[The Editors' Verdict]AI Doctors Four Times More Accurate Than Humans...A Solution to Overtreatment?

'Overtreatment' is a chronic issue in the medical field. Just a few years ago, it was common for excessive fees to be charged to elderly patients nearing the end of life under the pretext of various tests. Reports continue to emerge about doctors recommending unnecessary surgeries, such as for conjunctivochalasis or cataracts, which has led to ongoing conflicts with the insurance industry. The government is now pushing to introduce a 'selective rider for indemnity insurance,' which would lower insurance premiums if patients exclude excessive non-covered items such as manual therapy or MRI scans.

Against this backdrop, a study on an artificial intelligence (AI) diagnostic system announced by Microsoft (MS) on June 30 (local time) has raised hopes for solving the problem of overtreatment. The 'MAI-DxO' disease diagnosis system developed by MS was able to identify difficult-to-diagnose diseases with up to 85.5% accuracy. This result comes from experiments conducted on 304 challenging disease cases published in the world-renowned New England Journal of Medicine (NEJM).

In the 'MAI-DxO' system, a team of five AI doctors work together to make a diagnosis. The first AI repeatedly asks questions based on the patient's symptoms to estimate what and where the problem is. The second AI recommends tests that should be performed for a more precise diagnosis. The third AI challengingly refutes the disease estimated by the first AI and raises the possibility that the patient may have a different illness. The fourth AI examines whether there are more affordable testing methods than those previously suggested, and the final AI checks for any erroneous reasoning throughout the process before making the final diagnosis.

Human doctors attempted to diagnose in the same way, but their accuracy rate was only 20%. The AI doctor group achieved an accuracy rate nearly four times higher than that of human doctors. The 21 human doctors compared to the AI included family medicine (general practitioners) and internal medicine (specialists) physicians with an average of 12 years of experience. While general practitioners cover a wide range of medical fields, specialists have in-depth expertise in a particular area. The researchers analyzed that AI was able to achieve superior results because it combined the strengths of both general practitioners and specialists.

What is particularly noteworthy is that MS's research addressed the 'cost' issue in depth. Currently, healthcare spending in the United States is approaching 20% of GDP, and up to 25% of this is estimated to be unnecessary costs. According to MS, millions of diagnostic tests are conducted every year at enormous expense, yet these have little impact on patient care. 'MAI-DxO' repeatedly checks for tests in the diagnostic process that may incur unnecessary costs. It is also significant that the system sought a balance between diagnostic accuracy and the costs involved. No matter how accurate a diagnosis may be, if the cost is excessively high?or conversely, if the cost is low but the accuracy is poor?patients will not seek medical care.

Of course, this research has not yet received clinical approval, and further safety testing and validation are needed in the future. Even if the system's level of completion improves significantly, there remain barriers such as differing healthcare systems and regulations in each country. As we witness the rapid development of medical AI, it is natural to wonder whether AI could one day replace doctors. However, MS draws a clear line on this issue: "AI can only complement doctors and medical professionals; building trust with patients and their families remains the responsibility of humans."