Analysis of biases in classification algorithms of pulse oximetry for detection of sleep disorders

Abstract

The use of machine learning algotithms and techniques within the field of medicine has proven to be very useful in diagnostic assistance tasks. At the same time, in recent years, important issues have been revealed that affect the performance of these methods in specific subpopulations, presenting uneven performance in certain demographic groups. This uneven performance is often associated with the underrepresentation of these populations in the data used during training, resulting in biased models. This is a preliminary work that seeks to explore the existence of biases in the classification of apnea from pulse oximetry considering different ethnic groups. To do this, a database with ethnic information on patients and an algorithm to detect sleep disorders such as apnea or hypopnea are used. For the detection of such events, only the peripheral blood oxygen saturation (SpO2) signal is used together with the algorithm based on DAS-KSVD dictionary learning. The experiments consist of analyzing the MESA polysomnography database, which contains four ethnic groups, using the SpO2 signal for dictionary learning that improves the pathology detection process.