Artificial intelligence (AI) is rapidly transforming health care, offering significant potential to enhance access, efficiency, and outcomes, particularly in mental health. This collection of studies explores the effectiveness and acceptability of AI-driven and digital mental health interventions across varied populations and clinical settings. Sadeh-Sharvit et al. evaluated the Eleos Health AI platform in routine outpatient cognitive behavioural therapy (CBT), showing greater symptom reduction in depression (34% vs. 20%) and anxiety (29% vs. 8%) compared to treatment as usual (TAU), along with a 67% increase in session attendance. Lohani et al.’s national RCT on suicide prevention found that both clinician-delivered crisis response plans (CRP) and self-guided safety plans (SP) significantly reduced suicidality, though CRP led to higher plan adoption (77.42% vs. 51.72%) and perceived usefulness.
Chatbot-guided interventions were also effective. Schillings et al. demonstrated improvements in mindfulness and emotion regulation through a 3-week intervention using ELME, a psychoeducational chatbot, despite no significant change in perceived stress. Van Baal et al. found that Cory COVID-Bot significantly influenced health attitudes, with the compassion condition increasing the perceived importance of testing (M=4.38, (standard deviation: SD=4.91).
Across all studies, AI tools showed high user satisfaction and engagement, though challenges such as dropout rates, technical barriers, and digital inequities were noted. These findings emphasize the importance of ethical design, usability, and inclusivity in AI integration. Understanding stakeholder needs and tailoring interventions accordingly are crucial for leveraging AI’s potential to support preventive, accessible, and community-based mental health care.

Artificial intelligence, Digital mental health, Chatbot interventions, Cognitive behavioural therapy, Telehealth service.

AI is rapidly advancing, offering significant potential to improve health care delivery and patient outcomes. While widely adopted in sectors like finance and IT, its uptake in health care has been slower due to ethical and safety concerns. AI can assist in diagnosis, treatment, and self-monitoring, often outperforming human abilities in areas like medical imaging. It can also reduce the workload of physicians by automating routine tasks and supporting mental health through AI chatbots.[1-5]

Despite its promise, the perceptions and needs of key stakeholders such as patients, providers, and policymakers regarding AI use in community health care remain underexplored. Human factors like trust and privacy have historically influenced technology adoption in health care, and these must be considered for AI as well. Understanding user perspectives is essential to ensure AI’s successful integration, especially as health care shifts toward preventive models.[6-11]

AI refers to software systems designed to achieve complex goals by perceiving and interpreting data, reasoning, and making decisions. It includes machine learning, reasoning, robotics, and integrated cyber-physical systems.[12,13]

Digital technologies can support health promotion and disease prevention, but most current platforms focus on managing existing conditions. These tools are primarily used by younger, more affluent individuals with higher digital literacy. Older adults who could greatly benefit from preventive health interventions are often targeted by analog (nondigital) programs, despite evidence that digital tools can aid outcomes like physical activity in this age group.[14-18]

Digital interventions offer the advantage of home-based, independent use, yet age-related cognitive and motor decline can create barriers. Moreover, older adults are less often offered digital health services, even though they show a willingness to adopt and sustain the use of new technologies.[19,20]

Postpartum depression (PPD) is a common mental health condition affecting women after childbirth, presenting with symptoms like mood swings, sleep issues, and, in severe cases, suicidal or infanticidal thoughts. It can also negatively impact infant development. While psychotherapy and medications are effective treatments, breastfeeding mothers often hesitate to use antidepressants due to potential side effects in infants, making psychotherapy the preferred option for mild to moderate PPD.[21-24]

Barriers such as time, cost, childcare, and stigma can limit access to face-to-face therapy. Telehealth offers a promising alternative by delivering therapy and support through phones, websites, video calls, and apps. Already used in various medical fields, telehealth is increasingly applied in postpartum care, offering accessible, convenient, and stigma-free mental health support, especially for women in remote or underserved areas.[25]

Understanding why young people themselves avoid seeking help is key to closing this treatment gap. Identified barriers include stigma, embarrassment, lack of problem recognition, and a preference for self-reliance. Structural issues, lack of knowledge, and GP-related challenges also contribute. Addressing these youth-specific and systemic barriers is crucial for improving mental health support access.[26]

A systematic literature search was conducted using PubMed databases to identify relevant studies addressing AI-driven mental health apps, chatbots, and teletherapy platforms for accessibility.

The search strategy incorporated the terms: ‘AI in healthcare, Digital mental health, Chatbot interventions, Telehealth services’, and filters were applied to include clinical trials, randomized controlled trials, observational studies, and human Studies. The search included publications from January 1, 2020, to August 05, 2025, and was limited to studies published in English.

Inclusion criteria:

• Clinical trials, observational studies, and randomized controlled trials
• Studies involving human subjects
• Articles published in English
• Studies including male and female participants

Exclusion criteria:

• Books, commentaries, editorials, letters, documents, and book chapters
• Case reports, case series, and literature reviews
• Articles published in languages other than English
• Animal studies and in vitro (laboratory) studies

Effectiveness and acceptability of AI-driven CBT in routine outpatient care
In Sadeh-Sharvit S et al., this study adhered to the Consolidated Standards of Reporting Trials involving AI (CONSORT-AI) guidelines with written informed consent obtained in person. Ethics approval was granted by the Ozark Center’s Research Committee, Freeman Health System Institutional Research Board, and the Missouri Department of Mental Health under the project “Optimizing behavioural healthcare delivery through technology.”

Forty-seven adults (mean age 30.64, SD 11.02; 34 women, 72%) diagnosed with depressive or anxiety disorders and referred for individual outpatient CBT participated, along with their therapists. Participants were randomized to either the Eleos AI platform group (n=23) or TAU (n=24). Most participants were White (94%), and 6% identified as Hispanic or Latino.

Participants were recruited at the Ozark Center in Missouri and randomized post-consent at the therapist level. All therapists were master’s-level clinicians. Primary outcomes included AI platform feasibility and acceptability; secondary outcomes were patient health questionnaire-9 (PHQ-9) and generalized anxiety disorder-7 (GAD-7) scores, session attendance, treatment satisfaction, and perceived helpfulness. Depression and anxiety were assessed at baseline, month 1, and month 2. Treatment feedback was collected at trial end; therapist feedback was collected only from the AI group. All assessments were digital.

Therapists in the AI group used the Eleos Health platform, which captured and analysed session data and provided automated feedback and progress notes. They received a 45-minute training. TAU therapists delivered standard care without restrictions. Power analysis required 23 participants per group (Cohen’s d=0.8, α=.05). Mixed effects models analysed symptom changes using intent-to-treat, with missing data treated as random. Analyses were conducted in Statistical Package for the Social Sciences (SPSS) v27.

Seventy-two adults recommended for outpatient CBT for depression or anxiety were invited to participate; 47 (65%) consented. Of these, 16 in each group (Eleos AI and TAU) completed clinical outcomes (70% and 67%, respectively). Six therapists participated in the AI group (treating 1-6 clients each), and eight in the TAU group (treating 1-4 clients each), with no significant group differences in their professional background.

Participants in the AI group attended an average of 5.24 sessions (SD 2.31) versus 3.14 sessions (SD 1.99) in the TAU group during the first two months, a 67% increase. Baseline PHQ-9 and GAD-7 scores were similar across groups, with no significant differences in depression (t=0.34; P=.86) or anxiety (t=0.37; P=.60). Depression reduced by 34% with Eleos (d=0.82) vs. 20% with TAU (d=0.34). Anxiety reduced by 29% with Eleos (d=0.78) vs. 8% with TAU (d=0.14). One TAU participant was hospitalized; no other adverse events occurred.

High satisfaction and perceived helpfulness were reported in both groups, with no significant differences. Therapists using the AI platform praised its insights and efficiency. AI group therapists submitted notes faster (mean 14 hours, SD 38) than TAU (mean 69 hours, SD 73). AI notes were shorter (263 vs. 318 words) but met documentation standards, with fewer than two grammar errors on average.[27]

Treatment outcomes and engagement: CRP vs. Self-guided SP
Lohani M et al., this national and fully online RCT was conducted between May and December 2021, following institutional review board (IRB) approval from the University of Utah. Participants were recruited via online platforms (e.g., Reddit, Facebook) and screened through an online survey. Eligible participants (N=82) were aged 18-50, English-speaking, and reported suicidal ideation in the past week (Scale for Suicidal Ideation: SSI) or had a lifetime history of suicidal behaviour (SITBI-R). Exclusion criteria included medical inability to provide consent or current involvement in alternative suicide prevention therapies. All participants provided digital informed consent and received a $20 Amazon gift card for completing the 45-day follow-up.

Participants were randomized using stratified randomization based on sex and suicide attempt history into either the clinician-delivered CRP or the self-guided SP intervention. Both interventions were delivered via videoconferencing. CRP involved collaborative planning with a clinician, while SP was completed by participants using a digital Qualtrics form with clinician support as needed.

All clinicians were trained in both interventions. Fidelity was ensured through weekly supervision and intervention review by a licensed clinician. Assessments included baseline and 45-day follow-up SSI scores, therapeutic alliance, plan recall and adoption (binary items), and perceived treatment usefulness (0-10 scale). All measures were completed online. Data were analysed in RStudio. Mixed analysis of variance (ANOVA) assessed changes in suicidality. One-way ANOVAs evaluated alliance and perceived usefulness. Logistic regressions analysed binary outcomes of plan recall and adoption.

Suicidality was assessed using the SSI at baseline and follow-up. A mixed ANOVA was conducted with time (baseline, follow-up) as the within-subject factor and treatment (CRP vs. self-guided SP) as the between-subjects factor. A significant main effect of time was found, F (1, 56) =10.21, p=.002, ηp²=0.154, indicating reduced suicidality over time. However, no significant effect of treatment was observed, F (1, 56) =0.141, p=.708, ηp²=0.003, and no time × treatment interaction was found, F (1, 56) =0.098, p=.755, ηp²=0.002. Mean SSI scores decreased in both groups: CRP (from M=16.97 to M=13.79) and self-guided SP (from M=17.79 to M=13.93).

Therapeutic alliance was measured via the Working Alliance Inventory. A one-way ANOVA showed significantly higher scores in the CRP group (M=51.32, SE=1.74) than in the self-guided SP group (M=43.45, SE=1.62), F (1, 52) =10.934, p=.002, η²=0.174.

Treatment recollection did not differ significantly between groups, χ² (1) =0.439, p=.508, with 83.87% in CRP and 89.65% in self-guided SP recalling their plans.

Adoption of the treatment plan was significantly higher in CRP, χ² (1) =4.408, p=.036, with 77.42% adherence vs. 51.72% for self-guided SP. The odds of adoption were 3.20 times higher for CRP. Perceived usefulness ratings were significantly higher in CRP (M=6.87) than in self-guided SP (M=5.07), F (1, 58) =5.962, p=.018, η²=0.093.[28]

Behavioural and attitudinal outcomes following chatbot interaction
van Baal ST shows a pilot single-blind randomized controlled study used a between-group design with three conditions: exponential growth, compassion, and control. Conducted across four days in October 2020, 59 participants were recruited via random volunteer sampling through community advertisements in Melbourne, Victoria. Ethics approval was obtained from Monash Health (HREC/69725/MonH-2020-237291[v3]). Eligible participants were 18-29 years old, temporary visa holders, or Vietnamese nationals, fluent in English or Vietnamese. All participants received an AUD 50 voucher upon completing the exit survey.

Participants completed a pre-test survey assessing baseline attitudes toward testing and staying home, followed by a supervised and then unsupervised 30-minute interaction with Cory COVID-Bot via Facebook Messenger. The chatbot, designed as a middle-aged librarian, delivered public health information using simple language, emojis, and animations in two intervention groups. Animations highlighted COVID-19’s exponential spread or a family’s emotional experience. All participants were asked about their testing intentions and evaluated chatbot engagement via the digital behaviour change intervention (DBCI) Engagement Scale.

Data were collected on a visual analogue scale and through structured questions. Of the 59 participants, 46 completed all study phases. Eleven missed the pre-test, and two missed the post-test survey.

Primary outcomes included changes in perceived importance of testing, certainty in stay-at-home decisions, and intention to test. Cumulative link models and continuous ordinal regressions analysed ordinal outcomes, controlling for age, sex, and intervention group. Due to convergence issues, Wilcoxon rank sum tests were used for pairwise comparisons. Analyses were performed in R, with significance determined at a false discovery rate corrected alpha of 0.05.

The study included 59 participants: 18 males (M age=24.8, SD=3.4), 40 females (M age=25.9, SD=5.3), and one unreported. Sample composition varied across test conditions. Post-test evaluations of the Cory COVID-Bot were positive, with an average likelihood of recommending it rated 8.9/10. Participants reported strong perceived learning (M = 6/7 on a Likert scale). DBCI Engagement scores also reflected high engagement.

Experimental condition significantly predicted perceived importance of COVID-19 testing, χ²=15.978, p<0.001. The compassion condition yielded the largest increase (M=4.38, SD =4.91) compared to the control (M = -2.64, SD=7.65), b = -2.381, t (2) =3.236, p=0.01. No significant effect was found for the exponential condition (M= -0.30, SD=4.64), p=0.92.

The exponential condition showed a higher median likelihood of testing (Mdn=3) vs. control (Mdn=2.29), χ²=6.146, p=0.042, though not significant after correction (p=0.056). No difference was found between compassion (Mdn=2.59) and control, p=0.20.

Certainty about leaving home during lockdown was significantly influenced by chatbot interaction (pre vs. post), χ²=21.733, p<0.0001, and risk level, χ²=24.558, p<0.0001, with a significant interaction effect, χ²=36.255, p<0.0001. Certainty increased post-test in low- and minimal-risk scenarios but not in high-risk ones.[29]

Behavioural and emotional impact of the digital mental health intervention
In Schillings C, this two-arm, parallel-design randomized controlled trial evaluated a 3-week chatbot-guided intervention using ELME, a rule-based web app. The intervention group received interactive psychoeducational content, real-time dialogues, audio exercises, and individualized feedback twice daily (10-20 minutes each), with flexible timing and SMS reminders. The control group received TAU, completing only questionnaires and ecological momentary assessment (EMAs). Ethics approval was granted by Ulm University (401/20), and the trial was registered at DRKS (DRKS00027560). Written informed consent was obtained from all participants.

Assessments occurred at five time points: baseline (T0), pre-intervention (T1), daily during intervention (T1-T2), post-intervention (T2), and follow-up (T3). Participants received free access to the intervention, a chance to win a €25 gift card or five course credits, and relaxing exercises and feedback post-T3.

The primary outcome was perceived stress, measured via the perceived stress scale (PSS-10) (T0) and PSS-4 (T1-T3). Secondary outcomes included mindfulness (Freiburg mindfulness inventory), interoceptive sensibility interoceptive accuracy scale (IAS), and body perception questionnaire (BPQ) “Awareness” subscale, subjective well-being world health organisation (WHO), and emotion regulation (Emotion Regulation Questionnaire). EMAs assessed momentary perceived stress and interoceptive sensibility twice daily. Usability was evaluated with a German version of the 18-item mental health app usability questionnaire. Adherence was tracked by module completion rates; dropout reasons and user feedback (1-10 liking scale; 1-12 extent rating) were collected.

Data were analysed using hierarchical linear models for nested data (time within participants), applying random-intercept, random-slope models with REML estimation. Analyses were conducted in R (lme4, r2mlm), with significance set at P ≤ 0.05.

A total of 118 participants were randomized into an intervention group (n=59; 72% female) and a control group (n=59; 81% female). Baseline characteristics showed no significant differences between groups across demographics, stress, mindfulness, interoception, well-being, or emotion regulation measures.

No significant effects were found over time, between groups, or in the time × group interaction (p=0.13 to 0.96), indicating no measurable change in stress across the study period. While time and group had no significant main effects, a significant time × group interaction was found (β=1.130, p=0.03), suggesting mindfulness increased more in the intervention group over time.

No significant changes were observed over time or between groups in measures using both the IAS and BPQ. Momentary interoceptive sensibility did increase over time in both groups, but without significant group differences. Well-being improved significantly over time across both groups (β=4.237, p=0.005), but no significant group effect or interaction was observed (p=0.99 and p=0.29, respectively).

For the reappraisal subfacet, a significant time × group interaction was found (β=0.223, p=0.02), indicating improvements in reappraisal for the intervention group. No significant effects were found for the suppression subfacet. Mean module completion in the intervention group was 58%. Main dropout reasons included lack of time (n=19) and technical issues (n=22).[30]

Table 1: Summary of included studies

This collection of studies underscores the growing role and diverse applications of AI-enhanced digital interventions in mental health care. Across different populations and contexts, AI-supported tools demonstrated feasibility, acceptability, and, in several cases, clinical effectiveness.

Sadeh-Sharvit et al. found that integrating AI (Eleos Health platform) into CBT significantly improved depression and anxiety outcomes compared to TAU, with a 67% increase in session attendance. Notably, AI-supported therapy also enhanced documentation efficiency without compromising quality. Similarly, Lohani et al.’s national RCT revealed that both clinician-led and self-guided suicide prevention interventions reduced suicidality over 45 days. While no difference was observed in symptom reduction, clinician-led planning improved treatment adoption and perceived usefulness.

Chatbot-guided mental health interventions also showed promise. Schillings et al. observed improvements in mindfulness and emotion regulation in their 3-week ELME chatbot intervention, despite no significant changes in perceived stress.

Complementary findings from van Baal et al. highlighted that public health chatbots can significantly influence health attitudes and behavioral intentions. The compassion-focused condition notably increased the perceived importance of COVID-19 testing.

Collectively, these findings support AI’s potential in enhancing access, engagement, and clinical outcomes, particularly in underserved or remote settings. However, as AI tools are further integrated into community health, attention must be given to ethical design, usability across age groups, and reducing digital disparities. Tailoring interventions to user needs, especially among older adults and marginalized populations, is essential for maximizing public health impact.

The reviewed studies collectively highlight the potential of AI and digital interventions in transforming mental health care by improving accessibility, engagement, and clinical outcomes. AI-enhanced tools like the Eleos Health platform not only supported greater symptom reduction in depression and anxiety but also increased therapy session attendance and documentation efficiency. Similarly, digital suicide prevention strategies demonstrated effectiveness in reducing suicidality, with clinician-led plans showing higher adoption and perceived usefulness than self-guided versions.

Chatbot-based interventions, such as ELME and Cory COVID-Bot, also proved effective in enhancing emotional regulation, mindfulness, and public health attitudes. These tools were well-received, demonstrating high engagement and satisfaction across various demographics.

Tailoring AI-driven mental health solutions to user-specific needs and contexts is crucial, especially as care models shift toward prevention and community-based support. The research should continue to focus on inclusivity and sustainability to ensure AI fulfils its potential in advancing public mental health.

Not reported

No funding

Corresponding Author:
Samatha Ampeti, PhD
Independent Researcher, Department of Content
medtigo India Pvt Ltd, Pune, India
Email: ampetisamatha9@gmail.com

Co-Authors:
Shubham Ravindra Sali, Mansi Srivastava, Raziya Begum Sheikh, Sonam Shashikala B V, Patel Nirali Kirankumar
Independent Researcher, Department of Content
medtigo India Pvt Ltd, Pune, India

All authors contributed to the conceptualization, investigation, and data curation by acquiring and critically reviewing the selected articles. They were collectively involved in the writing, original draft preparation, and writing-review & editing to refine the manuscript. Additionally, all authors participated in the supervision of the work, ensuring accuracy and completeness. The final manuscript was approved by all named authors for submission to the journal.

Not applicable

None

None

https://doi.org/10.63096/medtigo3084233

Shubham RS, Samatha A, Mansi S, Raziya BS, Sonam SBV, Patel NK. AI-Driven Mental Health Apps, Chatbots, and Teletherapy Platforms Transforming Therapy Accessibility. medtigo J Neurol Psychiatry. 2025;2(3):e3084233. doi:10.63096/medtigo3084233 Crossref