Artificial intelligence tools built into the EHR may help health systems implement strategies such as targeted overbooking; however, the technology presents various layers of potential bias toward vulnerable patient populations, according to a Jan. 31 Health Affairs report.
For the study, a team of UC San Francisco researchers analyzed built-in AI models for the EHR that help clinicians predict the likelihood that a patient will no-show for their appointment. The strategy, called targeted overbooking, allows hospitals to schedule an additional patient in the same time slot in the event of a no-show. Targeted overbooking has the potential to harm patients if both the originally scheduled patient and overbooked patient arrive for their appointments, leaving the provider to see both patients during a time slot meant for one individual, according to the researchers.
The research team examined Epic's built-in AI tool, which displays a numerical estimate of the likelihood a patient will no-show based on the patient's personal information, clinical history, patterns of healthcare use and features of the appointment, such as time and day of the week.
Before implementing the model, the researchers identified several layers of potential bias with the technology that may harm vulnerable patient populations. The model included personal characteristics, such as ethnicity, financial class and body mass index, which "if used for overbooking could result in healthcare resources being systematically diverted from individuals who are already marginalized," according to the report. The model also presented issues with socioeconomic status.
Researchers addressed these potential bias issues by rebuilding the Epic model to exclude all personal information; however, the team realized that the new model would still not eliminate issues with societal inequity. For example, prior no-shows could be the result of a patient's inability to take off work or a patient with obesity who struggles with mobility may make it to their appointment on time to find it overbooked.
Researchers chose to implement a revised AI model that used only patient-positive interventions in response to the results of the no-show predictor. The model was launched in 12 clinics, which saw a 9 percent mean reduction in no-show rates compared to a year prior to implementing the scheduling tool.
Researchers concluded that multidisciplinary oversight is needed in the process of developing clinical decision support tools to ensure AI does not automate discrimination against vulnerable patients. They recommend healthcare institutions and software vendors partner when evaluating new AI tools.