Preliminary Results of Phenotype Characterisation for Cerebellar Ataxia using Automated Lower Limb Assessment

This study investigates the utilisation of an instrumented lower limb test alongside automated feature extraction and machine learning for the classification of common phenotypes in Cerebellar Ataxias (CAs) pure CA, CA with Bilateral Vestibulopathy (CA+BV) and CA with Bilateral Vestibulopathy and Somatosensory Impairment (CA+BV+SS). Participants completed the Heel-To-Shin test (HST) for CA evaluation while using an Inertial Measurement Unit (IMU) incorporated in the BioKin™ system affixed to their leg to capture the movement. The resultant kinematic parameters were quantitatively analysed to identify movements that are characteristic of each phenotype. The significant features were identified using one-way ANOVA and Recursive Feature Elimination, with features related to pronation and supination of the heel showing the most separation. Several feature subsets were explored for two sets of phenotypic groupings to generate a classification model to correctly identify each phenotype. The performance was evaluated based on macro-averaged F1-score and Receiver Operating Characteristic Area Under the Curve. The MLP classifier model performed best in classifying the CA+BV+SS phenotype, with a class-specific F1 score of 0.67 and an overall F1 score of 0.74 and ROC-AUC score of 0.81. The model that most correctly identified CA+BV phenotype with or without SS was the Extra Trees model, with a class-wise F1 score of 0.80, and also a macro-averaged 0.78 F1 and 0.89 ROC-AUC scores overall.Clinical relevance- Neurological disorders, such as CAs, exhibit various clinical features, including gait instability, limb incoordination, dysarthria, ocularmotor abnormalities, among many others. Phenotypic classification in CAs often requires integrating clinical, genetic, neuroimaging, and biochemical data. This paper offers an objective, instrumented approach to facilitate CA phenotype characterisation using only a simple lower limb test while demonstrating the possibility of using kinematic data to delineate complex disease presentations in cases of movement disorders.