A Two-Step Machine Learning Framework for Wearable Sensing Systems in Personal Healthcare
Master of Science
Wearable sensing systems can support a wide range of real-world applications. In the past years, a lot of research in this field explored how to design machine learning models using wearable sensor data for personal healthcare usage. There are two challenges in dealing with wearable sensor data for personal healthcare: 1) how to incorporate data sampled at different sampling rates into one model, e.g., daily sampled data and frequently sampled data, and 2) how to deal with the data that contain long-missing patterns. For overcoming these two challenges, we propose a two-step machine learning framework, where the first step extracts features from the data before a predefined time point, while the second step combines this summary with the rest part of data for the machine learning task. For investigating the first problem, we implement our framework for predicting dim light melatonin onset (DLMO) that uses daily sampled sleep parameters and frequently sampled sensor data (light exposure, skin temperature, and physical activity). For the second problem, we predict momentary stress state using sensor data that contains some long-missing segments. The experiment shows that the two-step framework has better performance on both tasks than traditional one-step models, which suggests that this framework is applicable to addressing the above two challenges.
sensor data; machine learning