Page - 33 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

commercially available sleep trackers to identify stages of sleep in test subject but aims to use and explain the root causes of the quality of sleep using alternative data captured from the environment where the test subject is resting. 4.4. Machine Learning Model Performance and Discussion The assessment of each machine learning model is made by the execution of each machine learning workflow. Each model assessment required a workflow as exemplified in Figure 6. The output and prediction were stored in the same database where sensor data was extracted. Scorer nodes may also be used to obtain direct model performance. Figure 6. Knime MultiLayer Perceptron workflow The results from the prediction of sleep phases are compiled in Table 4. These results are obtained from the machine learning workflows defined in the workflow program knime, which are trained, and stored as simplified and with models trained as PMML machine learning model. Later when a live system requires the analysis of sleep phases with data provided by our mutlisensor architecture, the stored PMML model can be used to make prediction based on the pre-trained models. Table 4. Machine Learning models prediction of sleep stages Algorithm Accuracy (%) Error (%) Accuracy Awake (%) Accuracy REM (%) Accuracy Light (%) Accuracy Deep (%) Decision Tree 89 11 88 87 93 82 Naïve Bayes 51 49 61 22 64 27 MLP 69 31 59 63 71 73 C.Gonçalvesetal. /MultisensorMonitoringSystem toEstablishCorrelations 33