Jakarta, INTI - Indonesia’s National Research and Innovation Agency (BRIN) is developing Chatbot LISA, an artificial intelligence (AI)-powered virtual assistant designed to help patients with chronic kidney disease undergoing hemodialysis access reliable health information. The chatbot combines Large Language Model (LLM) technology, Retrieval-Augmented Generation (RAG), and medical taxonomies to ensure that its responses are grounded in validated medical sources.

Elvira Nurfadhilah, Junior Research Engineer at BRIN’s Research Center for Data Science and Information (PRSDI), presented the development of Chatbot LISA during the 10th edition of the Data Science and Information Exploration Dialogue (DESAIN) webinar on October 21. The session, titled “From Emergency Departments to Hemodialysis: AI Innovations for Patient Care and Support,” explored how AI can contribute to healthcare services and patient assistance.

LISA is designed for hemodialysis patients, their families, and the general public. Its knowledge resources cover basic information about hemodialysis, daily care, fluid and nutritional management, and vascular access maintenance. The chatbot is intended to complement existing patient education rather than replace the role of doctors and nurses.

RAG Technology Helps Deliver Reliable Medical Information

One of the main challenges in developing healthcare chatbots is ensuring that responses are not only convincing but also supported by traceable and reliable information. To address this issue, LISA uses a RAG approach that retrieves information from reference documents before generating an answer.

Elvira explained that, unlike conventional chatbots that generate responses based primarily on knowledge learned during model training, RAG allows the system to retrieve relevant information from an external knowledge base.

“With RAG, we do not always need to retrain or fine-tune the LLM. We can simply update the RAG database,” she explained.

This approach enables the team to update the chatbot’s knowledge by revising its reference documents without repeatedly training the underlying language model.

LISA’s knowledge base incorporates national medical guidelines, including Indonesia’s Ministry of Health National Guidelines for Medical Services (PNPK) and consensus recommendations from the Indonesian Society of Nephrology (PERNEFRI). International references include Kidney Disease: Improving Global Outcomes (KDIGO), the Kidney Disease Outcomes Quality Initiative (KDOQI), and the National Kidney Foundation (NKF).

The team also uses selected healthcare articles, scientific journals, and questions collected from health portals, patient communities, and medical professionals to broaden the system’s knowledge resources.

Medical Taxonomy Helps LISA Understand Patient Language

Another challenge in developing the chatbot is recognizing the everyday terminology patients use when discussing their treatment. For example, Indonesian expressions such as “cuci darah,” “HD,” and “dialisis darah” refer to the same medical procedure: hemodialysis.

To address these variations, the team has developed a medical taxonomy covering 350 nephrology concepts and 1,841 synonyms in Indonesian and English. These terms are mapped to established medical classification and terminology systems, including the International Classification of Diseases (ICD) and Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT).

This approach is intended to help LISA interpret patient questions more accurately, even when users employ informal expressions or different terminology for the same medical concept.

In testing involving 71 questions validated by medical professionals, with each question tested ten times, the LISA approach achieved a reference retrieval accuracy score of 0.88, up from 0.80 using the baseline search method.

Answer relevance also improved from 0.89 to 0.92, while the answer faithfulness score remained at 0.90. According to Elvira, the improvements in retrieval accuracy and answer relevance were statistically significant.

These results indicate the potential of combining medical taxonomy and RAG to improve the relevance and traceability of information provided by healthcare chatbots.

Patient Safety Remains a Key Development Priority

BRIN is also prioritizing patient safety by limiting LISA’s responses to the information available in its reference sources. Questions outside the chatbot’s scope, such as those about earthquakes, are not answered when relevant information cannot be found in its knowledge base.

Questions involving individual treatment decisions, specialized therapies, emergency situations, and personal clinical judgments are instead directed to doctors or nurses for further consultation.

The team is also preparing a golden dataset, a collection of question-and-answer pairs accompanied by references and reviewed through clinical validation. The process involves five healthcare professionals: two senior hemodialysis nurses, two general practitioners trained in hemodialysis, and one internal medicine specialist.

The validation effort targets 531 question-and-answer pairs and approximately 825 medical terms. Elvira noted that the number of terms remains provisional, as the validators must confirm each term’s relevance to hemodialysis or other medical fields.

This validation process is expected to strengthen the chatbot’s reliability and help identify areas requiring further refinement before broader implementation.

BRIN Expands RAG Research Beyond Hemodialysis

In addition to Chatbot LISA, Elvira highlighted the development of RAGLab, a platform designed to build, test, and compare different RAG-based chatbot configurations without requiring developers to rewrite program code.

The team has also secured copyright for RAGLab version 1.0 and published scientific research on taxonomy-based query processing for Indonesian-language hemodialysis chatbots.

Future development plans include further validation by medical experts and patients, as well as research into cross-language hallucination detection through collaboration with France’s Laboratoire Interdisciplinaire des Sciences du Numérique (LISN), the National Centre for Scientific Research (CNRS), and Université Paris-Saclay.

Beginning in 2027, the team also plans to develop a Knowledge-Grounded RAG approach to strengthen the relationship between generated responses and verified knowledge sources.

Beyond kidney disease care, BRIN aims to adapt its RAG platform for other healthcare areas, including diabetes and hypertension.

Through Chatbot LISA and its broader RAG research initiatives, BRIN is exploring how AI can improve access to health information while maintaining clear boundaries around clinical decision-making. By combining validated medical references, specialized terminology, and professional oversight, the initiative aims to support patient education and advance the responsible use of AI in Indonesia’s digital healthcare ecosystem.

Conclusion 

BRIN’s AI-powered Chatbot LISA demonstrates how artificial intelligence can support chronic kidney disease patients by providing accessible, reference-based health information. By combining RAG technology, medical taxonomies, and clinical validation, the initiative aims to improve the reliability of digital health education while maintaining patient safety. BRIN’s ongoing research could also pave the way for similar AI-powered healthcare solutions for conditions such as diabetes and hypertension.

Read more: BRIN Identifies Nine Mangrove Species in Aceh with Potential for Medicinal and Healthcare Applications