AI Clinical Decision Support for Physicians

CDS falls into two broad families. Knowledge-based CDS runs explicit rules written by clinicians and pharmacists, such as drug-drug interaction checks, allergy checks, dose ranges, preventive care reminders and order sets. Predictive CDS uses statistical or machine learning models trained on past patient data to estimate risk of events such as sepsis, deterioration, readmission or no-shows. Delivery can be interruptive (a pop-up that stops the workflow), passive (a flag, a banner or a default choice in an order), or routed to someone else, such as a pharmacist or a rapid response team. Informaticists often use the "five rights" framework: the right information, to the right person, in the right format, through the right channel, at the right time. Alerts that break these rules produce alert fatigue. Many studies of drug interaction alerts have found that clinicians override the large majority of them, often well above 90 percent in some settings, largely because so many alerts are clinically irrelevant. Once clinicians learn that most alerts are noise, they also dismiss the rare one that matters. Predictive models add another risk: they can perform worse in practice than their developers reported. In 2021 a University of Michigan team published an external validation of the Epic Sepsis Model in JAMA Internal Medicine. It found much weaker discrimination than advertised (an AUC of about 0.63). The model missed about two thirds of sepsis cases while generating alerts on many patients who did not have sepsis. On regulation, the 21st Century Cures Act of 2016 excluded some CDS from FDA device regulation, mainly tools whose basis a clinician can independently review, and the FDA issued final CDS guidance in 2022. ONC's HTI-1 rule, finalized in late 2023, requires certified EHRs to expose information about decision support interventions, including predictive ones. The common misconception is that more alerts mean safer care. Past a point, more alerts make care less safe.