How “Doctor AI” Increases Hospital Survivability

A team led by Clemens Heitzinger of the “Center for Artificial Intelligence and Machine Learning” (CAIML) at the Vienna University of Technology used “extensive data from intensive care units in several hospitals” to teach a computer program how people with blood poisoning (sepsis) can be treated most successfully in intensive care, says a broadcast from the university. “In our project, we used a form of machine learning known as reinforcement learning,” says Heitzinger.

Computer works with reward system
The computer makes treatment decisions that are applied to virtual patients in the hospital bed. If they do well, the computer is rewarded. If the condition worsens or death occurs, he will be punished. “The computer program has the task of maximizing its virtual reward in every possible way,” says the expert: “In this way, a strategy can be automatically determined from extensive hospital data with a particularly high chance of success.”

Computer can consider more parameters
The medical staff make their decisions based on well-founded rules based on certain parameters to ensure the best patient care, the researchers wrote in the journals “Plos One” and “Journal of Clinical Medicine”: “But the computer can easily do many other things do Take into account parameters that a human might ignore, and that can lead to even better decisions in some cases.”

“In simulations, we treated virtual patients with the doctor’s strategy and with the calculated strategy and compared the results,” explains Heitzinger: without the use of artificial intelligence, 85 out of 100 patients survived 90 days, with the AI ​​strategy the there were three more, namely 88.

“Of course that doesn’t mean that you should leave the medical decisions in an intensive care unit to a computer alone,” he says: “But you can run artificial intelligence as an additional device at the bedside.” The medical staff can take advice from it and compare their own assessment with the AI’s suggestions. It would also be extremely helpful in training, he says.

Liability issues need to be clarified
Before “Doctor AI” is allowed to go to the clinic, however, legal questions need to be cleared up: first, who is liable for any mistakes made by the AI. “The problem also occurs the other way around: what if the artificial intelligence had made the right decision, but the human had chosen a different option and the patient had suffered damage as a result?” Heitzinger says. Clear rules are urgently needed here, because technically artificial intelligence could already be used in clinical practice for the benefit of patients.