This retrospective study analyzed secondary data from Yuna, an AI-powered DMHI. The dataset included message-level AI-derived emotional state data and session metadata from real-world Yuna users collected as part of routine platform operations. Users were included if they (1) were adults (aged 18 years or older), (2) were located in a country where Yuna is available, and (3) initiated at least one session with the platform. Eligible participants were identified through a retrospective cohort sampling approach using back-end app use records collected between April 17, 2025, and January 11, 2026. As this study was retrospective in nature, no direct recruitment took place. A formal power analysis was not conducted; the analytic sample reflected the full population of eligible users available within the observation window. This study was conducted and reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines for observational cohort studies (Multimedia Appendix 1) [22].

The study protocol was reviewed by Solutions IRB (study ID 1001) under expedited review procedures and was conducted in accordance with US human subject research regulations. The study involved a retrospective analysis of preexisting, deidentified data with no direct interaction or intervention with users. Investigators did not have access to identifiable private information, and informed consent was waived by the institutional review board due to minimal risk. User data were collected during routine app use, and users agreed to the platform’s terms and privacy policy, which permit the use of deidentified and/or aggregated data for service improvement and research purposes. Participants were not compensated for using the app. All data were deidentified before analysis.

Yuna is a commercially available mental health and wellness platform offered through workplace partnerships. Yuna’s primary feature is a clinician-informed, evidence-based AI-powered mental health coach that provides users with unlimited, on-demand conversational support to improve well-being. The platform incorporates personalization features and is designed to teach users skills from evidence-based frameworks, including dialectical behavior therapy, cognitive behavioral therapy, and acceptance and commitment therapy. During conversations, the system dynamically applies discrete intervention strategies based on the user’s expressed needs and conversational context. All therapeutic approach prompts and clinical protocols were developed by a licensed clinician and undergo iterative refinement through ongoing clinician review. High-priority safety protocols continuously monitor conversations for high-risk content; upon activation, standard coaching is suspended, and users are provided with crisis de-escalation support and emergency resources (eg, 988 Suicide and Crisis Lifeline in the United States).

Yuna’s emotion detection system generates message-level scores across 36 fine-grained emotional states using a machine learning model grounded in semantic space theory [23,24]. Semantic space theory conceptualizes emotion as a high-dimensional, continuous space in which dozens of distinct affective states, such as anxiety, contemplation, and calmness, are organized along gradients of meaning rather than being collapsed into a small number of discrete categories [24]. Consistent with this framework, the emotion detection system produces a continuous score for each of the 36 emotional states at the message level, reflecting the degree to which linguistic patterns associated with each state are present in user-generated text within a given conversational turn. This approach enables granular, continuous characterization of emotional expression across the arc of a session rather than relying on binary or categorical emotion labels. A complete listing of fine-grained emotional states and their assigned categories is provided in Multimedia Appendix 2. Yuna’s implementation of this approach is proprietary, and implementation-specific performance metrics such as classification accuracy or agreement with human raters have not been independently established or published. The emotional states analyzed in this study should therefore be understood as model-generated linguistic classifications rather than validated measures of users’ internal emotional experience.

When users sign up for Yuna, they may optionally provide their age and gender within their account profile. Country was derived from back-end platform data.

Descriptive statistics characterized the analytic sample, session-level emotional patterns, and message-level AI-derived emotional state distributions. Frequencies and proportions were reported for categorical variables; means and SDs were reported for continuous variables.

For each message, the dominant emotional state was defined as the AI-derived emotion with the highest score. Session-level emotional trajectory types were then operationalized based on the dominant AI-derived emotional category at session start and session end. Five mutually exclusive trajectory classes were derived from start-to-end category transitions. Reduced distress trajectories were defined as sessions beginning in a distress-associated state and ending in any nondistress state (cognitive processing, regulated positive, or social-interpersonal), reflecting movement away from aversive affect by the end of the session. Persistent distress trajectories were defined as sessions beginning and ending in distress-associated states, reflecting no movement away from aversive affect over the session. Escalating distress trajectories were defined as sessions beginning in any nondistress state and ending in a distress-associated state, reflecting movement toward aversive affect by session end. Stable trajectories were defined as sessions beginning and ending within the same nondistress emotional category (ie, cognitive processing, regulated positive, or social-interpersonal; sessions beginning and ending in distress-associated states were classified as persistent distress), reflecting affective consistency across the session. Transitional processing trajectories captured all remaining transitions (ie, sessions that began and ended in different nondistress categories), reflecting movement across affective states without a clear distress arc in either direction. Session-level proportions of each emotional category were computed as the mean proportion of messages assigned to that category within a session.

Group comparisons between reduced distress and persistent distress sessions were conducted using linear mixed-effects models on session-level engagement variables (message count, AI-derived emotional switches, unique AI-derived emotional states, and session duration) and affective composition (cognitive processing proportion and regulated positive proportion). A mixed-effects logistic regression model was then fit to identify conversational predictors of reduced distress trajectories relative to persistent distress trajectories among sessions beginning in distress-associated states, with AI-derived emotional switches, unique AI-derived emotional states, and session duration entered simultaneously. Every model included a random intercept for user to account for the nonindependence of sessions nested within users. Model fit was evaluated using a likelihood ratio test comparing the full model with a null model containing only the intercept and the random intercept for user.

Missing data were not imputed. Analytic models were fit using complete cases for each relevant outcome and predictor, and the proportion of missing data for each variable is reported. All analyses were conducted in R (version 4.6.1; R Foundation for Statistical Computing) with the lme4 (version 2.0.6), lmerTest (version 3.2.1), and broom.mixed (version 0.2.9.7) packages.

Fit Minded uses a standardized bias mitigation protocol across all industry-partnered research, applied prospectively and independent of study outcomes. The following safeguards were implemented for this study. First, research questions were developed by the Fit Minded scientific team based on scientific merit and were not prescribed by the client. Second, study outcomes were not guaranteed as contractual deliverables. Third, the analysis plan was prespecified prior to data access. Fourth, data analysis and interpretation were led by Fit Minded authors operating independently of Yuna. Fifth, findings were reviewed collaboratively across the full author team.

The initial dataset included 35,687 conversation sessions (unique user-session pairs) and 375,251 message-level AI-derived emotional states. Message records with a missing session identifier (565/375,251, 0.15%) were excluded before analysis as these messages could not be linked to a session. Sessions containing fewer than 3 messages (6817/35,687, 19.1%) were additionally excluded because a minimum of 3 sequential observations is required to detect a directional change in AI-derived emotional state (ie, an initial state, a transition, and an end point) [26]. The final analytic sample consisted of 28,870 sessions (28,870/35,687, 80.9%) contributed by 4463 unique users and included 365,088 message-level AI-derived emotional observations (365,088/375,251, 97.3%).

Demographic data availability varied across variables (Table 1), which is common in real-world digital health research [27]. Gender data were available for 84.3% (3761/4463) of users, country and language data were available for 94.3% (4207/4463) of users, and age data were available for 4.7% (209/4463) of users. More users who reported gender identified as female (2304/3761, 61.3%; 2304/4463, 51.6% of the total sample), followed by male (1457/3761, 38.7%; 1457/4463, 32.6% of the total sample). Most users were located in the United States (2639/4207, 62.7% of respondents; 2639/4463, 59.1% of the total sample), although the sample included users from 104 countries. English was the primary language for most users with available data (3880/4207, 92.2%; 3880/4463, 86.9% of the total sample), with 30 languages represented across the sample. User-level differences between those with and without demographic data are provided in Multimedia Appendix 3.

Session-level descriptive statistics, including message count, unique AI-derived emotional states, and AI-derived emotional switches, are reported in Table 2. AI-derived emotional states were organized into 4 broad categories: cognitive processing, distress associated, regulated positive, and social-interpersonal. At the message level, cognitive processing states were the most frequently observed category, followed by distress-associated, regulated positive, and social-interpersonal states (Table 2). The most frequently observed AI-derived emotional states included calmness, contemplation, disappointment, interest, and sadness. Frequencies for all AI-derived emotional states are provided in Multimedia Appendix 4.

Session-level analyses examined how AI-derived emotional categories shifted across the arc of each session, yielding 5 emotional trajectory types. Stable trajectories, defined as sessions beginning and ending within the same broad nondistress AI-derived emotional category, were the most common, followed by reduced distress trajectories, which reflected movement from distress-associated states toward cognitive processing or regulated positive states. Transitional processing trajectories, reflecting movement between nondistress categories, were the next most frequent, followed by persistent distress trajectories, defined as sessions beginning and ending within distress-associated states. Escalating distress trajectories, reflecting movement from nondistress states toward distress-associated states by session end, were the least common. Proportions for each trajectory type are reported in Table 2.

Among sessions beginning with distress-associated AI-derived emotional states, session duration data were available for 13.8% (1597/11,610) of sessions. Sessions with missing session duration data (10,013/11,610, 86.2%) were excluded from aim 2 analyses as session duration was a required predictor in the mixed-effects logistic regression model. Sensitivity analyses comparing complete and incomplete session data and re-estimating the model on the full available sample of sessions beginning in distress-associated states (n=11,610 sessions from 2925 users) are provided in Multimedia Appendix 5. The pattern of missingness in session duration was consistent with broader data availability limitations observed across the full analytic sample. Among sessions with complete session duration data (n=1597), 73.3% (1170/1597) transitioned toward nondistress states by session end. In comparison, 26.7% (427/1597) remained in distress-associated AI-derived emotional states throughout. Relative to persistent distress sessions, reduced distress sessions contained more messages, more AI-derived emotional switches, more unique AI-derived emotional states, longer session duration, and greater proportions of cognitive processing and regulated positive AI-derived emotional states (P≤.02 in all cases; Table 3). Adjusted intraclass correlations (intraclass correlation coefficients) for these models ranged from 0.13 to 0.27, and no model had a singular fit.

A mixed-effects logistic regression model examined whether the number of AI-derived emotional switches, the number of unique AI-derived emotional states, and session duration were associated with movement away from distress-associated AI-derived emotional states (ie, cognitive processing, regulated positive, and social-interpersonal) among sessions beginning with AI-derived distress (Table 4). More unique AI-derived emotional states were associated with significantly greater odds of reduced distress trajectories relative to persistent distress trajectories (odds ratio [OR] 1.16, 95% CI 1.09-1.24; z=4.41; P<.001; Table 4), indicating that sessions containing a broader range of unique AI-derived emotional states were more likely to transition away from AI-derived distress-associated states. Longer session duration was also positively associated with AI-derived reduced distress trajectories (OR 1.01 per additional minute, 95% CI 1.00-1.02; z=2.01; P=.04; Table 4). More AI-derived emotional switches were associated with lower odds of AI-derived reduced distress trajectories (OR 0.97, 95% CI 0.95-0.99; z=−2.95; P=.003; Table 4). The adjusted intraclass correlation coefficient for this model was 0.10 (latent scale).

The purpose of this study was to explore patterns in AI-derived emotional states and characterize session-level emotional trajectory types within real-world AI-powered DMHI sessions. Across a large, geographically diverse sample of real-world sessions, 5 session-level emotional trajectory types were identified, with stable and reduced distress trajectories accounting for the largest shares of sessions. Among sessions beginning in AI-derived distress states, nearly three-quarters (1170/1597, 73.3% of those with available session duration data) demonstrated reduced distress trajectories by the end of the session, and these sessions were characterized by more unique AI-derived emotional states, more AI-derived emotional switches, and longer session duration relative to sessions with persistent distress trajectories. Unique AI-derived emotional states and session duration were positively associated with reduced distress trajectories, while AI-derived emotional switches were negatively associated with reduced distress trajectories.

Yuna users were primarily female, consistent with broader patterns of female overrepresentation observed in real-world DMHI samples [28]. Most users were located in the United States, although the sample spanned 104 countries and 30 languages, consistent with evidence that AI-powered DMHIs may appeal to a broad, globally distributed user base [29]. Cognitive processing states were the most frequently observed AI-derived emotional category, accounting for over one-third of all message-level observations, followed closely by AI-derived distress-associated states. This distribution is consistent with theoretical accounts of AI-powered DMHI engagement, in which users are expected to alternate between expressing distress and engaging in reflective, meaning-making processes. The relatively higher proportion of AI-derived cognitive processing states may suggest that sessions were characterized more by active processing than by sustained AI-derived distress [7], although this interpretation should be made with caution given that the AI-derived emotional states are model derived rather than self-reported. At the session level, stable and reduced distress trajectories were the most common AI-derived trajectory types, together accounting for approximately half of all sessions, broadly consistent with evidence that engagement with AI-powered DMHIs may be associated with positive emotional and symptomatic outcomes [7,30]. While the present study cannot establish clinical improvement as trajectory types reflect AI-derived behavioral patterns rather than validated symptom outcomes, the prevalence of stable and reduced distress trajectories across sessions suggests that most users did not end the sessions in a worse AI-derived emotional state than they began. Of those who presented with escalating distress trajectories, it is important to note that this classification reflects a session-level categorical transition toward distress-associated linguistic patterns by session end rather than an indication of clinical deterioration or crisis-level distress. Users beginning sessions in neutral or positive states may naturally shift toward distress-associated states as they engage with emotionally salient content, which may reflect adaptive emotional processing rather than worsening [31,32]. Future research should examine whether AI-derived trajectory types correspond to user-reported emotional outcomes within the same session, which would strengthen the reliability of AI-derived emotional trajectories.

Among sessions beginning in AI-derived distress states, approximately three-quarters (1170/1597, 73.3% of those with available session duration data) demonstrated reduced distress trajectories by the end of the session. This suggests that AI-powered DMHIs such as Yuna may be associated with movement away from AI-derived distress, which is encouraging for the potential utility of these tools in real-world mental health contexts [7]. However, it is important to note that sessions beginning in AI-derived distress-associated states have a constrained range of possible trajectories. By definition, they cannot begin in a worse AI-derived emotional state than distress, making upward movement more likely regardless of intervention. This directional constraint is analogous to regression to the mean and should be considered when interpreting the proportion of reduced distress sessions. Moreover, given that trajectory types reflect AI-derived behavioral patterns rather than validated self-reported outcomes, direct conclusions about user emotional experience or clinical benefit cannot be drawn from these data alone. Sessions characterized by reduced distress trajectories presented with more unique AI-derived emotional states, more AI-derived emotional switches, and longer session duration relative to persistent distress sessions, although effect sizes for all 3 engagement variables were small. Notably, the direction of the association between AI-derived emotional switching and trajectory type differed across analyses. While reduced distress sessions showed more switching on average in the unadjusted group comparisons, switching was negatively associated with reduced distress trajectories in mixed-effects logistic regression after accounting for emotional diversity and session duration, suggesting that the relationship between switching and trajectory outcome is not independent of other session characteristics. Higher proportions of AI-derived cognitive processing states such as contemplation and interest may reflect movement from distress toward reflective engagement, consistent with theoretical accounts in which cognitive engagement is considered an active component of distress reduction [33,34]. Similarly, higher proportions of AI-derived regulated positive states such as calmness may reflect movement toward affective stabilization within the session. These patterns are consistent with session-level emotional conditions that have been theorized to accompany distress reduction, although the observational design and AI-derived nature of the emotional states preclude any direct mapping onto established therapeutic mechanisms. Whether these AI-derived patterns reflect genuine therapeutic processes, user behavioral tendencies, or artifacts of the emotion detection model cannot be determined from the present data alone.

Mixed-effects logistic regression analyses identified 3 conversational features associated with reduced distress trajectories among sessions beginning in AI-derived distress states. A greater number of unique AI-derived emotional states was positively associated with reduced distress trajectories, consistent with theoretical accounts in which access to a range of emotional states is considered a marker of productive emotional processing rather than sustained focus on a single state [33]. Longer session duration was also positively associated with reduced distress trajectories, consistent with evidence that greater engagement with DMHIs is associated with fuller realization of therapeutic benefits [35]. The low-barrier, on-demand nature of AI-powered DMHIs such as Yuna may support this kind of sustained emotional engagement as users can interact at moments of heightened emotional salience without the scheduling or stigma barriers of traditional care, although causal claims cannot be made from observational data. In contrast, a greater number of AI-derived emotional switches was negatively associated with reduced distress trajectories after accounting for unique AI-derived emotional states and session duration. The interpretation of this finding is less straightforward. More AI-derived emotional switching could reflect emotional instability or dysregulation but may alternatively reflect active emotional processing that did not resolve within the session window. Accounts linking rumination and emotional reactivity to sustained distress offer one plausible theoretical frame [36,37], but this explanation remains speculative in the absence of concurrent self-report data [38,39]. The clinical meaning of emotional switching in AI-powered sessions requires prospective examination. Future research should examine whether patterns of AI-derived emotional switching correspond to self-reported emotional experience or clinician-rated markers of dysregulation, which would clarify the clinical relevance of these AI-derived emotional states.

This study contributes to an emerging body of work demonstrating that AI-derived emotional states may capture meaningful behavioral patterns in mental health contexts without requiring clinical outcome data [40,41]. Rather than treating AI-derived emotional states as ground-truth representations of internal experience, the present study operationalizes them as behavioral proxies that vary systematically with session-level conversational features. This framing positions AI-derived emotional trajectory analysis as a process-level complement to traditional pretest-posttest evaluation, addressing a widely noted gap in DMHI research. As AI-powered platforms scale to serve large, geographically diverse populations, the capacity to characterize emotional dynamics at the session level without burdening users with additional self-report measures is a meaningful methodological advance. These findings suggest that AI-derived emotional trajectory data may offer a scalable, low-burden signal for understanding engagement quality in real-world mental health contexts. For platforms such as Yuna operating across large, geographically diverse populations, this capacity has practical implications for platform monitoring, early safety surveillance, and the design of future process-level research that moves beyond pretest-posttest symptom end points toward understanding how and when change occurs within sessions.

This study had several strengths. First, the use of AI-derived emotional states to characterize session-level emotional trajectory types is a methodological advance in process-level evaluation of AI-powered DMHIs, providing a novel empirical foundation for understanding how emotional patterns unfold within real-world mental health sessions rather than relying solely on pretest-posttest symptom outcomes. Second, the large, naturalistic sample of 28,870 sessions contributed by 4463 users across 104 countries enhances the generalizability of these findings and reflects the real-world diversity of individuals who engage with AI-powered DMHIs. Third, the AI emotion detection framework underlying Yuna’s emotional state classifications is grounded in a peer-reviewed, empirically validated emotion detection approach [24], although independent validation of Yuna’s specific implementation remains an important gap.

This study also had limitations. First, the observational design lacked a control group, precluding causal inference about the role of AI-powered engagement in producing the observed trajectory patterns. Future research using experimental designs is needed to isolate platform contributions to session-level AI-derived emotional patterns. Second, AI-derived emotional states are model-generated signals that have not been independently validated against self-reported emotional experience in this sample. Without concurrent self-report data, it is not possible to confirm that AI-derived trajectory classifications correspond to how users actually experienced the session emotionally. Future research should incorporate within-session self-report measures to establish the correspondence between AI-derived and user-reported emotional experience. Third, the 5 trajectory types were operationalized based on session-level starting and ending AI-derived emotional categories, which, while transparent and reproducible, may not capture the full complexity of AI-derived emotional dynamics within a session. Alternative operationalizations such as latent trajectory modeling or time-series approaches may yield different AI-derived emotional trajectory types, and replication using different classification approaches is warranted. Fourth, missing data, particularly for session duration and demographic variables, were extensive. Missing data are a common limitation of retrospective, real-world studies of commercial DMHIs; data collection often varies across platform type, back-end structure, and user priorities, resulting in incomplete data capture across variables [42]. Sessions missing duration data differed systematically from complete sessions, showing fewer messages and less emotional variability. Although the direction and significance of regression results were robust to inclusion of sessions with missing duration data, the associations were larger in magnitude in the full sample available than in the complete-case sample. This suggests that restricting the primary analyses to sessions with duration data may have attenuated the magnitude of these associations relative to all distress-starting sessions. Similarly, demographic data were substantially missing. As a result, the ability to characterize the sample fully or examine whether trajectory patterns varied across demographic subgroups was limited. Missingness analyses revealed small to moderate differences in key variables between those with and without complete demographic data. Additionally, response options to “What is your gender?” aligned with biological sex (eg, male or female) rather than gender identity; this conflation of sex and gender limits the interpretability of this variable. Future research should distinguish sex and gender as constructs; use response options appropriate to whichever is being measured; and examine whether AI-derived trajectory patterns differ across age, gender, and location. Fifth, this study could not account for a range of potentially confounding user-level characteristics. Baseline mental health severity, prior treatment history, length of time using the platform, and prior engagement history were not available in the dataset, and each could plausibly influence both how users engage conversationally and the emotional trajectories their sessions follow. The associations reported here should therefore be read as descriptive rather than as evidence of a specific mechanism. Sixth, and relatedly, the trajectories analyzed here are outputs of a single proprietary classification model. Apparent shifts between emotional categories may partly reflect properties of that model, including its sensitivity to particular lexical cues, its behavior on short or ambiguous messages, and any systematic drift in classification across the course of a conversation rather than genuine psychological change in users. Distinguishing model behavior from user experience will require validation against concurrent self-report.

This study provides a novel, exploratory characterization of AI-derived emotional state patterns and session-level trajectory types across a large, geographically diverse sample of real-world users of Yuna, an AI-powered DMHI. Most sessions were characterized by stable or reduced AI-derived distress trajectories, and sessions beginning in AI-derived distress most commonly moved toward nondistress states by session end. A greater number of unique AI-derived emotional states and longer session duration were positively associated with reduced distress trajectories, while more AI-derived emotional switching was negatively associated with reduced distress trajectories. Importantly, the trajectory types described here are AI-derived behavioral indicators generated by a proprietary emotion classification model; they are not validated clinical outcomes and should not be interpreted as evidence of symptom change or therapeutic benefit. With that constraint in mind, these findings lay an empirical foundation for process-level evaluation of AI-powered DMHIs and point toward session-level AI-derived emotional patterns as a potential avenue for understanding how these tools may support users in distress. Future research incorporating self-report measures, longitudinal designs, and experimental methods is needed to evaluate whether AI-derived emotional trajectory patterns carry a meaningful signal for user outcomes in real-world mental health contexts.