Machine Learning to Disentangle Neurobiological Heterogeneity in the Schizophrenia Spectrum Disease Course

About This Grant

PROJECT SUMMARY Schizophrenia is a neuropsychiatric disorder which lacks clinically actionable biomarkers to guide diagnosis, treatment, and prognostication. There are, however, well-replicated neurobiological findings in schizophrenia – namely, decreased hippocampal volume, widespread cortical thinning, and ventricular enlargement. Decades of work from our lab and others have established the hippocampus as central to the pathophysiology of schizo- phrenia, with structural and functional deficits associated with illness severity and cognitive dysfunction. Despite this, the clinical utility of hippocampal imaging remains limited, largely due to unresolved questions about the origin, trajectory, and heterogeneity of hippocampal pathology. Recent advances in machine learning have enabled innovative approaches to address these core questions and disentangle the known individual- level heterogeneity in schizophrenia, revealing potential disease subtypes with distinct progression patterns. This work both identifies several possible epicenters of disease (including the hippocampus) and suggests a network-based progression model via which pathology spreads along the brain’s white matter tracts (i.e., the connectome). Notably, our work suggests an important structural and functional differentiation between the anterior and posterior hippocampus, implicating the former as crucial during the early stages of psychosis. Given this context, I hypothesize that schizophrenia is not a unitary disease, but comprises multiple subtypes with distinct spatiotemporal patterns of gray matter degeneration, including a subtype in which pathology originates in the anterior hippocampus and propagates through its structural connectome. To test this hypo- thesis, we will integrate advanced machine learning with longitudinal modeling using MR imaging, uniquely leveraging our lab’s rich data and technical expertise. In Aim 1, I will characterize group-level patterns of gray matter volume (GMV) change over the first decade of psychotic illness, stratified by clinical trajectory, using longitudinal structural modeling. In Aim 2, I will apply a cutting-edge machine learning algorithm – Subtype and Stage Inference (SuStaIn) – to a large, cross-sectional discovery sample to identify latent disease subtypes based on inferred patterns of GMV progression. In Aim 3, I will evaluate the external validity of these subtypes using our in-house longitudinal cohort of psychosis patients (see Aim 1). This will be the first study to utilize SuStaIn to specifically probe the differential roles of the anterior and posterior hippocampus in the patho- physiology of psychosis and will directly test the hypothesis that disease emerges and propagates along structural brain networks. By integrating machine learning with longitudinal neuroimaging, these experiments will identify and validate biologically grounded subtypes of schizophrenia, probing key features of hippocampal pathology and offering insight into hippocampal dynamics across illness stages. This innovative approach addresses a longstanding barrier in schizophrenia research – its vast clinical and neurobiological heterogeneity – and has potential to transform diagnosis, monitoring, and treatment following a first episode of psychosis.