AI-assisted anti-seizure medication selection? A qualitative study of the views of neurologists and epilepsy patients

Abstract

OBJECTIVE: To understand the factors that influence acceptance of a machine learning-based clinician decision support (MLCDS) tool to assist selection of a first antiseizure medication (ASM) in patients newly diagnosed with epilepsy. METHOD: A qualitative descriptive approach using interview and focus group methods was conducted in Australia. Using purposive sampling we conducted nine semi-structured interviews with people with epilepsy (PWE) as well as nine neurologists, and a focus group with five people with recently diagnosed epilepsy (within one year of diagnosis). A thematic analysis using a framework approach was performed. RESULTS: Results revealed three main themes: 1) Patient-clinician interaction is essential to MLCDS acceptance in the process of ASM selection; 2) Opportunities for MLCDS involvement in clinical decision-making; and 3) Apprehensions about the use of MLCDS in ASM selection. PWE and neurologists emphasised that the selection of ASM was complex and multifactorial and must be made with clinician oversight, input and authority. PWE highlighted the importance of trust and transparency in the patient-clinician relationship. CONCLUSION: PWE and neurologists were supportive of the prospects for MLCDS systems to improve ASM selection for people with epilepsy. However, their support was not unqualified and was often predicated on claims about the nature and role of these systems that are highly contested in the larger literature on the use of machine learning in medicine. In particular, the idea that systems that perform better than clinicians will remain sources of "advice", that machine learning will free up clinicians' time for longer conversations with patients, and that medical artificial intelligence will be "explainable", are all controversial. Our results suggest that much work remains to be done to discover how best to introduce MLCDS into clinical settings without jeopardizing stakeholders' support for these systems.

Ähnliche Arbeiten