A Mixed Methods Comparison of Artificial Intelligence-Powered Clinical Decision Support System Interfaces for Multiple Criteria Decision Making in Antidepressant Selection

BACKGROUNDArtificial intelligence-powered clinical decision support systems (AI-CDSS) have recently become foci of research. When clinicians face decisions about treatment selection, they must contemplate multiple criteria simultaneously. The relative importance of these criteria often depends on the clinical scenario, as well as clinician and patient preferences. It remains unclear how AI-CDSS can optimally assist clinicians in making these complex decisions. In this work we explore clinician reactions to different presentations of AI results in the context of multiple criteria decision-making during treatment selection for major depressive disorder. METHODSWe developed an online platform for depression treatment selection to test three interfaces. In the probabilities alone (PA) interface, we presented probabilities of remission and three common side effects for five antidepressants. In the clinician-determined weights (CDW) interface, participants assigned weights to each of the outcomes and obtained a score for each treatment. In the expert-derived weights interface (EDW), outcomes were weighted based on expert opinion. Each participant completed three clinical scenarios, and each scenario was randomly paired with one interface. We collected participants impressions of the interfaces via questionnaires and written and verbal feedback. RESULTSTwenty-two physicians completed the study. Participants felt that the CDW interface was most clinically useful (H=10.29, p<0.01) and more frequently reported that it had an impact on their decision making (PA: in 55.5% of experienced scenarios, CDW: in 59.1%, EDW: in 36.6%). Clinicians most often chose a treatment different from their original choice after reading the clinical scenario in the CDW interface (PA: 26.3%, CDW: 33.3%, EDW: 15.8%). CONCLUSIONClinicians found a decision support interface where they could set the weights for different potential outcomes most useful for multi-criteria decision making. Allowing clinicians to weigh outcomes based on their expertise and the clinical scenario may be a key feature of a future clinically useful multi-criteria AI-CDSS.