Closing the accessibility gap to mental health treatment with a conversational AI-enabled self-referral tool

The study addresses the global challenge of limited access to mental health care, exacerbated by funding and staffing constraints and by individual-level barriers such as stigma, negative attitudes, and low perceived need for treatment. These barriers are often stronger for minority and disadvantaged groups (e.g., ethnic and sexuality minorities). In the UK, NHS Talking Therapies (formerly IAPT) relies heavily on self-referrals, yet existing referral pathways (phone calls, static webforms) can be inconvenient, stigmatizing, and difficult to navigate. The authors hypothesized that an AI-enabled, conversational self-referral tool integrated into service websites would lower the threshold to seek help by offering flexible, human-free, guided referral, thereby increasing overall referrals and disproportionately benefitting minority groups who face higher barriers. The purpose is to evaluate, in real-world clinical settings, whether such a tool increases the number and diversity of referrals and to understand mechanisms behind any observed gains via analysis of patient feedback.

The paper situates the work within evidence that mental health disorders are prevalent and burdensome worldwide, with the COVID-19 pandemic intensifying needs. Barriers to care include stigma, low perceived need, and structural issues, particularly affecting minority groups. Digital technologies and AI have been proposed to improve mental health service delivery, reduce workload, and enhance efficiency, with emerging findings that technology can facilitate help-seeking, improve comfort in sharing sensitive information, and support self-reporting. However, there is limited evidence on the marginal impact of AI tools in clinical settings and on diverse demographics. Chatbots in healthcare have gained attention but remain under-studied in real-world clinical deployment; thus, this work contributes by evaluating an AI-enabled self-referral chatbot within NHS Talking Therapies and examining demographic impacts.

Design: Multi-site real-world retrospective observational study in NHS Talking Therapies services across England. Fourteen services implemented the AI-enabled self-referral tool (Limbic Access), and 14 matched services (using webforms as a referral option) served as controls during the same timeframes. Tool: The chatbot is embedded on service websites, proactively visible to visitors, and collects referral-required information (eligibility, contact, demographics) and clinical data via standardized questionnaires (PHQ-9, GAD-7, WSAS) and additional screening. Data are attached to the electronic health record to support assessment. Frontloading measures aims to reduce administrative burden and allow private, flexible completion. Timeframes: For each implementing service, the closest three-month quarter before launch (pre) and after launch (post) were used. Launch months varied, enabling a staggered design to mitigate seasonal effects. The total sample encompassed approximately 129,400 patients across 28 services. Data sources: Publicly reported NHS Digital datasets (annual reports for total referrals; quarterly reports for demographics), which undergo quality checks. Matching: Each implementing service was matched to a similar service using Euclidean distance on normalized features, including total referrals in the pre-implementation quarter and time period alignment, selecting control services with webforms to ensure comparable referral modes. Analyses: - Total referrals: Summed total referrals per service for pre- and post-periods. Compared changes between tool and control services via Chi-squared test. A sensitivity analysis repeated the comparison for self-referrals only to rule out confounding by other referral routes. - Demographics: Examined percentage change in referrals by gender identity (female, male, non-binary), sexual orientation (heterosexual, bisexual, gay/lesbian), and ethnic group (Asian or Asian British, Black or Black British, Mixed, Other Ethnic groups, White). Logistic regression predicted whether a referral occurred pre (0) or post (1) implementation using sociodemographic group as predictor, with the most common group as reference; reported odds ratios (ORs) with confidence intervals (CIs). - NLP topic classification of feedback: From Sep 2021 to Mar 2023, 157,416 individuals used the tool; 29% (N=46,166) provided free-text feedback at referral completion. Excluding <10-character entries yielded N=42,332. A researcher open-coded 4,000 entries; the team agreed on nine themes (positive: Convenient; Provided hope; Self-realisation; Human-free; neutral: Needed specific support; Other neutral feedback; negative: Expected support sooner; Wanted urgent support; Other negative feedback). A supervised text classification pipeline used SBERT embeddings, PCA for dimensionality reduction, and multi-class logistic regression. Training and test data comprised 657 labeled entries (≥50 per theme). Performance via 100-fold cross-validation achieved micro-average F1=0.64. The trained model labeled all 42,332 entries. Group differences in theme proportions were assessed via Chi-squared tests with Bonferroni correction. - Assessment access: To assess downstream capacity effects, a two-way ANOVA tested interaction between time (pre/post) and tool usage on the percentage assessed, comparing implementing vs control services. Ethics and availability: Data are routine administrative datasets reported publicly; qualitative feedback data are not publicly shareable due to privacy; code/data for referral analysis to be made available upon acceptance.

- Overall referrals: Services using the AI-enabled self-referral tool saw a 15% increase in total referrals from pre to post, versus a 6% increase in matched control services. The increase for tool services was significantly greater (χ²(1)=86.3, p<0.0001). - Gender identity: Non-binary referrals increased by 235%, compared to 18% for females and 16% for males. Logistic regression indicated higher increases for non-binary individuals relative to females (OR=2.83, 95% CI [2.264, 3.545], p<0.0001) and males (OR=2.90, 95% CI [2.314, 3.629], p<0.0001). - Sexual orientation: Bisexual referrals increased by 30%, gay/lesbian by 19%, heterosexual by 14%. Bisexual vs heterosexual increase was greater (OR=1.14, 95% CI [1.074, 1.219], p<0.0001). Gay/lesbian vs heterosexual difference was not significant (OR=1.05, 95% CI [0.922, 1.189], p=0.216). - Ethnicity: Ethnic minority referrals increased by 31% vs 15% for White. Minority vs White increases were greater (OR=1.14, 95% CI [1.081, 1.202], p<0.0001). Fine-grained: Asian or Asian British (39% increase; OR=1.21, 95% CI [1.116, 1.318], p<0.0001); Black or Black British (42% increase; OR=1.24, 95% CI [1.108, 1.378], p<0.0001); Other Ethnic groups not significantly different (OR=1.08, 95% CI [0.913, 1.272], p=0.377); Mixed not significantly different (OR=1.01, 95% CI [0.917, 1.107], p=0.877). - Assessment access: No evidence that increased referrals reduced access to assessment; two-way ANOVA showed no significant interaction between time and tool usage (F(1,13)=0.09, p=0.764). - Feedback themes (N=42,332): 89% positive, 7% neutral, 4% negative. Common positives included convenience, hope from taking the first step, human-free interaction reducing stigma, and self-realisation of need for treatment. • Gender minority individuals referenced the human-free nature more often than females/males (χ²(1)=22.7, p<0.0001); marginally less mention of “Provided hope” (χ²(1)=8.84, p=0.079). • Bisexual individuals more frequently cited the human-free aspect than heterosexuals (χ²(1)=62.4, p<0.0001) and less often cited “Provided hope” (χ²(1)=27.8, p<0.0001). • Asian and Black groups more often mentioned self-realisation of need for treatment than White (χ²(1)=46.6, p<0.0001), and less often cited convenience (χ²(1)=24.2, p<0.0001). • Bisexual and Asian/Black groups provided more neutral feedback, notably around needing specific support (sexuality: χ²(1)=30.8, p<0.0001; ethnicity: χ²(1)=14.6, p=0.0036). Collectively, the tool increased overall access and disproportionately benefitted minority groups, with qualitative data suggesting mechanisms related to reduced stigma via human-free interaction and increased self-recognition of treatment need.

The findings address the primary hypothesis that an AI-enabled, conversational self-referral tool can increase access to mental health services and improve equity. Compared with matched services, implementing sites experienced a significantly larger rise in referrals without reducing progression to clinical assessment. Disproportionately larger increases among non-binary, bisexual, and ethnic minority groups suggest the tool helps bridge access gaps for populations that face higher barriers in traditional pathways. The qualitative feedback analysis indicates that mechanisms may differ by group: for gender and sexuality minorities, the human-free nature likely reduces anticipated stigma and judgment, fostering help-seeking; for Asian and Black individuals, the structured, reflective process may enhance self-realisation of need, counteracting low perceived need or stigma-related hesitancy. These results align with prior literature on digital tools facilitating disclosure and help-seeking and suggest AI-enabled referral pathways can complement system-wide access goals (e.g., NHS Long Term Plan). Importantly, no negative impact on assessments was detected, alleviating concerns about overloading downstream services. The work highlights policy and practice implications for embedding inclusive, low-threshold digital entry points into mental health pathways, with attention to tailoring features for diverse groups.

This large-scale, real-world evaluation shows that an AI-enabled conversational self-referral tool integrated into NHS Talking Therapies significantly increased overall referrals and enhanced diversity among those accessing care, particularly benefitting non-binary, bisexual, and Asian/Black ethnic groups. Qualitative analyses point to key mechanisms—reduced stigma via human-free interaction and improved self-recognition of treatment need—that help explain disproportionate gains among minority groups. The tool did not diminish access to subsequent assessments, supporting its viability within existing care pathways. Future research should employ designs that strengthen causal inference (e.g., stepped-wedge or randomized rollouts), examine long-term clinical outcomes and service efficiency impacts, evaluate generalizability beyond the UK context, and explore tailored enhancements to address specific needs highlighted by minority groups (e.g., specific support requirements), ensuring equitable and culturally sensitive deployment.

- Observational, retrospective design limits causal inference despite matched comparisons and staggered implementation. - Potential selection/cultural differences between services: sites adopting the AI tool may be more open to innovation, potentially influencing referral behavior independent of the tool. - NLP feedback analysis used a classifier with moderate performance (micro-F1=0.64) and relied on self-selected respondents, which may introduce bias in thematic distributions. - Conducted within NHS Talking Therapies in England; generalizability to other health systems or referral models may be limited.