Abstract

Wearable sensors have become more commonly used in everyday basis and powerful in terms of computational capacity and sensing resources, including capability to collect data from different bio-signals. The data collected from everyday wearables offers huge opportunities to monitor people’s everyday life without expensive laboratory measurements, including also behaviours and conditions only rarely seen in controlled laboratory environments. So far wearable sensors have mostly been used to monitor motion, but bio-sensor powered wearables can do a lot more: they can be used to monitor physiological reactions inside the human body as well as some psycho-physical reactions such as affection and stress. This development enables multiple interesting and important applications, such as early detection of diseases, seizures, and attacks. With stock wearables worn in everyday basis, one of the biggest challenges for such applications is the sensing data itself. In order to train reliable recognition and prediction models, high quality training data with labels needs to be collected. This paper focuses on lessons learned of challenges in data quality and quantity when such data sets are gathered. We discuss our own experiences when collecting data using wearable sensors for early detection of migraine attacks, but the same lessons learned can be generalized to other studies utilizing wearables for recognition medical symptoms and users’ everyday behaviour.