AI News

Four general-purpose large models average 64.35 points on medical evaluation; AI consultations can only serve as an auxiliary

Shanghai ObserverUpdated Nov 16 21:06

AI summary

Shanghai Observer reported that a journalist at the Medical Intelligence Center of Shanghai Artificial Intelligence Laboratory tested four general-purpose large models, domestic and foreign, using 36 evaluation sets and 1 million test questions, with an average score of 64.35. The head of the center said that a score in the 60s is mid-range, and that large models cannot be used directly in clinical practice and can only assist doctors. The report also mentioned that a parent refused an examination on the grounds that an AI diagnosis differed from the doctor's, and the child subsequently developed severe pneumonia; for the same case, when the input information differed in level of detail, the conclusions given by the AI varied greatly.

Why read this

AI score 59

The report uses the medical evaluation results of four general-purpose large models and cases of input differences to illustrate the capability boundaries and risks of AI consultations.

Topics

Tags

Body · AI translation

Have you noticed that ever since the elders in the family learned to use AI, it's almost like opening a door to a new world? At the first sign of a headache or a slight fever, they don't rush to the hospital to see a doctor, but instead ask AI to give them a "diagnosis and treatment plan." Parents trust AI over doctors, child develops severe pneumonia. Recently, a parent took a 3-year-old child to the pediatric department of the First Affiliated Hospital of Zhejiang Chinese Medical University. The doctor considered the child to have asthma complicated by a lung infection and recommended that the child undergo a chest X-ray and other examinations. However, the parent, on the grounds that the diagnosis given by AI differed from the doctor's, refused to accept further examinations, and then the parent took the child and left on their own. But just a few days later, this parent brought the child to the hospital again. The doctor's examination found that the child had by then developed severe pneumonia. The doctor said that during consultations nowadays they often encounter some parents who, holding the results of consulting AI, question the doctor: "Why is what you say different from what AI says?" So, can general-purpose large AI models actually treat patients directly? How reliable is AI consultation? Four general-purpose large models averaged 64.35 points, can only serve as an aid. With these questions in mind, the reporter went to the Medical Intelligence Center of Shanghai Artificial Intelligence Laboratory (the center has established China's first evaluation and verification platform for medical large model applications) and gave an "exam" to four general-purpose large models from home and abroad. The evaluators specially designed a standardized question bank covering 36 evaluation sets and 1 million test questions, capable of testing the medical application capabilities of large models from multiple dimensions such as medical knowledge Q&A, complex medical reasoning, and medical safety and ethics. The evaluation results showed that the four general-purpose large models had an average score of 64.35. The person in charge said: "A score of 60-plus is mid-range, not particularly bad among general-purpose models. But first, large models cannot be directly applied in clinical practice and can only assist doctors. Second, for the public, the leaderboard is just an open-source evaluation leaderboard; what it gives is the model's current technical professional capability, not clinical application capability." The medical knowledge and Q&A capabilities that large models possess do not equal the ability to diagnose and treat independently, nor do they mean that large models can replace doctors. AI consultation results depend on your description. How much does AI's answer "depend on the input"? Here's a very intuitive example: Input ①: 35-year-old male, liver space-occupying lesion found on ultrasound two weeks ago. Input ②: 35-year-old male, liver space-occupying lesion for two weeks, with a 15-year history of hepatitis B, a 10-year history of alcohol consumption, and alpha-fetoprotein 450 (ng/ml). With these two input methods, the conclusions given by AI differ greatly. In the first case, AI reminds that benign lesions are mostly likely. In the second case, AI reminds to be alert to the risk of cancerous change. Experts say that the answers given by AI are highly dependent on the descriptive information provided by the user. However, in medical scenarios, limited by the lack of a medical background, ordinary patients often find it difficult to accurately describe their condition, and omissions and misstatements are very common in actual consultations. This kind of error in information expression can be gradually corrected through "observation, listening, questioning, and pulse-taking" during a doctor's face-to-face consultation. Yan Feng, a professor in the Department of Chinese Language and Literature at Fudan University, also wrote in a post: "When asking AI about medical questions, it often has a tendency to cater to people." Netizens have also discussed this heatedly: "AI's statements cannot be fully trusted; do not avoid seeing the doctor for fear of illness. When seeing a doctor, you should still follow the doctor's professional judgment." In addition, the reporter's investigation also found that the AI "digital doctors" in many apps on the market falsely use experts' identities to provide users with health consultation services, and the doctors themselves are unaware of it. For large models, only by clarifying the boundaries of rights and responsibilities between AI and doctors, drawing a clear red line between health consultation and diagnostic consultation, and clarifying the attribution of responsibility can the technological potential of AI healthcare truly be transformed into public health benefits while ensuring patient safety and regulating industrial development. For ordinary users, Xiaonong kindly reminds: AI can be our helper for looking up common health knowledge, but it must never replace a face-to-face doctor. There is no room for trial and error in health; when you feel unwell, seeking medical attention promptly is the safest choice. Have you ever tried AI consultation? Do you think the judgments it gives are reliable? [Nonghao Shanghai · Han Jiayan, comprehensive editor] Nonghao Shanghai (ID: helloshanghai2013) comprehensive editing, please indicate the source when reprinting. Material sources @CCTV News @China Newsweek, and netizens from various channels on Weibo. If there is any offense, please contact us. Copyright statement: Reproduction without authorization is declined.

Source

Shanghai Observer