Assessing Clinicians' Reliance on Computational Aids for Acute Stroke Diagnosis.

The rapid rise of computational aids for stroke diagnosis have led to important concerns about clinicians developing an over-dependence on technology. Other studies have assessed reliance on clinical decision support systems in fields like diabetes, but no such study exists for stroke diagnosis. In this work, we developed a high-fidelity user interface for a computational aid designed to support acute ischemic stroke diagnosis. Engaging with stroke practitioners at the UCSD Stroke Center, we conducted an experiment to determine how technology for identifying stroke symptoms may affect their diagnostic decision-making processes. By assessing how clinicians changed their video-based diagnosis of stroke when provided with data visualizations and predictions from a machine learning tool, we observed that such computational aids do in fact affect clinicians' decisions but only in cases when the aid directly supports or contradicts their prior beliefs. Future computational aids for stroke diagnosis should focus on helping clinicians solidify their decisions rather than only providing them with overly quantitative information that may impede or confuse their judgement.