JMIR Publications

ABSTRACT

Background:

In health systems, AI is being increasingly implemented as a form of clinical decision support (CDS). Research on patient perspectives, however, lags behind clinician-focused studies. Before the introduction of AI-based tools, clinicians often used and continue to use rule-based guidelines and risk calculators to inform decision-making. Prior research has shown that risk calculators for CDS have mixed effects on patient adherence to treatment recommendations, while for AI-based tools the effect is still not clear.

Objective:

To assess differences in patients’ behavioral intentions, trust, and preferences for clinical recommendations provided by clinicians alone, clinicians supported by risk calculators, and clinicians supported by artificial intelligence-based risk predictions. We also evaluate how sociodemographic characteristics affect these perceptions.

Methods:

We surveyed 1,185 U.S.-based adults. Participants rated their likelihood of taking recommended clinical actions (behavioral intentions) across six clinical scenarios where they were randomized to a different recommendation source (clinician, clinician + risk calculator, clinician + AI). Participants also rated their trust in the recommendation, preferred source for recommendations, and provided qualitative feedback.

Results:

Behavioral intentions to take clinical recommendations did not differ by recommendation source (clinician, clinician + risk calculator, clinician + AI) but were affected by the clinical condition. Across clinical conditions, participants tended to trust clinician decision-making alone over those supported by risk calculators or AI. However, a plurality of participants ultimately preferred a clinician using a risk calculator for making recommendations. Several sociodemographic differences were detected across outcomes, but there were few clear patterns in perceptions. Qualitative findings revealed skepticism across all decision sources and potential misunderstanding on AI’s ability to personalize risk predictions.

Conclusions:

These findings offer novel, nuanced patient preferences shaped by task and context, offering guidance for aligning AI-supported care with patient values. Future work should support general patient-facing education regarding strengths and weaknesses of AI. There is also a need to develop standardized templates for AI-based risk predictions that may provide information to support shared decision-making.