Acceptability and implementation feasibility of a mobile application powered by generative artificial intelligence for mental health screening and referral in Uganda

Technical Summary: Acceptability and Implementation Feasibility of a Generative AI Mobile Application for Mental Health in Uganda

Problem Statement

Uganda faces a critical mental health crisis, with an estimated 14 million people affected by depression and anxiety disorders. This burden is exacerbated by a severe shortage of specialists (approximately one psychiatrist per million people), limited funding (1% of the national health budget), and pervasive stigma. While digital health tools and AI-driven interventions show promise globally, there is a lack of evidence regarding their acceptability and implementation feasibility in low-income African settings. Existing AI mental health tools are predominantly developed for high-income, English-speaking Western populations, raising concerns about cultural appropriateness, algorithmic bias, and performance in local languages. Furthermore, infrastructural constraints such as unreliable internet, high data costs, and limited smartphone ownership threaten equitable access.

Methodology

This study employed a mixed-methods proof-of-concept design conducted between January and May 2026 in Kampala, Uganda.

• Participants: A purposive sample of 50 participants was recruited, consisting of 25 members of the general public and 25 healthcare workers (including psychiatrists, psychiatric clinical officers, nurses, and medical officers).
• Intervention: The study evaluated a generative AI-powered mobile application designed for mental health screening, self-care, and referral.
  • Technical Architecture: The app utilizes a Large Language Model (LLM) integrated with validated screening tools: the Patient Health Questionnaire (PHQ-4, PHQ-9) and the Generalized Anxiety Disorder scale (GAD-7).
  • AI Mechanism: Unlike traditional rule-based chatbots, the system uses a document-based knowledge system. It retrieves and interprets pre-fed protocol information (PHQ/GAD guidelines) to generate dynamic, context-sensitive, and natural language responses. While the system is designed to construct responses based on a curated set of files rather than relying on the internet to generate the text itself, the application still requires internet connectivity for data transmission and interaction, a dependency explicitly identified as a principal implementation barrier.
  • Safety Features: The application includes automatic referral logic for severe symptoms, including an automated dialing function for nearby healthcare facilities in cases of suicidality. It is explicitly not a diagnostic or prescription tool.
  • Security: Data is protected via Transport Layer Security (TLS 1.2+) and AES-256 encryption at rest. User identification is minimized, using mobile numbers as unique identifiers with optional name registration.
• Data Collection: The study utilized a pre-post implementation assessment.
  • Quantitative: Baseline and endline surveys measured perceptions, acceptability, and usage patterns.
  • Qualitative: Key Informant Interviews (KIIs) and open-ended feedback were conducted to explore user experiences and barriers.
• Analytical Frameworks:
  • Acceptability: Evaluated using the Technology Acceptance Model (TAM), focusing on Perceived Usefulness (PU) and Perceived Ease of Use (PEOU).
  • Feasibility: Assessed using Bowen's feasibility framework, covering five domains: demand, implementation, practicality, integration, and expansion.

Key Results

The study demonstrated high acceptability and feasibility, though significant infrastructural barriers remain.

1. Acceptability (TAM Findings):

• Perceived Usefulness: Post-implementation, 95.6% of general public participants found the information relevant to their needs, and 90.9% stated they would recommend the application. Healthcare workers (HCWs) viewed the tool as a valuable decision-support mechanism that reduces workload and improves diagnostic accuracy for non-specialists.
• Perceived Ease of Use: 90.9% agreed the features were easy to use, and 77.3% could use the app without third-party assistance. However, technical glitches (repetitive questions) and language barriers (English-only interface) were noted as friction points.

2. Feasibility (Bowen Framework Findings):

• Demand: There was substantial latent demand; 100% of baseline participants expressed willingness to use the app. Post-implementation, 56.7% of the general public used the app for screening or self-care within one month.
• Implementation: The primary barriers identified were internet dependency, high mobile data costs, and lack of multilingual functionality. 40% of endline participants cited data costs as a major hurdle.
• Practicality: The app was deemed practical for clinical and community settings, with 45.45% of users completing a screening session in 5–10 minutes.
• Integration: 96% of participants at baseline preferred the app to be linked to healthcare facilities. At endline, 36.4% of users had followed up on a referral recommended by the app.
• Expansion: 100% of participants were willing to recommend the app to others. However, 76% of baseline participants indicated that bilingual functionality (English + local languages) is essential for broader reach.

Key Contributions

• Contextual Validation: This is the first study to evaluate a generative AI-powered mental health tool specifically within the Ugandan context, providing evidence that such technologies can be acceptable and feasible in low-resource settings.
• Generative AI Application: The study demonstrates the utility of LLMs in delivering structured, protocol-based mental health screening (PHQ/GAD) with the flexibility of natural language conversation, moving beyond rigid decision-tree chatbots.
• Task-Sharing Support: The findings suggest the tool can effectively support task-sharing by empowering non-specialist providers and the general public to conduct initial screenings and self-care, potentially alleviating the burden on scarce specialist resources.
• Implementation Insights: The research identifies specific, actionable barriers to scale-up in sub-Saharan Africa, namely the need for offline capabilities, zero-rated data access, and localization into local languages.

Significance and Claims

The paper claims that a generative AI-powered mobile application is a feasible and highly acceptable solution for mental health screening, self-care, and referral in Uganda. The authors assert that the application shows promise for:

• Reducing Stigma: By offering anonymity and privacy, the tool encourages disclosure among individuals reluctant to seek face-to-face care.
• Improving Access: It addresses structural barriers such as travel costs, waiting times, and specialist shortages.
• Strengthening Non-Specialist Care: It serves as a capacity-building resource for healthcare workers, providing structured guidance and reducing cognitive burden.

However, the authors maintain a modest stance regarding scalability. They conclude that while the proof-of-concept is successful, successful scale-up is contingent upon addressing structural barriers related to internet connectivity, data costs, language accessibility, and smartphone ownership. The paper explicitly states that larger studies are required to evaluate the tool's long-term effectiveness, cost-effectiveness, and integration into routine health systems before widespread national rollout.