Monitoring motor symptoms in Parkinson’s disease via instrumenting daily artifacts with inertia sensors

Abstract

Daily monitoring of Parkinson's disease is important since clinical assessments can only provide a brief and limited view of a patient's condition. However, traditional approaches rely heavily on patients' self-reports or diaries, which are subjective and often lack of necessary details. In this work, we instrument a handle that can be attached to cutlery with inertial sensors to collect motion data unobtrusively. By analyzing the data of patients and normal people collected in the clinic, we demonstrated that our machine learning based model can not only distinguish between patients and normal people, but also identify the disease levels in a fine-grained manner. To further understand how the self-tracking data is used in clinic, we conducted a semi-structured interview with several clinicians. Through the interpretation from the perspective of both physicians and patients, we found that our handle can help the physicians better understand disease progression and promote patients' engagement in tackling the disease.

Internet of Things Technologies and Machine Learning Methods for Parkinson’s Disease Diagnosis, Monitoring and Management: A Systematic Review

“…In concordance with the previous case, various ML classifiers, including NB, LR, kNN, SVM, adaptive boosting (AdaBoost), DT, RF, gaussian mixture model (GMM) and deep neural networks (DNN) classifiers, have been proposed in the literature to address this problem. To elaborate, tremor measurements have been obtained with the help of inertial sensors embedded in smartphones [55], Wii Remote [56] and spoon handles [57] to differentiate PD patients from healthy controls by exploiting conventional ML algorithms. RF and SVM models perform very well, achieving 0.94 and 0.98 AUROC, respectively.…”

“…Unexpectedly, the DL approach [72] performs a bit worse than other conventional ML approaches, but the different placement of the sensors may have influenced the results. Finally, H&Y scale scores have also been estimated with SVM models based on tremor measurements from inertial sensors embedded in everyday objects, achieving accuracy up to 77% and correlation coefficient up to 0.97 [56,57]. Consequently, in the considered studies, PD patients are classified according to their H&Y scores with slightly lower accuracy than when they are classified according to their UPDRS scores.…”