Smart Home Healthcare Figure 2. Prototype of Glutrac, the wearable device tested in this study. Figure 1. Schematic figure of the procedure for a measurement. Device Prototype Glutrac is an innovative wearable health device with a broad light spectrum. It has two sets of optical sensors, one on the top and the other on the back of the wearable housing. The wearer is advised to place his/her index finger on the top sensor while the back one senses signals from the wrist during the signal retrieval (see Figure 2). Data Processing PPG and associated invasive BGL data were grouped, analyzed, and modeled for each participant to build the personalized models. Data derived from all the participants were used to fit a global model. The number of visits to the trial center differed among participants and, therefore, the number of measurements differed as well. Data Transmission and Storage Concept The signal data were encoded in a proprietary format designed to preserve signal quality. A trial testing mobile application was used to 32 receive signal data from Glutrac via Bluetooth communication 5.0 and to send the data to a NoSQL database for storage and computing in the cloud. Feature Extraction The signals were denoised before feature extractions. More than 30 variables were derived from the time domain and the derivative of the processed signals. Majority of the features were related to the amplitude values and statistical indices of the signals. Additional features were extracted from the frequency domain of the signals using the fast Fourier transformation. Machine Learning Random forest regression modeling uses multiple decision trees for data learning, introducing a controlled bias and variance to reduce overfitting in order to achieve a low global error rate. This method has been widely applied in modeling PPG signal regression with relevant health indices.7 Adaptive boosting (adaboost), another popular method for complex signal classification, is an iteration algorithm that changes the sample distribution by modifying the weights of misclassified data to train a weaker learner.12 IEEE Consumer Electronics Magazine