This study is a prospective longitudinal observational study comprising two parallel cohorts: an MDD patient cohort (n=202) and a healthy comparison cohort (n=100). It focuses on patients with depression who are treated in outpatient settings; inpatients are excluded because their mobility and daily activities are substantially restricted and differ markedly from those of outpatients. Outpatients with depression and healthy participants are enrolled and followed over time. For the depression outpatient group, we observe the course of illness, including changes in symptom severity, the dynamic characteristics of symptom manifestations, and patterns of disease progression. For the healthy group, we collect digital behavioral and physiological data over the same 12-month period, with clinical assessments conducted only at baseline, to provide a comparative reference. In this study, all treatments, including medications and psychotherapy, are provided as part of routine clinical care at the discretion of the treating clinicians. The study does not introduce or modify any treatment interventions.

The study is being conducted in the psychiatric outpatient department of Peking University Sixth Hospital, a tertiary psychiatric hospital in Beijing, China.

For patients with depression, the inclusion criteria are as follows: (1) aged between 18 and 60 years; (2) meeting the diagnostic criteria for MDD according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), as assessed by the Mini-International Neuropsychiatric Interview (M.I.N.I. 7.0); (3) confirmed to be in a current depressive episode through the M.I.N.I. 7.0; (4) providing written informed consent; (5) having sufficient literacy and smartphone operation skills to complete the study assessments; and (6) a baseline Hamilton Depression Rating Scale-17 (HAMD-17) total score of 14 or higher (33). The exclusion criteria are as follows: (1) pregnant or lactating women; (2) severe or unstable physical illnesses (judged by the investigators to potentially interfere with participation in the study); (3) a history of neurological disorders (such as epilepsy, traumatic brain injury, multiple sclerosis, motor neuron disease, Parkinson’s disease, cerebellar ataxia); and (4) a diagnosis of other severe psychiatric disorders (excluding anxiety disorders) according to the M.I.N.I. 7.0 (34, 35), including organic mental disorders, substance use disorders, schizophrenia spectrum disorders, or bipolar disorder. For healthy individuals, the inclusion criteria are as follows: (1) aged between 18 and 60 years; (2) providing written informed consent; (3) having sufficient literacy and smartphone operation skills to complete the scale assessments; (4) a baseline HAMD-17 (36) total score of less than 14; and (5) using an Android-based smartphone. The exclusion criteria are as follows: (1) pregnant or lactating women; (2) severe or unstable physical illnesses (judged by the investigators to potentially interfere with participation in the study); (3) a personal history of psychiatric disorders or a first-degree family history of psychiatric disorders; (4) any current psychiatric disorder according to the M.I.N.I. 7.0, and (5) a history of traumatic brain injury or other neuropsychiatric disorders.

Participant recruitment was completed between August 2023 and February 2025, with 12-month follow-up assessments currently ongoing. Patients with MDD were recruited consecutively from the outpatient psychiatry clinic during routine clinical visits. Healthy participants were recruited concurrently from Beijing residents (with ≥1 year of local residence) through social media advertisements and community-based announcements. All participants must be able to remain in Beijing throughout the 12-month study period to ensure follow-up feasibility. Written informed consent is obtained from all participants prior to enrollment (Figure 1).

The study employed a longitudinal assessment protocol with seven evaluation timepoints: baseline (T0), and follow-up assessments at 2 weeks (T1), 4 weeks (T2), 8 weeks (T3), 6 months (T4), 9 months (T5), and 12 months (T6) post-enrollment. Patients with MDD completed clinical and psychometric assessments at all timepoints, while healthy participants completed assessments at baseline (T0) only. The temporal schedule of clinical assessments is detailed in Table 1. Data collection is expected to be completed by February 2026.

Data collection was organized into three main components based on data source and acquisition device: (1) clinician-administered clinical and psychometric assessments (Table 1); (2) Android smartphone–based digital assessments and passive sensing, including GPS-linked environmental exposure (Table 2); and (3) wearable monitoring using a Xiaomi smart band (Table 3).

Clinician-administered assessments were conducted during on-site visits at the psychiatric outpatient department. A self-designed demographic questionnaire was used to collect sociodemographic information, including age, gender, ethnicity, education level, marital status, employment status, dietary habits, and reproductive health history. A separate clinical characteristics registry documented body mass index (BMI), medication allergies, age of onset, date of diagnosis, number of prior hospitalizations, current psychotropic medications, tobacco and alcohol use, and physical and psychiatric comorbidities. Standardized clinical and psychometric instruments included the 17-item Hamilton Depression Rating Scale-17 (HAMD-17) (36), Hamilton Anxiety Rating Scale (HAMA) (37), Pittsburgh Sleep Quality Index (PSQI) (38), Biological Rhythms Interview of Assessment in Neuropsychiatry (BRIAN) (39, 40), International Physical Activity Questionnaire-Short Form (IPAQ-SF) (41), Sheehan Disability Scale (SDS) and Chinese Cognitive Assessment Scale (CBCT) (42). These measures were administered face to face by psychiatrists or trained research staff at baseline and scheduled follow-up visits. The timing of each clinician-administered assessment is detailed in Table 1.

Smartphone-based data were collected via a dedicated Android application. Only participants owning a compatible Android smartphone and consenting to digital data collection were enrolled in the smartphone-based component. Environmental exposure data were obtained by integrating the Seniverse Weather API (https://www.seniverse.com/api) with GPS-derived geographic coordinates from the participant’s smartphone. The following environmental variables were retrieved: air quality index (AQI), daily average temperature (°C), relative humidity (%RH), fine particulate matter (PM_2.5, μg/m³; PM₁₀, μg/m³), ozone (O₃, μg/m³), nitrogen dioxide (NO₂, μg/m³), and sulfur dioxide (SO₂, μg/m³). Passive sensing captured behavioral and contextual information from the Android device. App usage logs recorded application opening and closing events and usage duration. Screen activity logs included screen-on, unlock, lock, and screen-off events. Microphone data sampled 15-second audio segments every 10 minutes at a frequency of 10Hz, recording sound intensity (decibels, dB). Sensor data were obtained from built-in motion and orientation sensors of the Android smartphone (including accelerometers, gyroscopes, orientation sensors, magnetometers, gravity sensors, linear acceleration sensors, and light sensors), sampled in 15-second windows every 10 minutes at 10 Hz. Android smartphone–based active self-reports were used for high-frequency symptom monitoring between clinic visits. At predefined app-based assessment timepoints, participants completed the Patient Health Questionnaire-9 (PHQ-9) (43), Generalized Anxiety Disorder Scale-7 (GAD-7) (44), Patient Health Questionnaire-15 (PHQ-15) (45), Perceived Deficits Questionnaire-5 (PDQ-5) (46), and Insomnia Severity Index (ISI). For daily dynamic monitoring, participants were asked to complete a brief mood and symptom check once per day via the smartphone application. To maximize ecological validity and long-term adherence, no fixed assessment time was imposed; instead, participants received one daily reminder (8 am) and could choose a convenient moment during the day to respond. At each assessment, mood state was measured using a 0–10 Visual Analog Scale (VAS) (47) with the prompt “Right now, how is your mood?”(0 = very bad, 10 = very good). In addition, a modified momentary version of the two-item Patient Health Questionnaire depression screener (PHQ-2) (48) was administered, in which the original 2-week time frame was adapted to assess current state (e.g., “Right now, how much are you bothered by feeling down, depressed, or hopeless?”). All responses were time-stamped as part of the digital log.

Wearable device monitoring was conducted using Xiaomi Mi Band 7 Pro smart bands (Xiaomi Corporation, Beijing, China) for continuous biometric and behavioral tracking throughout the 12-month study period. All participants were provided with a study device at enrollment and instructed to wear it continuously, including during sleep, except when charging or during water-related activities that might compromise device function. The bands recorded four main categories of data. First, sleep parameters (49) included total sleep time (TST, hours), wake after sleep onset (WASO, minutes), sleep onset latency (minutes), wake-up time, bedtime, sleep efficiency (%), and sleep architecture (duration of light, deep, and REM sleep stages in minutes). These metrics were automatically derived by the device’s proprietary sleep-detection algorithm, which integrates triaxial accelerometry and photoplethysmography (PPG)-based heart rate signals to identify sleep-wake transitions and classify sleep stages. Second, cardiovascular and autonomic parameters comprised resting heart rate (beats per minute, measured during sleep and prolonged sedentary periods), continuous heart rate monitoring throughout wear time, and peripheral oxygen saturation (SpO₂, %, measured via a dual-wavelength optical sensor). Cardiovascular signals were sampled continuously or at regular intervals (approximately every 10 minutes) using the built-in PPG sensor, and minute-level summaries were exported via the Xiaomi Health mobile application. Third, physical activity and sedentary behavior (50) included daily step count, estimated energy expenditure (kcal), and sedentary events, defined as continuous sedentary periods exceeding 30 minutes as identified by sustained low movement on the accelerometer, with all metrics derived from triaxial accelerometer data processed using Xiaomi’s proprietary algorithms. Fourth, the device provided a proprietary stress index calculated from heart rate variability and heart rate patterns as an indicator of autonomic nervous system activity and perceived stress. Stress scores ranged from 0 to 100 and were categorized by the vendor as relaxed (0–29), mild (30–59), moderate (60–79), and high (80–100); daily stress scores and the distribution of time spent in each stress category were extracted from the Xiaomi Health app. All wearable data were synchronized to participants’ smartphones via Bluetooth and uploaded to the Xiaomi Health cloud platform, from which they were retrieved by the research team through authorized API access or manual export in standardized formats (e.g., CSV). Device wear time (hours per day) and the number of valid wear days (defined as ≥10 hours of continuous monitoring per day) were calculated as adherence and data-quality indicators and used in subsequent analyses. The Xiaomi Mi Band 7 Pro is equipped with a 3-axis accelerometer, a PPG optical heart rate sensor (green LED for heart rate; dual red/infrared LEDs for SpO₂), an ambient light sensor, and onboard data-processing capabilities, and provides a typical battery life of approximately 10–12 days, thereby minimizing charging-related gaps in data collection.

The study’s custom digital phenotyping application (MOOD app) was developed exclusively for Android devices due to technical and operating system–level constraints on data access in iOS. For patients with MDD, smartphone operating system was not used as an inclusion or exclusion criterion. Patients using Android devices (n = 114, 56.4%) contributed the full suite of digital data (passive smartphone sensing, EMA, GPS-based environmental exposures, and wearable monitoring), whereas patients using iOS devices (n = 88, 43.6%) contributed clinical assessments and wearable-derived data only. Healthy participants were required to use Android smartphones so that complete digital reference data could be obtained.

Quality control was ensured through an integrated, multimodal data collection and management system. A bidirectional data flow between participants’ devices and the central data management team supported continuous monitoring and feedback. Data were obtained through three main modalities: (1) Clinical assessments: Paper-based instruments administered by trained, certified clinicians during face-to-face evaluations, with subsequent dual independent data entry into electronic case report forms (eCRFs) using EpiData software (version 3.1), maintaining an error rate below 0.2%. (2) Physiological monitoring: Continuous physiological signals (heart rate, step count, sleep patterns, activity levels) collected via Xiaomi Mi Band 7 wristbands and automatically synchronized through the Xiaomi Health mobile application. (3) Digital behavioral data: EMA responses and passive smartphone usage patterns captured by a custom Android application developed by the Beijing Institute of Technology (BIT).

A structured, multi-tier storage approach ensured compliance with regulatory requirements and maintains scientific integrity. Real-time data collected via mobile devices——including EMA responses, smartphone sensor data, and wearable device metrics——were encrypted using AES-256 algorithms and stored on secure cloud servers during the active data collection phase. Upon study completion, the complete dataset will be transferred to secure, offline local servers at Peking University Sixth Hospital (PKUH6) for long-term storage and analysis. In accordance with Chinese national standards (GB/T series), all paper-based clinical health records were stored in access-controlled, locked cabinets and retained for a minimum of five years. Access to both digital and physical records was restricted to authorized research personnel only, with all data handling procedures logged and audited regularly.

Quality assurance is implemented across three distinct phases to ensure data reliability and scientific rigor. Pre-implementation phase: Protocol standardization was achieved through the development and implementation of a Research Operations Manual (version 2.3). All clinical assessors completed mandatory training and certification, achieving inter-rater reliability (Cohen’s κ ≥ 0.85) for MINI 7.0 diagnostic interviews. Wearable devices and smartphones underwent technical calibration to ensure measurement consistency across the study cohort. Active monitoring phase: Real-time compliance monitoring combined automated systems and human oversight. Participants received daily push notifications (8:00 AM) for the first 8 weeks for daily mood assessments. After week 8, cascading SMS reminder notifications were sent within a 3-day window (± 3 days) surrounding biweekly EMA assessment dates to maximize response rates. Automated algorithms flagged potential data quality issues, including wearable device discontinuities exceeding 7 consecutive days. Research staff maintained consistent participant-assessor pairings throughout the study period to minimize inter-rater variability. Post-collection phase: A three-tier data validation workflow is implemented using EpiData’s verification engine. First, automated algorithms flag anomalies based on pre-defined logical checks (e.g., out-of-range values, temporal inconsistencies). Second, trained data managers review flagged entries against source documentation. Third, senior data administrators adjudicate unresolved discrepancies. Prior to formal statistical analysis, processing pipelines are validated using synthetic datasets that simulate the structure and complexity of real study data.

Feasibility outcomes will be evaluated using descriptive statistics and reported with 95% confidence intervals where appropriate. We will estimate the recruitment success rate as the proportion of approached eligible patients who provide consent. Among enrolled participants, we will examine 8-week and 12-month retention rates stratified by baseline demographic and clinical characteristics, and characterize time-to-dropout patterns using Kaplan-Meier survival curves. For each data modality (EMA, passive sensing, wearables, clinical assessments), we will quantify the proportion of participants achieving adequate compliance thresholds [e.g., ≥70% for EMA and passive sensing data (42); ≥4 days/week for wearables (51)], mean and median compliance rates over time, and temporal adherence patterns including decay over time and weekend effects. Missing data patterns will be characterized by proportion and duration of missing episodes, temporal distribution across study phases and days of the week, participant-level factors associated with low compliance or data loss (including demographic, clinical, and platform-related variables such as Android vs. iOS), and the impact of platform constraints on multimodal data availability.

All secondary analyses are hypothesis-generating and will be conducted on participants with adequate data completeness. For MDD patients with complete multimodal data, we will systematically describe temporal patterns and inter-individual variability in key behavioral-physiological domains using visualization methods and summary statistics, including sleep parameters (duration, efficiency, timing, regularity), physical activity patterns (step counts, sedentary time, activity rhythms), cardiovascular dynamics (resting heart rate, heart rate variability), smartphone usage behaviors (screen time, app categories, diurnal patterns), and real-time mood states from EMA, with particular emphasis on the acute treatment phase (baseline to 8 weeks), when data availability and sample size are expected to be highest. Between-group comparisons will be conducted to examine differences in digital phenotypic features between MDD patients and healthy participants using appropriate statistical tests (t-tests, Mann-Whitney U tests, or linear regression models adjusted for demographic covariates including age, sex, education, and socioeconomic status), with effect sizes reported with confidence intervals; given the exploratory nature, we will not apply strict multiple comparison corrections but will interpret findings in the context of the number of tests performed. For participants with sufficient longitudinal data, linear mixed-effects models will be used to examine temporal trajectories of digital markers during treatment (with random intercepts and slopes to account for inter-individual variability) and associations between daily environmental exposures (photoperiod, ambient temperature, air quality [PM2.5], etc) and same-day or next-day fluctuations in mood, activity, and sleep, modeled as within-person effects with appropriate lag structures and adjusted for relevant time-varying and time-invariant covariates; seasonal effects will be explored where longitudinal coverage permits.

Preliminary analyses have been conducted to evaluate feasibility, compliance, and clinical outcomes. Between-group comparisons of sociodemographic and clinical variables were performed using independent samples t-tests for continuous measures and chi-square tests for categorical variables. Standardized mean differences (SMD) were calculated to quantify effect sizes. Data completeness was quantified across three modalities: (1) clinical assessment retention rates with 95% confidence intervals calculated using the Wilson score method; (2) EMA completion rates with ≥70% defined as adequate compliance (42) and (3) wearable device wear compliance with adequate compliance operationalized as ≥4 days/week, ≥10 hours/day wear time (51). Within-group changes in HAMD-17 and HAMA scores from baseline to weeks 2, 4, and 8 were evaluated using paired t-tests at each timepoint. Cohen’s d effect sizes with 95% confidence intervals were computed to quantify clinical significance of symptom change. All analyses were conducted in R version 4.5.1, with statistical significance set at 0.05 (two-tailed). Missing data were handled using available case analysis, with sample sizes reported at each timepoint.

We will prioritize conducting analyses using available data without filling in missing values when dealing with data from clinician-administered face-to-face assessments and EMA data. Additionally, we intend to conduct Full Analysis Set (FAS) and Per-Protocol Set (PPS) analyses. In addition to these methods, we will also employ multiple imputation to handle the data comprehensively. Regarding passive behavior monitoring data, we will record the timestamps of collected behavioral events without filling in missing values to preserve the original distribution of the data. For objective environmental data like weather, we will either get and fill the data within a specific time frame based on the participants’ latitude and longitude coordinates or keep the incomplete information empty for comparison analysis.

As the primary aim of this study is to assess feasibility and operational characteristics of intensive multimodal digital phenotyping, the target sample size was determined based on precision requirements for estimating key feasibility parameters rather than power calculations for effect size detection (52, 53). We aim to recruit 200 MDD patients, justified by the following considerations. First, using confidence interval–based approaches recommended for feasibility and pilot studies, a sample of 200 MDD patients provides an approximately 95% confidence interval margin of error of ±7 percentage points when estimating a proportion around 50% [e.g., recruitment success, retention, or the proportion achieving adequate EMA adherence ≥70% daily compliance), which was judged to offer adequate precision for feasibility assessment. Second, given that MOOD app data collection is technically feasible only on Android devices, we will implement platform-aware recruitment targeting approximately 150 Android users and 50 iOS users (reflecting the ~75:25 distribution in the Chinese smartphone market (54)]. Accounting for anticipated 20-25% attrition over the 8-week intensive monitoring period and 80% data completeness among retained participants, this approach is expected to yield approximately 90–100 participants with complete multimodal sensing data for secondary exploratory analyses, with an estimated 55–60 participants retained for extended 12-month longitudinal modeling. A target of 100 healthy participants provides adequate reference data for descriptive characterization and basic between-group comparisons.