Modeling the Progression of Speech Deficits in Cerebellar Ataxia Using a Mixture Mixed-Effect Machine Learning Framework

Background: Accurate and reliable prediction of changes in the severity Cerebellar Ataxia (CA) will be helpful trials disease-modifying therapies. This study demonstrated that objective acoustic measures were more sensitive than perceptive analysis through Scale for Assessment Rating (SARA) assessing CA progression, within a time window two years (mean). Method: Thirtyseven people with tested at baseline (time point 1, TP1) later 2, TP2). A machine-learning framework robust three-step feature selection criterion Bayesian data-driven clustering technique based on multivariate mixture extension generalized linear mixed model (GLMM) was used. The outcomes included cepstral-based) speech parameters recorded TP1 TP2. For testing subjects, dynamic conducted using samples from posterior distributions parameter estimates random effects. further employed penalized expected deviance (PED) comparison number groups procedure. Results: First, selected metrics individual patients showed significant worsening impairment (p<0.001, Kolmogorov–Smirnov test) Second, cluster divided into distinct subgroups showing strong association between disease duration, ~96% observed values falling 95% credible intervals. Third, training data, our (PEDFea1+Fea2=5175; groups=2) performed reliably univariate models (PEDFea1=4225, PEDFea2=3850; discriminating patients. Fourth, individual-level predictions change profiles over data. Conclusion: Such indeed holds promise to predict rate clinical progression Ataxic Speech individuals CA.