Microsoft Says Its New AI Is Better at Diagnosing Patients Than Real Doctors
"This is just going to change the way care is delivered in a big way." The post Microsoft Says Its New AI Is Better at Diagnosing Patients Than Real Doctors appeared first on Futurism .
A recent Microsoft study claims its AI model, the Microsoft AI Diagnostic Orchestrator (MAI-DxO), outperforms human doctors in diagnosing patients. Published in the respected journal Science, the research revealed that MAI-DxO was more than twice as accurate as generalist physicians in diagnosing patients. Unlike prior studies, this one required the AI to diagnose patients in a manner akin to a real doctor, rather than with all the facts presented upfront.
The tests took place in a controlled benchmark setting, which makes its effectiveness in a genuine medical setting uncertain. Microsoft CEO of its AI lab, Mustafa Suleyman, expressed optimism, stating that the era of AI improving lives through healthcare "is finally here" and predicting substantial deployments of intelligent systems in the near future.
The study analyzed 304 clinicopathological conference cases and 325 emergency department-to-admission encounters, representing a nearly even split between case reports and emergency department admissions. MAI-DxO achieved an 83.2 percent diagnostic accuracy with the former, compared to 36.1 percent for human physicians who had access to reference materials.
With emergency department cases, MAI-DxO scored 80.3 percent, while human doctors scored 46.8 percent. The AI model not only excelled in accuracy but also managed to order tests at half to one-third the cost of human doctors. If validated, these findings highlight AI's potential in medical applications. However, the role of a doctor extends beyond symptom analysis, including engaging with patients to understand their experiences fully.
The study attempted to address this by having MAI-DxO engage in a sequential process, working through a patient's initial complaint, formulating a diagnosis, and determining follow-up questions and tests while considering the associated costs. Previous studies had not tested diagnosis as an active, sequential information-gathering process, instead providing the complete case in a single prompt.
Critics argue that the study only imperfectly simulated a doctor-patient interaction by having MAI-DxO interact with a "gatekeeper agent" that mediated all information requests using the underlying case record. While AI models excel in analyzing and synthesizing information, human doctors have already gathered the detailed case files for the AI to analyze.
Written by urgent.news from Futurism's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
