This scoping review was conducted based on Arksey and O’Malley’s methodological framework and the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines (Checklist 1) [17,18]. The review was guided by the population-concept-context framework (population: not restricted; concept: passively collected digital data from smartphones or wearable devices; and context: suicide-related outcomes). The study was conducted through collaboration among 3 research teams, with external validation performed by separate research teams. Fourteen studies were analyzed based on the LHSF using a structured data charting and synthesis approach [16], and Google Sheets was used as a data management tool. The protocol was registered in the Open Science Framework [19].

Textbox 1 summarizes the query terms used in the search strategy. The search strategy was developed iteratively using the population-concept-context framework to ensure alignment with the review objectives. Four databases (PubMed, CINAHL, PsycINFO, and IEEE Xplore) were searched for studies published between January 1, 2015, and December 31, 2025. The search combined suicide-related terms with those related to digital sensing, as summarized in Textbox 1 (see full search strings in Multimedia Appendix 1). The final search was conducted in December 2025. Gray literature was excluded from the scope of this review. This boundary was established because the primary objective was to map granular digital signals onto the structured LHSF framework, which requires fully detailed and finalized descriptions of raw sensor processing, feature extraction, and statistical verification—details that are often omitted or partially reported in non–peer-reviewed literature. The reference list screening was conducted to identify additional relevant studies. A librarian at the affiliated medical library was consulted to review and refine the search strategy to enhance comprehensiveness and accuracy.

We included studies that reported correlational or predictive analyses that explored digital markers with suicide-related outcomes. Studies that relied solely on active data collection methods or did not assess suicide-related outcomes were excluded. No language restrictions were applied. Textbox 2 summarizes the detailed inclusion and exclusion criteria. The publication period was restricted to studies from January 2015 onward, corresponding to the introduction of digital phenotyping in clinical research and wearable devices [20].

Two authors (SK and JK) independently screened the titles and abstracts of potentially eligible studies, followed by full-text review. If disagreements arose, they were resolved through discussion between 2 reviewers, and those not resolved were adjudicated by the principal investigator (HK).

Two authors (SK and JK) initially extracted the data, which were reviewed by the principal investigator (HK). Subsequently, 2 separate teams (WTC and YZ; and TTL and ZYAT) provided external validation and contributed to refining the analysis. Discrepancies were resolved through team discussion. Google Sheets was used to organize a structured data matrix for data charting. The data charting form was developed by the research team and refined through discussion prior to full data extraction. Study characteristics, device and sensor types, and low-level features and high-level behavioral markers were charted. Narrative mapping was conducted according to the LHSF [16]. This scoping review followed Joanna Briggs Institute recommendations for scoping reviews [21]. Consistent with these guidelines, methodological quality assessment was not performed.

The LHSF was applied to organize data from raw sensor inputs to inferred behavioral patterns related to suicidality [16]. On the basis of the 4 hierarchical layers, the extracted features were grouped:

To synthesize findings across the 4 levels, we organized and categorized extracted features according to the LHSF structure. While the original framework was retained, specific digital features and behavioral markers were refined and expanded based on the included studies to reflect recent advances in digital phenotyping research. This approach enabled us to map contemporary empirical evidence onto the existing framework.

To ensure consistent operationalization across layers, explicit decision rules were applied during data charting. A feature was classified as a low-level indicator (layer 2) if it represented a directly quantifiable, sensor-derived measurement requiring no inferential interpretation. A construct was classified as a high-level behavioral marker (layer 3) if it represented a psychosocial or clinical construct that required aggregating or theoretically interpreting one or more low-level features. When the original study explicitly used a term corresponding to a layer 3 construct, it was retained as source defined. When no such label was present, 2 authors (SK and JK) independently derived a conceptually equivalent label based on the analytical context of the original study, indicated by superscript “a” in Table 1. Discrepancies in layer assignment or label derivation were resolved through discussion. A detailed overview of the operationalization criteria applied to each study is provided in Multimedia Appendix 2.

To summarize the strength of evidence across feature domains, each study was classified according to the statistical results reported in the original article. A study was counted as showing a significant association for a given feature domain if at least one indicator in that domain was significantly associated with, or predictive of, at least one suicide-related outcome, such as suicidal ideation, planning, attempts, risk, or events. This classification was retained even when other indicators or outcomes within the same study were nonsignificant. Domain-level evidence was then described by the number of studies reporting significant associations, rather than by a single subjective judgment. Predictive performance was interpreted against a conventional discrimination threshold of an area under the curve (AUC) of 0.80 or above.

We conducted the final search in December 2025 and identified 626 studies across the 4 databases (PubMed: n=374; CINAHL: n=130; PsycINFO: n=111; and IEEE Xplore: n=11). After removing 185 duplicate records, 441 (70.4%) articles were screened based on their titles and abstracts. Of these, 365 (82.8%) records were excluded, and 76 (17.2%) full-text articles were assessed. Subsequently, 62 (81.6%) articles were excluded (Figure 1). Ultimately, 14 (18.4%) studies were included [22-35]. Reference list screening identified no additional eligible articles.

Table 2 presents the characteristics of the included studies. Publication years of the 14 included studies ranged from 2020 to 2025, with most of the studies published in 2025 (n=5, 35.7%). Most studies were conducted in the United States (n=11, 78.6%). All included studies used observational designs, in which 12 (85.7%) were longitudinal studies, and 2 (14.3%) were cross-sectional. Furthermore, 11 (78.6%) studies were in prospective design, and 3 (21.4%) were retrospective ones. Regarding characteristics of participants, 3 (21.4%) studies included adolescents, and 11 (78.6%) studies included only adults. Most studies targeted high-risk populations (n=12, 85.7%), such as individuals who visited emergency departments for suicide-related reasons, psychiatric inpatients or outpatients, and those with a history of suicidal ideation or behavior. Duration of passive data collection varied across studies, ranging from 24 hours to 6 months.

Regarding demographic composition, most study samples were predominantly White or Caucasian, with White participants comprising 57% to 90% of samples across studies reporting race or ethnicity. Female participants comprised the majority across all studies, with female ratios ranging from 56% to 84.8%. Although most studies recruited high-risk adult clinical populations, 3 (21.4%) studies included adolescents. Among studies reporting clinical diagnoses, depressive and anxiety disorders were the most represented conditions. Socioeconomic status was inconsistently reported; where reported, samples tended to include participants with different levels of college education or current employment. Detailed demographic and clinical sample characteristics are presented in Multimedia Appendix 3.

Two (14.3%) studies reported nonsignificant or limited predictive performance for sleep- and activity-related passive features [26,28]. In predictive contexts, passively collected sleep and physical activity features had limited or no incremental predictive value for suicidal ideation when compared with, or added to, clinician-rated or self-reported measures [26,28]. In one study, passive sensor data alone demonstrated poor predictive performance for next-day suicidal ideation (AUC=0.56) and did not improve prediction when combined with EMA features [28].

Across the included studies, several directions for future research were identified (Table 3). First, many studies emphasized the need for replication in larger and more diverse samples, along with the use of longitudinal designs, and more frequent assessments to improve the robustness and generalizability of findings [22-24,27,28,31,33,35]. Second, a consistent focus was placed on improving data quality and methodological approaches, including refinement of passive sensing features, optimization of temporal resolution, and the application of more advanced analytic techniques such as machine learning and natural language processing [22,25-28,33-35]. Third, these findings collectively highlight the critical need to bridge the gap between passively collected digital data and clinically interpretable indicators of suicide risk [30,32]. Although diverse data streams and analytic approaches are increasingly used, there remains a lack of structured approaches to translate low-level digital features into meaningful behavioral or clinical constructs. This underscores the importance of developing frameworks that can systematically organize and represent the relationships between raw data and clinically relevant suicide-related outcomes. Further details are provided in Table 3.

Figure 2 presents the LHSF-based conceptual mapping synthesizing the reported associations between passive digital markers and suicide-related outcomes. Low-level features extracted from passive sensors consisted of 7 domains, including location type and variability, movement intensity, sleep parameters, autonomic physiological indices, respiratory rate, smartphone use metrics, and linguistic features.

Across the 4 layers of the LHSF, physiological sensor-derived features were linked to autonomic dysregulation markers in all 4 relevant studies and showed associations with very diverse outcomes, including suicidal ideation, suicide severity, risk, and attempts. Behavioral and mobility-related features derived from the global positioning system and accelerometers mapped to high-level markers that captured social withdrawal, physical inactivity, and sleep disturbance. These markers showed associations with suicidal ideation and suicide-related events. Smartphone use metrics mapped to smartphone use patterns and were linked to suicide-related events and risk indicators, including suicidal ideation and planning. Linguistic features extracted from text messages mapped to high-level markers that reflected expression related to suicide risk and psychological and behavioral patterns. These markers were associated with suicidal ideation, planning, and suicide attempts. Overall, mapping revealed that passive digital markers were represented across multiple hierarchical levels of the LHSF and linked to diverse suicide-related outcomes.

This scoping review proposed a suicide-specific adaptation of the LHSF to integrate fragmented evidence from digital sensing studies into clinically meaningful behavioral markers in suicide research. The adapted LHSF organized diverse digital features into higher-order constructs, such as autonomic dysregulation, sleep disturbance, and psychosocial dysfunction, and provided a conceptual bridge between technical sensor outputs and clinically interpretable suicide-related states.

Autonomic indicators, particularly HRV-related features, were significantly associated with suicide-related outcomes in all 4 relevant studies and achieved the highest predictive performance among the included studies (AUC=0.89) [33]. Beyond prediction, HRV and autonomic measures were repeatedly associated across the broadest range of suicide-related outcomes in correlational studies, including suicide severity, current suicidal ideation, and attempt history [24,29]. However, they should be interpreted alongside other behavioral and contextual markers [24,36]. This pattern is consistent with prior evidence linking autonomic dysregulation to suicide risk [36,37]. Suicidality was associated with dysregulation of the autonomic nervous system, reflected in reduced parasympathetic activity and altered HRV [36]. Electrodermal activity improved model performance for predicting the severity of suicidal ideation; however, it worsened model fit when the outcome was the presence of suicidal ideation [25].

Passive physiological monitoring, including HRV assessment using wearable devices, might complement conventional suicide risk assessments, particularly for individuals who did not accurately disclose their suicidal ideation or behavior. Passively collected physiological data could be processed into HRV indices that reflect autonomic nervous system regulation and dysregulation [38,39]. These physiological changes could be clinically meaningful, as substantial alterations in HRV might be interpreted as indicators of psychological distress. However, as predictive performance varied across studies and outcomes, such measures should be appropriately considered as objective correlates associated with suicide-related risk rather than stand-alone clinical markers.

Mobility- and geolocation-derived indicators suggest that changes in spatial behavior may reflect proximal signals of suicide-related risk, particularly when considered as within-person deviations rather than absolute behavioral levels. Increases in homestay and reductions in mobility were associated with subsequent suicidal events, indicating that disruptions in daily movement patterns might capture short-term risk dynamics [30,32]. These behavioral shifts might represent early manifestations of psychosocial withdrawal and loss of daily structure, both well-established precursors of suicidal crises, rather than reflect isolated symptoms. Therefore, abrupt deviations from an individual’s routine can be interpreted as signals of functional deterioration difficult to capture using conventional self-reporting methods.

However, the predictive value of geolocation features differed across indicators; measures such as entropy and distance traveled were not significantly associated with suicide outcomes [32]. These findings could suggest that mobility-based markers might be sensitive to contextual and individual variability and thus should be interpreted in relation to personalized baseline patterns. From a clinical perspective, this supports an approach that prioritizes within-person changes over static thresholds, reflecting a broader shift toward dynamic and temporally sensitive models for suicide risk assessment [28].

Although smartphone use patterns captured dynamic behavioral changes associated with short-term suicide-related outcomes, their interpretation remained highly context dependent. When individuals spend more time on their smartphones, they are more likely to experience near-term suicidal ideation and planning, suggesting that heightened device engagement might reflect acute psychological states preceding suicide risk [35]. Research indicated significant associations between specific app usage patterns and suicide risk, with certain types of media use being linked to elevated risk under specific conditions [23]. In addition, behavioral changes derived from passively collected smartphone usage data demonstrated the ability to predict suicide-related events over short time frames, including suicide attempts and psychiatric emergency visits [30].

As a result, smartphone-derived behavioral data could serve as real-time proxies for the underlying cognitive and emotional processes. However, digital behaviors were inherently ambiguous and thus might reflect multiple psychological functions, including coping, avoidance, and social interaction. Hence, smartphone-based indicators should be interpreted with caution and combined with other data sources such as EMA or clinical assessments, to improve their reliability and clinical interpretability.

Few sleep-related indicators were significantly associated with suicidal outcomes [31]. Only one actigraphy-derived metric, wake after sleep onset, was significantly associated with next-day suicidal ideation, and only in 1 of 7 participants at the individual level; other sleep parameters were not significantly associated [31]. A likely explanation is methodological rather than substantive. As actigraphy infers sleep from movement, it detects sleep with high sensitivity but poorly identifies wakefulness, so periods of quiet wakefulness are often misclassified as sleep, and metrics such as wake after sleep onset are often underestimated [40]. Self-reported sleep, in contrast, may capture the subjective distress more proximally linked to suicidal states, which could explain why passively sensed sleep metrics may show fewer significant associations. These findings suggest that the associations observed here may reflect current measurement constraints rather than the absence of a meaningful link. Methodological refinements, such as validating movement-based estimates against polysomnography and averaging across multiple nights to obtain more stable parameter estimates [41], together with larger samples and more consistent measurement strategies, may strengthen future findings and clarify the role of sleep-related digital markers.

Linguistic and communication-based indicators derived from smartphone data provided additional insights into suicide-related cognitive and emotional states, although findings varied across specific features. Text-based patterns that reflected suicide-related content were significantly associated with both passive and active suicidal ideation and planning [34]. Furthermore, longitudinal analyses demonstrated that temporal changes in emotion-related language, such as increases in anger and decreases in positive emotions, could differentiate suicide attempts from lower-risk states [22]. These findings suggest that language use may capture shifts in affective and cognitive processes preceding suicidal behavior.

However, not all linguistic features were significantly associated with suicide-related outcomes. Self-focus and sleep-related language patterns did not reliably distinguish between suicide risk states in some analyses [22,27]. Overall, linguistic markers may provide valuable indicators of internal psychological states. However, their clinical utility depends on integrating multiple features and considering variability across individuals and contexts.

These findings highlight the potential of passive digital markers for clinically meaningful suicide risk monitoring. In practice, the continuous monitoring of physiological, behavioral, and digital signals helps expand our assessment beyond static, retrospective viewpoints toward more dynamic and temporally sensitive models of suicide risk [12]. Approaches that prioritize within-person deviations from individual baseline patterns may enable earlier detection of subtle changes preceding suicidal crises, particularly among those who may not disclose suicidal ideation or behavior [14,34].

However, the predictive performance reported across the included studies requires careful interpretation. Among the predictive studies, AUC ranged from 0.56 to 0.89. In clinical screening contexts, AUC values of 0.80 or above are generally considered the minimum threshold for adequate discrimination [42]. By this standard, only one study achieved excellent performance (AUC=0.89) [33], whereas one study approached but did not reach this threshold (AUC=0.78) [30]. The remaining studies reported AUCs below this threshold (range 0.56-0.74) [26,28,32], and one predictive study reported only relative model performance improvements rather than an AUC [25]. These findings indicate that passive digital sensing for suicide risk prediction is still at an early stage and does not yet meet the performance levels required for clinical screening tools. Nevertheless, these AUC values should be interpreted in light of the inherent difficulty of predicting rare events such as suicidal behaviors and the reliance on passive data alone without integration with clinical assessments.

From a clinical perspective, these findings suggest that passive digital monitoring systems may function as continuous and low-burden safety nets and identify emerging risks in everyday contexts. In addition, integrating these digital markers into intervention frameworks, such as Just-In-Time Adaptive Interventions, may facilitate timely and context-sensitive responses, supporting early intervention before suicidal crises intensify [43]. These approaches may be particularly useful in community and outpatient settings, where continuous monitoring using only conventional assessment methods is unfeasible. Specifically, our findings suggest that physiological markers, such as HRV, which showed the strongest predictive performance and the broadest range of associations in our review, combined with behavioral and linguistic indicators, may provide a more context-sensitive approach to suicide risk assessment.

This review has several limitations. First, the relatively small number of included studies (n=14) limits the generalizability of our findings and the ability to establish a robust consensus on digital markers. Second, the lack of standardized data collection and extraction of suicide-related digital features across studies may have contributed to inconsistencies in the findings and limited comparability between studies. This issue is further compounded by simplified device classifications that do not account for variability across device brands, models, or operating systems (eg, iOS vs Android). Third, as a scoping review, this study did not formally assess the methodological quality of the included studies or perform a quantitative synthesis, such as a meta-analysis. Fourth, the exclusion of gray literature may have introduced publication bias. Although peer-reviewed publications were targeted to ensure complete technical descriptions for LHSF mapping, relevant negative or null results may exist outside these databases, potentially narrowing our synthesized evidence base. Fifth, most studies (78.6%) were conducted in the United States, and included samples were predominantly White or Caucasian samples and mostly female participants. This demographic and geographic homogeneity limits the generalizability of the present findings to culturally and demographically diverse populations. These limitations reflect the emerging nature of research in passive digital sensing for suicide-related outcomes and highlight the need for more rigorous and standardized methodological approaches. However, this review provides a structured framework for selecting digital markers for suicide prevention and offers a foundation for future research in digital suicide phenotyping.

Several directions for future research have emerged based on this review. First, future studies should adopt prospective longitudinal or real-time designs to examine multimodal integration, combining passive sensor data with EMA and electronic health record data to enhance clinical context and interpretability. Second, large-scale validation studies using advanced analytic approaches, such as machine learning and natural language processing, are needed, along with rigorous external validation, to improve the discrimination between general psychological distress and suicidal crises. Third, future research should move beyond correlational and predictive modeling and focus on intervention-focused research, such as clinical trials to develop and test actionable digital phenotypes within intervention frameworks, such as Just-In-Time Adaptive Interventions frameworks. These approaches may enable real-time clinical alerts or timely responses when specific digital markers are detected. Fourth, future reviews in the field of digital suicide phenotyping should consider incorporating gray literature to mitigate potential publication bias and establish a more comprehensive evidence base. Fifth, future research should prioritize the inclusion of demographically and geographically diverse samples to strengthen the generalizability and clinical translatability of passive digital markers across cultural contexts.

This scoping review provides a suicide-specific adaptation of the LHSF to organize fragmented evidence on passive digital sensing into clinically interpretable behavioral markers. Physiological indicators related to autonomic regulation were associated with suicide-related outcomes across all relevant studies and showed the strongest predictive performance (AUC=0.89). Furthermore, smartphone-derived behavioral markers demonstrated additional potential for capturing proximal changes in suicidal states, whereas linguistic markers showed more variable findings. In contrast, sleep- and activity-related indicators showed few significant associations. This review clarifies how passive digital features can be translated into higher-order behavioral constructs and advances the conceptual foundation for digital suicide phenotyping. These findings highlight the potential of integrating multimodal digital data to enable continuous, real-time monitoring of suicide risk in natural settings. Future research should focus on improving the reliability, interpretability, and external validity of digital markers to develop clinically actionable systems for early detection and timely intervention in high-risk populations.