A Hierarchical Architecture for Multisymptom Assessment of Early Parkinson's Disease via Wearable Sensors

• Published in: IEEE transactions on cognitive and developmental systems Vol. 14; no. 4; pp. 1553 - 1563
• Main Authors: Wang, Chen, Peng, Liang, Hou, Zeng-Guang, Li, Yanfeng, Tan, Ying, Hao, Honglin
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.12.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Abnormalities; Accelerometers; Algorithms; Correlation coefficients; Diseases; Gyroscopes; Heterogeneity; Hidden Markov models; Machine learning; Markov chains; Monitoring; multilevel fusion; multisymptom assessment; Parkinson's disease; Parkinson’s disease (PD); Self assessment; Sensor systems; Signs and symptoms; Telemedicine; Wearable computers; Wearable devices; wearable sensor system; Wearable technology
• ISSN: 2379-8920; 2379-8939
• DOI: 10.1109/TCDS.2021.3123157

Parkinson's disease (PD) is the second most common neurodegenerative disorder and the heterogeneity of early PD leads to interrater and intrarater variability in observation-based clinical assessment. Thus, objective monitoring of PD-induced motor abnormalities has attracted significant attention to manage disease progression. Here, we proposed a hierarchical architecture to reliably detect abnormal characteristics and comprehensively quantify the multisymptom severity in patients with PD. A novel wearable device was designed to measure motor features in 15 PD patients and 15 age-matched healthy subjects, while performing five types of motor tasks. The abnormality classes of multimodal measurements were recognized by hidden Markov models (HMMs) in the first layer of the proposed architecture, aiming at motivating the evaluation of specific motor manifestations. Subsequently, in the second layer, three single-symptom models differentiated PD motor characteristics from normal motion patterns and quantified the severity of cardinal PD symptoms in parallel. In order to further analyze the disease status, the multilevel severity quantification was fused in the third layer, where machine learning algorithms were adopted to develop a multisymptom severity score. The experimental results demonstrated that the quantification of three cardinal symptoms was highly accurate to distinguish PD patients from healthy controls. Furthermore, strong correlations were observed between the Unified PD Rating Scale (UPDRS) scores and the predicted subscores for tremor <inline-formula> <tex-math notation="LaTeX">{(R = 0.75,\;P = 1.40e - 3)} </tex-math></inline-formula>, bradykinesia <inline-formula> <tex-math notation="LaTeX">{(R = 0.71,\;P = 2.80e - 3)} </tex-math></inline-formula>, and coordination impairments <inline-formula> <tex-math notation="LaTeX">{(R = 0.69,\;P = 4.20e - 3)} </tex-math></inline-formula>, and the correlation coefficient can be enhanced to <inline-formula> <tex-math notation="LaTeX">{0.88}\,\,{(P = 1.26e - 5)} </tex-math></inline-formula> based on the fusion schemes. In conclusion, the proposed assessment architecture holds great promise to push forward the in-home monitoring of clinical manifestations, thus enabling the self-assessment of disease progression.