‘Global first’: the ‘discharge software’ indicates whether a patient can be removed from the IC

Liesbeth Vorstemans-Van Empelen (82) looks up at her doctor through her glasses. She is sitting on a chair next to her bed. There is an empty plate on a table, she has just had breakfast. Behind her is the respirator that she used to breathe through until recently. She was wheeled into intensive care four days ago. She doesn’t remember. She was in a coma.

“You look a lot better,” says Patrick Thoral, doctor in the intensive care unit of the Amsterdam UMC. “The chance is not zero that it will go wrong again. Let’s hope you don’t have to stay in the hospital much longer. Are you confident that you will go to the regular ward?”

“Yeah,” she says.

“If you feel something you don’t know, don’t be afraid to call the doctor again.”

The conversation is a seemingly normal hospital scene, but there is something special going on here. In deciding that it is safe to send Vorstemans-Van Empelen to another nursing ward, Thoral received advice from artificial intelligence, or AI for short. The Amsterdam UMC will announce on Thursday that it is using a new AI program. Nickname: ‘dismissal software’. It gives IC patients a score: if doctors send someone away, what are the chances that he or she will come back (ie become seriously ill again)?

The system looks at hundreds of parameters, such as the development of the heart rate and the oxygen level. Another important factor are notes from nurses in the electronic file, for example about how aware a person is of the environment. The patient is compared with about 25,000 people who were previously in the ICU. The software also provides a brief explanation of how the advice is created. According to the doctors, who developed the system with tech company Pacmed, it is a “world first”.

Day in IC costs thousands of euros

Doctors make decisions early every morning in a conference room. Which patient is going to leave the intensive care unit (be ‘discharged’)? Letting patients go too early is risky: if the patient does have to go back to the ICU, the risk of death is greater. But people can’t stay too long either. A treatment day in the ICU costs between two and three thousand euros.

Morning consultation about the ICU patients.

Information about the patient is projected on a large screen. Before the advice of the computer program is shown, the doctors themselves give a percentage. For example, what is the chance that Mrs. Vorstemans-Van Empelen will return to the ICU if they send her away now?

“15 percent,” Thoral thinks. “That is higher than you would prefer, but not extremely high either.”

A colleague clicks the mouse a few times. Then the computer advice appears. 15.5 percent. “Almost right!” someone shouts.

“We’re not going to get much better on her score,” says Thoral, “because of her age and medical history.” She can go to the lung ward.

The next patient is in his fifties, recovering from a brain haemorrhage and having a dangerous clot. Can he leave?

“No,” say several people.

“I think the chance that he will come back is 30, 35, 40 percent,” says Thoral. “Mostly because he has more problems than any other department can have. The burden of care is too high.”

The percentage of the software appears on the screen: 7.5 percent.

“There is a huge discrepancy between your prediction and that of the software,” says Paul Elbers, also an intensivist (specialized IC doctor). “Could the computer be right, and not you?”

“In this case, I doubt that,” says Thoral. “We use a medication here that is not intended for departments where there is no cardiac monitoring. We have to realize that the model is not trained for this kind of patient. Of course I know how it was developed.”

Data from five hospitals

That is the limitation of the system, Thoral says later in his room. “Not all patients are caught in it. Certainly rare patients from which the system has been able to learn less. If even more hospitals in the Netherlands provide data, we could improve the software.” Now the program works with data from five hospitals.

Thoral and Elbers came up with the idea in 2017 to do more with the thousands of data collected from patients in intensive care. Could they use it to develop an AI program that they would advise? “We kept running into the fact that we didn’t have enough beds,” says Thoral. “Since corona, the whole of the Netherlands has had this problem, but that was already an issue before that.”

The doctors collaborated with the company Pacmed, which developed several AI programs for the healthcare sector. The approval process to use it (inside and outside the hospital) took years. “Many people were afraid of cold feet,” says Thoral, “When you say AI, they think of self-driving cars taking over everything.”

Isn’t it a risk that doctors will rely too much on the software, that they no longer think for themselves enough?

“It is very emphatically still true,” says Thoral, “that the doctor bears the ultimate responsibility.”

That is why it has been agreed that doctors first state a percentage themselves before they see the computer advice. “Then the software only has the option to make adjustments,” says Thoral.

Of software estimate the probability of returning to the ICU.

A decision to fire someone from the ICU, says the doctor, has always been a matter of consultation. “I listen to advice from fellow doctors, from nurses, and now also from this software system. It’s an extra check to see if I’m overlooking something: we doctors don’t all have the numbers in the ICU in our heads.”

Also in the future, the software program will not make the decisions itself, Thoral expects. “Even in medicine, there are still things that we don’t put in the computer. I also look at the color of the face, how someone looks. We don’t record subjective things enough to make the model completely independent.”

The decision when someone can leave the ICU must be made at the right time. That is, especially not too early. Two days earlier, some doctors had the idea that Vorstemans-Van Empelen could safely go to the lung ward. Both Thoral and the computer program gave negative advice. “She was still breathing a little fast and her blood carbonation was higher than you’d hope for at this stage,” Thoral said. “Then I decided: no, it doesn’t seem wise to me. A few hours later it went wrong and we reconnected her to the ventilator.”

Now Vorstemans-Van Empelen is happy that she can leave. “In the regular ward you belong to the people again.”

“Here you are a bit isolated,” says Thoral, “between all the big machines.”

“Yes, this is real isolation. But thank God you can also watch television here.”

Doctors have decided not to tell patients about the discharge software. “I personally think most patients would understand,” says Thoral, “but a lot of our patients won’t be able to get a good grasp of it. Many are confused or not even conscious when they leave the ward.”

Vorstemans-Van Empelen does get to hear it, when Thoral asks her if the reporter is welcome in her hospital room. How does she feel that a computer program has advised her on whether she can go to the other department?

“I think it’s very good. Why not?”

She worked for years, since 1958, in the hospital herself, as a nurse. Then there was no discharge software, no heart-lung machines, hardly any technology at all in the hospital.

“So much can be done with it, why not?”