Ensemble Deep Learning on Wearables Using Small Datasets

This article presents an in-depth experimental study of Ensemble Deep Learning techniques on small datasets for the analysis of time-series data generated by wearable devices. Deep Learning networks generally require large datasets for training. In some health care applications, such as the real-time smartwatch-based fall detection, there are no publicly available, large, annotated datasets that can be used for training, due to the nature of the problem (i.e., a fall is not a common event). We conducted a series of offline experiments using two different datasets of simulated falls for training various ensemble models. Our offline experimental results show that an ensemble of Recurrent Neural Network (RNN) models, combined by the stacking ensemble technique, outperforms a single RNN model trained on the same data samples. Nonetheless, fall detection models trained on simulated falls and activities of daily living performed by test subjects in a controlled environment, suffer from low precision due to high false-positive rates. In this work, through a set of real-world experiments, we demonstrate that the low precision can be mitigated via the collection of false-positive feedback by the end-users. The final Ensemble RNN model, after re-training with real-world user archived data and feedback, achieved a significantly higher precision without reducing much of the recall in a real-world setting.