Up-Sampling Active Learning: An Activity Recognition Method for Parkinson’s Disease Patients

Abstract

Parkinson’s Disease (PD) is the second most common neurodegenerative disease. With the advancement of technologies of big data, wearable sensing and artificial intelligence, automatically recognizing PD patients’ Physical Activities (PAs), health status and disease progress have become possible. Nevertheless, the PA measures are still facing challenges especially in uncontrolled environments. First, it is difficult for the model to recognize the PA of new PD patients. This is because different PD patients have different symptoms, diseased locations and severity that may cause significant differences in their activities. Second, collecting PA data of new PD patients is time-consuming and laborious, which will inevitably result in only a small amount of data of new patients being available. In this paper, we propose a novel up-sampling active learning (UAL) method, which can reduce the cost of annotation without reducing the accuracy of the model. We evaluated the performance of this method on the 18 PD patient activities data set collected from the local hospital. The experimental results demonstrate that this method can converges to better accuracy using a few labeled samples, and achieve the accuracy from 44.3% to 99.0% after annotating 25% of the samples. It provides the possibility to monitor the condition of PD patients in uncontrolled environments.