Machine Learning Classification of Schizophrenia Using Neuroanatomical and Functional Markers to Support Individualized Treatment: A Literature Review

Abstract

Introduction: Heterogeneity in the clinical presentation and pathophysiology of schizophrenia presents challenges for effective diagnostic and therapeutic practices. Machine learning (ML) methods have emerged as a promising tool for addressing this, due to their ability to integrate neuroimaging and clinical data. This paper aims to review the literature on ML and its uses for identifying subtypes based on structural and functional brain features.  Additionally, the paper will discuss how these insights can be used to target pathophysiological features and individualize treatment. 

Methods: A literature search was conducted across various research databases for ML neuroimaging studies published in the past 10 years that included only schizophrenia samples. From this search, 38 were screened; 18 met criteria for full-text review and were included. Additionally, a secondary search was conducted to review potential clinical applications and individualized treatments. From this search, 6 articles were added, resulting in a total of 24 papers. 

Results: Evidence supported the existence of distinct neurobiological profiles amongst patients with schizophrenia. Across 7 studies, two recurring neurobiological subtypes were identified: one defined by widespread cortical gray matter loss and greater cognitive and negative symptoms; and another marked by an intact cortical structure but subcortical enlargement, associated with positive symptoms. Results across the remaining 11 studies varied, ranging from 2 to 6 subgroups. Several of these neurobiological patterns corresponded to differences in symptom severity and cognitive performance. Further, predictive-modelling studies demonstrated moderate to high accuracy in forecasting treatment response.  

Discussion: ML can be used to identify neurobiological subtypes of schizophrenia that align with differences in symptoms and potential treatment targets. Additional multimodal biomarkers further highlight substantial heterogeneity. While multimodal ML approaches may help integrate this variability, current methods face challenges. These findings suggest that ML can capture variability in schizophrenia, providing a basis for predicting treatment responses. 

Conclusion: ML can be used to identify neurobiological subtypes of schizophrenia and additional multimodal biomarkers which further highlight substantial heterogeneity and treatment targets. Hence, ML approaches may help integrate this variability, but current methods face challenges which require further research.