Please note: this event has passed
Title: Uncertainty Awareness and Trust in Cardiac DL Models
Abstract: In recent years, DL models have dominated medical research, but they are often developed without consideration of how the models will be used in clinical practice. In most cases where a DL model could potentially be used for complex predictive problems, it is unlikely that the model will be used as a stand-alone “black box” tool to replace clinicians. Rather, it will act as a decision-support tool to aid clinical decision-making. This consideration raises the important issue of clinical trust in the model. Therefore, research into more principled and reliable methods to implement DL models has gathered interest and is an ongoing topic of discussion for the adoption of DL-based healthcare applications. One way to develop clinical trust, and develop reliable predictive outcomes, is to quantify the confidence of a model’s automated decision using uncertainty estimates. Uncertainty estimation methods may provide the ability to understand the uncertainty of predictions but to build models that are trustworthy and uncertainty estimates should be incorporated as knowledge when training a DL model. This technique has been termed uncertainty-aware training and differs from uncertainty estimation of a prediction, with the latter having no feedback into the training of the model.
Therefore, considering the recent research and attention in the field it indicates the importance of uncertainty estimation in DL applications, more so in a highrisk area like healthcare. It is envisaged then that these estimations will become a necessary component to create reliable, trustworthy DL prediction models, that can learn to consider consequences of incorrect predictions.
Thus, I will present my work focusing on clinical trust and enhancing the reliability of cardiac DL models by looking at uncertainty. Furthermore, I will present results from research and experiments I have performed, that incorporates uncertainty aware training in cardiac DL models.
Speaker: Tareen Dawood, King’s College London
Biography: Tareen Dawood received her BSc degree in Electrical Engineering from the University of Witwatersrand in South Africa and then began her career in the corporate world and worked in a variety of industries. However, in 2018 she decided to move into academics after her own illness made her want to use her technical background to improve healthcare with a focus on the field of medical imaging. From 2018-2020, Tareen went to study her MSc in Biomedical Engineering at the University of Cape Town. Her thesis focused on automated feature detection from ultrasound to improve the diagnostic pathway for Hodgkin’s Lymphoma as it is often mistaken for TB and HIV within Africa. In 2021, Tareen began her DRIVE-Health studentship supervised by Dr Andrew King, Prof Reza Razavi and Dr Esther Puyol Antón. Her PhD project aims to develop an artificial intelligence (AI) decision support tool that can use big data to assist cardiologists in making better (trustworthy) decisions about diagnosing/treating cardiovascular patients.