Artificial intelligence is already widely used in health care — to enhance cancer detection, reduce paperwork and tailor treatments. The question is, does it actually improve patient outcomes?
Yes. Kaiser Permanente has demonstrated how AI help can improve the quality and efficiency of patient care and even save lives.
For more than a decade, Kaiser Permanente — the largest nonprofit integrated health-care delivery system in the country — has been developing an AI tool that detects signs of clinical deterioration in hospital patients to identify those at risk of getting much sicker. This is a fundamental skill that doctors hone throughout their medical training. An increase in heart rate and decrease in blood pressure could mean an infection is spreading. A repeat chest X-ray showing more fluid in the lungs could signify weakening heart function. A rising creatinine level could indicate worsening kidney failure.
The problem is that clinicians receive many of these inputs for all the patients they care for, and worrying trends can get missed until a patient becomes very ill. Hospitals have tried a variety of methods, including beeps and flashing alerts for each abnormal vital sign or lab test. But these can create “alarm fatigue,” which means many alerts end up being ignored. Also, when clinicians are continuously bombarded with new data, it can be difficult to discern the patterns that suggest there is a problem.
Another challenge is the workflow itself. Results from labs and imaging studies usually come in piecemeal. A physician might check for updates every few hours, but this might not be often enough to pick up on an acute worsening. And if nurses or other members of the care team find the emerging problem first, there can still be a delay in the physician arriving to the patient’s bedside.
Kaiser Permanente’s AI tool addresses several of these obstacles. Predictive algorithms have been built to account for a patient’s preexisting medical conditions, vital signs, laboratory tests, bedside nurse reports and other factors. And the tool receives hourly input from electronic medical records. If all this data reveals a significant risk of decline, an alert is issued.
The key difference in Kaiser’s use of AI is what happens next. First, the alert is reviewed by an off-site team of nurses who examine what triggered it. Then, if the patient needs to be evaluated in person, they have the patient assessed by the hospital’s rapid-response team, which then works with the patient’s physician to determine next steps.
From 2016 to 2019, this AI-powered alert system was rolled out to all 21 of Kaiser Permanente’s Northern California hospitals. Researchers then examined the outcomes of patients it flagged vs. those who would have triggered an alert if the system had been active at the time of their hospitalization. Their results, published in the New England Journal of Medicine, show there was a 16 percent lower mortality rate among patients who benefited from the AI tool. That’s equivalent to 520 deaths prevented per year.
“What AI does really well is to integrate a lot of different inputs,” Andrew Bindman, Kaiser Permanente’s chief medical officer, told me, “to identify patterns that are sometimes less visible at an early stage to our clinicians.” AI replicates clinicians’ pattern-recognition abilities but also pulls in information more frequently, enabling it to identify trends faster.
Like other health-care leaders who are enthusiastic about AI’s use in the field, Bindman stresses that it is not meant to replace human clinicians. Rather, it picks up data that humans might miss. The AI alert system still requires ongoing assessment by nurses and doctors. Because this tool deploys two additional teams, the remote nurses who conduct the initial review and the in-person rapid-response team, it is both high-tech and high-touch.
Kaiser Permanente’s results are not unique. A series of studies published last year in Nature Medicine analyzed patient outcomes at five hospitals, including Johns Hopkins Hospital in Baltimore, after an AI tool was deployed to identify patients who might have sepsis. This whole-body infection is a leading cause of in-hospital death, but it can be treated with antibiotics if it’s caught early enough. The AI tool resulted in lower mortality as well as less organ failure from sepsis and shorter hospitalizations.
The more often we see ChatGPT parodies or alarms about AI being used to create “deepfakes” and other bad things, the easier it becomes to either dismiss AI as a gimmick or fear it as an existential threat. But hospitals are demonstrating the good it can do to improve health and make care safer.