JMIR Mental Health

Interpretable Topic Modeling of Spontaneous Speech in Depression Using Large Language Models: Multilingual Four-Cohort Study

2026-08-17T16:00:19-04:00

Background: Depression is underdiagnosed worldwide, and clinicians rely on interpreting patients’ subjective speech. Qualitative analysis of patient language does not scale, and existing computational approaches describe topics with keyword lists that miss clinical nuance. Objective: We evaluated whether clustering spontaneous speech transcripts with large language models (LLMs) yields clusters whose membership is associated with validated clinical scales across multilingual cohorts, and whether LLMs can render those clusters human-readable through fine-grained natural-language descriptions. We further examined which interview questions yield clusters most strongly associated with clinical status, and how sociodemographic factors relate to cluster membership. Methods: We analyzed spontaneous speech transcripts from 4 independent cohorts: a French general population sample (1338 participants) and 3 clinical samples in Italian (n=116), Chinese (n=52), and Spanish (n=90). Responses to open-ended questions were transcribed, embedded with a multilingual language model, dimensionally reduced, and grouped by density-based clustering. An LLM then summarized each cluster into a natural-language description. Cluster membership was tested for association with validated clinical scales (Patient Health Questionnaire-9, Beck Depression Inventory, Generalized Anxiety Disorder 7-item scale, Athens Insomnia Scale, Multidimensional Fatigue Inventory, and Columbia Suicide Severity Rating Scale), clinician-assigned depression diagnoses, and sociodemographic factors (age, education, and sex). Results: Unsupervised clustering yielded clusters significantly associated with clinical scores in the French, Italian, and Chinese cohorts, with an exploratory association in the smaller Spanish cohort. In the French general population, Patient Health Questionnaire-9 depression scores differed across clusters (η²=0.19, 95% CI 0.17 to 0.24, <.001), as did anxiety (Generalized Anxiety Disorder 7-item scale), insomnia (Athens Insomnia Scale), and fatigue (Multidimensional Fatigue Inventory) scores (η²=0.14 to 0.16, all <.001). In the clinical cohorts, cluster membership was associated with clinician-diagnosed depression in the Italian sample (Cramér =0.74, 0.65 to 0.85, <.001) and with major depressive disorder diagnosis in the Chinese sample (=0.55, 0.36 to 0.78, =.001). A suicide-risk association in the smaller Spanish sample (Columbia Suicide Severity Rating Scale) was unstable and is reported as exploratory (mean =0.27). The feelings-and-sleep question yielded the strongest clinical associations, whereas questions about past or future events yielded small effect sizes (η²≤0.05). Age (η²=0.27, 0.24 to 0.32) and sex (Cramér =0.22, 0.21 to 0.30) were each associated with cluster membership for the “describe your last 24 hours” question (both <.001). Conclusions: Fine-grained LLM-based topic modeling automates the labor-intensive early stages of thematic analysis, in minutes of compute, and recovers core psychiatric constructs across languages. It produces human-readable cluster descriptions. Certain interview questions yield clusters more strongly associated with clinical status than others, and sociodemographic factors shape topic content independently of clinical status, an often-overlooked confound. As the cohorts differ, cross-cohort observations are preliminary without generalizability. Screening, treatment-planning, and decision-support uses remain to be established prospectively.

Technology-Themed Persecutory and Related Presentations of Delusions in the Digital Age: Narrative Review

2026-08-14T16:30:12-04:00

Background: Persecutory delusions have long mirrored prevailing cultural and technological concerns. Beliefs involving implanted devices, internet surveillance, hacked smartphones, algorithmic targeting, and AI-mediated control are increasingly visible in contemporary psychosis. However, we identified no prior review dedicated specifically to synthesizing technology-themed delusional content across historical and contemporary clinical contexts. Objective: This narrative review aimed to trace the historical evolution of technology-themed delusions, characterize their contemporary clinical phenomenology, examine relevant neurocognitive and sociocultural mechanisms, and discuss implications for psychiatric assessment and intervention. Methods: PubMed/MEDLINE, Scopus, and APA PsycInfo were searched without date restrictions, supplemented by backward citation tracking and targeted identification of historical primary sources. Eligible sources included original research, case reports and case series, reviews, and relevant conceptual or historical publications addressing delusional content involving technology. A narrative synthesis approach informed by the Scale for the Assessment of Narrative Review Articles (SANRA) framework was used. The final search was completed on May 1, 2026. Results: Forty-three sources were included in the narrative synthesis, comprising 9 case reports or case series, 14 empirical studies, 8 reviews, and 12 historical or conceptual sources. Technology-themed delusions appeared as pathoplastic variants of enduring persecutory, control, and referential motifs, clustering into 5 overlapping categories: surveillance and hacking beliefs, implant and control delusions, Truman Show–type broadcast phenomena, social media–specific referential ideas, and emerging algorithmic and AI-mediated themes. In the largest contemporary cohort (228 patients with psychosis), technology-themed delusions were described by 104 of the 201 (51.7%) patients with delusional thought content, and the odds of technology-themed delusions rose by approximately 15% per admission year (odds ratio 1.15, 95% CI 1.01-1.31; =.04). Cognitive models of persecutory ideation, the aberrant salience hypothesis, technological unfamiliarity, digital immersion, and wider sociocultural narratives all appeared relevant to these presentations. Pandemic-era conspiracy material, including 5G and vaccine-microchip narratives, further illustrated how culturally available technological explanations may scaffold delusional elaboration in vulnerable individuals. Conclusions: Technology-themed delusions represent contemporary expressions of enduring delusional motifs shaped by digital culture. Their assessment requires psychiatric expertise combined with sufficient technological literacy to distinguish proportionate privacy concerns, overvalued conspiracy beliefs, and fixed delusional convictions. Future research should use prospective cohort designs, cross-cultural comparisons, and targeted evaluation of cognitive behavioral and digital literacy–informed interventions, while monitoring emerging AI-mediated phenomena.

Beyond Usage Time: Rethinking Therapeutic Dose and Effect in Digital Therapeutics

2026-08-14T15:15:11-04:00

Digital therapeutics (DTx) have become increasingly prominent in mental health care, offering scalable, evidence-based interventions. A common assumption underlying their design and evaluation is that greater usage time leads to superior therapeutic outcomes, reflecting an implicit linear dose-response model. However, accumulating evidence challenges this simplified perspective and suggests that the relationship between usage time and clinical benefit is substantially more complex. In this viewpoint article, we critically examine the assumption that higher usage time is a necessary prerequisite for therapeutic success in digital mental health interventions. Drawing on psychotherapy dose-effect research, findings from digital intervention trials, and our own empirical work, we highlight several factors that complicate linear exposure models. These include rapid early response and plateau effects, substantial heterogeneity in user trajectories, motivational and contextual determinants of use, episodic patterns of engagement, and the distinction between on-platform activity and off-platform skill application. The linear dose metaphor, implicitly inherited from pharmacotherapy models, appears conceptually mismatched with learning-based digital interventions. Importantly, equating greater use with greater effectiveness risks conflating exposure with therapeutic mechanism, and may inadvertently promote evaluation frameworks that prioritize duration over meaningful change. Together, these insights suggest that usage time is an imperfect proxy for therapeutic engagement, and that high or continuous use should not be treated as a universal indicator of intervention quality. We argue for a shift toward mechanism-informed and individualized evaluation frameworks that prioritize the quality, timing, and functional impact of intervention use over cumulative exposure.

Digital Markers for Passive Remote Monitoring of Bipolar Disorder: Systematic Review

2026-08-13T16:15:11-04:00

Background: Bipolar disorder (BD) features episodic shifts among mania, hypomania, depression, mixed states, and euthymia. Timely detection of mood transitions is difficult due to infrequent clinical touchpoints. Digital health technologies, including wearables and smartphones, offer a unique opportunity to passively and continuously monitor behavior and physiology that could reflect underlying mood dynamics in real-world settings. Objective: This study aimed to systematically review passively collected digital markers for BD mood states, characterize devices/modalities and analytic approaches, appraise risk of bias, and identify design gaps and priorities for clinical translation. Methods: Following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, we searched MEDLINE, PsycINFO, Scopus, IEEE Xplore, and ACM Digital Library (February 16, 2026). We included peer-reviewed studies of adults with bipolar I disorder or bipolar II disorder (BDI or BDII) that measured passively collected markers and related them to depressive, manic, hypomanic, mixed, or euthymic states. Studies that relied exclusively on active measures (eg, lab tests and ecological momentary assessment) were excluded. Two independent reviewers screened studies, extracted study characteristics and results, conducted narrative synthesis, and assessed risk of bias. Results: Of 23,727 records, 57 studies met criteria. Most enrolled ≤50 participants (n=34, 60%) and monitored ≤365 days (n=46, 81%); 11 out of 57 studies (19%) collected data only in the clinic. Eight digital marker domains emerged: physical activity, heart rate (HR), electrodermal activity (EDA), geolocation, smartphone use, light exposure, sleep, and speech. Consistent patterns linked depression to reduced mobility and social interaction, later/variable sleep, and lower daytime light; mania and hypomania were associated with higher and more variable activity, shorter/advanced sleep, and increased communication. Circadian features derived from sleep/activity repeatedly aided prediction. EDA tended to be lower in depression; HR variability findings were mixed across settings and methods. Keyboard and speech features (eg, timing and prosody) showed associations and performed well in classification models. Twenty-one studies used machine learning; several reported strong performance for episode prediction/classification. However, external validation was usually absent, samples were small, monitoring windows were often short relative to episode timescales, clinical labels were infrequent/misaligned, and missingness was rarely modeled despite likely informativeness. Conclusions: Passive digital markers for BD show promise, with the most robust signals aligning with ( [Fifth Edition]) diagnostic features (sleep-wake patterns, activity, socialization, geolocation, and speech). To move from promise to practice, future studies should adopt longer within-subject monitoring, align label cadence with sensing granularity, standardize features/reporting, preregister analyses, externally validate models, minimize data collection to protect privacy, and expand physiological measurement beyond HR and EDA. These steps are essential to develop reliable, actionable tools for earlier detection and management of BD mood episodes. Trial Registration: PROSPERO CRD42024607765; https://www.crd.york.ac.uk/PROSPERO/view/CRD42024607765

Patterns of Engagement With an AI Conversational Agent for Mental Health and Associations With Anxiety and Depression: Cross-Sectional Study

2026-08-10T14:00:21-04:00

Background: Digital mental health interventions using conversational AI agents are increasingly being adopted as scalable alternatives to traditional care. Engagement is typically measured using volume-based metrics (eg, session counts and total time on a platform). However, these metrics overlook engagement patterns over time, which are not well understood in relation to mental health outcomes. Objective: The purpose of this cross-sectional study was to explore how different patterns of engagement with Mental’s AI conversational agent relate to self-reported depression and anxiety. We aimed to (1) identify and describe engagement profiles based on patterns of interaction depth and temporal consistency, (2) compare depression and anxiety symptoms across engagement profiles, and (3) explore whether engagement profiles were associated with mental health symptom severity. Methods: This cross-sectional observational study linked survey responses to back-end app usage data from 112 Mental app users who completed at least 5 sessions with the conversational AI agent. Engagement profiles were derived using median splits on interaction depth (α parameter) and temporal consistency (Gini coefficient). Depression was assessed using the Patient Health Questionnaire-8 (PHQ-8), and anxiety was assessed using the Generalized Anxiety Disorder-7 (GAD-7). One-way ANOVAs compared symptoms across profiles. Linear regression models examined associations between profiles and symptom severity, adjusting for age, gender, and total duration of use. Results: We identified 4 distinct engagement profiles based on interaction depth and temporal consistency: extended and episodic (profile 1; n=25), extended and consistent (profile 2; n=31), brief and episodic (profile 3; n=31), and brief and consistent (profile 4; n=25). Users in profile 1 (extended and episodic) reported the lowest anxiety (mean 2.68, SD 2.43) and depression (mean 3.48, SD 4.06), while profile 4 (brief and consistent) reported the highest anxiety (mean 10.00, SD 7.03) and depression (mean 10.60, SD 8.75). Significant differences were observed for anxiety (=8.07, <.001, ²=0.18) and depression (=5.47, =.002, ²=0.13). In adjusted models, engagement profile was significantly associated with depression (²=0.16, =2.87, and =.009) and anxiety (²=0.21, =4.04, and <.001). Compared to profile 1, users in profiles 2 and 4 reported significantly higher depression and anxiety. Profile 3 differed from profile 1 for anxiety only (β=3.11, =.047). Conclusions: Users with longer, clustered sessions reported the lowest symptoms, whereas those with brief, evenly distributed use reported the highest symptom levels, suggesting that the structure of engagement may be associated with symptom levels in ways that aggregate usage metrics do not capture. These findings are preliminary and hypothesis-generating, highlighting the importance of considering how engagement unfolds over time and suggesting that pattern-based measurement may improve the understanding of user outcomes in AI-powered mental health care. Future work should examine the directionality of these associations and whether distinct engagement patterns reflect meaningfully different modes of interacting with AI-powered care.

Momentary Mood and Affiliation Following Social Interactions in the Digital Age: Longitudinal Study Investigating Associations With Anxiety and Depression

2026-08-10T11:30:12-04:00

Background: Two-fold increases in the prevalence of youth anxiety and depression over the last two decades have mirrored exponential growth in opportunities for adolescent online social interaction via social media, short messaging service (SMS), and internet text messaging apps on smartphones. However, studies to date of self-reported online social interaction time have produced conflicting results. Understanding the role of dispositional and developmental differences in individuals’ responses to online versus offline social interactions may help elucidate whether and how online social interaction is related to anxiety and depression. Objective: This study aimed to investigate the relationship between older adolescents’ and emerging adults’ (18‐24-year-olds) mental health and (1) objectively measured time spent on smartphones and online social interaction apps, (2) momentary affective and affiliative responses to online and offline social interactions, and (3) the moderating role of developmentally and dispositionally elevated social sensitivity. Methods: Smartphone, social media (eg, Instagram), SMS, and internet (eg, WhatsApp) text messaging app time from participants’ screen use settings, as well as symptoms of anxiety and depression, and social sensitivity, were measured in 190 older adolescents and emerging adults (mean age 20.4, SD 2.2 years). Participants then completed a novel ecological momentary assessment (EMA) capturing affective and affiliative responses to recent online or offline social interactions 3× daily for 1 week. Symptoms of mental health were assessed again after 1 month. Results: Total online social interaction (combined social media and text messaging) app time, but not total smartphone time, was associated with greater anxiety, at both baseline and one month later. Affective and affiliative responses were less positive for online social interactions compared to in-person interactions. Anxiety, but not depression, was associated with feeling less happy, but not less included, after social interactions. Affective and affiliative responses to in-person, but not online, social interactions were negatively associated with depression across the 1-month study period. Finally, social sensitivity moderated the relationship between affective and affiliative responses to social media interactions and depression at baseline. Overall effect sizes were small. Conclusions: These findings emphasize the need to investigate individual factors influencing for whom online social interaction is harmful or beneficial. To do so, this study provides a novel, ecologically valid tool for understanding young people’s momentary responses to online and offline social interactions, as well as initial evidence for stronger associations between in-person than online social interaction responses and mental health for older adolescents and emerging adults. It also introduces evidence of social sensitivity as a potential, developmentally relevant vulnerability to the effects of online social interaction. Further research is needed in younger adolescent populations over longer timeframes.

Should I and Can I Use AI for Forensic Psychiatry Report Writing?

2026-08-07T16:15:11-04:00

The rapid evolution of AI, particularly large language models (LLMs), has renewed interest in their potential role in forensic psychiatry report writing. Recent evidence demonstrates that contemporary LLMs perform well in selected medical knowledge, documentation, and information management tasks and may reduce the administrative burden when deployed under appropriate clinical supervision. However, forensic psychiatric reports differ fundamentally from routine clinical documentation. They constitute expert evidence prepared for legal proceedings and therefore require transparent reasoning, explicit weighing of competing evidence, a robust factual foundation, and personal professional accountability. This viewpoint examines whether AI can and should be used in forensic psychiatry report writing by integrating recent empirical evidence, forensic psychiatry guidance, legal and regulatory frameworks, and emerging governance recommendations. Rather than comparing AI with an idealized human evaluator, the manuscript argues that the appropriate comparison is between 2 imperfect systems of reasoning. Human experts remain susceptible to cognitive biases, omission errors, and disagreement, whereas contemporary LLMs exhibit distinct vulnerabilities, including hallucinations, hidden omissions, probabilistic reasoning, and limited explainability. Although the mechanisms differ, both may ultimately compromise the reliability of expert evidence if left unchecked. Current evidence supports AI for bounded, reversible, and independently verifiable tasks, such as document organization, chronology construction, indexing, transcription, and structured summarization, particularly within secure and validated environments. By contrast, there remains insufficient evidence to support AI-assisted generation or material shaping of psycholegal reasoning, credibility assessments, or final forensic opinions. Because these activities require interpretation, accountability, and reasoning that can withstand judicial scrutiny, they remain fundamentally human responsibilities. The most defensible implementation model is, therefore, one of AI around the report rather than AI writing the report, in which AI serves as a supervised productivity tool while the forensic psychiatrist retains full authorship, accountability, and justification of all substantive conclusions.

Piloting a Digital App Based on the Friendship Bench for Depression and Anxiety in Palestine: Mixed Methods Study

2026-08-05T16:45:10-04:00

Background: The burden of mental disorders is high in conflict-affected populations. In Palestine, we piloted Inuka Coaching, a digital intervention adapted from the Friendship Bench delivered by trained and supervised lay coaches. This paper documents the implementation of the intervention in this highly volatile context after October 7, 2023. Objective: This study aimed to describe the implementation of Inuka Coaching, a digital mental health tool based on task shifting, in Palestine and examine contextual challenges, fidelity to the coaching model, and lessons learned regarding recruitment, retention, and delivery during escalating ethnic cleansing. Methods: Two Palestinian mental health professionals were trained and certified in the Inuka method as head coaches, and subsequently trained 5 lay coaches. Palestinian adults in Gaza and the West Bank were recruited primarily through social media and received up to 4 structured, text-based coaching sessions typically delivered over 4 weeks depending on participant availability and preference. Standardized mental health screening questionnaires (Self-Reporting Questionnaire–20 [SRQ-20] and Posttraumatic Stress Disorder Checklist for the [PCL-5]) were collected at baseline, immediately after the first coaching session, and 3 months after the final session. Session transcripts were reviewed to assess coaches’ fidelity to the Inuka method, and a focus group discussion explored coaches’ experiences with training, delivery, and contextual challenges. Results: Between August 2023 and February 2024, a total of 70 participants were enrolled. Baseline assessments indicated high levels of psychological distress: 95.7% (67/70) scored above the PCL-5 threshold of 31, suggesting likely posttraumatic stress disorder, and 69.4% (43/62) scored in the “at risk” range on the SRQ-20. The effectiveness of the method could not be determined as retention was low, with only 7.1% (5/70) completing the program. Coach fidelity was high, with 94.6% (35/37) of transcripts adhering to all 5 steps of the intervention. Coaches reported positive experiences with the method but identified challenges related to recruitment, session continuity, platform usability, and the need for flexibility during acute crises. Key implementation learnings included the importance of early in-person collaboration and training, culturally sensitive framing, flexible delivery and session structures, and robust support for lay coaches. Conclusions: While digital, task-shifted mental health interventions can be delivered with fidelity in conflict settings, sustaining engagement during escalating violence remains challenging. Future implementations require flexible design, context-sensitive recruitment and retention strategies, adaptive delivery models, and strong support for lay coaches.

Effects of Remotely Delivered and Web-Based Interventions on Depression Severity During the COVID-19 Pandemic: 3-Arm Randomized Controlled Trial

2026-08-04T16:15:12-04:00

Background: The COVID-19 pandemic highlighted a critical need for effective population mental health approaches to target the most prevalent disorders (eg, depression) during periods of elevated community distress. The effectiveness of remotely delivered and web-based interventions should be investigated to identify and innovate high-quality models for population mental health service delivery. Objective: The primary objective investigated the effectiveness of adding Mindfulness-Based Cognitive Therapy for Resilience (MBCT-R)—a live, online, synchronous, remotely delivered, group-based intervention—to Cambridge Health Alliance MindWell (CHA-MW), a web-based population health screening and stratified support program, compared with CHA-MW alone, on depression symptom severity. The secondary objective evaluated adding internet Cognitive Behavioral Therapy (iCBT)—an asynchronous, web-based, individual, digital intervention—to CHA-MW, compared with CHA-MW alone. Methods: Participants (N=97) were randomized in a 2:2:1 ratio to receive MBCT-R+CHA-MW (n=37), iCBT+CHA-MW (n=41), or CHA-MW alone (n=19) in a 3-arm randomized clinical trial, from May 2021 to September 2022 in an urban public safety net hospital outpatient setting. CHA-MW served as a low-intensity control condition. For the MBCT-R+CHA-MW arm, MBCT-R was an 8-session program mildly adapted from MBCT to address COVID-19–related risks for depression. For the iCBT+CHA-MW arm, iCBT was a 6-session curriculum added to CHA-MW. All study procedures, including regular mental health symptom screenings, were conducted remotely or via a web-based platform. The primary outcome was change in depression symptom severity during the 24-week study period using an intention-to-treat approach that used generalized linear mixed-effects models to evaluate the comparative effectiveness of MBCT-R+CHA-MW vs CHA-MW over time. A secondary analysis compared iCBT+CHA-MW vs CHA-MW on depression severity. Completer analyses were conducted (per-protocol 6+ sessions). The secondary outcome was mental health visit utilization frequency during the study period. Results: Both MBCT-R+CHA-MW (mean difference −14.1, 95% CI −21.0 to −7.2) and CHA-MW (mean difference −15.2, 95% CI −21.8 to −8.6) had significant reductions in depression symptom severity, with no statistically significant between-group differences. iCBT+CHA-MW (mean difference −12.7, 95% CI −17.4 to −8.1) also reduced depression symptoms but without between-group differences when compared with CHA-MW. Intervention completion rates were low (MBCT-R: 30% and iCBT: 24%), and completers demonstrated significantly greater reductions in depression severity than noncompleters (mean difference −8.5, 95% CI −16.2 to −0.8). Overall mental health clinician visits by group had no statistically significant differences. CHA-MW had the largest increase in participants with new psychopharmacology treatment visits during the 24-week study (CHA-MW +21%, MBCT-R +10%, and iCBT −5%). Conclusions: MBCT-R+CHA-MW, iCBT+CHA-MW, and CHA-MW were each effective in treating depression, without any intervention demonstrating superiority in intention-to-treat analyses. CHA-MW was as efficacious during the COVID-19 pandemic as more resource-intensive interventions that demanded greater time and effort from participants. Low completion rates for MBCT-R and iCBT during the COVID-19 pandemic may have contributed to these results. Trial Registration: ClinicalTrials.gov NCT04595084; https://clinicaltrials.gov/study/NCT04595084

Correction: The Effectiveness and Mechanisms of Action of App-Based Interventions for Improving Mental Health and Workplace Well-Being: Randomized Controlled Trial

2026-07-31T17:15:11-04:00