Weak Supervision for Time Series: Wearable Sensor Classification with Limited Labeled Data

Abstract

Using modern deep learning models to make predictions on time series data from wearable sensors generally requires large amounts of labeled data. However, labeling these large datasets can be both cumbersome and costly. In this paper, we apply weak supervision to time series data, and programmatically label a dataset from sensors worn by patients with Parkinson's. We then built a LSTM model that predicts when these patients exhibit clinically relevant freezing behavior (inability to make effective forward stepping). We show that (1) when our model is trained using patient-specific data (prior sensor sessions), we come within 9% AUROC of a model trained using hand-labeled data and (2) when we assume no prior observations of subjects, our weakly supervised model matched performance with hand-labeled data. These results demonstrate that weak supervision may help reduce the need to painstakingly hand label time series training data.

Tasks

Reproductions