AI helps ICU clinicians ‘find the signal in the background’

Key takeaways:

• Algorithms can help ICU clinicians save time in making care decisions.
• AI can also predict deterioration and cardiac arrest, which allows clinicians to intervene ahead of time.

AI saves time for pediatric cardiac ICU clinicians by identifying patients who need more attention, according to a presenter at the AIMed24 Annual Meeting.

“For any of you clinicians that were in an ICU at any point, at some point you were asked to go to the bedside of an unwell patient with this much data staring you in the face with these texts that are ever more incessant in our world and were asked to make a clinical decision that mattered,” Michael P. Goldsmith, MD, director of cardiac informatics at The Children’s Hospital of Philadelphia (CHOP) and University of Pennsylvania Perelman School of Medicine, said during his presentation while the sounds of beeping and chatter played to simulate the average ICU environment.

“This is an ever more challenging task in a care environment where messages are more intrusive, there [are] more data, and ultimately... a lot more noise to try to find the signal in the background.”

During his talk, Goldsmith described four algorithms that have helped ICU physicians improve their decision-making.

The first was a clinical decision support algorithm that functioned as a cardiac arrest prevention bundle with a list of steps for physicians to take, including a safety huddle, discussing acceptable vital signs at patients’ bedsides, having medications at the bedside and offering presedation for aggravating procedures. To test the algorithm, 15 cardiac ICUs were assigned the bundle, and 16 continued with their standard practice, Goldsmith said. In all the hospitals with the cardiac arrest prevention bundle, the relative risk for cardiac arrest was 30% lower than control hospitals. Goldsmith said it led to 11 fewer cardiac arrests each month at all of the intervention hospitals.

The second algorithm developed at CHOP flagged certain patients in Epic who were at a higher risk for cardiac arrest and instructed physicians to conduct a bedside huddle. At Goldsmith’s hospital, he said huddle compliance started strong but fell to 10% after a few years. He and his colleagues spoke to the clinicians and asked how to improve compliance.

“We came up with a number of impactful changes that would get the right information to the right person at the right time,” he said.

They added new flags that were more visible, and a timer for when patients at risk for cardiac arrest would need their next huddle. Adding these features increased compliance from 10% to almost 60%, and hospital staff performed more rescue events, Goldsmith said.

Another algorithm that Goldsmith and colleagues developed at CHOP was a machine learning model that identified infants with single-ventricle and shunt-dependent congenital heart disease who were at risk for deterioration. According to the study, the algorithm could predict deterioration up to 8 hours ahead of time, which allowed clinicians to intervene. In addition to the algorithm, Goldsmith and colleagues designed a user interface that explained what factors contributed to the patients’ elevated risk.

The last algorithm he described was developed to help physicians determine whether patients could be weaned off vasoactive infusions after cardiac surgery. When patients met certain criteria, the algorithm would display a flag suggesting the patient be weaned off of sedation. Any clinician, even junior clinicians, could probably make that decision after standing at the patient’s bedside for 5 minutes, Goldsmith said.

“But in a big, busy clinical ICU setting, getting that dedicated time to do that is challenging,” he said.

From before implementing the algorithm in his hospital to after implementation, Goldsmith said the length of vasoactive infusions decreased by 25%.

AI is making real impacts in ICUs, Goldsmith said, but there are still a lot of questions about how to implement these tools in the clinical setting.

“I think that is something we are going to tackle in the next 5 to 10 years — making them usable, interpretable and ultimately making that something you can do about it, and not necessarily a big thing, but a gentle thing that encourages clinicians to take the best care of their patients.”

References:

• Goldsmith MP. Considerations for successful implementation of inpatient AI-based clinical risk algorithms. Presented at: AIMed24 Annual Meeting; Nov. 17-19, 2024; Orlando.
• Goldsmith MP, et al. Crit Care Explor. 2021;doi:10.1097/CCE.0000000000000563.
• Ruiz VM, et al. J Thorac Cardiovasc Surg. 2022;doi:10.1016/j.jtcvs.2021.10.060.