Evaluation of stabilometry descriptors for human balance function classification using diagnostic and statokinesigram data

Abstract

Balance is the human ability to keep an upright position. Several conditions affect the balance function, impacting the subject’s quality of life. Objective balance assessment is performed by quantitative methods such as stabilometry. Recent works address the problem of extracting reliable information from stabilometry measures using spatial, kinematic, spectral, or structural descriptors. However, identifying an adequate representation of the subject balance condition is still an open problem. The individual evaluation of the discriminant ability for each of the selected features shows that even those descriptors exhibiting better statistical behaviors could be insufficient to provide the discriminant properties required by this problem. Nevertheless, using ML-based classification methods on a proper subset of descriptors yields promising balance condition classification results. Thus, this work evaluates the performance of two classification models (Gradient-Boosting models and Artificial Neural Networks, ANN), trained with statokinesigram descriptors and subject-related data, according to their ability to discriminate between “preserved” and “altered” balance conditions, labeled by an orthopedist on the public dataset of Santos and Duarte. To assess such an ability, a summarized representation of the balance performance is obtained by PCA-based feature selection. Then, two different classifiers were trained using a combination of the selected features. Different Gradient-Boosting and ANN models were evaluated using a grid search strategy. The best model (an ANN) reached a testing accuracy of 0.7187 and a recall value of 0.7222, proving that the balance condition classification is feasible, yet more comprehensive datasets could help to improve the classification performance.