Technical Summary: CARE-link

1. Problem Statement

The management of chronic conditions, particularly gestational diabetes, is hindered by a disconnect between clinic visits and patient behavior. Existing Electronic Health Records (EHRs) typically function as one-way data repositories that lack mechanisms to interpret patient-generated data (PGHD) in context or to provide timely, actionable feedback to patients. Consequently, clinically relevant changes often go unnoticed until the next visit, missing opportunities for proactive intervention. Furthermore, in low- and middle-income countries (LMICs), traditional digital health interventions often fail to reach broad socioeconomic groups due to literacy barriers or lack of access to specialized apps. While mobile messaging platforms like WhatsApp are ubiquitous in these regions, they are not natively integrated with clinical workflows to support continuous, context-aware care.

2. Methodology and System Architecture

CARE-link addresses these gaps by integrating an EHR with a WhatsApp-based messaging interface, mediated by Large Language Models (LLMs). The system employs a service-oriented, modular architecture designed for scalability and resource-constrained environments.

Technical Stack and Infrastructure

• Frontend: A React-based web dashboard for clinicians.
• Backend (EHR): A Laravel (PHP) web service managing user roles, patient records, and clinical visits, utilizing a PostgreSQL database hosted on Azure.
• Chatbot Service: A Python FastAPI application handling asynchronous WhatsApp interactions via the Meta Cloud API.
• AI Integration: The system utilizes the OpenAI ChatGPT API. The FastAPI service constructs domain-specific prompts by pulling structured patient data (demographics, vitals, history) from the database to generate context-aware responses.
• Deployment: The EHR and Chatbot are deployed as decoupled microservices (Azure App Service and Azure Functions, respectively) communicating over secure HTTP/2 with TLS encryption. Data persistence is managed via Azure Cosmos DB and Azure Database for PostgreSQL.

Core Functionalities

1. Dual-Facing Interface:
   • Patient Side: Patients interact via WhatsApp using text or voice (supporting English and Twi). The LLM provides education, medication reminders, emotional support, and emergency detection.
   • Clinician Side: A web dashboard allows clinicians to manage patient records and view AI-generated summaries.
2. Closed-Loop Workflow:
   • Data Ingestion: Patient messages are processed by an AI agent that extracts symptoms, behavioral patterns, and red flags.
   • Clinical Decision Support: The system pre-fills "Current Complaints" and "History of Presenting Illness" fields in the EHR based on chat logs. It also offers draft recommendations (e.g., dosage adjustments) for clinician review, ensuring human oversight.
   • Actionable Feedback: Clinical decisions and treatment plans are stored and used to trigger automated, personalized reminders and guidance back to the patient via WhatsApp.
3. Security and Privacy: The system adheres to HIPAA/GDPR principles through encrypted channels, role-based access control, and audit trails. No raw patient identifiers are sent to the LLM API; only prompt text is transmitted.

3. Key Contributions

• Integration of PGHD into EHRs: CARE-link bridges the gap between unstructured patient messaging and structured clinical records, automating the extraction of relevant clinical signals from WhatsApp conversations.
• LLM-Mediated Closed-Loop Care: Unlike static rule-based systems, CARE-link uses generative AI to facilitate adaptive, human-like engagement, allowing for real-time interpretation of patient data and dynamic generation of clinical summaries.
• Accessibility via Ubiquitous Platforms: By leveraging WhatsApp and supporting local languages (Twi) and voice input, the system lowers barriers to entry for patients with varying literacy levels in LMICs.
• Modular Extensibility: The microservice architecture allows for the independent scaling of components and the addition of new agents (e.g., nutrition coaches) without disrupting the core EHR.

4. Results and Illustrative Examples

The paper presents the system's implementation through illustrative use cases rather than large-scale clinical trial data. Key demonstrations include:

• Priority Triage: The AI successfully flags high-risk patients (e.g., reporting bleeding or dizziness) on the clinician dashboard, enabling rapid identification of emergencies.
• Workflow Efficiency: The "Add New Record" interface demonstrates how AI-generated summaries of patient chats can reduce manual data entry for clinicians during visits.
• Contextual Continuity: The system maintains a "memory" of past interactions, allowing the AI to reference previous glucose logs or diet diaries when generating new advice, thereby supporting longitudinal care.

5. Significance and Claims

The authors position CARE-link as a significant step toward enhancing clinical oversight and patient compliance in resource-constrained settings. Its significance lies in:

• Operational Impact: Reducing the documentation burden on clinicians and enabling continuous monitoring between visits, which is critical for conditions like gestational diabetes.
• Research Platform: The system serves as a testbed for investigating critical questions regarding LLM-mediated trust, ethical AI deployment in healthcare, and the sustainability of AI tools in LMICs.
• Scalability: The architecture is designed to be adaptable to other chronic conditions requiring behavioral support and longitudinal tracking.
• Economic Viability: The authors propose a dual-market strategy (free for public systems, premium for private) to ensure financial sustainability, potentially reducing long-term healthcare costs by preventing complications through better adherence.

The paper concludes that CARE-link effectively demonstrates the feasibility of integrating EHR workflows, AI agents, and widely accessible messaging platforms to improve the continuity of care.