Artificial Intelligence Models Improve Clinicians' Diagnostic Accuracy

TUESDAY, Dec. 19, 2023 (HealthDay News) -- Standard artificial intelligence (AI) models improve diagnostic accuracy, but systematically biased AI models reduce this accuracy, according to a study published in the Dec. 19 issue of the Journal of the American Medical Association.

Sarah Jabbour, from the University of Michigan in Ann Arbor, and colleagues examined the impact of systematically biased AI on clinician diagnostic accuracy in a randomized clinical vignette survey study. Clinicians were shown nine clinical vignettes of patients hospitalized with acute respiratory failure and were asked to determine the likelihood of pneumonia, heart failure, or chronic obstructive pulmonary disease as the underlying cause. Clinicians were shown two vignettes without AI model input to establish baseline diagnostic accuracy and were then randomly assigned to see six vignettes with AI model input: three standard-model predictions and three systematically biased model predictions.

Overall, 457 clinicians were randomly assigned: 231 and 226 to AI model predictions without and with explanations, respectively. The researchers found that for the three diagnoses, clinicians' baseline diagnostic accuracy was 73.0 percent. Clinician accuracy increased over baseline by 2.9 and 4.4 percentage points when shown a standard AI model without and with explanations. Clinician accuracy was reduced by 11.3 percentage points with systematically biased AI model predictions compared with baseline; providing biased AI model predictions with explanations reduced accuracy by 9.1 percentage points, representing a nonsignificant improvement of 2.3 percentage points compared with the systematically biased model.

"Although the findings of the study suggest that clinicians may not be able to serve as a backstop against flawed AI, they can play an essential role in understanding AI's limitations," the authors write.

One author reported receiving royalties from a patent from Airstrip.

Abstract/Full Text (subscription or payment may be required)

Editorial (subscription or payment may be required)