A comparison between Machine Learning models for OSA detection based on ECG signal

OSA is a one of the most common sleep disorders nowadays, which is diagnosed by a Polysomnography (PSG) study. Even though this is the golden standard to diagnose OSA, it is time consuming, very expensive and there are not many specialized centers to conduct it, this implies that fewer patients are diagnosed. The development of new solutions at a lower cost and in less time would allow more patients to be diagnosed and treated promptly. There are solutions that enable the diagnosis of OSA through monitoring signals from the human body, including an auto-diagnosis. However, these solutions do not aim to perform screening on the most significant parameters along with the best model for making predictions. The main objective of this tesis is to make a comparison between 27 Machine Learning (ML) models in order to find the best model to diagnose OSA. It also aims to find which are the most representative parameters in OSA detection. By doing a frequency-domain, time-domain and non-linear domain analysis to extract them from the RR intervals, and with a wilcoxon test and correlation matrix, select the most useful ones. The results showed that the best model was Support Vector Machine (SVM) with an accuracy, balanced accuracy, ROC AUC and F1 Score of 0.97. The most significant parameters found were: RR tri index, LF/HF ratio, alpha 2, HF\%, SDNN and RMSSD. This solution can be integrated into current clinical cases for a quick OSA diagnosis. Proposal does not aim to replace PSG for a complete and accurate diagnosis, but to be a pre diagnosis accessible to a larger number of patients. Health providers can implement this solution and reduce the number of patients in the waiting list. Also, this proposal would make research in OSA diagnosis more accessible and provide a framework that can be the starting point to other researchs.