0
0
March 23, 2024
Among the myriad applications of artificial intelligence (AI) in healthcare, one particularly promising area is its potential to alleviate the burden on healthcare providers by assisting in responding to patient inquiries submitted through online portals. This notion has been reinforced by a recent study conducted by researchers at Stanford Health Care, led by Dr. Patricia Garcia, published online in JAMA Network Open.
During the COVID-19 pandemic, the volume of patient messages surged by a staggering 157% from pre-pandemic levels, significantly contributing to provider workload. Each additional message translated to an extra 2.3 minutes spent on the electronic health record (EHR) per day, underscoring the urgency for solutions to mitigate this strain.
The Stanford study, conducted over a 5-week period from July 10 through August 13, 2023, aimed to evaluate the efficacy of an AI response system in addressing this challenge. The system integrated a Health Insurance Portability and Accountability Act (HIPAA) compliant large language model into the EHR, generating draft responses for providers to review and potentially utilize.
Involving 162 participating clinicians from divisions of primary care and gastroenterology and hepatology, the pilot program assessed the extent to which providers embraced AI-generated drafts and whether such a system alleviated their workload or enhanced their overall experience.
Results indicated an overall average utilization rate of 20% per clinician, with notable variations between groups. Nurses in gastroenterology and hepatology exhibited the highest utilization at 29%, followed by physicians and advanced practice providers (APPs) at 24%. In contrast, clinical pharmacists in primary care demonstrated the highest utilization rate at 44%, compared to 15% among physicians.
Despite not yielding time savings, the AI system demonstrated improvements in task load and work exhaustion scores. Notably, there was a significant reduction in physician task load scores (pre-survey mean [SD], 61.31 [17.23]; post-survey mean [SD], 47.26 [17.11]; paired difference, −13.87; 95% CI, −17.38 to −9.50; P < .001), accompanied by a decrease in work exhaustion scores by a third (pre-survey mean [SD], 1.95 [0.79]; post-survey mean [SD], 1.62 [0.68]; paired difference, −0.33; 95% CI, −0.50 to −0.17; P < .001).
While feedback on the AI response messages’ voice and tone was mixed, comments regarding message length predominated the negative spectrum. However, concerns regarding accuracy were relatively balanced, with no significant safety issues reported.
Dr. Garcia and her colleagues underscored the potential of AI-generated responses in mitigating cognitive burden and burnout among providers. Despite not achieving time efficiencies, the authors suggest that editing AI-generated responses may be less taxing than crafting responses from scratch.
The study’s findings carry implications for addressing burnout, which is associated with adverse outcomes such as turnover and diminished quality of care. Even modest improvements, the authors contend, could yield substantial benefits in clinician well-being and patient care.
Disclosure statements revealed that one coauthor received grants from Google, Omada Health, and PredictaMed, while another coauthor disclosed a patent and honoraria from Mayo Clinic. No other conflicts of interest were reported.
In conclusion, the study provides valuable insights into the potential of AI in alleviating healthcare provider burden and underscores the importance of continued research in this evolving field.
