Neurologists mostly have positive attitudes towards machine learning-based EEG seizure detection tools and are willing to be trained on their use, however obstacles including lack of availability stand in their way, a survey suggests.
Researchers surveyed 31 medical professionals involved in the management of epilepsy to better understand the perceived obstacles preventing them from using automated seizure detection tools in clinical practice to help inform future tools.
The team from the Applied Artificial Intelligence Institute, Deakin University, Melbourne, aimed to gather insights on various factors including the clinical utility of the tools, professional sentiment and doctors’ benchmark requirements.
Respondents were mostly female (68%) and from Australia or Singapore with more than half stating that they did not use any automated detection tool.
The majority of the respondents (63%) reported they would use seizure detection tools for all types of seizures if the tools were available to them. However, 10% of them said they would not use any seizure detection tools if they were available.
Over 80% of respondents said they would use seizure detection tools to alarm nurses and physicians of a potential seizure, annotate potential seizure segments for review and record vital signs and information during potential seizures.
Only 37% would use the tool to administer treatments on patients with epilepsy.
Regarding sentiment, most clinicians were receptive with 77% agreeing that they would use automated detection tools that were available in their clinical workflow and the same proportion were willing to spend time to be trained on their use.
Of those who disagreed, most had more than eight years of clinical experience and worked in paediatric units. Some of these respondents raised concerns about the lack of evidence around clinical value, with one clinician stating that the “hype in technology” would result in overdiagnosis and excess escalation on EEG findings.
When it came to benchmark requirements, most respondents said they would only accept an automated seizure detection tool if it could detect a minimum of 80% of seizures. They also said they would tolerate less than two false alarms per hour.
The top three barriers in clinical practice for neurologists related to availability followed by training and the blackbox nature of machine learning algorithms which prevented them from receiving information on the methods the tools used to detect a seizure.
Based on the findings, the researchers created a guideline on developing automated seizure detection tools that would meet the requirements of medical professionals and increase adoption of machine learning in clinical practice.
“We highlight the need for seizure detection algorithms that are not only transparent but also interpretable or explainable for the medical communities. However, this is a challenge, particularly for seizure detection and prediction algorithms based on deep learning,” the authors wrote in Epilepsy and Behavior [link here].
The authors declared no conflicts of interest.