Seite - 34 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

The analysis of results demonstrates that in our approach, the decision tree algorithm performed better, not only at global accuracy but also individual sleep phase detection. The multilayer perceptron could be improved by experimenting with different hidden layers and more iteration per row, however, they lack the ability to give clues about which features are more important from environment conditions, so they were disregarded. Focusing on the analysis of the decision tree, it becomes clear that moving average input are determinant as they appear in the upper nodes of the decision tree along with accelerometer data. Surprisingly, gas sensor data from the MQ6 and MQ135(CO2) sensors are the first to appear in the upper nodes of the decision tree. These observations motivate further studies to access the influence of such gases on sleep patterns. Aside from being able to identify sleep stages by time series data, this approach is able to characterize how active a person is by estimating the number of steps a person made analyzing time series sleep data (Figure 7). Using the same approach, with the number of steps extracted from the Fitbit wearable and with sleep data as input in moving average windows, it was possible to estimate daily number of steps with as mean absolute error of 87 steps and a mean squared error of 288. This provides evidence that is possible to estimate the activity of people by analyzing only sleep data of about 30 min with our multisensor architecture and sleep classification from the Fitbit wearable. The next stage is to remove the need for classified sleep patterns, which came from the Fitbit wearable, making this solution to use only accelerometer data and environmental data. Figure 7. Machine Learning workflow for step estimation from sleep data 5. Conclusions and Future Work The study here presented infers quality of sleep tracking using environmental and accelerometer data from people. The results demonstrate that it is possible with a good C.Gonçalvesetal. /MultisensorMonitoringSystem toEstablishCorrelations34