MetaParkinson: A Cyber-Physical Deep Meta-Learning Framework for <i>n-</i>Shot Diagnosis and Monitoring of Parkinson's Patients
M. Shamim Hossain, Mohammad Shorfuzzaman
Abstract
Parkinson's disease (PD) is a neurodegenerative condition that affects the motor system and can result in tremors, limb rigidity, and gait and balance disorders. Currently, conventional tests consist of a series of spiral drawings, with the notion that the most prevalent PD symptoms directly impact the visual appearance of handwritten geometric dynamics. Current disease diagnosis methods rely heavily on a large amount of preexisting clinical data to train their models. However, the lack of sufficient labeled data samples in many diseases, including PD, makes it challenging to build robust predictive models, especially deep-learning ones. With measuring devices becoming more common and the continuous increase in processing power, industrial cyber-physical characterization is emerging as a key enabling technology for a wide range of fields. This article proposes an industrial cyber-physical deep meta-learning framework for diagnosing and monitoring PD patients based on metric learning that aims to train a reliable PD classifier with limited data. In particular, we present a component-based cyber-physical system architecture to combine contrastive learning with a CNN encoder to obtain unbiased feature embeddings and then use a Siamese network to finalize PD classification. Our technique outperforms pretrained models with few training instances on a publicly available PD dataset of hand-drawn spiral and wave images.