Keywords: Neural networks, machine learning, signal processing
The L-Dopa Challenge was a study to study actigraphy from wearables in PD patients at two clinical sites. Patients were asked to wear 1 GeneActiv wristwatch on their most affected side, and 1 Pebble smartwatch on their least affected side. Acceleration data was collected from the sensors over 2 visits. Acceleration from individual axes are reported for the GeneActiv wristwatch and Pebble smartwatch. In addition to the acceleration data, clinical labels of symptoms severity and symptom presence have been provided for each motor task (see below) by a clinician. Limb-specific tremor severity score (0-4) are provided as well as upper-limb presence of dyskinesia (yes or no) and bradykinesia (yes or no). The goal of this challenge is to develop feature extraction methods for mobile sensor data which can be used to predict Parkinson's Disease state.
Recent development of research suggests that deep convolutional neural networks are suited to automate feature extraction from raw sensor inputs. We used CNN layers combined with LSTM to capture the temporal dynamics of the human activity data. The network was trained one-to-end using the Dyskinesia, Tremor and Bradykinesia score, with the given cost function and optimizer. Then, the output of the LSTM layer was used as a feature generator. We first transform the sensor data to a spectrogram of an accelerometer signal that demonstrates four dimensional representation of changes in the acceleration energy content of a signal as a function of frequency and time[Alsheikh,2016]. We trained the CNN LSTM on the spectrogram. merged with the demographic features given in the dataset in the last layer for membership predictions.
The accuracy is higher for Dyskinesia variable (Accuracy :88%) compared to Tremor which has five classes (Accuracy : 72%). Our tremor predictions have the highest Area under the curve value.