Population aging has brought renewed attention to late-life depression. Growth in single-person households, shrinking family networks, and weaker community ties can intensify the risk of depression in older adults, particularly in highly urbanized settings, such as South Korea [1,2]. These shifts also strain clinic-centered care models, which are often inaccessible to older adults living alone and facing mobility barriers, underscoring the need for remote, daily-life–responsive geriatric mental health strategies.

Ecological momentary assessment (EMA) involves the real-time, in situ measurement of behaviors, emotions, and thoughts in everyday contexts [3,4]. By capturing experiences as they occur, EMA reduces recall bias and is well suited to symptoms that fluctuate with context [4]. Advances in mobile and sensor technologies have enabled unobtrusive, scalable EMA deployment, positioning it as a potential infrastructure for practical routine mental health support methods [5,6]. Repeated self-monitoring via EMA may also confer therapeutic benefits, particularly when combined with standard treatments [7-11]; however, effects appear heterogeneous, with reduced responsiveness among individuals with more severe baseline symptoms, highlighting the need to identify who benefits most from EMA-based approaches [12].

Recent EMA research has moved beyond self-reported content to “side data” such as response time (RT), which may index underlying cognitive processes. Chung et al [13] reported a nonlinear (inverted U-shaped) association between EMA RT and depression severity, suggesting that both unusually fast and slow responses may reflect dysfunction (eg, impulsivity or cognitive slowing). Hernandez [14] further showed that EMA RTs correlate with Symbol Search performance, indicating that RTs capture general processing capacity even in noncognitive EMA tasks.

EMA-derived RTs may therefore help identify individuals vulnerable to depressive symptoms. Consistent with Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) descriptions of major depressive disorder, which include observable psychomotor retardation [15], Hernandez et al [16] applied drift diffusion modeling to binarized EMA responses and found that RT-derived parameters (eg, drift rate and boundary separation) systematically tracked traits such as neuroticism and depressive symptoms [16]. Together, these findings position EMA RT as a candidate marker linking cognitive efficiency with affective symptomatology.

This approach may be particularly relevant in geriatric populations, where aging and depression jointly exacerbate psychomotor slowing [17-19]. Older adults with major depressive disorder (MDD) show prolonged initiation and movement times relative to controls, implicating combined executive and motor delays [17], and reviews report especially large effects for psychomotor slowing in late-life samples [20]. Depressed older adults also appear less adaptable under repeated cognitive demands: greater symptom severity has been associated with slower reaction-time learning curves and reduced improvement across training blocks, with sluggish adaptation predicting later functional decline [21,22]. Some interventions may even backfire when effort exceeds coping capacity; for instance, one randomized controlled trial reported worsening depressive symptoms following speed-of-processing training [23]. Collectively, these results motivate EMA-based RT metrics as scalable markers for monitoring cognitive-motor slowing and adaptation in late-life depression.

Despite its potential usage, analyzing RTs from ordinal-scale EMA items presents distinct hurdles. While ordinal items, such as 5- or 7-point Likert scales, offer richer granularity than binary formats, their integration with RT data poses additional analytical considerations and requires further preprocessing steps.

One complication arises from emotional clarity. As shown in Hernandez et al [16], individuals with clearer emotional awareness due to intense emotions tend to respond more quickly to EMA questionnaires. Thus, trait-level factors, such as baseline emotional reactivity, as well as situational factors, including bereavement, acute illness, and exposure to daily stressors, may affect the RT of EMA reports. This can mask or confound the expected slowing associated with cognitive impairment, making RT a less straightforward marker when emotional clarity varies across individuals and between occasions.

A further complication lies in interface variability. Mobile EMA apps often use different visual formats (eg, sliders, buttons, or scales), which can unintentionally influence how quickly users respond. For example, a response option that is most proximal to a frequently used finger can lead to faster responses. These app-specific differences, such as user interface composition, introduce another layer of noise that must also be accounted for when analyzing ordinal-scaled EMA-based RT.

In light of these issues, this study pursues the following objectives. First, we tested whether short-term EMA use (≤4 wk) alongside usual case management is associated with symptom improvement among older adults at elevated risk for depression, evaluating EMA as a potential adjunctive intervention rather than a monitoring-only tool. Second, we examined how multiple EMA-derived features (eg, symptom ratings, response speed, and RT adaptation) relate to symptom change over time, with a focus on RT dynamics as a behavioral signature of improvement. We also introduce a scalable analytic approach that reduces key ecological confounds in RT (eg, momentary affect intensity and interface-related heterogeneity) and is broadly applicable across EMA apps and Likert-type scales. Third, we evaluated the clinical relevance of RT adaptation by using a Bayesian multilevel model to compare adaptive patterns between participants who benefited from EMA-adjunctive care and those who did not. With a modest sample size (n=50), partial pooling in a multilevel framework improves estimate stability while accommodating missing and unbalanced EMA observations [24]. Additionally, Bayesian posterior inference provides interpretable uncertainty for group differences in adaptation parameters, thereby quantifying the evidence for between-group differences in adaptive patterns [24-26]. This research ultimately supports the objective, remote stratification of individuals who are most likely to benefit from EMA-enabled care.

Participants eligible for this study were aged 65 years or older, regardless of gender, and had expressed their willingness to participate through recruitment notices posted at the community-based Suwon Geriatric Mental Health and Welfare Center in Gyeonggi-do, Republic of Korea. Individuals who met the following eligibility criteria were included: (1) participants must have had a prior diagnosis of depressive disorder by a psychiatrist, with no change in medication for the last 3 months and without symptom aggravation; (2) absence of medical or neurological disorders such as dementia, stroke, or Parkinson disease that may affect the research; and (3) current recipients of case management services at the Suwon Geriatric Mental Health and Welfare Center.

All participants who met these criteria provided written informed consent prior to enrollment and completed demographic information and preassessment questionnaires. Those who completed these surveys were then given detailed guidance and support for installing and using the app, with additional assistance provided to those less familiar with digital tools. Given the specific population’s limited experience with digital devices, only individuals who demonstrated the ability to follow the mobile app (BIG4+) usage guidelines independently were enrolled. Participation was entirely voluntary, and participants were free to withdraw at any time without any disadvantage. All procedures were approved by the Public Institutional Bioethics Committee designated by the Ministry of Health and Welfare of the Korean Government (MOHW; P01-202405-01-019) and were supported by Digital Medic Inc. Written informed consent was obtained from all participants. Participant privacy was protected by analyzing all data in deidentified form and restricting access to authorized research staff only. Electronic data were stored in password-protected encrypted files, and physical records were kept in locked storage. Participants received KRW 100,000 (approximately US $67.50) for completing the full study protocol. For early withdrawal, compensation was provided on a prorated basis according to the week of completion.

The study used an observational, single-arm design. Participants were community-dwelling older adults enrolled through the Suwon Geriatric Mental Health and Welfare Center.

Case management was provided throughout the study, consisting of 1 monthly home visit and 1 to 2 interim telephone contacts, during which case managers monitored participants’ daily living conditions and mental health status. In addition to this standard case management, participants were asked to use the Big4+ mobile app as an adjunctive care tool over a 4-week period. The Big4+ app, developed specifically for research purposes, was available on both Android (AOS) and iOS platforms and collected daily self-reports on mood, appetite, sleep quality, and overall well-being using a 7-point Likert scale (1=minimum to 7=maximum).

Each participant visited the center twice: once prior to initiating app usage (baseline assessment, V1) and once at the end of the 4-week period (postassessment, V2). During both visits, participants completed a battery of self-report questionnaires administered in paper format by case managers, with each session lasting approximately 30 minutes.

Standardized psychological scales administered to participants included the Korean version of the Center for Epidemiologic Studies Depression Scale-Revised (CESD-R) [27-29], the Korean version of the 15-item Short Geriatric Depression Scale (GDS-15) [30,31], the 9-item Patient Health Questionnaire-9 (PHQ-9) [32,33], and the Beck Anxiety Inventory (BAI) [34,35].

A total of 50 older adults aged 65 years or older were recruited for the study (mean age 70.6, SD 5.8 y), of whom 72% (n=35) were female. Among the participants, 49 had a documented history of MDD, and 1 had a history of bipolar disorder. To ensure consistency and avoid confounding effects arising from the natural mood fluctuations associated with bipolar disorder, only participants with MDD were included in all statistical analyses (n=49). Baseline psychological assessments indicated clinically elevated symptoms across multiple domains. Preassessment scores were as follows: GDS-15 mean=9.37 (SD 4.42), CESD-R mean=33.94 (SD 18.24), PHQ-9 mean=13.04 (SD 7.40), and BAI mean=22.76 (SD 16.46). Following the study period, scores were as follows: GDS-15 mean=7.22 (SD 4.44), CESD-R mean=22.65 (SD 16.15), PHQ-9 mean=8.59 (SD 6.30), and BAI mean=13.31 (SD 11.52). All participants successfully installed and used the BIG4+ app over the 4-week period, with a mean compliance rate above 93%. All descriptive details are summarized in Table 1.

To evaluate the benefit of EMA-based self-monitoring, preassessment and postassessment scores were compared across 4 psychological scales (GDS-15, CESD-R, PHQ-9, BAI). The results indicated that CESD-R, PHQ-9, and BAI scores violated the assumption of normality (P values =.01, .01, and .01, respectively), while the GDS-15 difference scores did not (P=.33). Consequently, Wilcoxon signed-rank tests were applied to the CESD-R, PHQ-9, and BAI, and a paired-samples t test was used for the GDS-15 (Table 2).

All 4 measures showed statistically significant improvements following the intervention. CESD-R scores decreased by a mean of 11.50 points (SD 15.8; SE 2.26; W=1045.00; P<.001), with a rank biserial correlation of 0.78. PHQ-9 scores decreased by a mean of 4.50 points (SD 6.9; SE 1.00; W=947.00; P<.001), with an effect size of 0.75. BAI scores were reduced by 9.00 points (SD 11.7; SE 1.67; W=940.00; P<.001), with a rank biserial correlation of 0.82. GDS-15 scores showed a significant reduction of 2.14 points (SD 2.8; SE 0.40; t₄₈= 5.30; P<.001), corresponding to a Cohen d of 0.76.

Linear relationships between EMA-derived features and changes in depressive and anxiety symptoms were examined using correlational analyses. A full table of correlational analyses is presented in Multimedia Appendix 2.

Descriptive statistics of EMA scores showed generally small-to-modest associations with mental health change scores (Table 3), with no significant correlations for ΔGDS-15 across items. Significant associations emerged for other outcomes, suggesting item-specific links with symptom change: ΔCESD-R was correlated with the minimum of appetite (r=0.294; P=.04) and general evaluation (r=0.29; P=.04). ΔPHQ-9 was correlated with the median of feeling (r=0.291; P=.04), mean (r=0.325; P=.02), and minimum of appetite rating (r=0.291; P=.04), and the minimum of general evaluation (r=0.286; P=.046). For anxiety, ΔBAI was associated only with the minimum sleep quality rating (r=0.335; P=.02).

Descriptive statistics of raw EMA RTs showed limited associations with symptom change (Table 3). No RT features were significantly correlated with ΔGDS-15, ΔCESD-R, or ΔPHQ-9. The only significant RT associations were observed for anxiety change (ΔBAI), where mean (r=–0.301; P=.04) and maximum (r=–0.35; P=.01) of appetite RT were negatively correlated with ΔBAI, underscoring the challenge of using noisy, naturalistic raw RT descriptives to track depressive symptom change in at-risk older adults.

In contrast, exponential-decay parameters derived from Z-RT (within subject×response-option standardization) showed clearer associative patterns (Table 4). For the feeling item, both the decay rate and asymptote were significantly associated with ΔGDS-15 (θ_b: r=–0.398; P=.005, θ_c: r=–0.321; P=.03), representing the strongest ΔGDS-15 correlations among the RT-derived features and exceeding those observed for EMA score descriptives. Feeling-item parameters also showed significant associations with other outcomes: ΔCESD-R was related to model fit and dynamics (R²: r=0.358; P=.01, θ_b: r=–0.302; P=.04, θ_c: r=–0.305; P=.04), while ΔBAI was associated only with goodness-of-fit (R²: r=0.33; P=.02). Notably, no significant associations were observed for decay parameters fitted to appetite, sleep quality, or general evaluation across change scores (all P>.05), and no Z-RT decay features were significantly related to ΔPHQ-9.

This study provides preliminary evidence for the feasibility and potential efficacy of EMA-based self-monitoring in reducing psychological symptoms among older adults at elevated risk. Across 4 validated scales (CESD-R, PHQ-9, BAI, and GDS-15), participants showed statistically significant symptom reductions over the 4-week period, with medium-to-large effect sizes. These findings align with prior work suggesting that EMA-related benefits may be enhanced when used as an adjunct to treatment as usual, and they extend that evidence to a geriatric context. The high adherence to the EMA mobile app (>90%) further indicates that EMA-based self-monitoring is feasible and acceptable in older adults, including those with limited familiarity with mobile technologies, supporting its promise as a scalable adjunct to depression care in late life.

Correlational analyses also indicated that EMA-derived features can predict variation in symptom change, but with important limitations. Descriptive statistics of EMA scores and raw EMA RTs showed symptom-specific associations across measures, yet they were not significantly related to changes in the geriatric depression scale (ΔGDS-15), suggesting reduced sensitivity for tracking geriatric depressive symptom change using simple descriptives alone. In contrast, features derived from exponential-decay parameters fitted to standardized response times (Z-RT) showed consistent associations with depressive symptom change (ΔGDS-15 and ΔCESD-R). In particular, faster adaptation (higher decay rate), higher postadaptation RT level (asymptote/baseline), and better fit quality for the emotionally salient feeling item are associated with greater symptom improvement, suggesting that RT dynamics during EMA engagement may reflect clinically meaningful change during EMA-adjunctive care.

These patterns were further corroborated by the Bayesian multilevel modeling results. Compared with nonresponders, responders to EMA-adjunctive care exhibited markedly faster adaptation (higher decay rate) and modestly higher postadaptation RT levels. This suggests that potential responders show a distinct trajectory characterized by more rapid accommodation to repeated self-reflection, although the underlying mechanism cannot be determined from the current design (ie, whether EMA-adjunctive care enhances self-monitoring efficiency, or whether individuals with greater baseline capacity for such adaptation are more likely to benefit). Evidence for between-group differences in the postadaptation RT level was comparatively uncertain, and the direction (ie, slightly higher baseline RT in responders) may appear counterintuitive if faster responding is assumed to reflect greater efficiency. One plausible interpretation is that faster late-phase responding among nonresponders could reflect a tendency toward more superficial or less attentive responding rather than more efficient self-reflection, but this remains speculative and warrants direct validation. Additionally, sensitivity analyses across plausible alternative specifications yielded the same qualitative conclusions, supporting the robustness of the central group-difference inference. Collectively, these results highlight the potential clinical use of behavioral signatures embedded in EMA responding, particularly RT adaptation dynamics, for tracking depressive symptom improvement in older adults.

Interestingly, consistent with the correlational findings, group contrasts were strongest when the model was fitted to the feeling item among the 4 EMA domains examined. Moreover, although prior work suggests that averaging RT across multiple EMA items can increase measurement reliability [14,16], group differences in adaptive patterns were larger when using the feeling item alone. This finding suggests that EMA response dynamics may be item-specific rather than purely reflecting a global response-speed trait, raising the possibility that different symptom domains could exhibit distinct behavioral signatures in EMA trajectories.

These findings suggest that EMA-adjunctive care may be both feasible and clinically beneficial for older adults at risk of depression. Digital mental health interventions (DMHIs) are often considered difficult to implement in late-life populations because limited familiarity with digital tools can reduce sustained engagement. In contrast, adherence in this study remained high (on average >90%), indicating strong real-world acceptability. The EMA protocol used here, brief daily reporting on four items, appeared manageable even for participants who might be expected to show reduced willingness or capacity for sustained participation due to depressive burden. To support sustained engagement in real-world deployment, the intended implementation can incorporate an explicit feedback loop, such as clinician-facing dashboards, periodic review during EMA engagements, or automated triggers that prompt timely outreach when a missing response is detected. Importantly, symptom reductions observed across validated scales indicate that EMA use in this context may confer benefits beyond feasibility, supporting EMA-adjunctive care as a promising component of geriatric depression management.

A second implication concerns the objective monitoring of symptom change. Remote psychiatric care still relies heavily on self-reported symptoms, ranging from retrospective standardized questionnaires to daily EMA ratings, both of which remain vulnerable to recall bias and other reporting distortions (eg, mood-congruent evaluation, self-reflective bias). EMA-derived RT metrics are not fully “passive” because they require active responding; however, they provide a comparatively objective behavioral signal that reflects underlying cognitive and psychomotor processes and is less directly shaped by deliberative self-presentation or response framing. The observed association between RT adaptation dynamics and clinically meaningful symptom improvement, therefore, supports RT trajectories as a complementary marker for tracking depressive symptom variation in naturalistic settings, potentially improving reliability when subjective reports are noisy or inconsistent.

Moreover, EMA-RT adaptation features may help make digital mental health interventions more actionable and personalized. As noted in the Introduction section, the effects of ecological momentary interventions are heterogeneous and depend on individual characteristics. Consequently, a practical barrier to scaling digital care is identifying early on who will benefit from a given protocol and who will require additional support. Our results suggest that RT adaptation patterns within an initial monitoring window (approximately 1 month) may provide an early indicator of EMA responsiveness. Such early stratification could guide stepped-care decisions (eg, prompting intensified clinician contact, alternative interventions, or safety monitoring for likely non-responders) while allowing responders to continue with lower-intensity, self-guided support, thereby supporting more efficient allocation of constrained clinical resources and optimizing just-in-time adaptive intervention approaches. We also note that this early stratification window may be shortened through systematic investigation of the optimal decision period. Specifically, building on the observed faster adaptation among responders, future work could quantify within-person day-to-day EMA-RT gradients (or slopes over rolling windows) and evaluate candidate “decision days” (eg, 7, 10, 14, or 21 days) to determine an earlier, reliable point for stepped-care decisions.

Finally, the item-specific links observed between EMA domains and symptom changes suggest a pathway to broaden EMA monitoring beyond depression. Because psychiatric symptoms frequently co-occur, identifying domain-specific behavioral signatures (eg, the feeling-depression linkage observed here) may enable EMA to track multiple symptom dimensions simultaneously, supporting more nuanced, individualized monitoring and intervention planning.

First, regarding the observed effectiveness of the EMA-based self-monitoring, although significant reductions in depressive and anxiety symptoms were detected across multiple validated scales, the absence of a control group precludes definitive causal attribution. Participants were not randomized, and potential concurrent treatments, such as case management, pharmacotherapy, or psychotherapy, were not systematically recorded or controlled, introducing possible confounding influences. Additionally, postintervention assessments were conducted only once at the 4-week mark without follow-up measurements to evaluate the durability of treatment effects. As such, the long-term efficacy and sustainability of EMA-based self-monitoring in older adults with depression remain undetermined.

While several EMA-derived features were shown to be associated with symptom variation, some unexpected patterns emerged in the relationship between self-reported EMA scores and symptom change. Specifically, positive correlations were observed between mood-related EMA scores and symptom change in both depression and anxiety domains. These findings suggest that higher EMA ratings during the intervention period were paradoxically linked to less clinical improvement. One possible interpretation is that individuals with persistent depressive symptoms may have reduced self-reflective precision, leading to inflated or undifferentiated positive reports. Alternatively, healthier participants may have demonstrated more nuanced and accurate self-monitoring. However, this hypothesis requires further empirical validation.

Additionally, given the modest sample size (n<50), the correlational analyses were likely underpowered to detect small associations between symptom-change scores and EMA-derived features; thus, null or weak correlations should be interpreted cautiously. Larger cohorts will be needed to more precisely estimate small effects and confirm the robustness and generalizability of the observed associations, particularly those involving the exponential-decay parameters.

Finally, in examining dynamic EMA response trajectories, the most distinctive group-level differences between responders and non-responders emerged in the fitted parameters of the feeling item, whereas appetite, sleep quality, and general evaluation items yielded less pronounced contrasts. This discrepancy may reflect the unique psychological salience of the feeling item, which directly targets the immediate emotional state, in contrast to the more somatically anchored or abstract content of the other items. Alternatively, the fixed order of EMA item presentation, wherein the feeling item was always administered first, may have contributed to its stronger predictive utility. It is plausible that the first item elicits more cognitive effort or reflects the initiation speed of self-reflective responses, thus capturing more variance in related cognitive functioning. This potential order effect should be systematically examined in future studies to better delineate what EMA-based response latencies truly index in ecological settings.

This study demonstrates the feasibility of using EMA-based self-monitoring as both an adjunctive care and a behavioral phenotyping tool in older adults with a high risk of depression.

The implications for digital health care are manifold. First, the integration of response-time dynamics into mobile health platforms offers a nonobtrusive means of tracking cognitive and emotional engagement, even in populations with limited digital literacy. The method’s compatibility with brief, low-frequency EMA schedules (eg, daily reports, about a month) makes it especially suitable for aging individuals, who may be less tolerant of intensive digital protocols. Second, the use of within-subject standardization and curve-fitting models supports individualized tracking, allowing for the detection of meaningful intra-individual changes. This has significant implications for the early identification of symptom relapse, monitoring of treatment responsiveness, and tailored interventions based on dynamic behavioral signatures.

In sum, the introduced EMA-based modeling offers a novel, sensitive, and pragmatic approach for advancing mental health care in aging populations. Future work should aim to validate these findings in randomized controlled settings, investigate the mechanistic underpinnings of RT adaptations, and explore broader applications in digital psychiatry and geriatric care.