Sebastian Goodfellow of the University of Toronto and his team have partnered with researchers at SickKids to diagnose heart arrhythmias by using AI to analyze seismic data, typically used to detect seismic events such as earthquakes. This new project will use similar AI techniques on data that is generated from electrocardiograms (ECGs) to deploy a model that detects and diagnoses common pediatric hearth arrhythmias. Goodfellow and his team from UofT are partnering with Dr. Mjaye Mazwi of SickKids and the multidisciplinary research group at SickKids known as Laussen Labs.
Goodfellow joined Laussen Labs in 2017 while working on a PhD in applied seismology when the lab had just begun to acquire “physiologic waveform data” like ECGs. He suggested that the gap between AI in mineral engineering and AI in health care has become smaller.
When it comes to detecting and diagnosing hearth arrhythmias, there are “only two staff physicians on duty at any given time too service 42 ICU beds”. Resultantly, these arrhythmias can often go undiagnosed for a period of time. The idea behind Goodfellow’s research is to use the expert clinicians to train an AI machine to monitor all ICU beds at all hours of the day.
Goodfellow suggests that the main challenge faced with developing such a model is the “translation gap” that can result from “difficulties creating computational infrastructure that can reliably ingest data for real-time classification,” the need for a high quality MLO platform for monitoring and training AI models, regulatory challenges with integrating AI into clinical domains, and concerns regarding biases. In order to “close this gap,” the development team must consist of a wide range of expertise from those in software development, MLOps, cloud management, law, cognitive psychology, and more.
As is the case with the introduction of AI in various industries, there is apprehension. Goodfellow illustrates that thinking about an AI model as “a product from the very start” which involves doctors as end users will help to overcome fears of distrust with the ML tools. It is also crucial to consistently and clearly present the performance of the model to the end user in a transparent fashion.
Read more about the research being undertaken at the University of Toronto.
