Hand Tremor Characterization from a Spatiotemporal Convolutional Representation

Context: Parkinson’s Disease (PD) is a neurodegenerative disorder related to dopamine deficiency that mainly entails motor conditions such as slowness of movement, postural instability, limb tremor, rigidity, and a decreased range of motion. Tremor, defined as a rhythmic and uncontrolled movement of limbs, is the most prevalent symptom in PD. In the clinical routine, tremors are assessed and quantified by observing the hands following postural and resting patterns. These configurations include voluntary muscular contractions and tremor perception reduction, which leads to noisy signals. The assessments are also subjective and depend on the expertise of professionals to determine whether the tremor is associated with PD. Method: This work introduces a deep volumetric representation that characterizes PD tremor patterns in resting and postural recording conditions. The strategy includes a convolutional architecture that extracts spatiotemporal patterns correlated with tremor, propagated through different layers until discrimination between PD and control subjects is achieved. Moreover, a set of explainability maps is computed by backpropagating output gradients into convolutionally learned spatio-temporal maps. Results: The method was evaluated on 80 videos (five PD patients and five control subjects), reporting an average accuracy of 92.5% and a perfect sensitivity score in the postural configuration. As for the resting scheme, the proposed method obtained an average accuracy of 90% and sensitivity of 80%. Conclusions: This approach showed efficacy regarding the localization of tremor patterns, recovering movement information while preserving the spatial and temporal representation. The strategy allows visualizing movement patterns from explainability maps of control subjects and PD patients.