Teams of medical AI agents keep bias in triage

Researchers did a test of teams of AI agents on 120 emergency cases from the MIMIC-IV data set. In each one, they changed the demographic data of the patient. They used 5 model families and 4 team methods. In all 20 setups, the team gave Black patients more acute triage, and team work did not remove the bias.

How it works

Each agent gives an answer. The team then mixes the answers by one of 4 methods. The authors say that the bias goes into a decision that looks correct.

The numbers

Fairness instructions did not change the difference. Removed demographic data did not change it. When each agent removed its bias, the difference increased. A team with 5 different views cut the difference by 67 to 94% when all 5 views mix in each step.

What is not known

The report gives no test in a hospital. One method does not mix the views, and the team cannot repair that one with a sure result. Each model has a method that increases its difference, but the method is different for each model.

Insight: More agents did not bring more safety

In a report of 21 October, a team of agents had a lower score than one agent in 12 of 13 tests. This medical report finds that teams keep the bias of their agents. Teams are sold as a method to remove errors, but 2 reports find no such increase, and a team must show its value. Get that proof before you buy.