Seite - 32 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

was also considered. The timestamp values of each entry of the dataset were converted into two integer values, representing the hour and the minute of the corresponding timestamp. In addition, some other data were added to each dataset entry, such as the number of the experiment or session (1 to 7), the starting timestamp of the experiment, also converted into two integer values, total sleep time and the time spent in each sleep stage (Awake, REM, Light and Deep) for the corresponding experiment (in minutes). Finally, to conclude the preparation of the dataset, and considering the information obtained from the Fitbit wristband app, each entry of the dataset received the total number of steps taken by the subject during the corresponding day and, for the respective timestamps, the sleep stage estimation, converted to a numerical value: 0 for awake stage, 1 for REM, 2 for Light and 3 for Deep. Figure 5. Fitbit sleep stages estimation for the first night of the experiment. 4.3. Machine Learning Model Development Initial experiments are a means to demonstrate the validity of the model and the prototypes being developed. In this case, the data from 7 experiments was considered and used to develop machine learning models to estimate sleep phase from historic time series data. The data used, came from the environment multisensory, bedside accelerometer while sleep identification came from the Fitbit wearable. As presented in Figure 2, the data was stored in local database which is accessible by machine learning workflows developed in the Knime platform [15]. The machine learning workflow was developed using machine learning algorithms and is compatible with the PMML specification [14] which means it can be transported to other machine learning frameworks and retain the same characteristics. With this approach the research team is able to test different theories and allows the rapid development of decision support systems. The data was prepared following CRISP-DM methodology where, the data domain was studied as detailed in section 1 and 3, the data was prepared for each algorithm which implied the transformation of simple data rows into summaries of one minute for each sensors and then a moving average of the last 15 and 30 minutes. This implies that the decision support system requires at least 30 min input data before it can predict sleep phase during sleep sessions. In the model evaluation stage, specific data preparation was required due to algorithmic restrictions. The algorithms being tested were, Naïve Bayes, Multilayer Perceptron and C4.5 based decision tree. These experiments lay the foundation to study the effects on environments on the quality of sleep. The methodologies employed takes advantage of C.Gonçalvesetal. /MultisensorMonitoringSystem toEstablishCorrelations32