Description‘Automation Bias in Clinical Decision Support Systems – lessons to be learned for Radiology’
There is limited information on how clinicians will interact with Artificial intelligence (AI) in the radiology department. AI is already present in many forms, with new applications being developed at an astounding pace. Studies have shown the benefit to the clinician and patient through dose reduction, automated image quality improvement and increased accuracy through decision support. The value of these systems to support the UK health service is being recognised explicitly in the NHS Long Term Plan (2019). Radiology has traditionally been regarded as a technologically advanced field, staffed with clinicians who are comfortable with both patient and machine, who are able to leverage the advantage of computers to benefit patients. The first proof of concept of the use of computers to assist in diagnosis was in the 1960s where a programme which converted radiographic images into numerical data to assist in diagnosis of pathology was proposed. Increased computing power has permitted these initial forays into the potential of machines for the diagnosis of disease from radiographic images to become a reality. Recent developments of complex systems, using deep learning technologies, whose functionality is not easily understood, is proving to be the most effective. Some clinicians are reluctant to adopt these systems in their clinical practice due to issues such as the lack of explainability of these systems – when the machine goes wrong, it is not always clear why. Other issues with the integration of AI systems in radiographic image interpretation should be considered before and following implementation. An understanding of how the human will interact with the system should be considered and whether this interaction results in a positive or negative outcome. Automation bias (AB) refers to the propensity of the human to over-rely on the output of the machine, for instance a reliance on spell-check over our own spelling abilities, even when the checker is incorrect. The user can leverage the strengths of the computer assistance if they are also aware of potential shortfalls of the system. Studies have found there are several predictors to the likelihood of AB. Experienced clinicians are more likely to be fixed on their initial opinion and less experienced clinicians are more likely to change their mind to concur with that of the computer. Education of all clinicians in the potentials and pitfalls of AI as used in healthcare may be able to mitigate against some of these issues and ensure optimal, responsible use of these exciting technologies.
|Period||15 Nov 2022|
|Event title||Medical Imaging Convention|
|Degree of Recognition||International|
- Automation bias
- human computer interaction
- clinical error