Behavioral pattern changes are commonly observed in psychiatric patients and have become an area of research in psychiatric disorders, including major depressive disorder (MDD), bipolar disorder (BD), and schizophrenia [1-4]. In particular, depression is associated with mood changes that lead to corresponding behavioral alterations [5,6]. Traditional approaches in continuous monitoring and assessing depressive symptoms primarily rely on patient monitoring via staff patrols, Closed-Circuit Television (CCTV) surveillance, and subjective reports from patients and caregivers [7-9]. Such methods often face limitations owing to recall bias and difficulties in capturing real-time symptom changes, especially when a single provider oversees multiple patients [10-12]. Additionally, these conventional monitoring approaches may not detect or manage critical incidents such as violence, aggression, self-harm, and absconding if resources are insufficient [13,14]. The depression scales commonly used in traditional monitoring can be discontinuous, and their results may vary depending on who administers them, leading to inaccuracies such as over- or underestimation [15]. Consequently, digital mental health solutions that use digital phenotyping to continuously monitor a patient’s status and objectively capture depressive symptoms have gained traction as an alternative strategy, potentially mitigating these limitations and improving patients and healthcare staff [16].

Digital monitoring, often referred to as “remote sensing” or “digital phenotyping,” is an emerging field that provides new opportunities for improved management and treatment by enabling objective and continuous patient observation [17]. Previous studies in this area have identified actigraphy-based and location data as key approaches for assessing depression, deriving various phenotypes that serve as psychiatric monitoring tools. According to a meta-analysis, 29 studies collected data on sleep, 19 focused on physical activity, 13 examined circadian rhythms, and 11 investigated location patterns, with all these phenotypes showing significant associations with depression [18]. For example, time in bed and sleep fragmentation (indexed by the number of awakenings and wake after sleep onset) were strongly associated with high depression scores [19,20]. Common actigraphy-based mobility indicators include interdaily stability (IS), intradaily variability (IV), and relative amplitude, with findings indicating that individuals with MDD or BD often exhibit low IS, IV (suggesting more fragmented activity), and amplitude [21]. Meanwhile, location-based features—such as location variance (LV), entropy (EN), normalized entropy (NE), and homestay duration—have been associated with depression. Lower LV and EN (indicating reduced regularity in location patterns) and longer homestay duration are linked to more severe symptoms [22].

However, despite the extensive use of numerous phenotypes in various studies, methods of collecting, aggregating, and analyzing these data remain underdeveloped. A general lack of discussion exists regarding whether the devices used in these studies are valid or reliable for detecting the behaviors of interest. Although certain behaviors may seem relatively straightforward to infer from single sensors, such as Global Positioning System for location or accelerometry for physical activity, concerns about validity and reliability persist [23-25]. Moreover, because the association between each feature and depression tends to be diverse or somewhat opaque, ascertaining whether these features accurately capture depression and applying them effectively in clinical practice remains challenging [18].

Therefore, our analysis was based on the concept of “movement regularity” by integrating real-time data extracted from wrist-worn devices (accelerometer data, representing micro-level movements) and ward sensors (location data, representing macro-level movements). Activity level changes are a significant indicator included in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition criteria, and numerous studies have consistently demonstrated a robust link between such shifts in activity patterns and depressive symptoms [26]. By utilizing wearable and ward sensor data inputs from two distinct data sources, we aimed to provide a more comprehensive and objective interpretation of behavioral regularity.

Another limitation of current digital monitoring research is that most studies in this field are focused on outpatients, whose activities vary considerably owing to individual and environmental differences. However, this study targeted psychiatric inpatients with more severe symptoms in a controlled environment to address these constraints. By defining up to four specific locations where inpatients can move, we were able to monitor their movement precisely within a restricted setting. Additionally, because observations occurred continuously in a closed ward, we classified each measure into nocturnal and diurnal intervals on a 24-hour basis, allowing investigation on time-specific differences in behavioral regularity.

We hypothesized that digitally measured behavioral regularity could serve as a robust indicator of depression severity by adopting four non-parametric measures—IS, IV, EN, NE—along with the ratio of IS to IV (ISV) to examine behavioral patterns. We interpreted these digital phenotypes through micro-level movements (actigraphy data) and macro-level movements (location data), proposing a strong association with the severity of depressive symptoms. To our knowledge, no study has been conducted in a closed psychiatric ward that uses wearable devices and ward sensors to capture micro- and macro-level movements in inpatients, or has further subdivided these phenotypes into diurnal and nocturnal periods for analysis. Furthermore, by comparing these digital measures with self-and clinician-administered scales, this study demonstrates how digital mental health indicators can facilitate objective continuous monitoring of depression within psychiatric wards. We aimed to develop a more efficient framework for monitoring patients with depression, and we demonstrate that digital monitoring focused on behavioral regularity can serve as an effective tool for continuous functional monitoring in clinical settings.

In this preliminary feasibility study using exploratory models, digital markers of movement regularity were associated with clinician-rated depression. We investigated how behavioral regularity, as measured by digital phenotypes, reflects depression severity in psychiatric inpatients. We found that the diurnal micro-movements obtained from actigraphy-based data and the nocturnal macro-movements obtained from location data effectively explained disturbances in depression, suggesting their potential as digital correlates. Moreover, these digital indicators showed more pronounced differences in patients with moderate/severe depression, indicating that they may be particularly meaningful in an inpatient population with more severe depressive symptoms, rather than in outpatient settings. Consequently, assessing behavioral patterns through digital monitoring not only addresses the limitations of traditional monitoring but offers a more continuous and objective method to closely observe patients, supporting its role as a valuable explanatory measure for depression.

We collected data from 10 inpatients in a psychiatric ward, totaling 18 sets of samples. These participants presented with various psychiatric conditions, including schizophrenia, psychosis, brief psychotic disorder, bipolar I disorder, bipolar II disorder, and adjustment disorder. Focusing on inpatients with more severe symptoms allowed for a very in-depth investigation of their clinical status within a controlled environment, minimizing the influence of individual lifestyles and diverse external settings. Moreover, by encompassing patients with a range of diagnoses, this study captured processes that extend beyond a single diagnostic label, leveraging the strengths of transdiagnostic methodologies. Given that depressive symptoms and behavioral irregularities are frequently observed across multiple psychiatric disorders [35-37], our broad approach enhances the scope of mental health research.

While actigraphy and location data are considered the gold standard for assessing movement in digital health [18,38,39], our approach emphasized the regularity in each dataset by integrating both data types. This multimodal strategy addressed limitations inherent to studies that focused on a single source [40]. Beyond reliance on a single wearable device, we incorporated wireless ward sensors, thereby mitigating issues such as false alarms and discrepancies in wrist-actigraphy measurements. This approach facilitated more effective long-term monitoring [23,41,42]. Moreover, the digital indicators demonstrated stronger correlations with the HAMD. These findings illustrate the interoperability between device-derived data and data from clinical evaluation, highlighting the potential of digital devices to mitigate biases such as the overestimation and underestimation commonly associated with self-reported measures [12,43-45]. Since traditional monitoring methods rely on staff patrols and CCTV [9], remote sensing can offer a valuable alternative for managing patients and alleviating the constraints of conventional surveillance. Finally, by utilizing two distinct data sources, we could evaluate both macro- and micro-level movements. The ward sensors that collect location data via wireless sensors provided insight into patients’ macro-level movements (how they physically walk to specified ward locations), while wrist-worn accelerometers captured micro-level activity patterns. This two-tiered approach provided a comprehensive interpretation of movement behavior, which offered insights from a broad and granular perspective, further enhancing our understanding of depressive symptoms.

We focused on movement regularity for two primary reasons. First, as shown in Figure 1, no significant consistent movement patterns emerged, and individual data fluctuated continuously, indicating a lack of daily activity regularity. The six-day data extraction period varied among participants, resulting in differences in the number of weekend days included. Moreover, simply averaging activity levels over 6 days can be directly influenced by reduced activity due to medication, and baseline activity levels may vary substantially among patients. Hence, raw averages of SVM values can fail to accurately reflect each patient’s condition. To address these challenges, we adopted the IS and IV concepts to quantify the regularity and fragmentation of activity levels, rather than relying solely on raw averages.

For location data, unlike previous studies, we conducted our research in a closed psychiatric ward where movement was highly restricted, limited to four predefined clusters. To track behavioral changes within these clusters, we applied the EN concept, allowing an assessment of patients’ movement regularity beyond mere location proportions. Through these indicators of behavioral patterns, we aimed to determine how stable and regular activity levels influence depression severity. Additionally, because daytime and nighttime activity patterns significantly differed, leading to compensatory effects in overall measures, we divided the analysis into diurnal and nocturnal phases for a more nuanced evaluation.

We identified five key variables that effectively capture behavioral regularity and its relationship to depression severity: IS_day, ISV_day, IV_day, EN_night, and NE_night. ISV is of particular interest, as it represents the ratio between IS and IV, underscoring the significance of both stability and fragmentation in movement regularity. Through LMM analysis, we observed significant differences in these variables between the asymptomatic/mild depression group (HAMD ≤13) and the moderate/severe depression group (HAMD >13). Notably, all five variables showed consistent differences between these groups, suggesting that individuals with greater depression severity exhibit more fragmented, irregular daytime activity rhythms and increased randomness in nighttime activity. Therefore, movement regularity across temporal scales (day and night) showed a stronger association with individuals experiencing more severe depression. Furthermore, Pearson correlation analysis indicated that these five variables align with previous research linking them to depression severity [22]. The classification into daytime and nighttime showed that more severe depression correlates with greater daytime behavioral instability and heightened nighttime activity; however, DSAS scores did not significantly relate to these digital markers.

Additionally, our correlation analysis revealed that the digital phenotypes representing micro-movements are interrelated, indicating that regularity and fragmentation are intertwined elements of movement behavior. The correlation coefficients between IV_night and IS_day, as well as IV_day and IS_night, were particularly noteworthy. One tentative interpretation is that relatively stable sleep patterns may be accompanied by larger daytime activity fluctuations, whereas irregular nighttime behavior may co-occur with differences in the perceived structure of daytime patterns. However, this remains speculative, necessitating further research to clarify the precise relationship between IV and IS.

In multiple regression, a model incorporating EN_night and IS_day accounted for 63% of the variance in HAMD scores (in-sample explanatory capacity). These results suggest that nighttime activity variability and daytime mobility stability can serve as vital digital markers for assessing depression severity in inpatient psychiatric settings. When comparing our regression findings with self-reported measures, we found that digital phenotypes correlated more strongly with clinician-administered assessments than with self-report scales, suggesting that digital monitoring may provide a more objective evaluation of the depressive states of patients. This may in turn help compensate for the discontinuity inherent in clinician-associated assessments and serve as a valuable adjunct tool in clinical practice.

This study has some limitations. First, the small sample size (10 patients, 18 samples) substantially limits statistical power and reduces the stability of correlation and regression estimates. Consequently, these findings should be regarded as preliminary and exploratory, intended primarily for hypothesis generation rather than confirmatory inference. To illustrate uncertainty, 95% confidence intervals for key associations and regression coefficients are reported in Supplementary Tables 1 and 2. The limited sample also restricted the inclusion of potential confounders, such as diagnosis and medication use, and precluded a meaningful post hoc power calculation given the complexity of the mixed-effects modeling approach. These constraints underscore the need for larger, more homogeneous cohorts and longitudinal designs to confirm these findings and improve generalizability.

Second, the SVM values obtained from accelerometer data were averaged hourly for each patient and used to calculate IS and IV. As these hourly averages occasionally contained missing data, errors arose in calculating IS and IV, particularly when large gaps existed between measurements. Such gaps complicate the accurate depiction of movement regularity (including stability or variability) by obscuring potential fluctuations in the data.

Third, we defined four clusters and categorized all other locations as “others.” However, this classification caused several challenges. When “others” reached 100%, it was treated as missing data, since EN was calculated solely from the four predefined clusters. This approach made it difficult to accurately capture behavioral regularity. For instance, a patient who consistently visits non-clustered spaces, such as the bathroom or hallway, should theoretically display high EN, yet those locations were labeled as “others,” resulting in EN exclusions or a zero value for missing data. Furthermore, if a patient occasionally removed their watch while sleeping outside the designated patient room, thereby generating hours of “others” at 100%, those hours also became missing from EN calculations.

Nevertheless, defining “others” yielded practical benefits in resolving technical issues. Occasionally, sensor signals showed overlapping location errors within the same room, depending on the bed position, or recorded multiple areas simultaneously during nighttime when patients were expected to remain in their rooms, likely owing to technical or data transmission errors. As interpreting movement regularity under these conditions proved difficult, we excluded “others” from the predefined clusters and refrained from attaching specific meanings to each cluster. The calculation of EN as a digital phenotype remained unaffected because the majority of signals overlapping with the patient room were registered as “others.” When computing EN, we recalculated the percentage of time spent in each cluster (excluding “others”) and derived an adjusted EN value, thus addressing data overlap.

Although continuous monitoring of depressive symptoms through digital tools offers a longitudinal approach that captures evolving data over time and accommodates temporal relationships, it remains insufficient for definitively establishing causality or encompassing all facets of depression. Moreover, the best way to process raw digital data into a valid, explanatory phenotype remains unclear. Whether the current indicators, particularly those derived from an average of six days, truly reflect real-time movement patterns is also open to question. While correlations between digital phenotypes and depression severity have been identified, future studies should refine these indicators and employ more rigorous designs to discern causal relationships. Consequently, further research is warranted to validate continuous digital monitoring as a practical tool, particularly in larger, more homogeneous samples and longitudinal frameworks, to generalize its use in populations with depression.

Despite these limitations, the present findings provide a concrete foundation for translating digital phenotyping into real-world psychiatric care. Circadian rhythm has long been recognized as a central mechanism in mood regulation, with interventions such as light therapy and structured daily routines demonstrating therapeutic efficacy [46,47]. In this context, our digital indices of daytime macro-movement and nighttime micro-movement regularity may serve as behavioral proxies of circadian stability. Leveraging these features for clinical use could open pathways to rhythm-based interventions that are tailored to individual patterns.

Building upon these implications, our dual-level approach—capturing both macro- and micro-level regularity—offers a unique contribution to the field of digital psychiatry. In practical terms, these metrics could be incorporated into real-time monitoring platforms for psychiatric inpatients. For example, nursing staff could receive automated alerts when a patient exhibits a sudden drop in daytime regularity (e.g., IS_day) or an unexpected increase in nighttime EN (e.g., EN_night). Such threshold-based alert systems may facilitate earlier detection of clinical deterioration, reduce monitoring burden, and enable more consistent, personalized care. While specific cutoff values and predictive thresholds remain to be validated in larger studies, the implementation of such tools holds promise for enhancing inpatient psychiatric care through data-driven, individualized decision support.

Despite promising findings, challenges remain in translating digital phenotyping into psychiatric practice. In particular, future research should validate the reliability and clinical utility of behavioral regularity derived from macro- and micro-level movement data. This study underscores the clinical potential of digital phenotyping in psychiatry by integrating macro- and micro-level behavioral regularity. Together, these results underscore the clinical potential of dual-level digital monitoring and suggest a path toward future diagnostic precision and individualized psychiatric intervention.