Discriminative ability and responsiveness to rehabilitation of a set of trunk acceleration-derived gait indexes in subjects with gait disorders due to central nervous system diseases

This dissertation investigates the discriminative ability and responsiveness to rehabilitation of trunk acceleration-derived gait instability indexes in individuals with neurological conditions, specifically Parkinson’s disease and primary hereditary cerebellar ataxia. Using a single magneto-inertial measurement unit placed at lumbar level during steady-state walking, this research aims to enhance the assessment of gait abnormalities, fall risk, and rehabilitation outcomes. The resulting seven studies were conducted in collaboration with neurological and rehabilitative institutions, combining expertise in neurology, rehabilitation, and biomechanics. For Parkinson’s disease (PD), a set of frequency – based and nonlinear trunk acceleration-derived gait indexes, including the Harmonic Ratio (HR), Recurrence Quantification Analysis, and multiscale entropy measures, proved effective in characterizing gait abnormalities and instability. HR in the antero-posterior direction (HRAP) was identified as the most robust marker for detecting gait abnormalities and recurrent falls, showing strong correlations with clinical features such as pelvic kinematics and axial rigidity. HRAP ≤ 1.50 predicted recurrent falls with 77% accuracy. The responsiveness of HR to rehabilitation was demonstrated in subjects with PD at moderate disease stage. In a second study, significant improvements in HR and pelvic rotation were also observed following an outpatient rehabilitation intervention. HR normalized to healthy levels post-rehabilitation, while other indices, such as RQA and step length variability, showed limited sensitivity to rehabilitation. Machine learning techniques were applied in one study on PD. A support vector machine (SVM) model achieved 86% accuracy in distinguishing subjects with PD from healthy controls, after reducing the features to a minimal set, including HR and step length variability. The SVM model outperformed other algorithms, such as Random Forest and Decision Tree, highlighting the potential of machine learning for improving diagnostic accuracy in clinical settings. The research on primary degenerative cerebellar ataxia focused on assessing a variety of trunk acceleration-derived indexes to identify gait imbalance and identify falls risk. The HR, step length coefficient of variation (CV), and the largest Lyapunov’s exponent (sLLE) were among the most effective measures for distinguishing ataxic from healthy gait and identifying recurrent fallers. Unlike in Parkinson's disease, where trunk instability was correlated with axial rigidity, gait abnormalities in cerebellar ataxia were more closely associated with a loss of inter-limb coordination and increased variability and dynamic instability of trunk acceleration patterns. Notably, sLLE was the most responsive to rehabilitation, reflecting improvements in dynamic stability and trunk control post-intervention. Additionally, a study utilizing data augmentation techniques, specifically, generative artificial intelligence to balance datasets, demonstrated significant improvements in classifying gait patterns in cerebellar ataxia. By addressing the inherent imbalance in rare disease datasets, the findings of this study may have important implications for the development of diagnostic tools in rare diseases, particularly when clinical data is limited, such as in gait analysis studies. Overall, this research highlights the utility of trunk acceleration-derived gait indexes in clinical assessments of gait stability, fall risk, and rehabilitation outcomes. The application of machine learning algorithms further enhances the diagnostic potential of these measures, offering novel tools for clinicians managing gait disorders in neurological populations. The integration of data augmentation methods also opens new possibilities for improving the classification of rare diseases like cerebellar ataxia.