Unsupervised Machine Learning Approach to Discover Subtypes of Progressive Supranuclear Palsy

Progressive supranuclear palsy (PSP) is a rare, rapidly progressive, neurodegenerative disease characterised by ocular motor abnormalities, postural instability, parkinsonism and cognitive dysfunctions. Recently, eight PSP phenotypes have been identified, on the basis of expert opinion. However, a number of objective measures could be used to identify distinct PSP subgroups. Here, we used first order radiomic features through a clustering approach to identify two PSP subgroups. Subsequently, we performed a clinical characterization of such clusters using the PSP rating scale subscores. Forty patients were enrolled in the study and each of them underwent a T1-weighted magnetic resonance imaging (MRI). Eighteen first order radiomic features were extracted from cortical (cingulate and occipital) and subcortical (pallidum, putamen, hippocampus and amygdala) regions of interest (ROIs). The MRI was pre-processed before features extraction. An unsupervised machine learning (ML) algorithm, k-means, was implemented to cluster data into two groups based on radiomic features and a subsequent feature importance was employed. The algorithm found two clusters (Cluster 1 and Cluster 2) in differentiating subgroups of PSP. A univariate statistical analysis was applied on clinical and demographical variables after the identification of the clusters and statistical significant differences were obtained for bulbar (p=0.034, p=0.047, p=0.048, p=0.024, respectively for amygdala, cingulate white matter, hippocampus and occipital white matter ROIs) and gait and midline PSP-rs subscores (p=0.028, p=0.027, respectively for pallidum and putamen ROIs). Amygdala, hippocampus, cingulate and occipital white matter ROIs showed a higher disease severity for Cluster 1, vice versa for pallidum and putamen ROIs. These results suggest that first-order radiomic features reflecting the distribution of grey-level intensities in an ROI could be an add-on tool to clinical rating scales for the identification of PSP subgroups.