JMIR Mental Health

Incorporating Objective Behavioral and Biological Measures in a Large-Scale Decentralized Mobile Health Trial for Depression: Randomized Controlled Trial

2026-10-09T17:30:26-04:00

Background: Decentralized (ie, fully remote) clinical trials (DCTs) offer a scalable and ecologically valid approach to evaluating digital mental health interventions (DMHIs). Yet, incorporating objective measures of well-being within DCTs poses novel methodological and feasibility challenges that can limit the rigor and precision of data collection in the field. Objective: We recently conducted the Behavior, Biology, and Well-being (BeWell) Study, a large-scale, 3-arm randomized DCT evaluating the Healthy Minds Program app (Humin)—a self-guided DMHI that includes didactic well-being information and meditation components. The current paper’s objective is to describe the methodological approach to collecting and processing a multimodal battery of psychological (questionnaires and structured clinical interview), biological (blood and stool assays), and behavioral (stimulus-elicited affect [SEA] task, video log, and Mnemonic Similarity Task [MST]) measures, longitudinally across baseline, the 4-week intervention period, and 17-week follow-up. Secondarily, we evaluate its feasibility. Methods: The BeWell Study methodology is described. Descriptive statistics characterized recruitment and retention, as well as data completion and data quality. Binary logistic regressions evaluated whether recruitment sources, demographics, and depression severity predicted measure completion and data quality. Results: From June 1, 2022, through July 15, 2024, 52,052 individuals started the screening questionnaire. Of these, 2.22% (n=1156 participants; mean age 38.69, SD 12.14 years; women=75.50%; non-Hispanic White=68.88%) completed the screener, met eligibility criteria (including elevated depressive symptoms [Patient Health Questionnaire-9 [PHQ-9] score≥5]), completed baseline measures, were enrolled, and were randomized. Participants were recruited from Craigslist (694/1154, 60.14%), a US-based advertisement website, and university email messages (107/1154, 9.27%); participants were enrolled from every US state. Across the study, 84% (971/1156) of participants were retained (completed ≥1 measure at 17-week follow-up). Across time points, ≥81.40% (941/1156) of participants completed questionnaires and ≥75.17% (869/1156) completed biological samples. Of those completing questionnaires, ≥84.14% (796/946) completed SEA and video log tasks and ≥91.75% (868/946) completed the MST. Of completed assessments, high-quality or analyzable data rates across all time points were ≥97.89% (926/946) for questionnaires, ≥90.20% (773/857) for SEA, ≥84.60% (725/857) for video logs, ≥94.20% (1056/1121) for the MST, ≥98.84% (849/859) for dried blood spots, and ≥99.43% (865/870) for microbiome samples. Greater depressive symptoms at screening, but not age, predicted completion of baseline measures (odds ratios 0.96‐0.99; <.001). For enrolled participants, recruitment sources and baseline depressive symptoms were not predictive of data completion or quality. Older age and having a college degree predicted significantly higher rates of data completion across all data types (questionnaire, behavioral, and biological) at week 4 or week 17 follow-ups. Conclusions: This methodological approach offers one model of incorporating objective measures of well-being into DCTs testing DMHIs. Implications and limitations of our approach are discussed. Trial Registration: ClinicalTrials.gov NCT05183867; https://clinicaltrials.gov/study/NCT05183867

Ensemble of Domain-Specific Natural Language Processing and Large Language Models for Detecting Suicidal Ideation and Mental Health Conditions in Social Media Text: Development and Evaluation Study

2026-10-08T17:30:58-04:00

Background: Mental illness contributes substantially to global disability, and public adoption of AI for mental health support is accelerating without commensurate safety evaluation. General-purpose large language models can fail to recognize naturalistic text expressing suicidal ideation, with immediate clinical consequences. Domain-specific natural language processing methods offer a contrasting approach, but ensembles integrating the two have not been formally characterized. Objective: We aimed to (1) benchmark 5 modeling approaches for classifying mental health–related social media text, (2) develop and optimize ensembles integrating a domain-specific natural language processing classifier with a fine-tuned large language model, and (3) derive a closed-form framework constraining ensemble weighting in safety-critical classification. Methods: We analyzed 60,889 publicly available Reddit and Twitter texts spanning 9 categories (anxiety, bipolar disorder, depression, a normal baseline, personality disorder, stress, suicidal ideation, attention-deficit/hyperactivity disorder, and autism spectrum disorder). Labels were derived from the originating subreddit or a self-stated condition and are not clinical diagnoses. Five architectures were compared: prompt-engineered base GPT-4o-mini; fine-tuned GPT-4o-mini; a domain-specific classifier (support vector machine over symptom-informed lexical features); and 2 ensembles of these models, hybrid probability–indicator fusion and soft probability fusion. Data were split into 70%, 10%, and 20% at the post level (n=42,622, n=6089, and n=12,178); weights were selected on the validation split and all reported performance comes from the held-out test split. Closed-form bounds on permissible language model weights were derived for both strategies. CIs are paired bootstrap intervals (n=10,000 replicates); accuracy differences were tested with exact McNemar tests. Results: The domain-specific classifier achieved 91.9% accuracy (95% CI 91.4%-92.4%), exceeding the base (58.7%) and fine-tuned (88.6%) large language models. Both ensembles outperformed either constituent model, reaching 93.9% (soft fusion; 95% CI 93.5%-94.3%) and 93.8% (hybrid fusion; 95% CI 93.4%-94.2%) at a weighting of 60% classifier to 40% fine-tuned model, selected on validation (P<.001 for both against the classifier alone; the 2 ensembles did not differ from each other [P=.16]). The hybrid ensemble reduced the miss rate for the suicidal ideation category from 32.5% (650/2000) to 1.9% (37/2000), a 17.6-fold reduction relative to the base model, although its advantage over the classifier alone within that category was not significant (P=.68). Accuracy fell discontinuously above a 50% language model weight, where the hybrid ensemble reduced exactly to the fine-tuned model; the selected weight satisfies the derived instance-level bound of 44.0%. Conclusions: Domain-specific clinical grounding remained necessary for stable ensemble performance in this corpus, and probabilistic fusion outperformed both constituent models when language model weight was bounded by a derivable, model-agnostic constraint. Because labels were community-inferred and only one corpus was analyzed, these results are hypotheses requiring external validation on clinically characterized data before use in decision support.

Effectiveness of Chatbots in Mental Health Screening and Assessment: Systematic Review

2026-10-06T15:00:18-04:00

Background: Mental health disorders affect 970 million people globally; yet, over 50% do not access timely evaluation due to structural barriers and professional shortages. Chatbots and AI-based conversational agents have emerged as promising tools for mental health screening and assessment. Objective: This study systematically evaluated the effectiveness, accuracy, reliability, and acceptability of chatbots and AI-based conversational agents for mental health screening and assessment in adults. Methods: Systematic search conducted in May 2025 across PubMed/MEDLINE, PsycINFO, Scopus, and Web of Science (2019‐2025), following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines. Eligible studies evaluated chatbots or AI for mental health screening/assessment in adults (≥18 y). Risk of bias was assessed using appropriate tools (Risk of Bias 2, Quality Assessment of Diagnostic Accuracy Studies-2, Joanna Briggs Institute checklists, and Mixed Methods Appraisal Tool). This systematic review was registered with PROSPERO (International Prospective Register of Systematic Reviews; CRD420251072392). Results: Eighteen studies (2021‐2025) were included, with samples ranging from 20 to 3902 participants. Rule-based chatbots demonstrated high reliability (Cronbach α >0.85) and good acceptability (Acceptability of Intervention Measure >19/25). Generative models (large language models) achieved sensitivities of 0.84 to 0.93 and specificities of 0.80 to 0.96 for depression and anxiety, with correlations up to =0.96 with expert clinicians in suicide risk assessment. Hybrid approaches combining large language models with machine learning achieved exceptional performance for cognitive impairment (-score of 92.1%, specificity of 99.6%). Most studies reported high user satisfaction (≥70%), although barriers existed among older populations. Methodological quality was heterogeneous with a moderate risk of bias in critical dimensions. Conclusions: Chatbots and AI conversational agents demonstrate clinically relevant performance in mental health screening and assessment. However, safe implementation requires clear clinical protocols, professional supervision, integration with electronic health records, and active mitigation of algorithmic bias. These technologies should complement rather than replace clinical judgment. Trial Registration: PROSPERO CRD420251072392; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251072392

Electronic Health Record–Based Phenotyping for Obsessive-Compulsive Disorder: Algorithm Development and Multicenter Validation Study

2026-10-02T15:15:11-04:00

Background: Obsessive-compulsive disorder (OCD) is a common psychiatric disorder, with two-thirds of affected individuals reporting severe impairment. Despite its substantial burden and moderate heritability, the etiology of OCD remains poorly understood, and treatments are often suboptimal. Although recent genome-wide association studies (GWAS) have identified some risk loci, much of the genetic architecture of OCD remains undiscovered, underscoring the need for scalable approaches to identify large, well-defined patient cohorts. Objective: This study aimed to develop and validate a scalable electronic health record (EHR)–based phenotyping algorithm for identifying OCD cases to support large-scale genetic and translational research. Methods: We leveraged EHR-linked biobank data from 2 large hospital systems, namely Vanderbilt University Medical Center (VUMC) and Mass General Brigham (MGB), to develop a high-throughput phenotyping algorithm integrating diagnostic codes, medication records, and natural language processing (NLP) of clinical notes. Algorithm performance was evaluated through expert chart review, and genetic analyses were performed in individuals of European genetic ancestry using the polygenic scores (PGS) of OCD, major depressive disorder (MDD), and height derived from the most recent GWAS. Results: Expert chart reviews demonstrated our algorithm combining both () codes and NLP achieved the highest positive predictive values (PPV) for OCD case identification (0.84 at VUMC; 0.91 at MGB) compared to using either codes or NLP alone, albeit with reduced case yield. At both sites, algorithm-defined OCD cases of European genetic ancestry showed significantly higher OCD PGS than controls. In sensitivity analyses adjusting for MDD status, OCD PGS associations were more robust than MDD PGS associations, while height PGS showed no association, supporting the genetic plausibility and relative specificity of the phenotype. Conclusions: This study presents a scalable and cost-efficient EHR-based approach for identifying OCD cases across health systems. The algorithm achieves high PPV, and among individuals of European genetic ancestry, algorithm-defined cases show significant OCD PGS enrichment, supporting its utility for large-scale genetic studies and advancing understanding of the disorder’s complex etiology.

Generative AI Adoption in Psychotherapy Practice Among Mental Health Professionals: Multicountry Cross-Sectional Survey

2026-09-30T17:00:03-04:00

Background: Generative AI (GenAI) tools, such as OpenAI’s ChatGPT, Google’s Gemini, and Microsoft Copilot, are being adopted across health care settings. Despite the popularity of GenAI, empirical data on the extent of adoption by mental health clinicians in routine psychotherapy practice are scarce. Objective: This study aimed to characterize purposeful GenAI use among mental health professionals providing psychotherapy, including tools used, purposes, perceived workload effects, and the organizational context surrounding use. Methods: We conducted a cross-sectional online survey between January 13, 2026, and March 30, 2026. A convenience sample of trained mental health professionals who had provided psychotherapy within the previous 12 months was recruited through professional associations and academic and clinical mailing lists, with snowball dissemination by email. The survey assessed lifetime purposeful GenAI use in psychotherapy practice; tools and purposes of use; perceived work burden reduction; and, for the previous 12 months, employer or professional organization encouragement, prohibition, and participation in AI training. Descriptive statistics were calculated. Univariable logistic regression models were used to examine associations between GenAI use and candidate factors. Variables meeting an entry criterion of P<.10 were entered into a multivariable logistic regression; statistical significance was set at P<.05. Results: A total of 766 mental health professionals who provide psychotherapy in 30 countries completed the survey. Of these, 54.6% (n=418, 95% CI 51.0%-58.1%) reported having purposely used at least 1 GenAI tool in clinical psychotherapy practice. ChatGPT was the most frequently used tool (354/418, 84.7%, 95% CI 81.2%-88.2%). The most commonly reported clinical purpose was assisting with treatment planning (175/418, 41.9%, 95% CI 37.1%-46.7%), followed by managing administrative tasks (n=173, 41.4%, 95% CI 36.6%-46.1%) and generating psychoeducational materials for clients (n=166, 39.7%, 95% CI 35.0%-44.4%). Among AI users, 82.8% (346/418, 95% CI 79.1%-86.4%) reported that these tools reduced their overall work burden. Only 18.1% (139/766, 95% CI 15.4%-20.9%) of respondents reported institutional encouragement to use GenAI tools, while 81.1% (n=621, 95% CI 78.3%-83.9%) reported not having received any professional training on AI use. In multivariable logistic regression (α=.05), younger age was associated with GenAI use, while practice setting showed mixed evidence of association. The overall adjusted effect of practice setting did not reach statistical significance (P=.07), although urban practitioners had lower odds of GenAI use than rural practitioners (adjusted odds ratio 0.614, 95% CI 0.388-0.971; P=.04). Conclusions: This study extends previous predominantly national and profession-specific research by providing a multicountry, multiprofessional snapshot of reported GenAI use in psychotherapy practice. Use extended beyond administrative tasks into clinically consequential activities, including treatment planning and psychoeducation, while participation in professional AI training remained relatively uncommon. Findings highlight the need for professional bodies, health systems, and training programs to develop practical guidance and competencies for safe, privacy-conscious, and clinically appropriate GenAI use.

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.