Project 462863

Critically ill children suffer preventable injury and death as a result of abnormal heart rhythms. Detecting these abnormalities is challenging for doctors and nurses. Machine learning (ML) models have been shown to be highly accurate at detecting these abnormalities, but there has been no successful translation of ML models to the bedside where they could improve patient monitoring and prevent deterioration and death. This translation gap is due to a combination of technical and regulatory challenges. We intend to address both the regulatory and translation gaps in this proposal in a manner that will be broadly instructive for the community working on applied clinical ML by building upon extensive preliminary work to refine and deploy a machine learning model capable of highly accurate classification of heart rhythm in critically ill children. To accomplish these goals we will: 1) Make publicly available the first large paediatric heart rhythm dataset to spur collaboration, labeling and human benchmarking in this underserved area. 2) Develop a sharable, reproducible approach to development of an expert ML model capable of highly accurate classification of heart rhythm 3) Deploy this expert ML model using a technical strategy that permits us to both validate the model and monitor and continuously improve its performance at the point of care 4) Ensure in a silent (non-interventional) trial that the ML model performs well and equitably across patient groups as a way of ensuring that all patient groups benefit from these innovation. At the end of this work, models developed and validated using this approach will be ready for prospective interventional trials to see where and how they can augment human clinicians. As importantly, as part of a commitment to open science, our dataset, modelling approach, deployment approach and computer code will all be publicly available for others to reproduce our results and create other models that utilize waveform data.