Seminar: Machine Learning Seminar
Machine Learning for Screening and Diagnosis of Respiratory Pathologies
Lecturer:
Jeremy Levy
Research Areas:
| Respiratory diseases affect the lives of more than one billion people worldwide and are among of the leading causes of mortality and morbidity. There is a need to engineer new noninvasive, low-cost, and robust methods for identifying respiratory diseases. Modern advances in machine learning represent an exciting avenue for addressing this challenge. This thesis is devoted mostly to the analysis of lung sounds and oximetry signals. We present the development and validation of a deep learning algorithm, OxiNet, leveraging long-term continuous physiological recordings for the task of robust diagnosis of OSA by leveraging continuous physiological data recorded during sleep. Utilizing the to-date largest retrospective study, we used 12,923 polysomnography recordings from six independent databases, to develop and evaluate a deep learning model, called OxiNet, for the estimation of the apnea-hypopnea index, based on the oximetry signal. We evaluated OxiNet performance across ethnicity, age, sex, and comorbidity. OxiNet missed 0.2% of all test set moderate-to-severe OSA patients against 21% which the best benchmark. To address the challenge of generalization performance and reduce the gap between source domain and target domain performances, we developed an algorithm named Deep Unsupervised Domain adaptation using variable nEighbors (DUDE). DUDE enables to increase the generalization performance of OxiNet by up to 5% on target domain data. First, we developed a new approach to the representation and analysis of one-dimensional signals as surfaces, which involves embedding geometrical objects in Euclidean spaces. This new approach allows to define distortion measures based on these representations. This approach is shown to be effective, as demonstrated on lung sounds, in the context of diagnosing a total of eight respiratory conditions, achieving a recall of 0.89 on a dataset of 918 recordings. Furthermore, this method was extended to define a metric for assessing signal similarity and has been used in the context of face recognition. Ph.D. Under the supervision of Prof. Emeritus Josh Zeevi and Prof. Joachim Behar.
|