A convolutional neural network approach for quantification of tremor severity in neurological movement disorders

Tremor is the most common movement disorder which is observed in Parkinson’s disease, Essential tremor, Dystonia and similar neurological disorder patients. Tremor gets worst with increase in age and if left misdiagnosed from the begining. In the neuroclinics tremor is assesed using a manual clinical rating task which is time taking and cumbersome and mostly perception based. A need is ascertained by the neuro-clinicians to develop an automatic tremor severity quantification system which should be reliable, accurate in accordance to the manual rating scale, highly precise and with a good recall rate. Hence, an accurate tremor quantification system is essential in providing appropriate diagnosis to assist the clinicians in a clinical or home based setting. In this research, a tremor severity assessment system is proposed which adopts a convolutional neural network based approach to accurately quantify the severity of tremor as measured in data collected from a wearale device. 27 tremor patients PD(14) and ET(13) were assessed using an accelerometer based wearable device mounted on the upper limb of the patient. Neurologists provided clinical scores to the patients using Unified Parkinson’s Disease Rating scale. The convolved 2-D image of the frequency characteristic of the tremor signal with Kernels was used to train the CNN. The proposed CNN architecture outperformed (accuracy = 0.91, l.w.kappa = 0.91) the other machine learning method used for the same purpose.