Parkinsonian gait patterns quantification from principal geodesic analysis

Parkinson is a neuromotor disease caused by dopamine deficiency that produces progressive alterations in locomotion. Gait analysis is a primary alternative to observe, quantify and follow Parkinson's disease (PD). Nonetheless, these patterns are coarsely captured from marker-based setups which alter natural gestures and limit the sensibility to describe disease progression. This work, from a markerless video analysis, built a Riemmanian geometry gait representation to analyze principal geodesic variations and quantify alteration on locomotor patterns. The proposed approach project walking markerless videos into a bank of deep features. These deep activations are coded into compact covariance matrices, which coexist in a special Riemmanian manifold. A video gait descriptor is then formulated as the geometric mean and the principal geodesic directions with greatest variance. These descriptors are finally used in an automatic supervised PD classification task. Experiments were carried out in a dataset with 22 patients, equally distributed between control and Parkinson's disease. Following a leave-one-patient-out cross-validation, the proposed video gait descriptor achieves an average accuracy of 99% and a true positive rate of 97%. Besides, the resultant geometry descriptor space was projected in a low-dimensional version, as an alternative to carried out observational gait analysis. This space shows a robust clustering among the evaluated classes. The proposed approach takes advantage of greatest geodesic variance to quantify abnormal locomotion PD patterns. The validation suggests that the strategy could be potentially implemented as an alternative to support diagnosis and following of the disease.