Two-Stage Convolutional Neural Network for Classification of Movement Patterns in Tremor Patients

Abstract

Accurate tremor classification is crucial for effective patient management and treatment. However, clinical diagnoses are often hindered by misdiagnoses, necessitating the development of robust technical methods. Here, we present a two-stage convolutional neural network (CNN)-based system for classifying physiological tremor, essential tremor (ET), and Parkinson’s disease (PD) tremor. Employing acceleration signals from the hands of 408 patients, our system utilizes both medically motivated signal features and (nearly) raw data (by means of spectrograms) as system inputs. Our model employs a hybrid approach of data-based and feature-based methods to leverage the strengths of both while mitigating their weaknesses. By incorporating various data augmentation techniques for model training, we achieved an overall accuracy of 88.12%. This promising approach demonstrates improved accuracy in discriminating between the three tremor types, paving the way for more precise tremor diagnosis and enhanced patient care.