Toward Sustainable Healthcare : Implementing LLM Chatbots for Patient Support and System Efficiency

다국어 초록 (Multilingual Abstract)

This work proposes a bilingual healthcare chatbot based on RAG for medically accurate AI that is multilingual, accessible and computationally sustainable. The trained pipeline uses 182,822 clinical explanatory paragraphs sourced from the MedMCQA corpus and leverages the combined efforts of Dense Retrieval (MiniLM-L12-v2), Cross-encoder Reranking (ms-marco-MiniLM-L-6-v2), text generation (FLAN-T5) and bidirectional translation between English and Chinese (Helsinki-NLP) to mitigate hallucinations, false inferences and bilingual inaccuracies that plague large language models to date. A comprehensive evaluation of 300 held-out validation questions shows that the reranking significantly improves retrieval alignment (Recall@5 = 0.197 ± 0.398, MRR = 0.104 ± 0.267). An average cosine similarity score of 0.333 ± 0.245 substantiates clinically coherent multilingual results produced by the generator. Multilingual tests suggest that the results are coherent, clinically sound and semantically comparable in and across English and Chinese.
System latency (1961.3 ± 1574.7 ms) and small model parameters align with Green AI which suggests that it's adaptive and doesn't need intensive resources. The varied assessment across 21 domains suggests that performance differs which denotes the challenging nature of domain specific medical explanations on a large scale. These findings suggest that retrieval conditioned generation is a trustworthy method to advance factual citation and suppress unfounded statements even across medically diverse QA settings. Low retrieval efficacy and a dataset containing Q-A pairs like practical exams limitation results, however, this paper offers three novel contributions to the field: (1) A trustworthy end to end bilingual RAG framework for medical QA, (2) Empirical data championing the significant findings on reranking and retrieval depth, and (3) An integrative evaluation that brings together semantic, multilingual and Green AI considerations for clinical responsible AI. This research establishes the foundation for future domain specific retrievers, hybrid search algorithms, multilingual expansions, uncertainty-aware reasoning, and real-world trials with clinicians and patients. Finally, the proposed strategy provides a realistic, transparent, and morally sound pathway for developing AI technology capable of supporting modern global healthcare circumstances.
Keywords: AI in healthcare, LLM, Chatbots, Patient Support System, RAG, Hallucination, Privacy and Data Security, AI sustainability or Green AI.

목차 (Table of Contents)

• Table of Contents
• Chapter 1. Introduction……………………………………………………………………………1
• 1.1. Background of the Study…………………………………………………………………...2
• 1.2. Problem Statement…………………………………………………………………………3
• 1.3. Motivation and Research work…………………………………………………………….4
• 1.4. Purpose of Study…………………………………………………………………………...5
• 1.4.1. Research Gap Summary ………………………………………………………………5
• 1.5. Organization of thesis……………………………………………………………………...6
• Chapter 2. Literature Review…………………………………………………………………….9
• 2.1.The Feasibility and Level of Generative AI and LLMs in Healthcare …………9
• 2.2.Empathy, Efficacy, and Trust in LLM Chatbots …………………………………………...10
• 2.3. Workflow Efficiency and Clinical Documentation …………………………………..… 11
• 2.4. Safety, Assessment and Clinical Decision Making ……………………………….13
• 2.5. Present Challenges………………………………………………………………………..14
• 2.6. Overview of the Literature Review……………………………………………………….17
• 2.7. Research Questions ………………………………………………………………………18
• 2.8. Recent Advances in LLMs and Generative AI in Healthcare…………………………….24
• Chapter 3. Methodology…………………………………………………………………………26
• 3.1. Research Framework……………………………………………………………………...26
• 3.2. Mathematical Foundations of RAG………………………………………………………29
• 3.2.1. Retrieval Phase……………………………………………………………………….30
• 3.2.2. Phase of Reranking…………………………………………………………………...30
• 3.2.3. Generation Phase ………………………………………………………………….. 31
• 3.2.4. Handling Cross-Lingual……………………………………………………………...31
• 3.2.5. Evaluation Metrics…………………………………………………………………... 31
• 3.3. Dataset Collection and Description……………………………………………………….33
• 3.3.1. Layer Architecture of the RAG-Based Healthcare Chatbot………………………….34
• 3.4. Data Preprocessing Pipeline………………………………………………………………35
• 3.4.1. Preprocessing Workflow Overview…………………………………………………..36
• 3.4.2. Visualization of the Preprocessing Pipeline………………………………………….37
• 3.4.3. Alignment with Evaluation Framework……………………………………………...39
• 3.5. Evaluation Metrics and Model Performance ……………………………………………..39
• 3.5.1. Retrieval Evaluation Metrics ………………………………………………………...39
• 3.5.2. Generation Evaluation Metrics ………………………………………………………40
• 3.5.3. Integrated Performance Summary …………………………………………………...41
• 3.5.4. Ethical Consideration in Evaluation …………………………………………………41
• 3.6. Comparative Model Performance………………………………………………………...42
• 3.6.1. Quantitative Comparison……………………………………………………………..43
• 3.6.2. Qualitative Generation and Bilingual Evaluation ……………………………………44
• 3.6.3. Integrated Model Performance ………………………………………………………45
• 3.6.4. Ethical and Reliability Context……………………………………………………….46
• 3.7. Validity and Ethical Considerations………………………………………………………46
• 3.7.1. Quantitative and Qualitative Validity………………………………………………...47
• Chapter 4. Results………………………………………………………………………………..49
• 4.1. Retrieval Performance Evaluation………………………………………………………..49
• 4.1.1. Baseline Comparison Table (BM25 vs. Dense, Dense + Reranker) with p-values …………………………………………………………………………………………………...51
• 4.1.2. Reporting Annual Energy Consumption (Colab Estimate) ………………………... 52
• 4.1.3. Stratified Analysis …………………………………………………………………. 53
• 4.2. Ablation Study……………………………………………………………………………56
• 4.2.1. Reranking and Retrieval Depth………………………………………………………56
• 4.3. Visualization of Reranking Impact……………………………………………………….58
• 4.4. Cross-Lingual Question–Answer Evaluation……………………………………………..59
• 4.5. End-to-End RAG System Performance…………………………………………………...60
• 4.6. Summary of Findings……………………………………………………………………..61
• Chapter 5. Research Findings……………………………………………………………………63
• 5.1. Discussion ……………………………………………………………………………... 63
• 5.2. Contributions ……………………………………………………………………...…... 66
• 5.3. Limitations.……………………………………………………………………………. 66
• 5.4. Future Work …………………………………………………………………………… 69
• 5.5. Conclusion ……………………………………………………………………………. 71
• Chapter 6. References and Appendix…………………………………………………………… 73
• 6.1. References …………………………………………………………………………….73
• 6.2. Appendix …………………………………………………………………………….. 84