Prediction of optimum pressure in CPAP devices for osas patients by artificial intelligence

Obstructive Sleep Apnea Syndrome (OSAS) is a sleep disorder that manifests itself as breathing breaks due to recurrent upper airway constrictions or obstructions during sleep. The breathing breaks during sleep disrupt the continuity of individuals' sleep, prevent deep and restful sleep, and thus negatively affect their daily lives. In addition, breathing breaks often reduces oxygen saturation in the blood, causing many serious diseases, especially cardiovascular diseases, and may even lead to death. In order to eliminate all these negative effects, an effective treatment should be applied to patients diagnosed with OSAS. Continuous Positive Airway Pressure (CPAP) therapy is the gold standard treatment method for OSAS. In this treatment performed with CPAP devices, the upper airway of the patients is kept open with the determined constant pressure. The most effective therapeutic pressure (optimum CPAP level) for patients is determined by sleep specialists through a manual titration process in sleep laboratories. However, this process is both costly, time consuming and quite tiring. For this reason, many studies have been conducted in the literature to develop alternative solutions to titration in order to determine the optimum CPAP level. Most of the studies focused on the direct prediction of the optimum CPAP level to be applied to the patient, and they produced regression-based formulas based on similar features as alternative solutions. However, until today, an alternative solution that is fully accepted in the clinic, applicable and with high accuracy has not been developed yet. In this thesis study, the optimum CPAP levels of those who applied to Necmettin Erbakan University, Meram Medical Faculty, Department of Chest Diseases, Sleep Clinic with the suspicion of OSAS were predicted. Before the prediction process was performed, it was determined whether the subjects included in the thesis study were OSAS patients or not and required CPAP therapy or not by using Nonlinear Analysis and Rule-Based Algorithm approaches and nasal cannula airflow and oximetry signals, which are among the polysomnography signals. With the Nonlinear Analysis Approach, OSAS patients were differentiated from healthy individuals with 93.10% accuracy, 96.43% sensitivity and 81.82% specificity, and patients requiring CPAP therapy could be identified with 96.25% sensitivity and 89.53% precision. The Rule-Based Algorithm approach distinguished OSAS patients from healthy individuals with 98.62% accuracy, 98.21% sensitivity and 100% specificity, while it was able to identify patients requiring CPAP therapy with 100% sensitivity and 100% precision. After the identification of patients requiring treatment with OSAS therapy, different feature sets were created by taking into account the demographic and anthropometric information of these patients, nasal cannula airflow and oximetry signals, apneic attacks and oxygen reductions during the night. Afterwards, the most effective ones were selected among the features in different feature sets on the optimum CPAP levels and the optimum CPAP levels of the patients were predicted by using both all features and effectively selected features with linear regression analysis and various artificial intelligence methods. Finally, 40 different CPAP estimation models were developed with the features and methods (stepwise multiple linear regression analysis, artificial neural networks, support vector machine, random forest and k closest neighborhood) with high prediction performance. 36 of these 40 CPAP estimation models were obtained by different artificial intelligence methods, only 4 of them were obtained by linear regression analysis. As a result, Support Vector Machines with polynomial or radial basis kernel functions created 24 of 36 models developed with artificial intelligence methods and became the most successful method in prediction process. 25 model among the developed 40 models revealed a high correlation (0.6 ≤ r <0.8; 0.36 ≤ r2<0.64) between the CPAP levels predicted in the thesis and the optimum CPAP levels determined by the specialists in sleep laboratories, and 10 model defined the this relationship as very high (r ≥ 0.8; r2 ≥ 0.64). When all the results obtained in the thesis were evaluated in general, it was seen that the characteristics of airflow and oximetry polysomnography signals, apneic attacks experienced by OSAS patients and the decreases in oxygen levels due to the attacks had great effects on optimum pressures. In addition, this thesis study revealed that the linear and nonlinear relationships between the features used for prediction can be explained better with artificial intelligence methods, and thus, many artificial intelligence methods generally produce more successful results compared to linear regression analyzes, which are mostly preferred in the literature for prediction of optimum CPAP levels. The thesis study has brought developed new CPAP prediction models and new features used as prediction parameters in models to the literature.