IBS symptoms are subjective and a five-minute visit rarely captures enough of them. The hospital built an AI consultation system on MOI: patients complete a guided interview from home, while doctors get reply suggestions, subtype classification and auto-generated records.
Wuhan Union Hospital is a Class-A Tertiary teaching hospital affiliated with Tongji Medical College, Huazhong University of Science and Technology, and one of China's leading general hospitals.
IBS is a chronic condition whose symptoms are, in most cases, not life-threatening — and patients are correspondingly reluctant to make a hospital trip for it. Against a slow, cumbersome in-person pathway, they would rather consult online.
The existing in-person model is not only time-consuming; because consultation is not standardized it carries a risk of missed and mistaken diagnosis, and it offers little in the way of long-term, individualized follow-up. That is a poor fit for what IBS patients actually need.
The hospital had accumulated a great deal of valuable clinical case data, but it could not simply be exported for analysis and model training. Data silos and non-standard formats meant this material could not support large-scale data-driven research or AI development, keeping clinical research and clinical practice apart.
At the same time a very large IBS patient population presses on limited specialist clinic capacity. And because the consultation system is medical device software, its development had to satisfy regulatory requirements — demonstrating safety and effectiveness and passing clinical validation, the precondition for Class II medical device software registration and the point reviewers scrutinize most closely.
The hospital brought in MatrixOne Intelligence, a multi-modal AI data platform, to build an end-to-end IBS consultation system running from data processing through to the intelligent application.
The platform combines data governance, intelligent parsing, multi-modal search and a converged data foundation. It parses and structurally extracts from unstructured sources — PDFs, audio and video, images — and builds high-quality training datasets shaped to the requirements of a specific model, which is what gives the AI application a solid data base to stand on.
De-identified unstructured text was extracted into a predefined schema — patient information, chief complaint, history of present illness, diagnosis — as normalized JSON. This converted data that had been locked inside images as dead weight into material a machine can analyse and use.
From those structured records, a locally deployed large language model (Qwen3-32B) with prompts designed around the consultation workflow generated doctor-patient dialogue datasets closely modelled on real clinical encounters.
Two augmentation strategies addressed early-stage model weaknesses. Personas — patients with different temperaments and ways of expressing themselves — were introduced to regenerate dialogue, greatly widening diversity and improving the model's empathy and conversational naturalness. And negative samples were added: 10–20% non-IBS or deliberately confusable adversarial cases, correcting the model's tendency to diagnose IBS in every conversation and training genuine differential diagnosis.
Automated consultation and record generation free clinicians from repetitive work, returning significant time to them.
The system standardizes the consultation pathway: data collection follows current clinical guidelines and expert consensus, reducing the risk of missed diagnosis, and individualized evidence-based treatment plans are generated from the patient's specific situation — which supports both adherence and long-term management.
Online pre-consultation triages milder cases effectively, relieving pressure on the gastroenterology specialist clinic, while patients gain a more convenient and more private consultation channel.
And the project turned dormant hospital data into a research asset. The large body of structured real-world data the system accumulates supports frontier research into IBS mechanisms and subtype transitions in a way that was not previously possible.