Utilizing Structural MRI and Unsupervised Clustering to Differentiate Schizophrenia and Alzheimer's Disease in Late-onset Psychosis

Abstract

Late-onset psychosis (LOP) represents a highly heterogeneous and understudied condition, with potential origins ranging from atypically late onset of schizophrenia (SCZ) to Alzheimer's Disease (AD). Despite the clinical necessity of differentiating these conditions to guide effective treatment, achieving an accurate diagnosis remains challenging. This study aimed to utilize data-driven analyses of structural magnetic resonance imaging (MRI) to distinguish between these diagnostic possibilities. Utilizing publicly available datasets of MRI scans from 699 healthy control (HC) participants and 469 patients diagnosed with SCZ or AD, our analysis focused on bilateral subcortical volumetric measures in the caudate, hippocampus, putamen, and amygdala. We first trained an unsupervised K-means clustering algorithm based on SCZ and AD patients and achieved a clustering accuracy of 81% and an area under curvature (AUC) of 0.79 in distinguishing between these two groups. Subsequently, we calculated the Euclidean distance between the AD and SCZ cluster centroids for each of ten patients with unexplained onset of psychosis after age 45 from a clinical MRI registry. Six patients were classified as AD and four as SCZ. Our findings revealed that among LOP participants, those classified in the SCZ cluster exhibited significantly greater right putamen volumes compared to those in the AD cluster (p<0.0025). There were also intriguing clinical differences. While we do not have diagnostic biomarker information to confirm these classifications, this study sheds light on the heterogeneity of psychoses in late life and illustrates the potential use of widely available structural MRI and data-driven methods to enhance diagnostic accuracy and treatment outcomes for LOP patients.