JMIR Publications

ABSTRACT

Background:

With the potential and rapid development of artificial intelligence and related technologies, AI algorithms are being embedded into various health information technologies to assist clinicians’ decision making in clinician-patient encounters.

Objective:

The objective of this study is to explore how clinicians perceive AI assistance in their diagnosis decision making and suggest paths forward as to what necessitates to achieve AI-human teaming in healthcare decision making.

Methods:

This study uses a mixed methods approach utilizing hierarchical linear modeling (HLM) and sentiment analysis through natural language understanding (NLU) techniques.

Results:

A total of 114 clinicians who practice in family medicine and interact with AI algorithm to make patient diagnosis participated in online simulation surveys during 2020- 2021. Our qualitative results show a promise that clinicians’ overall sentiment toward AI-assisted patient diagnosis was positive and comparable to those of live patient encounters. However, it also showed that the process of diagnosis decision making by the given AI physiology algorithms did not align with the way clinicians make diagnosis decision. In the follow-up quantitative survey, clinicians perceive that current AI assistance was not likely to enhance their diagnostic capability and rather negatively affect their overall task performance (β=-0.421, p=0.016). Interestingly, clinician’s level of clinical diagnosis capability is rather associated with clinicians’ ex ante quality such as education (β=1.880, p=0.072) and age (β=2.428, p=0.071) on diagnostic capability as well as existing technology habit on both dependent variables (β=0.232, p=0.009 and β=0.244, p=0.003, respectively).

Conclusions:

This paper sheds light on clinicians’ current perception and sentiment toward AI-enabled diagnosis technology in healthcare decision makings. We showed here that while overall sentiment toward the AI assistance was positive, current form of AI assistance is not linked to efficient decision-making in that AI algorithms are not aligned with humans’ subjective clinical reasoning. We suggest that health policy makers and HIT developers need to gather behavioral data from clinicians in various disciplines and specialties to make clinical AI algorithms to be aligned with humans’ subjective and unique clinical reasoning patterns.