<p>Huntington’s disease and other fatal neurodegenerative disorders have a significant global impact, affecting millions and often leading to dementia or death. Dementia is a general term with multiple symptoms that reduce a person’s cognitive abilities. The common symptoms of dementia include memory loss, difficulty concentrating, mood swings, and a confused state. This paper aims to improve dementia severity class detection by managing two key phases. The first phase involves analyzing the significant predictors from two benchmark versions of the dataset. The second phase focuses on enhancing the classification accuracy using machine learning techniques. Our novel approach involves two methods. The first method is the selection of clinically and anatomically significant predictors out of the total predictors from the benchmark dataset using the Z-distributional test. The second method governs a joint conditional estimate-based distributional random forest, which is used to classify the different severity levels of dementia. The evaluated results indicate that the resultant dataset clinically and anatomically significant predictors are more closely aligned with the distribution test line for each dementia class. These significant predictors with high correlations are well visualized in the quantile-quantile plot and are considered to have greater significance for detecting dementia severity classes. The proposed approach effectively detects heterogeneity in multivariate data, achieving an accuracy of 93.20%. The performance of other measures is improved in terms of precision, recall, and F-measures, making it more robust than existing methods. This methodology enhances early dementia severity detection and effectively attains multivariate data heterogeneity.</p>