Pengembangan Model Hybrid Generatif Untuk Chatbot Call Center Medis Menggunakan Transformer Dan Large Language Model

Sanjani, Lukman Arif (2026) Pengembangan Model Hybrid Generatif Untuk Chatbot Call Center Medis Menggunakan Transformer Dan Large Language Model. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025232027-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Transformasi digital mendorong kebutuhan chatbot cerdas yang mampu memberikan respons cepat dan akurat, khususnya pada domain kesehatan yang kompleks. Salah satu tugas administratif penting dalam praktik klinis adalah penyusunan medical discharge instructions, yang sering kali memakan waktu dan rentan terhadap inkonsistensi. Meskipun kemajuan Natural Language Processing (NLP) melalui arsitektur Transformer dan Large Language Model (LLM) memungkinkan otomasi proses ini, penggunaan model generatif murni menghadirkan risiko hallucination yang berpotensi membahayakan secara klinis. Oleh karena itu, penelitian ini mengusulkan pendekatan Retrieval-Augmented Generation (RAG) yang dikombinasikan dengan klasifikasi klinis untuk menghasilkan teks medis yang lebih akurat dan terkontrol.Penelitian ini mengembangkan metode hybrid berbasis dataset MIMIC-IV dengan mengintegrasikan prediksi kode ICD-11 dan RAG untuk menghasilkan medical discharge instructions secara otomatis. Prediksi ICD-11 berfungsi sebagai struktur awal dan penentu konteks klinis pada tahap retrieval. Evaluasi menunjukkan skor faithfulness sebesar 0,95 pada RAG Triad, serta kinerja terbaik BERTScore pada RoBERTa-Large dengan nilai F1 sebesar 0,8254. Meskipun contextual relevancy masih moderat akibat variasi dan keterbatasan data klinis, hasil penelitian menunjukkan bahwa pendekatan ini efektif untuk mendukung pengembangan chatbot medis yang akurat dan kontekstual.
==================================================================================================================================
Digital transformation has driven the need for intelligent chatbots capable of providing fast and accurate responses, particularly in complex domains such as healthcare. One critical administrative task in clinical practice is the preparation of medical discharge instructions, which is often time-consuming and prone to inconsistency. Although advances in Natural Language Processing (NLP) through Transformer architectures and Large Language Models (LLMs) enable the automation of this process, the use of purely generative models poses a risk of hallucination that may be clinically harmful. Therefore, this study proposes a Retrieval-Augmented Generation (RAG) approach combined with clinical classification to produce more accurate and controlled medical text.This study develops a hybrid method based on the MIMIC-IV dataset by integrating ICD-11 code prediction with RAG to automatically generate medical discharge instructions. ICD-11 prediction serves as the initial document structure and as a determinant of clinical context during the retrieval stage. Evaluation results show a faithfulness score of 0.95 using the RAG Triad, as well as the best BERTScore performance achieved by RoBERTa-Large with an F1 score of 0.8254. Although contextual relevancy remains moderate due to variability and limitations in clinical data, the findings demonstrate that the proposed approach is effective in supporting the development of accurate and context-aware medical chatbots.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Clinical Text Geneation, Fine-tuning, Hybrid Model, Large Language Model, Retrieval-Augmented Generation, Transformer
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.76.E95 Expert systems Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) R Medicine > R Medicine (General) > R858 Deep Learning
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Lukman Arif Sanjan
Date Deposited:	30 Jan 2026 02:29
Last Modified:	30 Jan 2026 02:29
URI:	http://repository.its.ac.id/id/eprint/131167

Actions (login required)

View Item