One Diagnosis, Different Opinions: Russia Develops a Method to Evaluate Medical AI
Researchers at Lomonosov Moscow State University have developed a methodology for evaluating the responses of medical AI systems while accounting for disagreements among physicians. The approach makes it possible to compare language models with expert assessments more accurately. The findings were published in the international scientific journal Scientific Reports.
Making a diagnosis is always a complex and consequential task, and the answer is not always clear-cut. Different physicians may describe the same diagnosis in different ways. One may use formal medical terminology, another its abbreviated forms, while a third may provide a detailed description of the symptoms. All three assessments can be correct and describe the same underlying condition.
For medical AI, however, this variation creates a problem. AI is not a person and has no clinical experience. To evaluate a neural network's performance, its responses are typically compared with a physician's assessment, which serves as the reference standard. But if another physician formulates the diagnosis differently, the system may receive a lower score even when its answer is essentially correct.
Researchers at the Center for Artificial Intelligence at Lomonosov Moscow State University, working with colleagues from the Artificial Intelligence Research Institute (AIRI) and the Almazov National Medical Research Centre, have proposed a way to account for this variation. Their methodology compares AI-generated responses not with a single physician's assessment but with the opinions of several physicians, while also taking into account how closely the experts agree with one another.
How the Method Works
The researchers developed two relative metrics. These compare how closely a language model's responses align with physicians' assessments and how closely the physicians agree among themselves. This makes it possible to evaluate an AI system in light of the level of consensus among medical professionals in a particular case.
The researchers also developed a separate algorithm to compare diagnostic formulations. It helps determine which diagnostic hypotheses are equivalent in meaning, even when they are worded differently. The algorithm takes into account assessments generated by language models, information from the International Classification of Diseases (ICD), and the textual similarity of the formulations.
This helps prevent a neural network from receiving a low score simply because it uses different terminology. For example, a physician might write “acute gastritis,” while the AI system might describe “acute inflammation of the stomach lining.” The wording is different, but both refer to the same condition. The methodology helps identify this semantic equivalence.
Physicians Put the Method to the Test
The researchers tested the methodology on 360 Russian-language dialogues. Medical professionals acted out some of the conversations, while the rest featured actors playing patients. Each dialogue was independently evaluated by seven medical residents. The researchers then compared the physicians' assessments with the diagnostic hypotheses generated by eight language models.
The results showed that several models agreed with the experts, on average, at least as closely as the physicians agreed with one another. Moreover, the variation in agreement scores among physicians was greater than the variation between most of the AI systems tested and the doctors.
For the researchers, this is an important finding: evaluations of medical AI need to account for the natural differences in how physicians assess cases. Otherwise, testing may give a distorted picture of a technology's capabilities. The authors also stress that agreement between an AI system's responses and physicians' opinions does not, by itself, establish that a diagnosis is correct. Nor does this type of evaluation replace clinical trials or physician review.
What Could This Mean for Healthcare?
The new methodology could be useful in developing and comparing clinical decision support systems. These tools help healthcare professionals analyze patient information, assess symptoms, and formulate possible diagnoses. But before such systems can be used in clinical practice, it is essential to understand how reliably they perform.
A more objective evaluation would help developers identify the strengths and weaknesses of different models, compare competing systems, and improve their algorithms. For healthcare organizations, it could also mean clearer, more informative results when evaluating software systems.
The research also has international potential. Medical language models are being developed and tested in many countries, and differences among physicians' assessments are not unique to any one healthcare system or language. An approach that accounts for disagreement among experts could therefore be of interest to international developers of medical software and research centers.
