Deep learning for Parkinson's disease: a case study on Freezing of Gait

We propose a deep-learning method for feature extraction from gait data of Parkinson’s disease patients. Our goal is to verify whether a fine classification of gait between similar groups can be achieved. To this end, we refer as a case study to the Freezing of Gait (FOG), and we measure gait data from two groups of patients, which exhibit (respectively, do not exhibit) this symptom. Wearable inertial sensors are employed, and data are collected during activities similar to those performed by patients during their daily living. Moreover, most patients are in daily on state, hence the two groups are difficult to classify, as their gait does not exhibit evident differences. Whereas classical Machine Learning methods are not sufficiently robust to perform such a fine classification, if they are fed with features extracted by means of a deep network, the results are satisfactory also when a large dataset is not available and data present a mild degree of heterogeneity