Towards understanding spatio-temporal parkinsonian patterns from salient regions of a 3D convolutional network

Gait motion patterns such as step length, flexed posture, absent arm swing and bradykinesia, constitute the main source of information to describe and quantify Parkinson disease. Nevertheless, such quantification is commonly developed under marker based protocols, losing natural motion gestures, and only taking into account a limited description of the locomotion process. This work introduces a 3D convolutional gait representation, that uses markerless video sequences to automatically predict parkinsonian behaviours. A remarkable contribution herein presented is the quantification of spatio-temporal salient maps, that stand out body regions related with Parkinson disease, and result from activations that mainly contribute on the classification task. For doing so, a convolutional architecture is trained from a set of walking videos, recorded from parkinsonian and control subjects. Then, a prediction of disease is obtained according to motion patterns computed by convolutional learned scheme. Salience motion patterns are obtained by retro-propagating the output softmax network prediction over the video space. From a total of 22 patients, and a total of 176 video sequences, the proposed approach achieved an average accuracy score of 88%. Interestingly enough, the recovered salience maps focus the attention on relevant parkinsonian biomarkers such as the head motion and trunk posture, that namely is excluded on classical gait analysis.Clinical relevance- Salient spatio-temporal regions can potentially support and complement the diagnosis and the following of Parkinson's disease. Also, such complex relationships could potentially evolve in further understanding of this pathology.