Medical decision systems based on artificial intelligence are increasingly used in radiology, with over 350 systems approved by the US FDA. Despite their widespread adoption, the rapid pace of AI development has outpaced our understanding of their clinical value, creating what experts call an AI chasm. This gap arises from technical, logistical, and integration challenges in clinical workflows. Unlike aviation, where automation has been optimized, medical AI implementation lacks comprehensive research on its interaction with human users.
Radiologists rely heavily on cue utilization, using sensory signals from the clinical environment to guide diagnoses. This skill, known as ecological rationality, allows clinicians to make efficient decisions based on context. In contrast, AI systems often use all available information, including irrelevant cues, to optimize their decision-making processes. This debounding leads to decisions that might not align with clinical realities and can introduce vulnerabilities, such as difficulty in anticipating AI errors or recognizing biases.
A recent article published in the Lancet, argues for a cognitive approach to AI implementation in medicine, moving beyond observational studies to explore the mental mechanisms driving clinical decision-making.