Sleep-wake prediction with machine learning methods

Sleep is vital for our health and overall quality of life. Adequate and quality sleep increases our energy levels, improves our ability to focus and positively affects our mood. In addition, a regular and healthy sleep habit contributes to the prevention of many chronic diseases. On the other hand, short sleep duration or poor sleep quality can lead to various health problems in individuals. Therefore, regular monitoring and analysis of sleep quality is of great importance. Today, thanks to innovations in health technologies, the quality and duration of sleep can be measured and analyzed using various devices and techniques. However, these techniques are usually applied in sleep laboratories, making the process timeconsuming and costly. In addition, difficult access to sleep laboratories makes sleep analysis not suitable for everyone. These situations reveal the need for alternative and more accessible methods in sleep analysis. In this thesis, the effectiveness of five different machine learning algorithms in predicting sleep and wakefulness states is investigated. The study evaluated the algorithms' ability to model time series data and their suitability for problems with continuous, sequential data. According to the results, the Long Short-Term Memory (LSTM) algorithm achieved the highest success in sleep/wake prediction with 93.48% accuracy. In addition, the Random Forest algorithm achieved 81% and Decision Trees achieved 76%. These results show that the LSTM algorithm is an ideal choice for problems with continuous and sequential data. This finding shows that LSTM is more effective than other algorithms in predicting sleep and wakefulness states and enables the development of more advanced sleep analysis methods.