Wearable biosensors, as a key component of wireless body area network (WBAN) systems, have extended the ability of health care providers to achieve continuous health monitoring. Prior research has shown the ability of externally placed, non-invasive sensors combined with machine learning algorithms to detect intoxication from a variety of substances. Such approaches have also shown limitations. The difficulties in developing a model capable of detecting intoxication generally include differences among human beings, sensors, drugs, and environments. This paper examines how approaching wireless communication advances and new paradigms in constructing distributed systems may facilitate polysubstance use detection. We perform supervised learning after harmonizing two types of offline data streams containing wearable biosensor readings from users who had taken different substances, accurately classifying 90% of samples. We examine time domain and frequency domain features and find that skin temperature and mean acceleration are the most important predictors.
Keywords: addiction; biosensor; classification; data stream; feature extraction.