Exploring the Predictive Value of Structural Covariance Networks for the Diagnosis of Schizophrenia

1424 

Abstract Submission 

Clara Vetter^1,2,3,4, Annika Bender⁵, Dominic Dwyer⁶, Anne Ruef⁷, Lana Kambeitz-Ilankovic⁸, Linda Antonucci⁹, Stephan Ruhrmann⁸, Joseph Kambeitz¹⁰, Anita Riecher-Rössler¹¹, Rachel Upthegrove¹², Raimo K. Salokangas¹³, Jarmo Hietala¹³, Christos Pantelis¹⁴, Rebecca Lencer¹⁵, Eva Meisenzahl¹⁶, Stephen J. Wood¹⁷, Paolo Brambilla¹⁸, Stefan Borgwardt¹⁹, Peter Falkai⁷, Alessandro Bertolino²⁰, Daniel Rückert²¹, Nikolaos Koutsouleris^7,22,23

¹Max Planck Institute of Psychiatry, Munich, Germany, ²Department of Psychiatry and Psychotherapy, Ludwig-Maximilian University, Munich, Germany, ³Munich Center of Machine Learning (MCML), Munich, Germany, ⁴Helmholtz Association - Munich School for Data Science (MUDS), Munich, Germany, ⁵Department of Psychiatry and Psychotherapy, Ludwig Maximilian University, Munich, Bavaria, ⁶Orygen, the National Centre of Excellence for Youth Mental Health, Melbourne, Victoria, ⁷Department of Psychiatry and Psychotherapy, LMU University Hospital, LMU Munich, Munich, Germany, ⁸University of Cologne, Cologne, Germany, ⁹University of Bari Aldo Moro, Milan, Italy, ¹⁰Department of Psychiatry and Psychotherapy, University of Cologne, Cologne, Germany, ¹¹Faculty of Medicine, University of Basel, Basel, Basel, ¹²Institute of Mental Health, University of Birmingham,, Birmingham, United Kingdom, ¹³Department of Psychiatry, University of Turku, Turku, Finland, ¹⁴Melbourne Neuropsychiatry Centre, Carlton, Victoria, ¹⁵Institute for Translational Psychiatry, University Münster, Münster, Germany, ¹⁶Department of Psychiatry and Psychotherapy, Medical Faculty, Heinrich Heine University, Düsseldorf, Germany, ¹⁷Centre for Youth Mental Health, University of Melbourne, Melbourne, Australia, ¹⁸Department of Neurosciences and Mental Health, Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policli, Milan, Italy, ¹⁹Department of Psychiatry and Psychotherapy, University of Lübeck, Lübeck, Germany, ²⁰Department of Basic Medical Science, Neuroscience and Sense Organs, University of Bari Aldo Moro, Bari, Bari, ²¹Technical University Munich, Munich, Germany, ²²Max Planck Institute for Psychiatry, Munich, Germany, ²³Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King's College London, London, United Kingdom

Clara Vetter  
Max Planck Institute of Psychiatry|Department of Psychiatry and Psychotherapy, Ludwig-Maximilian University|Munich Center of Machine Learning (MCML)|Helmholtz Association - Munich School for Data Science (MUDS)
Munich, Germany|Munich, Germany|Munich, Germany|Munich, Germany

Annika Bender  
Department of Psychiatry and Psychotherapy, Ludwig Maximilian University
Munich, Bavaria

Dominic Dwyer  
Orygen, the National Centre of Excellence for Youth Mental Health
Melbourne, Victoria

Anne Ruef  
Department of Psychiatry and Psychotherapy, LMU University Hospital, LMU Munich
Munich, Germany

Lana Kambeitz-Ilankovic  
University of Cologne
Cologne, Germany

Linda Antonucci  
University of Bari Aldo Moro
Milan, Italy

Stephan Ruhrmann  
University of Cologne
Cologne, Germany

Joseph Kambeitz  
Department of Psychiatry and Psychotherapy, University of Cologne
Cologne, Germany

Anita Riecher-Rössler  
Faculty of Medicine, University of Basel
Basel, Basel

Rachel Upthegrove  
Institute of Mental Health, University of Birmingham,
Birmingham, United Kingdom

Raimo K. Salokangas  
Department of Psychiatry, University of Turku
Turku, Finland

Jarmo Hietala  
Department of Psychiatry, University of Turku
Turku, Finland

Christos Pantelis  
Melbourne Neuropsychiatry Centre
Carlton, Victoria

Rebecca Lencer  
Institute for Translational Psychiatry, University Münster
Münster, Germany

Eva Meisenzahl  
Department of Psychiatry and Psychotherapy, Medical Faculty, Heinrich Heine University
Düsseldorf, Germany

Stephen J. Wood  
Centre for Youth Mental Health, University of Melbourne
Melbourne, Australia

Paolo Brambilla  
Department of Neurosciences and Mental Health, Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policli
Milan, Italy

Stefan Borgwardt  
Department of Psychiatry and Psychotherapy, University of Lübeck
Lübeck, Germany

Peter Falkai  
Department of Psychiatry and Psychotherapy, LMU University Hospital, LMU Munich
Munich, Germany

Alessandro Bertolino  
Department of Basic Medical Science, Neuroscience and Sense Organs, University of Bari Aldo Moro
Bari, Bari

Daniel Rückert  
Technical University Munich
Munich, Germany

Nikolaos Koutsouleris  
Department of Psychiatry and Psychotherapy, LMU University Hospital, LMU Munich|Max Planck Institute for Psychiatry|Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King's College London
Munich, Germany|Munich, Germany|London, United Kingdom

Structural Covariance Networks (SCNs), reflecting coordinated neurodevelopmental processes, offer insights into brain reorganization associated with aging and diseases such as schizophrenia. Despite group-level alterations in SCNs among patients, individual-level analysis has been hindered by a lack of established methods. This study addresses this gap by comparing two individual SCN computation methods using structural MRI-derived gray matter volumes (GMV) for the classification of schizophrenia.

In this study, we investigated the predictive value of two single-subject SCN computation methods derived from regional gray matter volumes (GMV) measured by structural MRI for classifying patients with schizophrenia within a sample comprising patients and healthy controls (N_PAT = 154, N_HC = 366).

The first SCN method leveraged a reference sample (N = 627) and quantified a single subject's contribution to the reference group's SCN. The second approach defined a subject's SCN from the single image using a symmetric version of KL divergence.

To assess the additional predictive value of SCNs compared to regional GMV, we employed a stepwise analysis using linear support vector machines within a nested cross-validation framework. Initially, each modality (the two SCNs and regional GMV) was separately analyzed. Subsequently, we evaluated their complementary predictive value through stacked generalization.

To address various model design choices, we systematically varied the granularity of the cortical parcellation atlas used (100 vs. 200 parcels), the feature dimensionality reduction technique employed (LASSO-regularization vs. principal component analysis (PCA)), and, for the two SCN modalities, the type of network features used (pairwise structural covariance, i.e., SCN edges, vs. network summary metrics).

This resulted in a total of 28 models, all externally validated in an independent sample (N_PAT = 71, N_HC =74). For the best-performing model in each modality, global model explainability analyses were conducted to identify the most contributing features. Additionally, derived risk scores were analyzed for their differential relationships with clinical variables.

Machine learning models trained separately on individual SCNs and regional GMV demonstrated consistent classification capability for distinguishing patients with schizophrenia from healthy controls, regardless of cortical parcellations and dimensionality reduction techniques.

Among the unimodal models, the LASSO-regularized model trained on the edges of reference-sample-SCNs, computed using the 200-parcel cortical atlas, achieved the highest balanced accuracy (BAC) in the discovery sample (67.03%). Notably, this model outperformed all other unimodal SCN-based models but did not surpass the performance of LASSO-regularized regional GMV models.

Decisions in regional GMV models were driven by the somatomotor network, default mode, frontoparietal control, visual, and limbic networks. In SCN-based models, discriminative pairwise structural covariance primarily involved the ventral attention, default mode, frontoparietal control, visual, and somatomotor networks.

Stacked generalization revealed that combining SCN modalities and regional GMV significantly improved model performance compared to models trained on individual modalities alone (BAC = 69.96%). Similarly, the highest external validation performance was observed in the multimodal, stacked model using principal component analysis for dimensionality reduction (BAC = 67.10%).

No associations were found between models' decision scores and the clinical variables assessed.

In conclusion, individual SCNs, whether derived from normative samples or KL-divergence, contribute valuable information for schizophrenia classification beyond regional GMV. However, the study did not establish direct links between the identified structural information and clinical phenotypes.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ²

Classification and Predictive Modeling ¹

Connectivity (eg. functional, effective, structural)

Multivariate Approaches

Cortical Anatomy and Brain Mapping

Computational Neuroscience

Cortex

Machine Learning

Plasticity

Psychiatric Disorders

Schizophrenia

STRUCTURAL MRI

Other - Structural Covariance Networks