and enroll participants, leveraging existing partnerships with local stakeholders. We collaborated with community health workers, nongovernmental organizations, and TRANSFORM [47] community representatives, who had well-established relationships with the residents. These trusted individuals facilitated introductions and provided culturally appropriate explanations of this study to potential participants. Importantly, recruitment will not be limited to individuals already enrolled in TRANSFORM. Community representatives will help identify potential participants, introduce this study, provide culturally appropriate explanations, and facilitate trust with potential participants. Participants will progress through clearly defined stages: approached and screened for eligibility, enrolled if inclusion criteria are met, followed up monthly with both active and passive data collection, and retained in the analytic sample provided sufficient data completeness is achieved. Numbers at each stage will be reported to ensure transparency.

This study will recruit participants diagnosed with SMIs, including MDD, psychotic disorders (such as schizophrenia and bipolar disorder), and PTSD, as per the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition) [58]. These conditions were selected due to their high burden in low-resource settings, their recurrent nature, and the critical need for early detection of symptom exacerbation [42,59-61]. Inclusion criteria require participants to be at least 18 years of age, residents of the Korail slum, diagnosed with SMI, and in possession of an Android smartphone. Exclusion criteria include those younger than 18 years of age, those without a smartphone, and individuals diagnosed with substance-induced psychosis or mental illness.

Participants will be informed about this study through participant information sheets (PIS) and will be given 48 hours to provide written informed consent. The PIS will be provided in Bengali, the local language, and will clearly explain the purpose of this study, the type of data to be collected, what participation involves, risks and benefits, data security, and the right to withdraw at any time without affecting access to services. Participants will be contacted through community engagement representatives and ongoing recruitment activities under the TRANSFORM project [47]. These representatives, who have established relationships within the community, will distribute the PIS and introduce this study to potential participants. Trained research assistants will obtain written informed consent at the local field office with a witness present from participants willing to take part in this study. This will be done in a private setting to ensure confidentiality, and the information sheet will be explained in Bengali. After obtaining consent, the DataDoc app will be installed on participants’ smartphones to passively collect data. The research team will assist them in downloading the app, granting required permissions (eg, access to device usage), setting up the app to ensure proper functionality, and troubleshooting any technical issues. Participants will be trained on app installation and basic use. Monthly clinical assessments will be conducted either in-person or via telephone, while passive data will be captured continuously in the background. Technical support will be available through community fieldworkers to address issues such as app functioning, battery management, or data sync.

Participants will install DataDoc on their smartphones for data collection as outlined below. This data will be collected via the app automatically and passively once the user gives the app permissions. A link to the app will be sent to participants via email or text (based on their preference). This is an .apk file which they can install on their Android phone. We will support participants to do this (eg, in person in the Korail field office) if needed. Upon downloading the app, they will be asked to give the app permission to access their device usage. Please note, the app also has the potential to connect to music and health data, for example, music data (ie, permission to access their Spotify [Spotify AB] data) and health data (ie, permission to access their Google Health app data), but we will only be using app usage for this study.

Relapse events will be identified using a combined approach. First, thresholds on validated scales (active data) will trigger a potential relapse flag. These events will then be reviewed by 2 independent clinicians and a researcher with expertise in psychiatry and psychology. The researcher will be blinded to the sensor-based DP data and will adjudicate relapse status using only clinical and scale-based information. In case of disagreement, a third senior clinician will review the case to reach a consensus.

We have developed an app, DataDoc (Figure 1), to collect DP data, which we will use for data collection in this study. Some active data will be collected directly through the DataDoc app, while assessments requiring clinician input will be obtained separately. After giving permissions, the app will sync this data, and then the participant will be presented with a screen asking them to complete the questionnaires outlined below in the active data section. Participants will be asked to complete the questionnaire via the app only. The questionnaire must be completed within 2 weeks of syncing the device, music, and health data. Participants will be called with reminders if they have not completed the questionnaire after 1 week and again after 2 weeks. At this point, we will ask them to resync their data and complete the questionnaire, so all their data is synced to the same time point.

Participants will complete a set of baseline assessments and assessments at intervals as shown in Table 1.

As illustrated in Figure 2, data is first collected on the participant’s phone using the DataDoc app. Data are then securely uploaded to a General Data Protection Regulation–compliant cloud server, where they are stored in encrypted form at rest. From there, encrypted anonymized data are transferred to the University of Warwick’s secure server for long-term storage and analysis, in line with Warwick’s 10-year research data policy. No personally identifiable information (eg, phone numbers and message content) is collected at any stage. Access to both the cloud server and Warwick’s secure server is restricted to authorized study personnel using password-protected accounts.

This study was approved by the Biomedical and Scientific Research Ethics Committee at the University of Warwick (reference number Biomedical and Scientific Research Ethics Committee 98/23‐24). Data collected from participants are securely stored and will not be available for sharing due to privacy considerations. The smartphone app (DataDoc) used for data collection and the feature processing code are available upon request for research purposes.

Participants were informed about the nature, procedures, potential risks, benefits, and voluntary nature of the research before enrollment. Informed consent was obtained from all individual participants before data collection. Participants received detailed information sheets and provided written consent confirming their voluntary participation and understanding of their rights, including the right to withdraw at any time without penalty. The consent process ensured comprehension of this study’s aims, data collection procedures, privacy protections, and how data would be used and stored.

To recognize participants’ time and engagement in this study, compensation was provided at a rate of 1200 BDT per month for the duration of their involvement. This compensation was disclosed to participants during the consent process and was approved as part of the ethics review.

All data collected from participants are securely stored in accordance with local and international data protection standards and are not publicly available due to privacy and confidentiality considerations. Identifiable personal information has been protected and anonymized where possible to prevent unintended disclosure.

The results of this study will offer critical insights into the application of DP for predicting the relapse of SMIs in low-resource settings, specifically among slum residents in Dhaka, Bangladesh. By using passive data collected from participants’ smartphones, such as device interactions, sleep patterns, movement, and social engagement, this study aims to assess the ability of ML models to detect early signs of relapse. Our findings will contribute to the broader understanding of how DP can be applied in real-world settings to improve mental health care, especially in LMICs where access to traditional mental health services is difficult or limited [42,49].

This study’s innovative approach of combining passive data with clinical assessments allows for the development of personalized, dynamic interventions tailored to individuals’ behavioral patterns. This could represent a significant advance over static mental health monitoring methods, offering more timely and effective interventions. However, several challenges and limitations must be acknowledged. First, while DP offers a novel way to monitor behavioral changes, the accuracy and validity of these markers in predicting clinical outcomes require further exploration. The reliance on self-reported clinical assessments may introduce bias [4], as participants’ perceptions of their mental health may not always align with objective symptoms. Moreover, such assessments often lack ecological validity, as they are typically conducted in clinical settings and may not capture the nuances of daily life. In contrast, DP offers the potential to monitor individuals in real-world contexts, providing richer insights into behavior and symptom fluctuations outside the clinic [12,13]. Furthermore, privacy concerns surrounding the continuous collection of personal data must be addressed [12], even though strict data protection protocols are in place.

The results will also have implications beyond the scope of this study. The potential for DP biomarkers to serve as dynamic tailoring variables in adaptive interventions could reshape how mental health treatments are delivered, particularly in resource-poor settings. As smartphones become increasingly prevalent in LMICs, leveraging these devices for mental health care could provide a scalable, cost-effective solution that reaches underserved populations. The findings from this study will help guide future research and intervention strategies, paving the way for larger, more comprehensive studies in similar settings.

A key limitation of this study is the requirement for participants to own an Android smartphone, which may systematically exclude individuals with SMIs who are older, poorer, or have lower educational attainment. These subgroups may be among those most vulnerable to relapse, meaning our sample could underrepresent the most at-risk individuals [84,85]. To mitigate bias, we will document sociodemographic and clinical characteristics of enrolled participants and compare these with available data on the broader Korail slum population. In addition, subgroup analyses stratified by socioeconomic and demographic variables will be conducted to evaluate whether predictive performance varies across groups. These steps will help us interpret findings in light of potential exclusion and inform future efforts to reach harder-to-engage individuals with SMIs.

Finally, our sample size estimation was based on the national prevalence of mental disorders in Bangladesh (17%). We acknowledge that this may not accurately capture the epidemiology of SMI in slum settings, where prevalence may be higher or differently distributed due to socioeconomic disadvantage, overcrowding, and limited access to care. As such, our study may be under- or overpowered relative to the true prevalence in Korail. This limitation will be transparently acknowledged in reporting and highlights the need for more robust epidemiological studies of SMI within urban slum populations.

Several practical limitations may affect feasibility. First, requiring participants to install the app as an .apk file may present challenges for individuals with low technical literacy. To address this, in-person installation support and illustrated step-by-step instructions in Bengali will be provided. Second, continuous passive data collection, particularly from GPS sensors, has the potential to increase battery consumption. Given that there is potentially unstable access to electricity in slum environments, participants may disable or uninstall the app to conserve power. While the DataDoc app has been optimized for low-frequency sampling to reduce power demand, battery drain remains a possible barrier to adherence. Furthermore, to minimize battery drainage in a setting with intermittent electricity, the app will upload data only when connected to Wi-Fi or charging [86]. Third, concerns about data privacy and security may limit participant engagement. Although strict protocols are in place, including anonymization and secure server storage, participants may still perceive privacy risks, which could affect long-term acceptability. To help mitigate this, a prior qualitative study was conducted with the same Korail community to explore participants’ understanding, perspectives, and willingness to share digital data [87]. Insights from this study informed the design of our consent process and data collection strategy. Nevertheless, residual concerns about privacy remain a possible barrier. We will therefore monitor such issues throughout this study and systematically document dropout or data loss attributable to these barriers.

This study represents a step forward in understanding how DP and ML can be used to predict relapse in individuals with serious mental disorders living in low-resource settings. By combining passive smartphone data with clinical assessments, we aim to build a predictive model that can identify the EWS of relapse, allowing for timely interventions.

While there are challenges associated with the use of DP, such as data privacy and the interpretation of behavioral markers, the potential benefits for early detection and intervention in mental health are substantial. The findings from this study will not only contribute to the growing body of research on DP but also offer practical solutions for addressing the mental health treatment gap in underserved populations. Ultimately, this research could help improve the quality of mental health care in LMIC settings and impoverished communities and provide a model for scalable, technology-driven interventions in other low-resource environments. This research has the potential to inform both local and global strategies for improving access to mental health care in low-resource settings.