JMIR Mental Health

Perinatal Mental Health Detection and Prediction Using Mobile Sensing Data: Systematic Review

2026-09-25T14:45:10-04:00

Background: Perinatal mental health disorders affect approximately 20% of pregnant and postpartum individuals, and are associated with substantial maternal and infant morbidity. Traditional assessment relies on infrequent, subjective self-reports. Mobile devices, including smartphones and wearables, offer opportunities for continuous and objective measurement, but evidence on their assessment utility in perinatal populations remains fragmented. Objective: This review aimed to examine the application of wearable devices and smartphones for detecting and predicting perinatal mental health outcomes, with emphasis on predictive performance, informative features, and methodological rigor. Methods: We conducted a systematic review following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines (PROSPERO CRD420251249218). Six databases (PubMed, Web of Science, Scopus, PsycINFO, IEEE Xplore, and ACM Digital Library) were searched initially in January 2026 and supplemented by an amended search in June 2026, with no publication date restrictions. Evidence was synthesized narratively, and the risk of bias was assessed using PROBAST+AI (Prediction Model Risk of Bias Assessment Tool With AI extension). Results: The initial and supplementary searches yielded 1952 unique records after deduplication, of which 10 studies met the inclusion criteria. The included studies covered postpartum depression, prenatal stress, discrete emotions during pregnancy (eg, happiness, anxiety, and sadness), and maternal loneliness. High discrimination metrics were reported for postpartum depression in individual studies, including a multiclass area under the curve of 0.85, a binary area under the curve of 0.871, and an -score of 0.9872. Heart rate variability, GPS-derived mobility, physical activity, and sleep features were most frequently reported as useful, and their interpretation requires perinatal-specific contextualization. Methodological quality was limited, with 80% (12/15) of PROBAST+AI assessment units rated as having high overall quality concern or risk of bias, mainly due to small samples, limited validation, inadequate handling of missing data, and potential overfitting in the analysis domain. Conclusions: Mobile sensing shows preliminary potential for perinatal mental health assessment, but current evidence does not yet support clinical screening or decision-making, and independent external validation in perinatal populations is currently lacking. Progress toward clinical utility requires broader mental health outcome coverage, larger longitudinal cohorts, standardized analytical and reporting practices, adoption of modeling approaches better suited to perinatal trajectories, independent external validation, and human-centered monitoring designs.

Possible Role and Function of AI Conversational Agents in Dialectical Behavior Therapy for Borderline Personality Disorder: Qualitative Interview Study

2026-09-25T13:30:13-04:00

Background: Borderline personality disorder (BPD) is associated with substantial distress and a high risk for suicide. Individuals with BPD may be unable to access evidence-based treatments like dialectical behavior therapy (DBT). Artificial intelligence conversational agents (AI-CA) are increasingly discussed as scalable tools for mental health support, but little is known about how DBT clinicians understand the possible role of AI-CA in treatment. Objective: This study aimed to explore DBT psychologists’ perspectives on integrating AI-CA into DBT for BPD in the future. Methods: Seventeen psychologists in Sweden, each with at least 1 year of clinical DBT experience (mean 6.4 years, SD 5.6), participated in semistructured interviews as part of this qualitative study. Interviews were conducted in Swedish, transcribed verbatim, and analyzed using reflexive thematic analysis within a constructivist framework. Participants did not test a specific AI-CA. Results: Three main themes were developed from the data. The first main theme, “Who Are We in Therapy?” explored how participants defined AI-CA relationally, positioning it variously as a tool, team member, or supervisor, and a sometimes harmful competitor. How these positionings were configured shaped what AI-CA was seen as allowed to do. The second main theme, “The Stoic Helper,” captured how AI-CA was constructed as an extension of the ideal helper: available, competent, adaptable, and tireless, able to provide support in moments when human therapists could not or preferred not to be present. Participants’ hopes for what AI-CA could become often mirrored qualities they found difficult to sustain in their own clinical work. The third main theme, “The Well-Intended Accommodator,” captured concerns that AI-CA may reinforce dependency and function as a safety behavior by supporting reassurance-seeking rather than autonomy. A central concern was not whether AI-CA could generate validating responses, but whether it could know when validation supports change and when it becomes maladaptive accommodation (AI functional ambiguity). Conclusions: Perceived benefits mainly centered on accessibility and support for DBT skills generalization, whereas key concerns involved alliance disruption, reinforcement of behaviors that would ideally be targeted for change, dependency, and questions regarding responsibility in high-risk situations. Integrating AI-CA into DBT is not only a technical question but a relational and ethical one. How AI-CA is positioned in relation to the therapist, person in treatment, and team shapes which tasks are considered acceptable and what form integration can take. The findings highlight the need for implementation frameworks that account for relational dynamics, treatment-specific considerations, and AI functional ambiguity that may arise when AI-CA operates in complex therapeutic contexts.

Assessing the Need for Mental Health Support From Free-Text Responses: Development and Validation of Language-Based Assessments in Adults With Internalizing Symptoms

2026-09-25T11:30:13-04:00

Background: Machine learning and natural language processing have demonstrated significant potential for mental health assessment: describing your mental health in your own words can offer a more ecologically valid approach than traditional rating scales. However, most models focus on specific diagnoses, conditions, or symptoms, which may prematurely assign labels and potentially reinforce stigma in the context of early-stage mental health screening. Objective: This study develops a language-based assessment model that assesses the need for mental health support based on probed natural language and validates it against best-estimate assessments from multiple experienced psychotherapists. Methods: We analyzed an enriched online sample (n=600 for development and n=212 for validation), in which about half reported experiencing internalizing symptoms (depression or anxiety). Participants described their mental health using open-ended responses regarding (1) mental health, (2) suicidal thoughts, (3) medical history, and (4) depression. The responses were converted into contextual word embeddings using a large language model and entered as predictors in a ridge regression using nested cross-validation. Two to three experienced psychotherapists assessed each participant’s need for mental health support on a scale from 1 (no support needed) to 5 (potential crisis). Their assessments were based on longitudinal clinical data (natural language, validated scales, clinical interview, sociodemographics, and clinical history) and were averaged into a best-estimate assessment for model validation. We used the Sequential Evaluation With Model Preregistration framework, which separates model development from validation in a held-out set to support robust estimations and generalizability. Results: The language-based assessments closely aligned with the best-estimate assessments (=0.82) and showed strong correlations with established clinical rating scales for depression (Patient Health Questionnaire-9), anxiety (Generalized Anxiety Disorder 7-Item Scale), stress (Perceived Stress Scale 10), and suicidality (Inventory of Depression and Anxiety Symptoms; =0.62-0.77). Language-based visualizations of topics and word embeddings showed that low need for support assessments was associated with mentioning well-being and good health, while high assessments were related to depression, anxiety, and suicidality. Conclusions: This study demonstrates that natural language responses analyzed through large language models and machine learning can be used to assess individuals’ need for mental health support in close alignment with best-estimate assessments from experienced psychotherapists. Using less than 5 minutes of respondent time, this approach offers a practical tool for early-stage mental health screening in both clinical and self-guided screening contexts.

Practical Guide to Large Language Models for Information Extraction in Behavioral Health Notes: Tutorial

2026-09-24T16:30:12-04:00

Background: Mental health clinical notes contain decision-critical information often absent from structured electronic health record fields. Large language models (LLMs) can extract clinically relevant signals from narrative text; however, variability in output format, limited reproducibility, and inconsistent evaluation remain barriers to clinical deployment. Despite rapid advances in LLM-based information extraction, clear and reproducible guidance for interdisciplinary clinical teams is limited. Objective: This tutorial aims to present a structured workflow for zero-shot information extraction from mental health clinical notes using locally deployed open-source LLMs. It aims to reduce barriers for clinicians and researchers with limited familiarity with natural language processing (NLP) or LLM-based pipelines. Each stage includes key decision points and examples. The workflow is illustrated on two tasks using synthetic notes: (1) detection of self-injurious thoughts and behaviors (SITB) in pediatric emergency department (ED) notes and (2) antipsychotic medication nonadherence detection in outpatient notes, using schema-constrained outputs and standardized evaluation. Methods: We describe a five-stage zero-shot LLM pipeline: (1) infrastructure setup with local deployment via to prevent protected health information (PHI) transmission; (2) task definition specifying the clinical construct, output format, and evaluation; (3) dataset preparation using synthetic notes; (4) iterative prompt development using a hold-out development set with binary and Likert scale outputs constrained via JSON schemas; and (5) output parsing, normalization, and validation. We generated 300 synthetic notes per task using separate LLMs for generation and evaluation; 200 notes were used for evaluation, and 100 notes (50 positive and 50 negative) were used as a prompt-development set and excluded from final metrics. Evaluation used Large Language Model Meta AI (Llama) 3.2 and Llama 3.3 with deterministic decoding (temperature=0). Performance was assessed using accuracy, precision, recall, and -score; Likert thresholds were optimized using the Youden index with bootstrapped CIs. Results: We demonstrated the pipeline’s functionality using 2 example behavioral health detection tasks. Across both examples, the more capable model (Llama 3.3) performed better than the lighter model used earlier in development (Llama 3.2), and we described how the pipeline’s evaluation and error-analysis steps work in practice. These examples also illustrated 2 useful design choices: requiring the model to output in a fixed format reduced errors, and using a graded rating scale, rather than a simple yes/no format, allowed the detection threshold to be adjusted based on clinical risk tolerance. These results are meant to show that the pipeline works as intended, not to serve as a benchmark of real-world accuracy. Conclusions: A schema-driven, zero-shot LLM workflow can support reproducible extraction of clinically relevant information from narrative notes. Local deployment enables processing without transmitting PHI to external servers. This tutorial provides a transferable methodology for institutional adaptation and validation prior to clinical use. All prompts, code, and datasets are publicly available via Zenodo (European Organization for Nuclear Research [CERN]).

Trajectories of Persistent Exposure to Self-Harm–Encouraging Websites and Subsequent Mental Health Outcomes in US Youth: Longitudinal Cohort Study

2026-09-23T15:45:10-04:00

Background: Self-harm, including suicidal thoughts, self-injurious behavior, and disordered eating, is a major public health concern among adolescents and young adults. Youth increasingly encounter self-harm–related content online, including websites that explicitly encourage harmful behaviors. Although exposure to such content has been linked to poorer mental health, most research is cross-sectional and has not examined longitudinal exposure trajectories. Whether persistent exposure represents a distinct digital risk pattern remains unclear. Objective: This study examined longitudinal trajectories and correlates of exposure to self-harm–encouraging websites over 1 year and tested whether persistent exposure was associated with subsequent suicidal ideation, self-injury, and disordered eating. Methods: We analyzed data from 2197 US youth aged 13 to 22 years who completed waves 2 and 4 of Project Lift Up, a national longitudinal cohort study. Website exposure was categorized into 4 trajectories: none, desist (wave 2 only), new (wave 4 only), and persistent (both waves). Categories reflected reported exposure at the 2 assessment waves rather than the frequency or continuity of website use between assessments. Logistic regression models estimated associations between exposure trajectory and wave 4 suicidal ideation, self-injury, and disordered eating, restricting analyses to participants without the respective outcome at wave 2. Models were adjusted for demographics and wave 2 depression or anxiety symptoms and other mental health indicators. Results: Overall, 26.5% (583/2197) reported exposure to self-harm–encouraging websites at wave 2, and 18.5% (407/2197) reported persistent exposure. Privacy concerns, embarrassment, curiosity, and lack of offline support were associated with greater odds of persistence (adjusted odds ratios [aORs] 1.84‐3.23), whereas accidental exposure was associated with lower odds of persistence (aOR 0.28, 95% CI 0.19‐0.41). Compared with no exposure, persistent exposure was associated with higher odds of suicidal ideation (aOR 1.88, 95% CI 1.24‐2.86; =.003), self-injury (aOR 3.14, 95% CI 2.05‐4.80; <.001), and disordered eating (aOR 2.46, 95% CI 1.54‐3.93; <.001). Adjusted predicted probabilities showed a graded pattern, with persistent exposure associated with the highest predicted probabilities (suicidal ideation: 33.4% vs 21.3%; self-injury: 25.7% vs 10.2%; disordered eating: 11.7% vs 5.3%). New exposure was associated with increased odds of self-injury (aOR 2.32, 95% CI 1.30‐4.15; =.004) but was not significantly associated with suicidal ideation or disordered eating. The desist trajectory was not significantly associated with outcomes. Conclusions: Persistent exposure to self-harm–encouraging websites may represent a distinct digital risk pattern associated with higher subsequent odds of suicidal ideation, self-injury, and disordered eating. Assessing patterns of repeated website exposure and youths’ motivations for seeking this content—not merely whether exposure occurred—may improve the identification of youth at heightened risk and inform digital prevention and clinical intervention strategies.

Psychotherapeutic Competencies in Blended Care Psychotherapy: Development of a Competence Framework

2026-09-16T15:15:11-04:00

Background: Blended care psychotherapy (BCP) combines face-to-face psychotherapy with digital interventions, such as videoconferencing, apps, or virtual reality, which can be incorporated in various ways and at different stages of the therapeutic process. Despite its increasing implementation, conceptual and terminological inconsistencies continue to impede a shared understanding of the competencies required for its effective delivery. Objective: This theoretical paper introduces a framework of psychotherapeutic competencies in BCP, designed to provide conceptual clarity. Methods: The framework was developed in four steps. (1) An exploratory literature search was conducted to identify existing competence and educational frameworks relevant to psychotherapy, digitalization, and educational psychology. (2) These models informed a preliminary theoretical framework. (3) This preliminary framework was critically reviewed and refined in an expert workshop with 5 licensed psychotherapists and researchers experienced in BCP. (4) Lastly, findings from a systematic review of 35 studies on BCP competencies were integrated, resulting in the final 8-domain framework. Results: The resulting framework comprises of BCP (eg, face-to-face, videoconferencing, and apps), (knowledge, skills, and attitudes), as well as 8 (Creation of Setting and Infrastructure; Transfer of Communication and Methods; Therapeutic Relationship; Self-Regulation, Self-Confidence, and Self-Care; Data Security, Legal Issues, Ethics, and Safety; Technological Aspects; Scientific Knowledge and Methods; and Individual and Cultural Diversity). Conclusions: This conceptual contribution aims to strengthen the theoretical foundation of BCP, support competence-based education for psychotherapists, and guide future research toward the empirical validation and refinement of BCP-specific competencies.

Virtual Reality–Assisted Therapy Compared to Cognitive Behavioral Therapy for Patients With Treatment-Resistant Schizophrenia: Assessor-Blind Randomized Controlled Trial

2026-09-16T13:15:11-04:00

Background: Auditory verbal hallucinations (AVH) are among the most disabling symptoms of schizophrenia and often persist despite treatment. Virtual reality–assisted therapies (VRTs) are a new generation of relational interventions for AVH, but comparative evidence against active interventions is currently lacking. Objective: This study aimed to assess whether VRT (9 sessions) is superior to a short course of cognitive behavioral therapy (CBT) targeting AVH (9 sessions) in reducing AVH in individuals with treatment-resistant schizophrenia. Methods: In this assessor-blind, parallel-group randomized controlled trial conducted from 2019 to 2026 at an academic center in Montreal (Canada), adults with schizophrenia or schizoaffective disorder and persistent AVH were either referred by their health care team or self-referred. A total of 136 participants were randomly assigned 1:1 to VRT or CBT, stratified by sex and clozapine use status. Both 9-session interventions targeted maladaptive beliefs and relationships with voices and were administered by trained psychotherapists. The predetermined primary outcome was the evolution of AVH severity over time, measured at baseline, post treatment, and 3 months post therapy using the auditory hallucination subscale of the Psychotic Symptoms Rating Scale. Secondary outcomes notably included the general psychotic symptomatology measured using the Positive and Negative Syndrome Scale. Linear mixed-effects models were used to assess time-by-treatment interactions. Psychotherapy sessions and assessments were conducted primarily in person, with CBT being occasionally delivered via videoconferencing during the COVID-19 pandemic. Results: Participants had a mean age of 40.3 (SD 12.8) years, 63.2% (86/136) were male, and 56.6% (77/136) received clozapine. Intention-to-treat analyses (VRT, n=67; CBT, n=69) showed a significant time-by-treatment interaction favoring VRT (=.013) with a moderate effect size at 3 months post therapy (Cohen =0.614). Both therapies showed significant within-group improvements in the primary outcome, with large effect sizes for VRT (Cohen =0.81 post therapy and Cohen =1.17 at 3 months) and moderate for CBT (Cohen =0.58 post therapy and Cohen =0.39 at 3 months). While there were no between-group differences for secondary outcomes, within-group improvements were observed in psychotic symptoms, emotional regulation, and voice acceptance for both therapies, and VRT also reduced maladaptive beliefs about voices and improved self-esteem. Conclusions: VRT outperformed a targeted short course of CBT in reducing persistent AVH for up to 3 months after the intervention in a North American population with treatment-resistant schizophrenia. Improvements were also seen in some secondary outcomes, such as general psychotic symptomatology, and those were similar for both interventions. Overall, these findings support the value of VRT as a personalized and clinically effective intervention. Trial Registration: ClinicalTrials.gov NCT04054778; https://clinicaltrials.gov/study/NCT04054778

Psychological Therapy in the Age of Large Language Models: Framework for Therapist-Delivered and AI-Supported Functions

2026-09-16T13:00:19-04:00

Large language models are increasingly used within and alongside therapy. As large language models perform more therapy functions, questions arise about what the future might hold for therapists and what they will do. We argue that the enduring therapist role in the age of AI-assisted care is currently best understood through relational, adaptive, and accountability functions. These functions include therapeutic challenge, use of the therapeutic relationship as a mechanism of change, rupture detection and repair, bearing witness to suffering, calibration of pace and treatment burden, and clinical judgment under uncertainty across the broader care pathway. Drawing on psychotherapy theory, digital mental health research, the declarative-procedural-reflective model by Bennett-Levy, and our clinical experience, we propose a clinically informed, hypothesis-generating, relational-adaptive-accountability framework. This framework is intended to support further empirical testing and may have implications for workforce development, supervision, training, and service design.

Adverse Experiences in Brief Meditation Practices: Randomized Controlled Trial

2026-09-14T17:00:34-04:00

Background: Meditation has become increasingly popular in recent decades. However, relatively little remains known about the prevalence of and risk factors for adverse experiences related to a single meditation practice. Objective: The objective of our study was to examine adverse experiences associated with 3 brief, digitally delivered meditation practices (mindfulness, self-compassion, and gratitude) relative to using the internet as usual, as well as to investigate whether preintervention characteristics could predict such outcomes. Methods: In a secondary analysis of a randomized controlled trial using samples that were representative of the US and UK adult populations with regard to ethnicity, sex, and age, we examined adverse experiences associated with 3 brief (ie, 5 or 10 minutes) meditation practices (ie, mindfulness, self-compassion, and gratitude) relative to using the internet as usual. We also investigated the potential of using preintervention characteristics to predict such outcomes. Results: A total of 5049 participants completed all preintervention measures and were randomly assigned to meditation or control conditions. Across the sample, 4.1% (204/4925) of participants reported having a distressing experience during the intervention, and 7.1% (348/4908) of participants experienced an increase in negative affect from before to after the intervention. The results showed that participants who were randomized to a brief meditation intervention were no more likely to report a distressing experience than those who were randomized to use the internet as usual (odds ratio [OR] 1.05, 95% CI 0.76-1.47; P=.76). The results also showed that participants who were randomized to a brief meditation intervention were less likely to report clinically relevant increases in negative affect relative to using the internet as usual (OR 0.63, 95% CI 0.50-0.80; P<.001). Notably, participants in the 10-minute condition had a significantly higher likelihood of reporting a distressing experience than those in the 5-minute condition (OR 1.42, 95% CI 1.07-1.89; P=.02). Preintervention characteristics showed acceptable discrimination ability to predict a distressing experience (area under the curve=0.73) and slightly lower ability to predict increased negative affect (area under the curve=0.67). Conclusions: Taken together, we found that the brief, digitally delivered meditation practices tested in this study carry risks of adverse experiences that are comparable to or lower than those of typical activities on the internet; 10-minute condition was more likely to result in distressing experiences than 5-minute condition; and adverse responses to a brief meditation practice can, at least to a certain degree, be predicted using preintervention characteristics. Trial Registration: Open Science Framework 94HKS; https://osf.io/94hks/overview

Tool or Companion? Reframing Conversational AI to Prevent Psychological Harm

2026-09-10T15:30:13-04:00

People increasingly turn to conversational AI for companionship, emotional support, and well-being, using both purpose-built companion apps, such as Replika and Character.AI, and general-purpose assistants, such as ChatGPT and Claude. While some evidence suggests potential benefits, including short-term reductions in loneliness and mood improvement, several adverse outcomes have been reported in both clinical and nonclinical populations, including emotional dependence, exacerbation of symptoms, and self-harm. The fluent and apparently empathic responses from these models lead users to engage with them not only as tools but also as if they were social entities. This framing is conceptually misleading and may pose risks across different user profiles, particularly for vulnerable individuals. Drawing on research in AI, psychiatry, psychology, and network science, we highlight mechanisms through which emotional reliance develops and the boundary between tool and companion erodes. Design choices that evoke personality and warmth encourage users to anthropomorphize these systems. Simulated empathy, generated through probabilistic language patterns rather than genuine emotional experience, creates a structurally asymmetric interaction, in which the user discloses and the system responds, but without reciprocity, vulnerability, or accountability. Overvalidation and sycophancy can reinforce maladaptive cognitions, delusional ideation, and distorted perceptions of reality, as they tend to reinforce people’s beliefs, even at the expense of the accuracy of models’ responses. These mechanisms are not incidental: they emerge from alignment procedures that reward responses perceived as warm and empathic. The result is a self-reinforcing feedback loop between the model and the user that may amplify maladaptive beliefs, delusional ideation, and emotional distress, even in those who engage for largely functional purposes. Understanding these dynamics requires an examination of both what these agents can do—considering their technical limitations and implementations—and what humans believe they can do, including social and psychological impacts. We argue that conversational AI should be treated primarily as a tool supporting human systems rather than as a substitute for human relationships. Perhaps more importantly, reviewing the current hype surrounding AI interactions can help reformulate a paradigm that contributes to human well-being and societal value, while minimizing misconceptions, maladaptive interactions, or social disintegration.