Scalability of data analytics in Health related Activity recognition in Internet of Things
Ubiquitous presence of wearables and smartphones has enabled unobtrusive collection of human activity and vital signs related data in abundance from sensors such as accelerometer, photoplethysmogram, galvanic skin response, electrocardiogram among many others. Data analytics on such healthcare data has immense potential in improving quality of life outside hospital and nursing homes. In this work, we focus on developing learning approaches that can classify a wide variety of activities of daily living as well as assist in healthy living by early detection and prevention of chronic diseases. The intervariability among people’s health and activity behavior calls for complex subject specific models. Another part of the problem lies in the fact that training a subject specific model requires a tremendous amount of labelled data, which is not practically available. In order to make such personalized models from generic models, transfer-learning approach can be considered that uses a pre-built model trained in a similar domain on a different set of population and learns the specific user health and activity behaviors from this general model in a much faster scalable manner. Another important aspect is developing approaches that can address the problem of limited availability of labelled data. Active learning is one such approach that aims at identifying data points that, if labelled and used for training can improve the model. Labelling is then required for the most promising data points only and not for the complete training data. A combination of transfer based learning, and active learning approach will be looked upon that can help develop fast semi-supervised scalable models for health related applications by reducing the training and labelling costs at the same time.