Mental health issues have become a significant global public health challenge, affecting the quality of life and social functioning of hundreds of millions of people [1]. According to the World Health Organization, there are 322 million people with depression and 264 million people with anxiety worldwide, yet a large number of individuals still fail to receive timely diagnosis and intervention [2,3]. Traditional mental health assessments mainly rely on clinical interviews [4] and self-report scales [5], which, although widely used, have limitations such as high subjectivity [6], long assessment periods [7], and insufficient ecological validity [8], making it difficult to capture dynamic psychological changes in daily life [9]. In addition, many patients seek help only when symptoms become severe, hindering early intervention [10]. Therefore, there is an urgent need for an objective, continuous, and noninvasive mental health monitoring method to address the limitations of traditional approaches.

With the proliferation of smartphones, wearable devices, and other digital health technologies, passive sensing has emerged as a promising tool for mental health monitoring [11,12]. This technology uses embedded sensors such as accelerometers, GPS, and microphones to continuously and unobtrusively collect data on users’ behavior, physiology, and environment without requiring active participation [13]. These data reflect an individual’s daily activity patterns, social behaviors, voice features, and sleep quality, indirectly revealing their psychological state [14]. For example, GPS trajectories can be used to assess the level of social activity [14], accelerometer data can be used to analyze physical activity patterns [15], and voice analysis may detect emotional changes [16]. Compared to traditional methods, passive sensing offers significant advantages in terms of objectivity, continuity, and low intervention [12], showing great potential for monitoring and early warning of various mental disorders, including depression [17], anxiety [18], and bipolar disorder [19].

However, the data generated through passive sensing are typically high-dimensional, noisy, and unstructured, making it challenging to process using traditional data analysis methods [20]. The introduction of machine learning (ML) techniques provides a solution to this challenge. ML can automatically extract features from vast and complex sensor data and establish a mapping relationship between data and mental health status [11]. For instance, supervised learning algorithms (eg, support vector machine [SVM] and random forest) can be trained on labeled data to classify depression or anxiety levels [21,22], unsupervised learning (eg, clustering analysis) can uncover latent behavior patterns [23], and deep learning (eg, convolutional neural networks [CNNs] and recurrent neural networks) can extract deep features related to mental health directly from time-series data such as voice or motion sensors, overcoming the limitations of manual feature engineering and significantly improving recognition performance [24]. The continuous data collection by smart devices, combined with advanced algorithms, enables real-time monitoring and early warning of mental health status, providing strong data support for personalized interventions and precision medicine [25].

Although several reviews have addressed the application of passive sensing technologies in mental health monitoring, significant methodological gaps still exist. First, most reviews focus on the sensing technology itself rather than the integration of sensing with ML [26-28]. Second, there is limited discussion of key aspects of ML (such as data preprocessing and feature engineering), restricting the reproducibility and optimization potential of the technology [29-31]. Moreover, existing studies often focus on a single mental disorder (eg, depression [32,33] or anxiety [34]). Furthermore, they fail to include relevant databases such as PsycINFO [34-36], IEEE Xplore [37], and the ACM Digital Library [34,38]. On the basis of these gaps, this study aimed to conduct a scoping review to comprehensively assess the current application and development trends of ML techniques in mental health monitoring based on passive sensing data from wearable devices and smartphones. Specific objectives included to (1) synthesize and summarize current mainstream technical frameworks and commonly used methods for data acquisition, preprocessing, feature engineering, and algorithm selection; (2) evaluate association patterns between diverse behavioral features (eg, sleep, physical activity, and social interaction) and specific mental health conditions (eg, depression and anxiety); (3) identify the primary limitations and challenges of existing approaches with respect to sample representativeness, data quality, and model generalizability; and (4) explore potential future improvements in standardized data collection, algorithm optimization, and ethical guidelines.

This scoping review adhered to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [39] guidelines to ensure methodological rigor and transparency. The review protocol was preregistered on the Open Science Framework (registration 10.17605/OSF.IO/74ACP). The full PRISMA-ScR checklist is provided in Multimedia Appendix 1.

To ensure the systematic and reproducible nature of the literature search, we used an interdisciplinary database search strategy covering 7 authoritative databases across the fields of medicine, engineering, psychology, and computer science: Web of Science Core Collection, PubMed, IEEE Xplore, Embase, PsycINFO, Scopus, and the ACM Digital Library. In addition, we manually screened the reference lists of the included studies and used citation tracking to extend the search scope.

Before the official search, we collaborated with librarians, mental health experts, and health informatics specialists to develop a detailed search strategy based on the definitions in the Biomarkers, Endpoints, and Other Tools vocabulary [40]. The core keywords included “wearable devices,” “smartphones,” “mental health,” and “machine learning,” which were combined using Boolean logic. All search queries were pretested to ensure high relevance across different databases. To mitigate any bias introduced by daily updates to the databases, all searches were completed synchronously on February 3, 2025. After removing duplicates using the EndNote X20 software (Clarivate Analytics), the retrieved studies were systematically organized and analyzed. The specific search strategy is detailed in Multimedia Appendix 2.

Study eligibility was determined using the population, intervention, comparison, outcomes, and study design framework, as shown in Textbox 1, with additional methodological considerations detailed later in this section.

Two key methodological considerations guided our eligibility assessment. First, given the rapid evolution of wearable artificial intelligence (AI) technology and our focus on modern methodologies, we restricted the search to articles published after 2015. Second, while changes in mental health states (eg, emotions, mood, and stress) may serve as potential indicators of disorders such as depression or anxiety [41], examining these factors in isolation fails to reflect clinical diagnoses of mental health conditions [42,43]. Thus, we excluded studies focusing solely on these states to maintain research precision.

It should be noted that the comparison component of the population, intervention, comparison, outcomes, and study design was not applied as our scoping review focused on technological applications rather than therapeutic comparisons.

Before the formal screening process, the evaluators (SYS and JZ) randomly selected 30 articles and conducted a preliminary review of titles and abstracts followed by full-text evaluation based on the inclusion and exclusion criteria. A double-blind cross-validation approach was used. The pilot screening yielded a Cohen κ of 0.91, demonstrating a high level of agreement between the evaluators, and no adjustments to the screening process or criteria were necessary. Subsequently, SYS and JZ independently screened the remaining articles, and any discrepancies were resolved through discussion with a third-party evaluator (WQ), who made the final decision. On February 15, 2025, the screening was completed, and 42 studies meeting the criteria were included.

To explore the application of ML methods in the field of mental health and analyze how passive, noninvasive sensor data combined with ML techniques can be used to predict, detect, or monitor mental health disorders and outcomes, we followed the process of ML model development and validation. This process clarified the key stages of data acquisition, data processing, ML method application, and health outcome screening.

The first stage related to study characteristics. We used a standardized form to extract the key characteristics of each included study. Two independent reviewers collected data on study design, sample size, participant demographics, and the specific mental health issues addressed. For gender data, raw values were converted into percentages, with median values and IQRs subsequently calculated. In addition, we documented the duration of data collection and the sources of funding for each study. Using the R packages ggplot2, dplyr, and ggsci (R Foundation for Statistical Computing), we visualized the distribution of sample characteristics and funding sources.

In terms of data acquisition, we focused on the measurement environment, types of devices used, and specific configuration parameters reported in the studies. Detailed information, including device brands, sensor types, and body placement, was systematically documented. Using R packages such as ggplot2, tidyr, dplyr, igraph, fmsb, and tibble, we constructed an integrated network diagram illustrating the relationships among device types, sensor categories, and behavioral features. In addition, reshape2 and ggplot2 were used to analyze feature use patterns across studies and generate a heat map highlighting the prevalence of specific features under different mental health conditions.

Regarding data preparation, we systematically compiled the data preprocessing techniques used across the included studies. This encompassed a variety of signal denoising methods, including specific filtering algorithms and their parameter settings; strategies for handling missing data; and approaches to address class imbalance. These approaches were organized into comparative tables to highlight methodological variations among research teams. Special emphasis was placed on sensor-specific preprocessing techniques tailored to different data modalities.

Regarding feature engineering, we extracted features along three dimensions: (1) time window parameters, including window length and overlap ratio; (2) feature extraction techniques, spanning time-domain and frequency-domain analyses; and (3) feature selection strategies. We also compiled methods for assessing feature importance and, across 31% (13/42) of the studies, extracted quantitative metrics such as importance rankings and effect sizes. Data were processed using the tidyverse package in R, and we visualized the top 3 features’ associations with psychiatric disorders in each study via network graphs rendered using ggraph and igraph.

In terms of ML methods, we comprehensively documented the ML strategies used across the studies, including algorithm classes, model validation schemes, and performance metrics. Using the R packages dplyr, networkD3, htmlwidgets, and jsonlite, we generated an overview diagram summarizing the application of ML methods.

Figure 1 provides an overview of the key data items extracted from the included studies. In addition, we categorized the features into 8 main behavioral categories: sleep, physical activity, circadian rhythm, sociability, location, physiology, phone use, and environmental behavior [33]. Each behavioral category included several lower-level features that resulted from the integration of various individual features reported in each study. As a result, a single study may identify multiple associations with the same feature. Detailed definitions and classifications of these behavioral categories can be found in Multimedia Appendix 3.

The study selection process is detailed in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flowchart (Figure 2). A total of 42 studies were included in this scoping review. Among them, cohort studies accounted for the largest proportion (n=23, 55%), followed by cross-sectional studies (n=14, 33%). The median overall sample size was 60.5 (IQR 54-99) participants, whereas the median proportion of female participants was 57.5% (IQR 46.25%-67.34%). The study populations were primarily focused on depression (n=23, 55%) and anxiety disorders (n=9, 21%), followed by bipolar disorder (n=4, 10%) and schizophrenia (n=6, 14%).

In terms of control group design, 48% (20/42) of the studies used healthy or population-based control groups, 17% (7/42) compared patients with different mental health conditions, and only 2% (1/42) were based entirely on a healthy population.

Notably, as many as 76% (32/42) of the studies had sample sizes of <100 participants, with 24% (10/42) of the studies having sample sizes of <50 participants. In addition, 45% (19/42) of the studies collected passive data for <7 days, and 12% (5/42) of the studies did not report the specific duration of data collection. The distribution of sample sizes and data collection durations is shown in Figure 3A. Furthermore, the distribution of participant age characteristics can be found in Figure 3B.

Regarding the availability of datasets, 38% (16/42) of the studies did not provide explicit information, whereas the remaining studies included 31% (13/42) that provided open access datasets. A detailed list of the characteristics of the included studies is shown in Table 1, and the complete list of studies can be found in Multimedia Appendix 4 [14,15,17-19,24,25,44-79].

In addition, we conducted a detailed analysis of the funding sources for the included studies. The results revealed that 62% (26/42) of the included studies received funding, with these studies having a total of 69 funding instances. The breakdown of these funding instances showed that government agencies were the most prominent source, accounting for 57% (39/69) of instances. This was followed by funding from universities and research institutions (19%, 13/69) and nonprofit organizations or foundations (15%, 10/69). Funding from international organizations (6%, 4/69), private individuals (3%, 2/69), and corporations (1%, 1/69) was less common. This distribution indicates a high level of public sector interest and suggests that the field has relatively low appeal for industry and private capital.

Furthermore, as shown in Figure 4, the major funding agencies were the National Institute of Mental Health in the United States, which provided 10% (7/69) of the funding instances, followed by the National Institute on Drug Abuse, which contributed 4% (3/69) of the funding instances. These data not only highlight the concentration of funding sources but also provide a strong basis for further exploration of research trends and policy support in this field.

For a long time, mental health monitoring has relied on subjective assessment methods such as clinical interviews and self-report scales, which are often limited by ecological validity and recall bias [4,5]. With the widespread use of smartphones and wearable devices, passive sensing technologies now enable continuous, noninvasive collection of multidimensional data on user activity, heart rate, sleep, social interactions, and more in natural environments. This provides more objective, granular, and quantitative information on mental health status [14]. However, the multisource, time-series data generated by these sensors often exhibit characteristics such as high dimensionality, redundancy, and noise, making it difficult for traditional statistical methods to effectively capture the associations between underlying behavioral patterns and psychological indicators [20].

ML, particularly deep learning models, offers advantages in automatic feature extraction and nonlinear pattern learning. These models can uncover subtle emotional and cognitive features from complex sensor data, significantly improving the accuracy of mental state detection and prediction [24]. While several reviews have focused on passive sensing technologies themselves or provided single-technology summaries for specific mental disorders, there is a lack of a comprehensive review that systematically evaluates the combination of passive sensing and ML for mental health monitoring, including the entire technical framework and practical outcomes.

This study was conducted in this context and, for the first time, comprehensively reviewed and integrated research on the combination of passive sensing and ML for mental health monitoring from 2015 to 2025. It covered key aspects such as data acquisition, preprocessing, feature engineering, algorithm selection, and validation. The aim was to reveal the relationships between digital biomarkers and various mental disorders; assess the strengths and weaknesses of different ML models; and summarize existing challenges regarding sample size, data quality, model generalization, and ethical compliance. The following discussion will focus on methodological limitations, technical optimization pathways, clinical validation requirements, and ethical considerations.

Analysis revealed that cohort studies dominate the field (23/42, 55%), reflecting a consensus on the importance of longitudinal data collection. However, the sample sizes were generally small, with a median of only 60.5 (IQR 54-99) participants. Furthermore, 76% (32/42) of the studies included <100 participants, with 24% (10/42) of the studies having <50. This small sample size increases the risk of model overfitting, which undermines the robustness of the conclusions, especially in the validation of complex models such as bidirectional LSTM [44]. In addition, nearly half (19/42, 45%) of the studies had data collection periods of <7 days, whereas the pathological features of mental disorders typically require extended observation periods (eg, manic episodes in bipolar disorder often last several weeks [85]). Short-term monitoring may struggle to effectively distinguish transient mood fluctuations from true disease states.

Regarding data sharing, only 31% (13/42) of the studies provided open access data, and 38% (16/42) of the studies did not clearly specify data availability, which, coupled with a lack of standardized assessment protocols in the field, hinders result reproducibility [28]. It is noteworthy that only 2% (1/42) of the studies conducted external validation. The widespread absence of external validation, which is crucial for assessing model generalizability and avoiding overestimation of performance due to sample bias or overfitting, further highlights the limitations in model dissemination and validation in the current research [86].

The bottlenecks related to insufficient dataset size and lack of diversity are primarily constrained by the sensitivity of mental health data, the complexity of cross-study collaborations, and the challenges of long-term monitoring. The sensitivity of mental health data necessitates strict privacy protection requirements, significantly limiting the availability of data [87]. Moreover, conducting cross-center, cross-cultural studies requires coordinating multiple institutions and facing issues such as inconsistent technical standards and high data heterogeneity, which further increase collaboration costs [88]. More critically, the pathological features of mental disorders often require long-term monitoring for accurate capture, yet patient compliance, device battery life, and the continuous investment of research resources present practical barriers [89]. In response to these challenges, recent research has explored solutions from both system architecture and algorithm optimization perspectives.

At the system level, scalable digital data collection platforms (eg, Intelligent Sensing to Inform and Learn [90] and AwarNS [91]) have reduced the implementation barriers for multicenter studies through modular design and privacy protection mechanisms. Intelligent Sensing to Inform and Learn provides standardized sensor toolkits for iOS and Android devices, whereas AwarNS integrates a sense-analyze-act framework, supporting the closed-loop process from data collection to intervention and providing infrastructure for large-scale heterogeneous data collection.

At the algorithm level, researchers have reduced reliance on labeled data through unsupervised learning [92] and used data augmentation and generative adversarial networks to synthesize virtual samples, thereby expanding training datasets [93]. In addition, federated learning has enabled cross-institution data collaboration through distributed training, improving model generalization capabilities while maintaining privacy protection [94]. Although these methods have not fully resolved the data scarcity issue, they offer feasible paths to balance data scale and quality.

In terms of technological implementation, 76% (32/42) of the studies used a single device. While this reduces implementation complexity, it may overlook other potentially important signal dimensions. In addition, most studies (24/42, 57%) opted for wrist-worn devices, which, while convenient, have limitations in fully monitoring physiological and behavioral signals. For example, head-worn devices can capture electroencephalography activity [70], whereas chest sensors are better suited for monitoring cardiopulmonary function [25]. Future research should consider the integration of multilocation devices (eg, head, chest, and waist) to enhance the multidimensionality and consistency of the data, as well as explore multimodal fusion architectures (eg, CNN-LSTM combining physiological and GPS data [24]).

The quality of passive sensing data faces dual challenges in device compliance and technological reliability. Studies have shown that individual differences in device use compliance—such as irregular wear, forgotten charging, or voluntary discontinuation—often result in significant gaps and noise in the data. For instance, Mullick et al [52] reported a 67.74% data loss rate in their adolescent sample due to compliance issues such as forgetting to wear the device, charging requirements, and skin irritation. This issue may be further exacerbated by variations in user motivation. The intrusiveness of long-term monitoring may lead to “device fatigue” [95,96], whereas socioeconomic factors such as the availability of charging infrastructure, as well as privacy concerns across different cultural backgrounds, can also affect users’ willingness to continue participation.

From a technological perspective, consumer-grade wearable devices (eg, Fitbit) often use closed algorithms that may obscure the true quality of the data [55], and device malfunctions can exacerbate data sparsity [18,44,47,65]. To address these challenges, future research could focus on 3 main areas: the development of adaptive reminder systems based on individual use patterns, the establishment of effective gamification incentives, and the optimization of lightweight device designs (eg, smart rings) [97-99]. When combined with intermittent active feedback methods such as ecological momentary assessment [100], these solutions may help establish a more robust multimodal data validation system.

Regarding data preprocessing, noise reduction techniques such as outlier detection based on statistical features and time-domain filtering are effective at reducing interference. However, due to the differing signal characteristics of various sensors (eg, accelerometers, photoplethysmography, and ECG), there remains a lack of standardized practices in the field [101]. More importantly, few studies have systematically evaluated the actual impact of these preprocessing methods on the final model performance, resulting in a lack of empirical evidence to guide method selection.

The handling of missing data also presents challenges. Traditional methods such as mean imputation or KNN imputation can partially restore data completeness, but they may distort the true characteristics of physiological signals. On the other hand, directly deleting low-quality data may lead to the loss of critical clinical information [102]. Recent studies on motor activity data in schizophrenia have proposed a day-night difference-sensitive heuristic imputation method that preserves the key biological marker of nocturnal activity suppression while achieving a diagnostic accuracy of 94% [76]. This provides a reference framework for preprocessing similar data types.

In addition, techniques such as synthetic oversampling offer potential solutions for addressing class imbalance. However, the synthetic samples generated often fail to adequately reflect the complex temporal patterns in real data, leading to potential model overfitting or poor performance in practical applications [103,104]. A deeper challenge lies in the fact that many studies [15,19,45,46,48,51,57,73] overlook the details of time granularity, simplifying sensor data into daily summaries. This may miss micropatterns that signal symptom changes. For example, sleep fragmentation features in patients with bipolar disorder may be entirely masked in daily average sleep duration metrics [19].

The selection of window length, overlap rate, and feature selection strategies significantly impacts the performance and generalization ability of the model. The studies included in this review indicate that the window size in time-series analysis ranges from a few seconds to several days, and its choice directly affects signal integrity and model discriminability. For example, in high-frequency motion data analysis, a 2.5-second window effectively captures fine-grained behavior fluctuations such as walking and resting (with step count accuracy reaching 99.2%). However, it is also prone to interference from motion artifacts and other noise, requiring the application of a third-order Butterworth filter to enhance signal quality [17]. In contrast, a 48-hour window in circadian rhythm analysis can effectively capture periodic movement patterns. However, a long window may obscure short-term behavioral abnormalities in patients with depression—such as sudden decreases in activity or disruptions in movement patterns, which could be smoothed out by the 48-hour window, leading to the model missing transient features related to fluctuations in the patients’ condition [77]. This suggests that there is an optimal threshold for window length—shorter windows may capture fine-grained behavioral fluctuations but may amplify noise, whereas longer windows may mask transient psychological state changes, requiring dynamic adjustments based on the target behavior’s timescale.

The range of overlap rates also plays a critical role in model performance, balancing sample independence and data use efficiency. Khan et al [24] systematically analyzed the impact of overlap rates (0%-90%) on time-series model performance in a study on anxiety-related behaviors. The results indicate that a high overlap rate increases sample density through resampling, improving the model’s ability to capture brief behaviors, but it also introduces data redundancy, significantly increasing the risk of overfitting. A low overlap rate reduces sample correlation but may miss key state transitions. This research provides empirical evidence for the dynamic optimization of overlap rates in time-series analysis, highlighting the need to consider the target behavior’s duration and the model’s generalization requirements when adjusting parameters.

Differences in feature selection methods directly influence model complexity and interpretability. Variations in feature space search strategies, previous assumptions, and computational costs across different techniques can lead to significant divergence in the key variables captured by the model and their clinical relevance. For instance, brute force search, which exhaustively tests feature combinations (generating 14,892 models), can cover more potential patterns but may introduce redundant features, leading to overfitting and reduced interpretability [56]. In contrast, stepwise ML algorithms based on biological priors (such as circadian rhythm theory) progressively select features. While this sacrifices some model diversity, it allows the focus to remain on core variables with clinical mechanisms, and the resulting feature interpretability is stronger [15]. In cross-modal scenarios, embedded methods (eg, Depress-Deep Convolutional Neuro-Fuzzy combining CNN with statistical features) enhance accuracy to 85.1% by removing 88% of redundant features, whereas traditional SVM recursive feature elimination methods, reliant on manual feature engineering, perform more limitedly when multimodal data are missing [77]. Therefore, the choice of method should balance data-driven pattern discovery with theory-driven mechanism interpretation, optimizing based on the research goals (exploratory or confirmatory) and data characteristics (such as modality and completeness).

The included studies suggest underlying patterns linking passive digital features to mental disorders; however, these findings are predominantly exploratory and require further validation. Notably, the significant methodological heterogeneity across studies—in terms of study quality, data acquisition methods, and reporting standards—limits the interpretability of pooled associations between digital features and affective states. Nevertheless, we would be remiss in ignoring the growing consensus on the relationship between digital features and emotional states. Consequently, we provide a synthesis of these preliminary findings, emphasize their preliminary nature, and urge readers to consider these limitations when interpreting the results.

In depression-related research, heart rate (20/42, 48%), movement index (15/42, 36%), and step count (14/42, 33%) were the most frequently reported features; these metrics correlate strongly with core depressive symptoms such as psychomotor retardation associated with anhedonia [105]. However, although smartphone use behaviors ranked highly in feature importance analyses, only 17% (7/42) of the studies actually collected these data, likely reflecting data acquisition challenges—phone call records raise privacy concerns, whereas step counts can be easily extracted from commercial wearables [19]. In anxiety disorder research, physiological and sleep characteristics such as heart rate (5/42, 12%) and sleep duration (3/42, 7%) were commonly observed. However, the importance of environmental and behavioral factors such as time spent at home suggests that future studies should focus on enhancing the integration of environmental sensors. In schizophrenia research, the movement index (3/42, 7%) was predominant, which may be associated with motor impairments caused by antipsychotic medications [106].

Window segmentation lays the foundation for constructing stable data sequences by structuring inputs for feature extraction and, via overlapping windows, increasing the number of training samples to enhance model generalizability. However, there were variations in the selection of window sizes across the studies, and no unified guidelines have been established. In feature extraction, deep learning–based approaches have emerged as the dominant strategy because they automatically capture complex feature interactions and markedly improve predictive performance. CNNs in particular excel at automated feature extraction but rely on large-scale training datasets, increasing the risk of overfitting [107]. To mitigate this issue, previous studies have suggested that layerwise dropout regularization is an effective strategy, although the generalizability of such architecture optimization solutions across different datasets still requires further validation [47].

It is important to note that, although end-to-end deep models offer performance advantages, their black box nature limits their interpretability in clinical applications [108]. Some studies have introduced Shapley additive explanations [25,54] and integrated gradients [47] to enhance transparency, but core issues of model explainability persist. In some research, efforts to improve interpretability have led to oversimplified feature spaces (eg, retaining only 3 core features [65]). While this may facilitate clinical decision-making, it could result in the loss of multidimensional information essential for mental health monitoring. This highlights the ongoing challenge in model development to balance clinical interpretability with the need for sufficient representation of symptom heterogeneity.

In model selection, current studies retain the application of traditional ML methods (such as random forest, SVM, and logistic regression) while continuously exploring the potential of deep learning models. CNNs and their variants deliver superior performance for time-series data processing, nonlinear relationship modeling, and automated feature derivation; however, their substantial computational demands hinder deployment on edge and wearable devices, further intensifying the complexity-interpretability trade-off [24,47]. Autoencoders offer an unsupervised dimensionality reduction approach that opens new avenues for feature engineering. Notably, Baygin et al [72] developed a lightweight model based on a probabilistic binary pattern, which not only achieved >98.5% classification accuracy in ECG signal–based anxiety detection but also significantly reduced computational load. Their model, with a limited but highly interpretable feature set, offers a new paradigm for balancing performance and interpretability.

Furthermore, there is mounting evidence suggesting that personalized models yield superior predictive accuracy. For example, individualized LSTM models explain up to 39% of within-subject variance in anxiety symptoms—far exceeding the explanatory power of traditional cohort-based models [18]. These findings highlight the critical importance of modeling individual differences in mental health prediction.

Passive sensing data often involve highly sensitive personal information. However, only 14% (6/42) of the studies explicitly mentioned anonymization procedures [44,48,51,55,59,62] despite the fact that data anonymization and encryption are fundamental requirements for ensuring privacy and security. Moreover, as research may involve repeated monitoring and prolonged data collection, informing participants and obtaining their informed consent, as well as ensuring their right to withdraw consent for data use, are crucial issues that must be addressed when translating such technologies into clinical practice.

Insufficient sample representativeness is another significant limitation. Most studies (23/42, 55%) [53] relied on single-center or specific datasets (such as university student populations [14,18,45,49] or outpatient hospital patients [15,19,46,48,51,61,62,69]), which often exhibit high homogeneity in dimensions such as age, gender, and ethnicity, making it difficult to represent a broader population. This distribution imbalance may lead to decreased model performance in other populations and even trigger systemic biases. Future research should enhance model generalizability by expanding sample diversity and using statistical correction methods (such as multiple comparison corrections). In addition, insufficient model interpretability may undermine physicians’ trust in algorithmic results, potentially affecting clinical decision-making [109].

The generalizability of current research findings is further limited by the lack of representation across cultural and socioeconomic backgrounds. Due to differences in cultural norms, stigma, and access to technology, mental health symptoms and behavioral patterns (eg, social interactions and mobility) may differ significantly across cultures [110]. For example, in some Asian cultures, the concept of family honor may lead individuals to reduce their use of public social applications (eg, social media engagement) due to psychological distress while increasing the use of private communication apps (eg, one-on-one messaging) [111]. In African American communities, distrust of mental health services may manifest as a preference for avoiding mental health monitoring applications that require sensitive data authorization or increased reliance on anonymous communication platforms that do not require real name registration, resulting in systematic bias in passively collected data such as location tracking and voice features [112]. In addition, socioeconomic disparities such as lower device ownership and unstable internet access in low-income or rural communities often lead to sparse or noisy data, affecting data quality [113].

In addition, current research primarily validates the correlation between passive monitoring features and mental health, but due to the lack of causal inference, it remains unclear whether these features can serve as effective intervention targets or clinical prediction biomarkers [18,45,65]. Furthermore, most studies in this review (26/42, 62%) relied on government funding, with industry involvement at only 1% (1/69), which may delay the technology translation and practical application process. Moreover, most studies (39/42, 93%) focused on the analysis of a single disease, neglecting common comorbidities. For example, the comorbidity rate between depression and anxiety is as high as 25% [114]. Therefore, future algorithm designs should aim to identify cross-diagnostic features, thereby providing a more comprehensive assessment of an individual’s mental health risks.

In terms of ethical deployment, addressing user-centered challenges is critical. Long-term adherence is often hindered by concerns about data privacy or perceived utility, especially in clinical populations. Transparent, dynamic consent interfaces and the provision of personalized feedback can effectively alleviate privacy concerns and enhance technology acceptance [115]. Furthermore, when integrating passive sensing technologies into existing health care workflows, it is essential to consider the digital literacy differences between clinicians and patients. Therefore, developing specialized training programs is crucial to ensure equitable adoption [116].

Although this study systematically reviewed the potential of passive sensing technologies in mental health monitoring, it must be acknowledged that the field currently lacks unified device approval standards, which has become a key bottleneck in the clinical translation of these technologies. Current regulatory frameworks primarily address traditional medical devices, whereas mental health monitoring technologies such as smartwatches and smartphone sensors do not yet have established uniform regulatory requirements due to their passive data collection methods, device heterogeneity, and diverse application contexts [117,118]. For example, while the US Food and Drug Administration has launched a pilot program for digital health technology precertification, specific standards for continuous mental health monitoring are still absent [119]. Similarly, the European Union’s Medical Device Regulation includes AI algorithms but does not specify how to assess the reliability of passively collected data [120]. This situation leads to potential fluctuations in model performance due to device changes, making it challenging to meet the clinical diagnostic requirements for consistency and reproducibility.

To the best of our knowledge, this is the first scoping review assessing the current status and development trends of ML techniques in passive mental health monitoring based on wearable device and smartphone sensor data. Compared to previous similar studies, this review has the following advantages. First, it is the first to provide a comprehensive analysis of the entire technological process from data collection to model validation, addressing the gap in previous research that only covered partial technological aspects. Second, by integrating data from 42 empirical studies, we explored the associative patterns between different mental disorders and behavioral traits, covering a broader range of disease types and feature dimensions than most similar reviews. In addition, this study highlights the limitations of existing methods in areas such as sample representativeness, data quality, and model generalizability while proposing paths for future technological improvements, clinical validation, and ethical practice. These contributions provide a more comprehensive assessment of the current state of methodologies in the field and offer technical references and practical insights for further optimizing the application of passive sensing technologies in mental health monitoring.

However, this study also has some limitations. First, in terms of literature search, while we used a cross-disciplinary database co–search strategy, the language and literature selection criteria (only including English-language journal articles and excluding conference papers and other nonjournal literature) may have led to the omission of important research from non–English-speaking countries and time-sensitive technological solutions from conference reports. Moreover, although MeSH (Medical Subject Headings) terms were considered, a keyword-based approach was prioritized to capture emerging terminology and maintain consistency across databases; this may have introduced some noise but ensured broader coverage of interdisciplinary literature. Second, regarding the algorithm performance comparisons presented in our results, it should be noted that these quantitative findings reflect model performance under specific study conditions. Due to substantial methodological variations in experimental designs, evaluation metrics, and data characteristics across the studies, these results are more suitable for descriptive analysis of current research trends rather than definitive conclusions about algorithmic superiority. Third, our findings may be partially influenced by the broader methodological limitations of the included studies, such as small sample sizes, potential overfitting of models, unclear generalizability of results, and differences in measurement metrics and outcome definitions across studies. These factors may affect the robustness and applicability of the conclusions. Moreover, although we strictly adhered to the PRISMA-ScR guidelines, applied clear inclusion and exclusion criteria, and systematically recorded risk of bias and other relevant factors, the broad scope of health outcomes and the differences in study designs somewhat limit our ability to comprehensively assess the overall quality of the studies [121].

Despite the significant potential of passive sensing technologies in mental health monitoring, substantial breakthroughs are still needed to translate exploratory research into clinically practical tools. On the basis of a systematic evaluation of current limitations and opportunities, we propose a comprehensive development road map aimed at addressing methodological, technical, clinical, and ethical challenges while promoting the translation of research outcomes into practice.

This study provides a comprehensive assessment of the current status, technical pathways, and future directions of passive sensing technologies based on wearable devices and smartphones combined with ML for mental health monitoring. We found that passive features such as sleep, physical activity, physiological signals, and social behavior showed stable associations with various mental disorders, including depression, anxiety, bipolar disorder, and schizophrenia. Among these features, heart rate, movement index, and step count were the most commonly used indicators. At the algorithmic level, deep learning models (eg, CNN and LSTM) demonstrated excellent performance in mental health prediction due to their robust temporal data processing capabilities. In contrast, traditional methods such as random forest and XGBoost continue to play a significant role in clinical research due to their strong interpretability.

Despite these advancements, challenges such as small sample sizes, short monitoring durations, limited device types, data noise and missing values, lack of model interpretability, and insufficient external validation hinder the generalizability and clinical translation of these studies. To advance the field, future research should focus on large-scale, multicenter, longitudinal studies across diverse populations; establish standardized protocols for data collection and preprocessing; strengthen multimodal integration and personalized modeling; and introduce explainable AI and privacy protection mechanisms to ensure the safe and efficient application of passive sensing technologies in real-world health care environments.