by InTrieste
Artificial intelligence is increasingly finding a place in medicine — not as a replacement for physicians, but as a tool designed to support clinical decision-making. A new study led by researchers at the University of Trieste suggests that generative A.I., when carefully trained and validated, may significantly improve the management of hepatitis C, one of the world’s most persistent public health challenges.
The research, published in the journal Liver International, describes an A.I.-based clinical assistant developed by an international team coordinated by Mauro Giuffrè, a researcher in the Department of Medical, Surgical and Health Sciences at the University of Trieste and at Yale University School of Medicine. The system is built on GPT-4, the large language model developed by OpenAI, and has been specifically adapted to support clinical decisions related to hepatitis C virus (HCV) infection.
Hepatitis C remains a major global health concern. According to estimates from the World Health Organization, around 58 million people worldwide live with chronic HCV infection, which can lead to severe complications including liver cirrhosis and hepatocellular carcinoma. In response, the W.H.O. has set an ambitious goal: eliminating hepatitis C as a public health threat by 2030, through a 90 percent reduction in new infections and a 65 percent reduction in mortality. Achieving that target depends on early diagnosis, appropriate treatment and strict adherence to clinical guidelines — areas where A.I., the researchers argue, could offer meaningful support.
The system described in the study was “prepared,” in the researchers’ terminology, using the European clinical guidelines for hepatitis C treatment and then validated directly by the same experts who authored those guidelines. Rather than relying on a single approach, the team tested two distinct strategies for training the model.
The first approach, known as retrieval-augmented generation, or RAG, allows the A.I. to retrieve relevant sections of the European guidelines in real time when responding to clinical questions. Two versions were evaluated: one that incorporated a single key paragraph (RAG-Top1) and another that integrated ten relevant passages (RAG-Top10). The second approach, called supervised fine-tuning, involved directly training the model on the guideline texts themselves.
The differences in performance were striking. The base version of GPT-4, without any domain-specific adaptation, achieved an accuracy rate of 36.6 percent when evaluated by clinical experts. By contrast, the RAG-Top10 system reached an accuracy of 91.7 percent, followed by RAG-Top1 at 81.7 percent and the supervised fine-tuned model at 71.7 percent. The results suggest that structured, transparent access to authoritative scientific sources plays a decisive role in making A.I. outputs clinically reliable.
Equally notable was the study’s validation process, which the authors describe as unprecedented in this field. The A.I.’s responses were assessed by two independent groups. One consisted of four hepatologists selected from among the authors and chairs of the European Association for the Study of the Liver’s hepatitis C guidelines — clinicians who are among Europe’s leading experts on the disease and directly responsible for shaping international standards of care. The second group included physicians from a tertiary referral center, Humanitas Hospital in Rozzano, Italy.
This dual evaluation, combining guideline authorship and real-world clinical practice, allowed what the researchers describe as an assessment approaching a “gold standard” for evaluating A.I. accuracy in medicine. According to the authors, both the RAG and supervised fine-tuning approaches improved not only the correctness of responses but also their clarity and the appropriateness of treatment selection across different clinical scenarios.
The researchers emphasize that the system is not intended to replace clinicians. Instead, it is designed as a decision-support tool, particularly useful in complex cases where strict adherence to evolving guidelines is essential. In settings where specialist expertise may be limited, such tools could help standardize care and reduce variability in treatment decisions.
While further testing will be needed before such systems can be deployed in routine clinical practice, the study adds to a growing body of evidence that generative A.I., when rigorously trained and validated, may play a meaningful role in supporting physicians. In the case of hepatitis C — a disease for which effective treatments already exist, but are not always optimally applied — that support could prove especially valuable.
