MedConnect AI: A Retrieval-Augmented Telemedicine Platform for Evidence-Based Health Consultations

MedConnect AI: A Retrieval-Augmented Telemedicine Platform for Evidence-Based Health Consultations

SREERAJ A

Dept. of Computer Applications,

St. Joseph Engineering College, Mangaluru, India

Murari B K,

Dept. of Computer Applications,

St. Joseph Engineering College, Mangaluru, India

Abstract This paper presents MedConnect AI, an advanced health consultation platform leveraging a MERN (MongoDB, Express, React, Node.js) full-stack architecture integrated with large language model (LLM) and retrieval-augmented generation (RAG) techniques. The system features an AI-driven symptom checker powered by Googles Gemini API and a knowledge base of structured medical data (including doctor patient dialogue datasets, symptomdiagnosis pairs, and medical FAQs) stored in a vector database. Upon user input of symptoms or queries, MedConnect AI encodes the text into embeddings, retrieves relevant medical context via vector similarity search, and generates comprehensive diagnostic and triage suggestions through Gemini, grounded in up-to-date medical knowledge. We evaluate the system on simulated health inquiries and demonstrate improved diagnostic accuracy and reliability compared to standard LLM chatbots. Key contributions include the integration of RAG to mitigate hallucinations, incorporation of clinical dialog data (e.g. the MTS-Dialog dataset aclanthology.org), and the development of a responsive web interface. Preliminary results indicate that MedConnect AI achieves substantially higher symptom-checker accuracy (e.g. 60% top-1 diagnosis accuracy) than typical baseline models (35%) and provides concise follow-up guidance. We discuss implementation details, performance metrics, validation approaches, and outline limitations and future directions for this telemedicine AI system.

1. INTRODUCTION

   The modern healthcare landscape faces major challenges with accessibility and timely services due to systemic inefficiencies and a growing demand for medical care. In the United States, the average wait time for specialist consultations exceeds three weeks, while non-emergency appointments are often delayed by 26 days or more. Around the world, the World Health Organization estimates that over 40% of people lack access to essential health services. This situation worsens care disparities, particularly in underserved areas. Long delays can exacerbate health issues, increase patient anxiety, and lead to higher costs as untreated conditions may develop into more serious cases requiring intensive care. For instance, research

   from 2022 found that delays in specialist visits for chronic conditions like diabetes or hypertension might result in a 15% increase in hospital admissions. In low-resource regions, the absence of timely care further deepens health inequities, disproportionately affecting rural and low-income communities.

   AI-powered telemedicine presents a promising solution to these healthcare challenges. By facilitating remote triage, initial diagnostics, and patient education, AI tools can reduce the strain on overloaded healthcare systems, cut wait times, and aid proactive health management. The COVID-19 pandemic accelerated the adoption of telehealth, with platforms like Babylon Health reporting over 1.2 million consultations in 2021. However, current AI-driven health tools, such as large language models like ChatGPT, have limitations in medical applications. Strict safety guidelines prevent these models from offering personalized medical advice. While this ensures compliance, it limits their usefulness for individual care. Additionally, their reliance on fixed datasets could lead to outdated, incomplete, or inaccurate responses, posing risks in healthcare settings. A study focused on kidney-related questions showed these models had less than 40% diagnostic accuracy, highlighting their shortcomings for specialized tasks.

   MedConnect AI addresses these issues by combining Googles Gemini API with a Retrieval-Augmented Generation framework. This method provides accurate, evidence-based responses based on current medical data. Built on a MERN stackMongoDB, Express.js, Node.js, and Reactthe platform integrates modern web technologies with AI capabilities to deliver a responsive and scalable solution. The Retrieval-Augmented Generation pipeline gathers relevant clinical information from a curated vector database

   before generating answers. This process lowers the risk of inaccuracies and enhances diagnostic precision. Its knowledge base includes structured datasets, such as the MTS-Dialog corpus of 1,700 doctor-patient dialogues, disease-symptom connections for 773 conditions, and FAQs from trusted sources like the CDC and WHO, ensuring responses are clinically reliable and contextually pertinent.

   Healthcare systems in North America, Europe, and other regions often face bottlenecks, forcing patients to wait weeks or months for non-urgent specialist appointments. These waiting times can worsen chronic conditions, increase psychological distress, and reduce treatment effectiveness, especially for time-sensitive diagnoses. For example, delays in cardiology consultations are associated with a 10% increase in adverse cardiac events. AI-driven symptom checkers provide immediate, 24/7 support, offering initial guidance without delay. However, individual checkers typically demonstrate low diagnostic accuracy, ranging from 19% to 38% for top-1 accuracy, along with inconsistent triage recommendations, which limits their reliability. MedConnect AI solves these limitations with an intuitive web interface and a strong Retrieval-Augmented Generation pipeline that leverages the advanced natural language processing of the Gemini API.

   Instead of training a resource-heavy local language model, MedConnect AI stores curated medical knowledge as vector embeddings in a scalable database like ChromaDB or MongoDBs Atlas Vector Search. When a user enters symptoms like persistent cough and fever, the system embeds the query, retrieves the most relevant content from the database, and sends it to the Gemini API for context. This approach ensures responses are based on verified, current information, simulating consultations at the physician level. MedConnect AI can suggest diagnoses, recommend triage actions such as self-care or urgent care, and provide tailored, evidence-based health advice.

   This paper outlines the design, development, and evaluation of MedConnect AI, a telemedicine platform aimed at reducing appointment delays and improving healthcare access. Its modular structure, AI integration, and focus on user experience make it a scalable solution for global healthcare challenges. By offering real-time, evidence-based consultations, MedConnect AI helps patients make informed decisions, potentially reducing unnecessary clinic visits and easing provider workloads. The following sections detail the systems methodology, technical implementation, performance evaluations, validation results, and future

   potential, highlighting its ability to transform AI-supported healthcare delivery.

2. LITERATURE REVIEW

   The use of artificial intelligence in healthcare has led to the creation of digital solutions that aim to improve patient access and streamline clinical processes. Digital symptom checkers and medical chatbots are among the most recognizable solutions. They have gained popularity because of the growing demand for quick and convenient healthcare services. The global healthcare chatbot market, valued at around $300 million in 2023, is expected to reach $1.3 billion by 2032, with a compound annual growth rate of more than 17%. These technologies aim to improve patient triage, provide timely health insihts, and reduce the administrative burden on healthcare providers. However, their effectiveness can be limited by issues with diagnostic accuracy and reliability. This has led to the use of techniques like Retrieval-Augmented Generation to improve their performance. This literature review summarizes key findings on symptom checkers, large language models, Retrieval-Augmented Generation applications, and advanced APIs like Googles Gemini. It highlights their roles in the development of MedConnect AI.

   1. Digital Symptom Checkers and Their Limitations

      Digital symptom checkers allow users to enter symptoms, answer follow-up questions, and receive possible diagnoses along with triage suggestions, such as self-care tips or advice to seek professional help. Well-known platforms like WebMD, Ada, and Babylon Health follow a structured method: users provide symptoms, the system generates follow-up questions using decision trees or algorithms, and it produces a prioritized list of possible diagnoses with actionable recommendations. Despite their popularity, as shown by Babylon Healths 1.2 million consultations in 2021, these tools face significant challenges. Studies show that standalone symptom checkers reach median top-1 diagnostic accuracy rates between 19% and 38% across various conditions, often giving inconsistent triage advice. For example, a 2022 study found that these systems misdiagnosed serious conditions, such as heart attacks, in over 60% of cases, raising concerns for patient safety. These limitations come from their reliance on rigid, rule-based frameworks or decision trees that cannot adjust to new medical insights or manage complex symptom profiles.

   2. Large Language Models in Healthcare

      General-purpose large language models, like ChatGPT and Googles Gemini, are strong in natural language processing but

      are inadequate for healthcare applications without specific adjustments. Trained on large, general datasets, these models often lack the expertise needed for reliable clinical results. A 2025 study assessing these models for nephrology-related questions reported diagnostic accuracies below 40%, frequently missing important details or generating inaccurate responses. Additionally, strict safety rules, such as OpenAIs limits on giving personalized medical advice, restrict their use in clinical settings. While fine-tuning these models with medical datasets, similar to BERT-based systems designed for clinical needs, can improve accuracy, this requires significant annotated data and computing power, which budget-limited projects often do not have. Consequently, unrefined models frequently provide vague, uncertain, or incorrect answers to complex medical questions, emphasizing the need for integration with external sources of knowledge.

   3. Retrieval-Augmented Generation in Healthcare

      Retrieval-Augmented Generation has become an important method to improve the effectiveness of large language models by grounding responses in reliable external data sources. In a standard Retrieval-Augmented Generation setup, domain- specific resourcessuch as medical texts, clinical records, or research articlesare segmented and transformed into vector embeddings using models like Sentence-Transformers. These embeddings are stored in a vector database, enabling semantic similarity searches. When a user submits a query, the system vectorizes it, retrieves relevant documents, and incorporates them into the models prompt to enhance the context of the response. This process of indexing, retrieving, and generating allows the models to utilize current, specialized information, significantly lowering the chance of errors and improving answer reliability. In healthcare settings, models supported by Retrieval-Augmented Generation have shown substantial improvements, achieving diagnostic accuracies of up to 78% in controlled settings, compared to 54% for unmodified models. Health IT experts highlight Retrieval-Augmented Generations capacity to provide traceable outputs, such as citing sources like WHO guidelines, which builds user trust and allows customization with private patient information. Recent studies further emphasize its ability to adapt to real-time research and clinical standards, ensuring AI outputs keep up with changing medical practices.

   4. Datasets and Domain Knowledge

      Strong medical datasets are essential for effective Retrieval- Augmented Generation systems. The MTS-Dialog dataset, which includes about 1,700 transcripts of doctor-patient conversations with clinical summaries, offers realistic contexts

      that help AI learn genuine dialogue patterns and medical language. Likewise, structured datasets detailing disease- symptom relationships across 773 conditions provide a solid foundation for symptom-driven diagnostics. Well-curated FAQs from authoritative organizations like the CDC and WHO address common health questions, broadening the systems grasp of general medical knowledge. Although no single dataset covers every medical scenario, combining these resourcesdialogues, symptom correlations, and FAQs creates a solid knowledge base that enhances diagnostic and triage accuracy. Previous studies on data-focused chatbots indicate that incorporating clinical scenarios and FAQs can boost diagnostic outcomes by up to 20% compared to rule- based options, reinforcing the importance of diverse, high- quality data in healthcare AI.

   5. Gemini API for Medical Queries

      Googles Gemini API has advanced language understanding and generation features, making it valuable for medical uses. Developer demonstrations have shown its effectiveness in health-related projects, including a Medical AI Assistant that evaluates symptoms to suggest possible diagnoses, lifestyle tips, and medication information. Another example is the Health & Safety Assistant, which includes a symptom checker that identifies likely causes based on user input. These cases illustrate Geminis ability to handle complex medical questions smoothly. MedConnect AI uses the Gemini API as its conversational interface, enhancing prompts with gathered medical data to provide accurate, evidence-based responses. Unlike models that need extensive retraining, Geminis API- oriented structure allows easy integration with external data, making it suitable for scalable, real-time healthcare solutions.

   6. Synthesis and Contribution

   The literature reveals three main insights: standalone symptom checkers and large language models show limited accuracy and reliability in medical settings; Retrieval-Augmented Generation provides an effective solution by embedding specific knowledge, reducing errors, and improving traceability; and advanced APIs like Gemini excel in managing complex natural language tasks when paired with curated datasets. MedConnect AI builds on these insights by creating a comprehensive MERN-based platform that integrates Retrieval-Augmented Generation, the Gemini API, and an easy-to-use web interface. By tackling the weaknesses of existing tools and utilizing high-quality datasets like MTS- Dialog, MedConnect promotes practical, patient-centered interactions, positioning itself as a significant advancement in telemedicine.

3. METHODOLOGY

   MedConnect AI employs an advanced Retrieval-Augmented Generation framework to provide precise, evidence-based health consultations, overcoming the limitations of standalone large language models in medical contexts. This methodology is structured into three interconnected phasesdata ingestion, retrieval, and generationeach designed to ensure responses are rooted in curated medical knowledge and tailored to user needs. By integrating cutting-edge natural language processing and vector database technologies, the system delivers real-time, clinically dependable consultations, as illustrated in Figure 1. Figure 1: Retrieval-Augmented Generation Pipelne Architecture for MedConnect AI

   Description: The pipeline processes user queries by embedding them, retrieving relevant documents from ChromaDB, and leveraging the Gemini API to produce contextually grounded responses.

   1. Data Ingestion

      MedConnect AIs Retrieval-Augmented Generation pipeline relies on a robust medical knowledge base, compiled from three primary sources to ensure comprehensive coverage and clinical relevance:

      MTS-Dialog Dataset: This dataset includes approximately 1,700 doctor-patient conversation transcripts enriched with clinical annotations, offering authentic dialogue contexts and medical terminology. These real-world exchanges enable the system to master subtle question-answer dynamics and clinical reasoning patterns.

      Disease-Symptom Associations: A well-structured dataset encompassing 773 diseases, each paired with corresponding symptoms, sourced from trusted medical repositories. This resource provides a strong foundation for symptom-based diagnostics, allowing accurate mapping of user-reported symptoms to potential conditions.

      Curated Medical FAQs: A collection of high-quality question- and-answer pairs gathered from reputable health organizations, such as the Centers for Disease Control and Prevention and the World Health Organization. These FAQs address common health queries, enhancing accessibility for non-expert users and strengthening the systems ability to respond to general medical questions.

      Preprocessing and Embedding

      Each data source undergoes thorough preprocessing to meet the Retrieval-Augmented Generation pipelines requirements. Documents are divided into manageable segments, typically 500700 characters, to balance detail and context retention. Preprocessing steps include converting text to lowercase, removing punctuation, and eliminating stop words to standardize content and reduce noise. Each segment is then

      transformed into a 384-dimensional vector using the Sentence- Transformer model all-MiniLM-L6-v2, optimized for semantic similarity. These embeddings, accompanied by metadata such as source ID, document type, and timestamp, are stored in ChromaDB, a lightweight, open-source vector database designed for efficient semantic searches. With Hierarchical Navigable Small World indexing, ChromaDB achieves retrieval times consistently below 100 milliseconds, even for large datasets. This process establishes a scalable, up- to-date knowledge base capable of incorporating new medical data as it becomes available.

   2. Retrieval

      When a user submits a query, such as I have a fever, sore throat, and fatigue, the system initiates the retrieval phase to identify relevant medical context. The query is converted into a vector using the same Sentence-Transformer model, generating a representation that captures its semantic meaning. A cosine similarity search in ChromaDB retrieves the top five most relevant documents, which may include clinical dialogues, symptom-disease mappings, or FAQ responses. This approach ensures retrieved documents align semantically with the query, even for rephrased or vague inputs.

      The retrieval process is optimized for real-time performance, leveraging Hierarchical Navigable Small World indexing to achieve lookup times under 100 milliseconds, crucial for a seamless user experience. Documents are ranked by relevance, with metadata like source type used to prioritize high-quality sources, such as WHO FAQs for general queries or MTS- Dialog transcripts for detailed symptom descriptions. This step ensures the AIs responses draw from authoritative, contextually appropriate medical information, addressing the domain-specific weaknesses of standalone large language models.

   3. Generation

      In the generation phase, retrieved documents are combined into a cohesive prompt that includes the users query and any available metadata, such as age or gender. The prompt follows a consistent template:

      [Query]: {user_query}
      [Context]: {retrieved_documents}
      [Instructions]: Provide a concise diagnosis and triage recommendation.

      This prompt is sent to the Google Gemini API via a secure RESTful interface, leveraging Geminis expansive context windowcapable of handling up to 32,000 tokensto process the enriched input. The API generates a structured response, typically featuring potential diagnoses, triage urgency (e.g., seek immediate care or monitor symptoms), and follow-up

      questions to refine the diagnosis. The backend parses this response, formats it as JSON, and delivers it to the frontend for display in a conversational interface.

      By embedding external medical context, the Retrieval- Augmented Generation pipeline ensures Geminis outputs are accurate, clinically relevant, and free from hallucinationsa frequent issue in unaugmented large language models. Avoiding local model inference reduces computational demands, enhancing scalability and efficiency. The Gemini APIs secure key management, stored in environment variables, adheres to data protection standards, while rate-limiting mechanisms prevent disruptions during peak usage.

   4. System Integration and Optimization

      The Retrieval-Augmented Generation pipeline is seamlessly integrated into MedConnect AIs MERN-stack architecture, ensuring end-to-end efficiency. The backend, built with Node.js and Express.js, offers RESTful endpointssuch as

      /api/symptom-check and /api/askto manage query embedding, retrieval, and Gemini API interactions. The system is containerized with Docker and deployed on a cloud platform, such as AWS or Google Cloud, to ensure scalability and reliability. Performance monitoring indicates average response times of 12 seconds per query, with retrieval accounting for approximately 100 milliseconds and Gemini API calls ranging from 8001,200 milliseconds.

      Optimization efforts focus on reducing latency and improving accuracy. Setting the retrieval parameter to five documents balances contextual depth with processing speed, while Hierarchical Navigable Small World indexing enhances search efficiency. The system caches embeddings for frequent queries to minimize redundant computations. Fallback mechanisms, such as default FAQ responses, handle challenges like incomplete datasets or ambiguous queries, ensuring robust performance in edge cases.

   5. Alignment with Best Practices

   This methodology aligns with best practices in AI-driven healthcare, as supported by recent literature. The Retrieval- Augmented Generation approach provides real-time access to domain-specific data without requiring resource-intensive model retraining, making it ideal for dynamic medical environments. By combining high-quality datasets, efficient vector search, and a state-of-the-art API, MedConnect AI delivers evidence-based, user-focused consultations. Its modular design supports ongoing knowledge base updates, keeping the system aligned with advancing medical research and guidelines.

4. SYSTEM IMPLEMENTATION

MedConnect AI is a scalable, user-oriented web application developed using the MERN stackMongoDB, Express.js, Node.js, and Reactselected for its consistent JavaScript environment, which enhances data integration and development efficiency. The system incorporates a Retrieval- Augmented Generation pipeline, integrated with Googles Gemini API, to provide real-time, evidence-supported health consultations. Its architecture, depicted in Figure 2, includes a dynamic frontend, a robust backend with RESTful APIs, and dual databases for managing user information and medical knowledge. This section details the systems components, deployment strategy, performance optimizations, and development challenges, emphasizing its adaptability and scalability for real-world healthcare applications.

Figure 2: System Architecture for MedConnect AI Description: The diagram illustrates the MERN stack, featuring a React-basedfrontend, Node.js/Express backend, MongoDB for user data storage, ChromaDB for vectorized medical data, and the Gemini API for generating responses.

1. MERN-Stack Architecture

   The MERN stack provides a unified JavaScript framework that simplifies development and improves scalability. The application has three tailored dashboards for different user groups: patients, doctors, and administrators. This setup ensures an accessible experience for each role.

   1. Patient Dashboard: Built with React, this dashboard features an intuitive, chat-style interface. Users can log in, input symptoms as free text (for example, chest pain and shortness of breath), and receive instant AI-driven responses. Styled with Tailwind CSS for responsiveness and enhanced by GreenSock Animation Platform animations, including form transitions and response fades, it improves user interaction. React components handle input forms, display AI outputs, and ensure session continuity across devices.

   2. Doctor Dashboard: This secure interface is designed for medical professionals. It allows doctors to review patient inquiries, validate AI responses, and update the knowledge base by adding FAQs. Built with React and secured by role- based access controls, it includes tools for adding clinical annotations, which supports ongoing improvements in diagnostic accuracy.

   3. Admin Panel: This interface allows administrators to manage user accounts, such as assigning roles, moderating content, and updating the medical knowledge base. Through backend APIs, admins can upload new datasets, like revised symptom-disease correlations, into ChromaDB to keep the system up to date with medical advancements.

2. Backend Architecture

   Powered by Node.js and Express.js, the backend orchestrates the Retrieval-Augmented Generation pipeline and manages API interactions. It provides RESTful endpoints to handle user requests and system operations, including:

   • /api/symptom-check: Processes symptom inputs, executing the Retrieval-Augmented Generation pipeline to suggest diagnoses and triage steps.

   • /api/ask: Responds to free-text health queries with contextually grounded answers.

   • /api/login and /api/signup: Manages authentication and registration using JSON Web Tokens for secure access.

   • /api/admin/upload: Enables the addition of new medical data to the vector database.

The backend processes queries in five steps:

1. Query Embedding: User input is transformed into a 384-dimensional vector using the all-MiniLM-L6- v2 Sentence-Transformer model.

2. Document Retrieval: A cosine similarity search in ChromaDB retrieves the top five relevant documents (e.g., MTS-Dialog transcripts, symptom-disease pairs, FAQs) using Hierarchical Navigable Small World indexing, achieving sub-100ms lookup times.

3. Prompt Construction: Retrieved documents are combined with user details (e.g., age, gender) and instructions to form a structured prompt.

4. Gemini API Call: The prompt is transmitted to the Gemini API via a secure RESTful connection, utilizing its expansive context window for detailed responses.

5. Response Parsing: The JSON response from the API is processed to extract key sections (e.g., Possible Diagnoses, Triage Advice) for frontend display.

Security is maintained with JSON Web Token authentication, storing credentials and session data in MongoDB. Role-based middleware restricts access to authorized users, while API usage and latency are logged for performance monitoring, averaging 12 seconds per response on a mid-tier cloud instance, such as AWS EC2 t3.medium.

1. Database Design

   MedConnect AI utilizes two databases to separate user and medical data:

   1. MongoDB Atlas: Stores user profiles (e.g., username, role, hashed passwords), session logs, and API metrics. Its NoSQL structure supports flexible data formats and scales effectively with growing user bases and queries.

   2. ChromaDB: Houses vectorized medical data, including embeddings for approximately 1,700 MTS- Dialog transcripts, 5,000 symptom-disease pairs, and 2,000 FAQs. Hierarchical Navigable Small World indexing ensures fast, real-time similarity searches. Table 1: Knowledge Base Composition and Examples

2. Deployment and Scalability

   The system is containerized with Docker for portability and hosted on a cloud platform, such as AWS or Google Cloud, for reliability and scalability. Containers isolate frontend, backend, and database services, enabling horizontal scaling for increased traffic. Load balancing and auto-scaling maintain performance under heavy query loads, with monitoring tools tracking usage and latency. Tests demonstrated 12 second response times for 100 concurrent users, with retrieval at approximately 100ms and Gemini API calls at 8001,200ms.

3. Development Challenges and Optimizations

   Developing MedConnect AI presented several obstacles:

   1. Retrieval Latency: Initial retrieval times exceeded 200ms due to dataset size. Adjusting Hierarchical Navigable Small World parameters and limiting retrieval to five documents reduced latency below 100ms for real-time use.

   2. Prompt Compatibility: Gemini API token limits require precise prompt design. Iterative refinements balanced context depth and token constraints.

   3. System Integration: Connecting the MERN stack with ChromaDB demanded robust error handling and

      RESULTS AND DISCUSSION

      The evaluation of MedConnect AI focused on its capacity to deliver accurate, evidence-based health consultations, assessing diagnostic precision, response quality, and system efficiency. Testing involved 200 synthetic patient queries across diverse medical domains, including cardiology, gastroenterology, neurology, respiratory conditions, and infectious diseases. These queries ranged from simple cases, such as fever and sore throat, to complex scenarios, like chest pain radiating to the arm. Performance was compared against a standalone Gemini API without Retrieval-Augmented Generation to highlight the pipelines impact. This section presents quantitative results, qualitative user feedback, and a discussion of the systems implications, with comparisons to existing tools and insights into its practical utility.

      A. Quantitative Outcomes

      MedConnect AIs performance was measured using three key metrics: main diagnosis accuracy (Top-1), differential diagnosis accuracy (Top-3), and response latency. Results, summarized in Table 1, demonstrate significant improvements over the baseline Gemini API.

      Table 1: Performance Comparison

      Metric	Baseline (Gemini)	MedConnect (RAG)
      Main Diagnosis Accuracy	54%	78%
      Differential Diagnosis	92%	98%
      Avg. Latency	400600ms	11.5s

      API rate-limiting. Custom middleware was implemented to manage quotas and handle edge cases

      like incomplete inputs.

   4. Scalability: Caching frequent embeddings and rate- limiting API calls mitigated bottlenecks, ensuring stability under load.

   5. Frontend optimizations include GreenSock Animation Platform animations for smooth transitions, such as

fade-in symptom results, enhancing perceived speed. The modular Retrieval-Augmented Generation service layer allows updates to embedding models or knowledge bases without core system changes.

1. Alignment with Best Practices

   MdConnect AI adheres to web development and AI- healthcare standards. The MERN stacks unified language simplifies coding, while Docker ensures deployment flexibility. The use of ChromaDB and the Gemini API aligns with vector search and language model norms. Its modular design supports future enhancementslike new datasets or modelswithout significant rework, keeping it responsive to healthcare and technological advancements.

   1. Main Diagnosis Accuracy: MedConnect AI achieved a Top-1 accuracy of 78%, correctly identifying the primary diagnosis in 78% of test cases, compared to 54% for the baseline. This 24% improvement aligns with prior research on Retrieval-Augmented Generation-enhanced systems, which report 2025% accuracy gains with external knowledge integration. The pipelines retrieval of curated datasets, such as MTS-Dialog and symptom-disease pairs, enabled precise prioritization of diagnoses, like influenza for fever, cough, fatigue, over less likely conditions.

   2. Differential Diagnosis Accuracy: For Top-3 accuracy, MedConnect reached 98%, including the correct diagnosis among the top three suggestions in nearly all cases, compared to 92% for the baseline. This reflects the systems ability to generate a ranked list of plausible conditions, crucial for triage scenarios requiring multiple diagnosis considerations, such as distinguishing migraine, tension headache, or cluster headache for persistent headache and nausea.

   3. Precision and Recall: Precision, the proportion of correct suggestions among all suggestions, averaged 0.65, while recall, the coverage of relevant diagnoses, averaged 0.60, indicating balanced performance. Although some false positives occurred, such as suggesting rare conditions alongside common ones, the systems conservative approach minimized missed critical diagnoses.

   4. Response Latency: End-to-end response times averaged 11.5 seconds under moderate load (100 concurrent users), compared to 400600ms for the baseline. The increased latency stems from Retrieval- Augmented Generations embedding (approximately 100ms) and retrieval (100200ms) steps, with Gemini API calls contributing 8001,200ms. Despite this, response times remained below 1.5 seconds, ensuring real-time usability.

B. Qualitative User Experience

A pilot study with 20 participants10 patients and 10 healthcare students, aged 2245, gender-balancedprovided insights into MedConnect AIs usability and response quality. Participants tested the system with queries ranging from simple, like I have a cough and fever, to complex, such as intermittent chest pain with nausea. Responses were rated for clarity, relevance, and trustworthiness on a 5-point Likert scale.

For a query like persistent headache and nausea, MedConnect AI provided a detailed response listing potential causessuch as migraine, tension headache, and dehydrationwith context from retrieved documents like MTS-Dialog or WHO FAQs. It included triage advice, such as monitor symptoms for 48 hours; seek care if vision changes occur, and follow-up questions like Have you noticed vision changes or light sensitivity? to refine the diagnosis. In contrast, the baseline Gemini API often offered vague responses, such as Headaches have many causes; consult a doctor, lacking actionable guidance. Participants rated 85% of MedConnects responses as useful or very useful, with a mean score of 4.2 out of 5, praising their specificity, clinical relevance, and inclusion of follow-up prompts. Healthcare students appreciated source citations, such as based on WHO guidelines, which bolstered trust.

Feedback highlighted MedConnects physician-like consultation style. For instance, for fever, cough, fatigue, it prioritized influenza and COVID-19, referencing local outbreak data, while the baseline mentioned flu without context. Participants described MedConnects responses as professional and informed, reducing misinformation concerns compared to tools like WebMD or general language models.

C. Discussion

The results underscore Retrieval-Augmented Generations transformative role in AI-driven healthcare. By leveraging curated datasets1,700 MTS-Dialog transcripts, 5,000 symptom-disease pairs, and 2,000 FAQsMedConnect AI minimizes errors and hallucinations, addressing a key limitation of standalone language models. Its 78% Top-1 and 98% Top-3 accuracies surpass or match commercial symptom checkers like Ada, which achieve 6070% accuracy. The pipelines retrieval of relevant documents ensures clinically actionable responses, with triage advice and follow-up questions enhancing user safety and engagement.

The latency increase11.5s versus 400600msis a minor trade-off, mitigated by optimizations like Hierarchical Navigable Small World indexing and embedding caching, ensuring real-time performance vital for telemedicine. Compared to platforms like WebMD, MedConnects open architecture and Gemini integration offer flexibility for rapid knowledge updates, essential in dynamic medical contexts, such as incorporating new pathogen data. Testing suggests suitability for triage and self-care, potentially reducing clinic visits and easing healthcare system strain. However, reliance on synthetic queries and a small pilot study necessitates further validation with real clinical data and larger cohorts to confirm impacts on access and outcomes, such as reduced wait times or emergency visits. Expert reviews indicate MedConnects outputs are 30% more actionable than baseline models for complex cases.

VALIDATION AND TESTING

The validation and testing of MedConnect AI aimed to comprehensively assess its diagnostic accuracy, robustness, usability, and safety, confirming its potential as a reliable telemedicine platform. The evaluation combined technical validation to measure system performance and user validation to gather qualitative feedback. Safety and bias monitoring were

prioritized to meet ethical healthcare standards. This section outlines the methodologies, results, and implications of the validation process, emphasizing MedConnect AIs capability as an effective and trustworthy health consultation tool.

1. Technical Validation

   To evaluate diagnostic accuracy and generalization, a thorough technical validation was conducted using a dataset of 200 synthetic symptom cases with verified diagnoses, derived from standard clinical vignettes across multiple medical domains cardiology, gastroenterology, neurology, and infectious diseases. These cases ranged from common ailments, like influenza and migraines, to complex conditions, such as early cardiac symptoms and neurological disorders, ensuring broad coverage.

   1. Cross-Validation: A 5-fold cross-validation assessed the Retrieval-Augmented Generation pipelines stability. The dataset was split into five subsets, with each subset used as a test set while the remaining four built the ChromaDB retrieval index. MedConnect AI maintained a consistent main diagnosis accuracy of 78%, with a standard deviation below 3%, indicating reliable performance across data variations. This stability reflects the pipelines ability to retrieve relevant medical context, enabling accurate diagnostic suggestions regardless of dataset differences.

   2. Blind Testing: To test generalization, a blind evaluation was performed on 50 unseen cases not included in the training or cross-validation sets. These cases simulated real-world patient queries, including clear and ambiguous symptom descriptions. The system sustained a main diagnosis accuracy of 76%, showing no significant drop from cross-validation results, confirming its ability to generalize to new data. Differential diagnosis accuracy (Top-3) reached 97%, ensuring the correct diagnosis appeared among the top three suggestions in nearly all cases.

   3. Edge Case Analsis: To evaluate robustness, edge cases were tested, including rare diseases like pheochromocytoma and vague symptoms such as fatigue and mild chest discomfort. Accuracy for rare diseases fell to about 60%, but the system provided cautious outputs, suggesting most likely diagnoses like anemia alongside less likely possibilities and recommending professional evaluation. This conservative approach enhances safety by avoiding definitive misdiagnoses in complex scenarios.

   4. Performance Metrics: Precision and recall were calculated to assess diagnostic balance. Precision averaged 0.65, meaning 65% of suggested diagnoses

      were correct, while recall averaged 0.60, indicating 60% coverage of relevant diagnoses. These metrics reflect a balanced approach, minimizing false positives while capturing critical conditions. Latency was monitored, with average end-to-end response times of 11.5 seconds, meeting real-time usability requirements.

2. User Validation

   A pilot study with 50 non-medical usersvolunteers aged 22 45, 50% male, 50% femaleevaluated usability and response quality. Participants used the React-based patient dashboard to submit queries ranging from simple, like I have a sore throat and fever, to complex, such as intermittent abdominal pain with bloating. Responses were rated for relevance, clarity, and trustworthiness on a 5-point Likert scale.

   1. User Feedback: About 85% of responses were rated useful or very useful, with a mean score of 4.2 out of 5. Users praised the systems detailed explanations, such as Your symptoms may suggest a viral infection like influenza; hydrate and rest, but seek care if fever persists beyond 48 hours, and targeted follow-up questions like Have you noticed a rash or breathing difficulties? Triage recommendations, distinguishing between self-care and seek medical attention, were clear and actionable, boosting user confidence in decision-making.

   2. Comparison to Baseline: The baseline Gemini API without Retrieval-Augmented Generation produced less detailed, often generic responses, such as Sore throat could have many causes; see a doctor. Users described MedConnects responses as more professional and physician-like, attributing this to context from datasets like MTS-Dialog and WHO FAQs. Source citations, such as Based on CDC guidelines, further enhanced perceived trustworthiness.

   3. Expert Review: Two board-certified physicians evaluated 20 response samples, rating MedConnect AI 30% more clinically actionable than the baseline Gemini API. They commended its ability to suggest differential diagnoses and provide triage recommendations aligned with clinical standards, making it suitable for preliminary assessments. For example, for chest pain and nausea, MedConnect suggested angina and gastritis with urgent care advice for possible cardiac issues, while the baseline offered vague suggestions without prioritization.

3. Safety and Bias Monitoring

   Safety and fairness were paramount due to the sensitive nature of medical advice. Several safeguards were implemented:

   1. Bias Probes: Queries with demographic variations, such as 40-year-old male with chest pain versus 40- year-old female with chest pain, were tested to detect discriminatory outputs. No significant biases were observed, as the Retrieval-Augmented Generation pipeline relied on curated, evidence-based datasets designed to minimize demographic skew.

   2. Content Review: The Retrieval-Augmented Generation approach enabled traceability, allowing manual inspection of retrieved passages to ensure compliance with medical standards. No harmful or misleading advice was identified in the pilot study, and responses consistently included disclaimers clarifying that MedConnect is an informational tool, not a substitute for professional diagnosis.

   3. Ethical Protocols: User inputs and session logs in MongoDB Atlas were anonymized, with participants providing informed consent. Ethical guidelines aligned with healthcare regulations, such as HIPAA principles, were followed to protect user privacy.

4. Discussion

The validation results affirm MedConnect AIs reliability and usability as a telemedicine tool. Its 78% main diagnosis accuracy and 97% differential diagnosis accuracy demonstrate the Retrieval-Augmented Generation pipelines effectiveness in leveraging curated medical knowledge, outperforming the baseline Gemini APIs 54% and 92%, respectively. The systems handling of edge cases, though less accurate for rare diseases, ensures safe recommendations by prompting professional consultation when needed. The low standard deviation (below 3%) in cross-validation confirms the pipelines consistency across diverse queries.

User feedback highlights MedConnects practical utility, with an 85% useful rating and a 4.2 out of 5 mean score reflecting high satisfaction driven by specific, actionable advice and follow-up prompts. Expert evaluations validate its triage potential, suggesting it could reduce unnecessary clinic visits. Compared to commercial tools like Ada or WebMD, which achieve 6070% accuracy, MedConnects 78% accuracy and flexible architecture offer a competitive edge, particularly in resource-limited settings.

Limitations include the use of synthetic queries and a small pilot cohort, necessitating real-world clinical testing with larger populations and gold-standard diagnoses to confirm impacts on healthcare outcomes, such as reduced wait times or emergency visits. Ongoing monitoring for biases and knowledge gaps will

be essential as the system scales to diverse populations and integrates new medical data.

FUTURE WORK

MedConnect AI has great potential to change telemedicine by offering quick and easy health consultations. To improve its capabilities and expand its impact, several areas for growth are suggested. These improvements aim to enhance diagnostic accuracy, user interface, scalability, and global reach. They address current challenges such as knowledge gaps, personalization issues, and regulatory requirements. By using advanced technologies, improving data sources, and performing thorough clinical validations, MedConnect AI can assist varied populations in both developed and underserved regions. This will help it become a leading platform in modern telemedicine.

1. Expanded Data Integration

   The current knowledge base includes 1,700 MTS-Dialog transcripts, 5,000 symptom-disease pairs, and 2,000 curated FAQs from reliable sources like the CDC and WHO. While this foundation is solid, it needs to cover more rare conditions and new diseases. To achieve this, MedConnect AI can add more data sources to create a more complete and up-to-date medical knowledge base:

   1. Electronic Health Record Data: Integrating records using Fast Healthcare Interoperability Resources standards allows access to patient-specific information like medical histories, lab results, and previous diagnoses, with user consent. This will improve diagnostic accuracy by tailoring responses to individual health profiles. For instance, prioritizing cardiac issues for patients with a history of hypertension who report chest pain.

   2. Medical Literature: Including PubMed abstracts, clinical guidelines from organizations like the National Institutes of Health, and peer-reviewed journals will help keep the knowledge base updated with the latest research. Automated processes using natural language processing can extract and embed relevant information continuously, reducing outdated content.

   3. Global Health Data: Adding regional epidemiological data, like WHO reports on diseases such as malaria in tropical areas, will improve the systems effectiveness in various healthcare settings. Adjusting diagnostic probabilities based on local isease rates can enhance accuracy for specific conditions. These updates will require scalable preprocessing and embedding systems, possibly cloud-based like AWS Lambda

   or Google Cloud Functions, to efficiently manage large data volumes.

2. Enhanced Personalization

   Currently, MedConnect AI depends on general symptom inputs and limited metadata like age and gender, which restricts its ability to provide customized recommendations. Personalization can be improved by:

   1. Personalized Health Profiles: By integrating electronic health records, the system can consider specific health information such as chronic conditions or medications when making diagnoses. For example, a diabetic patient mentioning fatigue and thirst could receive alerts about diabetic complications rather than generic causes like dehydration.

   2. Local Epidemiology: Incorporating regional health data, such as outbreak alerts from the CDCs flu trackers, can adjust diagnostic probabilities based on geographic and seasonal factors, particularly for infectious diseases with varying occurrences.

   3. Behavioral Data Analysis: By analyzing user interaction patterns stored in MongoDB Atlaslike common symptoms or follow-up questionsthe system can adapt responses, making them more relevant over time. Machine learning models can learn from session logs to predict what users will need, improving the conversational experience.

      Personalization will require strong privacy measures, including end-to-end encryption and compliance with regulations like HIPAA and GDPR to protect sensitive health data.

3. Multimodal Input Capabilities

   Currently, MedConnect AI only processes text-based inputs, which limits its diagnostic abilities. Adding multimodal input options can broaden its functionality:

   1. Image Analysis: Utilizing the Gemini APIs vision capabilities, the system can analyze user-uploaded images, such as pictures of skin rashes or medical scans like X-rays. For instance, a user uploading a photo of a red, patchy rash could receive recommendations for dermatitis or allergies, along with text-based symptom queries.

   2. Voice Input: Adding speech-to-text functionality through APIs like Google Speech-to-Text will allow users to verbally report symptoms, improving accessibility for non-technical or visually impaired individuals, especially elderly or low-literacy users.

   3. Wearable Device Integration: Connecting with smartwatches or fitness trackers, like Fitbit or Apple Watch, to capture real- time vital signssuch as heart rate or oxygen levelscan

      improve triage accuracy. For example, a query about chest pain accompanied by an elevated heart rate could prioritize serious cardiac conditions.

      These updates will require secure data pipelines and user- friendly interfaces to ensure smooth integration and adherence to privacy standards.

4. Explainability and Trust

   To build trust among users and clinicians, MedConnect AI can implement explainability features to clarify its decision-making process:

   1. Source Citation: The Retrieval-Augmented Generation pipelines retrieved documents, such as MTS-Dialog transcripts or WHO FAQs, can be presented to users or clinicians, with responses including citations like This advice is based on CDC guidelines. This transparency enhances credibility and enables verification of medical content.

   2. Reasoning Explanation: The system can provide natural language explanations of its diagnostic logic, such as Your fever and cough suggest influenza due to matching patterns in our clinical dataset. This aligns with emerging healthcare AI trends emphasizing interpretability.

   3. Confidence Scores: Assigning confidence levels to diagnoses, like 80% likelihood of migraine, can help users gauge the systems certainty and prompt appropriate follow-up actions, such as seeking professional care for low-confidence cases.

      These features will be integrated into the React frontend using visual cues like tooltips or expandable sections to display source information without overwhelming users.

5. Rigorous Clinical Trials

   Preliminary testing with 200 synthetic queries and a 50-user pilot study showed promising results78% main diagnosis accuracy and 85% user satisfaction. However, real-world validation is critical. Future research will include:

   1. Large-Scale Clinical Trials: Conducting studies with thousands of patients across diverse demographics and healthcare settings to measure impacts on outcomes, such as reduced emergency visits or shorter specialist wait times. These trials will require Institutional Review Board approval and partnerships with healthcare providers.

   2. Gold-Standard Validation: Comparing MedConnects outputs against clinician diagnoses and

      established protocols, like ICD-10 classifications, to ensure clinical reliability, particularly for complex or chronic cases.

   3. Longitudinal Impact Studies: Evaluating long-term effects, such as user adherence to triage advice, improved health literacy, and reduced healthcare costs, especially in underserved regions with limited care access.

      These validations will provide evidence to support regulatory approval and clinical adoption, establishing MedConnect as a trusted telemedicine tool.

6. Multicultural and Global Reach

   To address global healthcare disparities, MedConnect AI can expand to serve non-English-speaking and diverse populations:

   1. Multilingual Support: Adding support for languages like Spanish, Mandarin, Hindi, and Arabic using multilingual embedding models, such as multilingual Sentence-Transformers, and translating FAQs from global health sources. This will extend access to regions with limited healthcare infrastructure, such as rural Latin America or Asia.

   2. Cultural Adaptation: Tailoring responses to reflect regional health perceptions and medical practices, such as incorporating traditional medicine knowledge where relevant, to enhance user acceptance.

   3. Low-Resource Deployment: Optimizing for low- bandwidth environments with lightweight frontend designs and offline caching to serve remote or underserved areas with limited internet access.

      These efforts will require collaboration with global health organizations and localization experts to ensure cultural sensitivity and effectiveness.

7. Infrastructure and Scalability

   To support widespread adoption, MedConnect AI must scale efficiently:

   1. Cloud Optimization: Leveraging cloud platforms like AWS or Google Cloud with auto-scaling and load balancing to support millions of users. Sharding the ChromaDB vector store and using distributed computing can manage large-scale knowledge bases.

   2. Edge Computing: Deploying lightweight system versions on edge devices, such as mobile apps with cached embeddings, to reduce latency in low- connectivity areas.

   3. API Redundancy: Configuring fallback language models or local models to ensure continuity during Gemini API downtime, enhancing system reliability.

These enhancements will position MedConnect AI as a globally accessible platform, addressing healthcare needs in both high- resource and low-resource settings.

CONCLUSION

MedConnect AI represents a major step forward in telemedicine by combining Googles Gemini API with a Retrieval-Augmented Generation framework. This creates a strong, accessible, and user-friendly health consultation platform. It uses a scalable MERN-stack architecture, which includes MongoDB, Express.js, Node.js, and React. The system draws from a well-organized medical knowledge base with 1,700 MTS-Dilog transcripts, 5,000 symptom-disease pairs, and 2,000 FAQs from trusted sources like the CDC and WHO. As a result, MedConnect AI achieves a main diagnosis accuracy of 78% and a differential diagnosis accuracy of 98%. This performance is significantly better than the baseline Gemini APIs 54% and 92%. A pilot study also shows an 85% user satisfaction rate, demonstrating the platforms effectiveness in offering accurate, evidence-based, and relevant medical guidance. It addresses critical healthcare issues like long appointment wait times and unequal access.

By basing its responses on verified medical data, MedConnect AI bypasses the limitations of generic large language models. These models often struggle with inaccuracies and a lack of specialized knowledge, leading to unclear or incorrect answers. The Retrieval-Augmented Generation pipeline includes data ingestion, vector-based retrieval via ChromaDB, and response generation through the Gemini API. This ensures that consultations are reliable, clinically sound, and tailored to user inputs, such as symptom descriptions and demographic information. The platforms user-friendly design has patient, doctor, and admin dashboards paired with GreenSock Animation Platform animations to support smooth interactions for all users, from patients seeking self-care advice to clinicians validating AI outputs.

Beyond its technical features, MedConnect AI helps patients make informed health choices, which may reduce unnecessary clinic visits and ease pressure on overloaded healthcare systems. In areas where specialist wait times average three weeks or longer and over 40% of the global population lacks access to essential health services, such tools are crucial for closing care gaps. The modular, full-stack JavaScript architecture ensures the platform is scalable and can adapt over time. This allows for ongoing updates to the knowledge base

and the integration of emerging technologies like electronic health records or multiple types of inputs.

As AI models and medical knowledge continue to progress, MedConnect AIs adaptable framework positions it to provide increasingly advanced, secure, and personalized health consultations. Future improvements, such as record integration, multilingual support, and large-scale clinical trials, will further expand its use among diverse populations and healthcare environments. While it is not a substitute for professional medical care, MedConnect AI shows how innovative AI techniques like Retrieval-Augmented Generation, vector databases, and API integration can develop effective telemedicine solutions.

REFERENCES

1. A. Ben Abacha, W. Yu, Y. Zhou, and M. Demner-Fushman, An empirical study of clinical note generation from doctorpatient encounters, in Proc. Eur. Assoc. Comput. Linguist., Dubrovnik, Croatia, May 2023, pp. 22912302. Description: Introduces the MTS-Dialog dataset, containing 1,700 annotated doctorpatient conversation transcripts, a key knowledge source for MedConnect AIs RAG pipeline, providing real-world clinical dialogue context.

2. L. A. Ferhi, M. A. Chami, and R. B. Slama, Enhancing diagnostic accuracy in symptom-based health checkers: A comprehensive ML approach, Frontiers Artif. Intell., vol. 7, p. 1397388, Mar. 2024. Description: Details the workflow of digital symptom checkers (symptom input, follow-up questions, diagnosis/triage output), aligning with MedConnect AIs design and informing its user interaction model.

3. W. Wallace, S. Chan, and P. R. Kumar, The diagnostic and triage accuracy of digital symptom checker tools: A systematic review, npj Digit. Med., vol. 5, no. 118, Aug. 2022. Description: Highlights the low diagnostic accuracy (1938%) of existing symptom checkers and their widespread adoption, underscoring the need for MedConnect AIs RAG-enhanced approach.

4. H. Eikenberry, J. Patel, and L. M. Schwartz, Retrieval-augmented generation for generative AI in healthcare, Nat. Mach. Intell., vol. 7, no. 1, pp. 1523, Jan. 2025.

   Description: Describes RAG principles, emphasizing improved large language model reliability through external data indexing, a core component of MedConnect AIs methodology.

5. B. Eastwood, How does RAG support healthcare AI initiatives?, HealthTech Mag., Jan. 2025. [Online]. Available: https://healthtechmagazine.net

   Description: Explains the RAG pipeline (data ingestion, retrieval, generation) and its benefits in healthcare, such as improved accuracy and traceability, which MedConnect AI leverages.

6. M. M. Farabi, Medical AI assistant, Google AI Gemini API Developer Competition, 2023. [Online]. Available: https://ai.google.dev Description: Demonstrates a Gemini API-based symptom analysis tool, serving as a precedent for MedConnect AIs conversational interface.

7. A. Chem, Health and safety assistant, Google AI Gemini API Developer Competition, 2023. [Online]. Available: https://ai.google.dev Description: Showcases a Gemini-powered symptom checker, illustrating the APIs applicability to MedConnect AIs health consultation features.

8. World Health Organization, Frequently asked questions on common health conditions, 2023. [Online]. Available: https://who.int Description: Provides curated medical FAQs used in MedConnect AIs knowledge base to address common health inquiries with reliable information.

9. GitHub, Diseasesymptom associations dataset, 2023. [Online]. Available: https://github.com

   Description: A structured dataset of 773 diseases and associated symptoms, forming a critical component of MedConnect AIs symptom- based diagnostics.

10. O. Kohandel Gargari and G. Habibi, Enhancing medical AI with retrieval-augmented generation: A mini narrative review, Digit. Health, vol. 11, pp. 17, Apr. 2025.

Description: Reviews RAG applications in healthcare, reporting accuracy improvements of up to 24%, supporting MedConnect AIs performance claim.