Imagine a pair of eyes that feel gritty, burn under the sun, or blur when you try to read a screen. This is dry eye disease, a common and often frustrating condition where the surface of the eye lacks sufficient moisture or the right kind of oil to stay healthy. It is not just a minor annoyance; it can damage vision and make daily life difficult. Diagnosing it, however, is tricky. A doctor cannot rely on a single test. Instead, they must piece together a puzzle made of many different clues: how the patient describes their pain, their medical history, their lifestyle habits like how much time they spend on computers, and what the doctor sees when looking through a microscope at the eye's surface. For years, artificial intelligence has been good at looking at pictures or reading numbers, but it has struggled to do both at once in a way that mimics a human doctor's careful reasoning.

A team of researchers from hospitals and research institutes across China has built a new kind of computer program designed to solve this exact problem. They created a system that acts like a digital expert, capable of reading a patient's written answers to questionnaires, understanding their medical history, and examining high-resolution photos of their eyes all at the same time. This system, which they call a multimodal large language model, was trained to recognize the complex patterns that signal dry eye disease. By combining these different types of information, the model aims to spot the disease early, even before it becomes severe, and to confirm a diagnosis with high accuracy. The researchers tested this tool on hundreds of patients and found that it could effectively distinguish between healthy eyes and those suffering from dry eye, performing well even when tested on data from different hospitals with different equipment.

The study began by gathering a large collection of real-world data from four different hospitals. The team enrolled 874 participants, dividing them into two groups. The first group, consisting of 743 patients, came from a single hospital in Beijing and was used to teach the computer how to recognize the disease. The second group, with 131 patients from three other hospitals in different cities, was kept separate to test whether the computer could apply what it learned to new, unseen situations. Every participant underwent a thorough examination. They filled out detailed surveys about their symptoms, such as how often their eyes felt dry or how much their vision fluctuated. They also provided their medical history, including information about their age, any other health conditions they had, and their daily habits, such as how much time they spent looking at electronic screens.

Alongside these written answers, the doctors took detailed photographs of the patients' eyes. These images showed the surface of the eye, the tiny oil glands in the eyelids, and the blood vessels. In some cases, the doctors used a special dye to highlight damaged areas on the eye's surface. The researchers fed all this information—the survey answers, the medical history, and the eye images—into their new computer model. The model was built on a powerful foundation of existing artificial intelligence technology but was specially tuned to understand medical language and images. A key part of the design was a method to translate the structured numbers from the medical records, like scores from the surveys, into natural language that the computer could process alongside the images. This allowed the model to see the connections between a patient saying their eyes burn and the specific redness visible in a photograph.

The researchers asked the model to perform two main tasks. The first was early warning, where the system tried to predict the risk of dry eye based only on the patient's survey answers and lifestyle factors, before any eye images were taken. The second task was diagnosis, where the model used all the available information, including the eye images, to determine if a patient definitely had the disease. The team tested the model against two different sets of rules used by doctors: one set of international guidelines and another set of guidelines specific to China. These rules define exactly what combination of symptoms and signs counts as the disease.

The results showed that the model was highly effective. When trying to predict the risk of dry eye using only the questionnaire data, the model was correct about 96 percent of the time according to the Chinese guidelines and about 82 percent of the time according to the international guidelines. It was particularly good at catching the disease when it was present, missing very few cases. When the model was given the eye images in addition to the survey answers to make a final diagnosis, its accuracy improved further. It correctly identified the disease in more than 92 percent of cases under international rules and over 93 percent under Chinese rules. The model was also very good at correctly identifying people who did not have the disease, avoiding false alarms.

To understand how the model made its decisions, the researchers looked at which pieces of information it relied on most. They found that the structured data from the questionnaires and medical records was the most important factor. When they removed this information and asked the model to rely only on the eye images, its performance dropped significantly. This suggests that for dry eye disease, the patient's own description of their symptoms and their medical history are the strongest indicators. However, the images were still valuable. The model used them to confirm its suspicions and to point out specific physical signs, such as blocked oil glands or tiny spots of damage on the eye surface. The computer learned to focus its attention on the exact areas of the eye that a human doctor would examine, highlighting the redness and the texture of the eyelid glands.

The team also tested the model on the separate group of patients from the other three hospitals to see if it would work in different real-world settings. These hospitals used different cameras and equipment to take the eye photos. Despite these differences, the model maintained its high level of accuracy. It successfully diagnosed the disease in the new group of patients, proving that it could adapt to variations in how the data was collected. This suggests that the tool could be useful in a wide range of clinics, not just the one where it was first created. The researchers noted that the model was especially good at identifying the disease in older adults, a group where the condition is most common, though they acknowledged that the data included fewer cases of very mild disease.

The study concludes that this new approach offers a promising way to improve how dry eye disease is detected and diagnosed. By combining the patient's voice with the doctor's visual evidence, the model creates a more complete picture of the patient's health than either method could provide alone. The researchers emphasize that while the tool is powerful, it is designed to assist doctors, not replace them. It provides a robust, scalable way to screen for the disease and support clinical decisions, potentially helping to catch the condition earlier and reduce the burden on patients who suffer from chronic eye discomfort. The work demonstrates that when artificial intelligence is trained to understand both the numbers and the images that doctors use every day, it can become a reliable partner in caring for eye health.