Objective: Our goal was to classify spontaneous movement using a commercially available motion sensor in a clinically meaningful 3-/9-step discrete scale (SMS-3/SMS-9).

Background: Brady-hypokinetic and dyskinetic spontaneous movement is the fundamental characteristic of the motor syndrome of patients with Parkinson’s disease (PwP).

Methods: Ethical approval from the Technical University of Munich was obtained prior to data collection. In total, we evaluated 27 PwP with a fluctuating motor disorder and recorded 219 hours of labeled raw movement data in a free-living setting acquired by a wrist-worn 6-axis accelerometer/gyroscope at a sampling frequency of 62.5 Hz. Clinical information was assessed every minute by an MDS certified rater. Additionally, activities of daily living were recorded. Sensor data was validated using an external reference frame, preprocessed and then was used to train deep neural networks (Long Short-Term Memory, LSTMs & Convolutional Neural Networks, CNNs).

Results: For training, short (4-12 second) segments were extracted from the data and used to train a CNN to perform a SMS-3/SMS-9 classification. The input observation vector to the CNN were 4.8 sec raw data segments; classification results for each 1 minute segment were generated by averaging across windows. Subject-specific classifiers achieved 89 – 100% testing accuracy. Partial generalization to novel participants could be achieved by incremental adaptation of a pre-trained network using a small subset of person-specific data, achieving SMS-9 accuracy of 94%. We also tested approaches with Recurrent Neural Networks (LSTM) and spectral analysis data, which showed inferior performance compared to the CNN. Ultimately, we were able to develop a continuous Spontaneous Movement Score (SMS), which reflects the current motion state of PwP.

Conclusions: We describe a novel approach for the objective classification of the core motor characteristic of PwP using deep learning and a low-cost commercially available sensor device. This method is not limited to a controlled test setup, but can be applied in free-living situations. The precision and temporal resolution of the measurements is unprecedented, and could be used for numerous clinical indications, such as monitoring of patients, diagnosis, and closed-loop therapy.

« Back to 2017 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/deep-learning-in-objective-classification-of-spontaneous-movement-of-patients-with-parkinsons-disease-using-large-scale-free-living-sensor-data/