Classification of schizophrenia from resting state connectivity: role of reduced dimension and noise

Abstract

INTRODUCTION: Schizophrenia is widely understood as a dysconnection syndrome, including disruption of functional connectivity (FC). Advanced machine learning techniques have been explored such as the use of permutation homology features (Caputi, 2021) and deep learning (Mahmood, 2021), but the improvements over classical methods remain incremental, potentially due to structured noise in fMRI data that can bias connectivity results (Kopal, 2020), and the imbalance between high-dimensional features and limited sample sizes (Poldrack, 2017). Thus, this study evaluates the role of denoising strategies combined with efficient robust classifiers (Rehák Bučková, 2023) using principal component analysis (PCA) and logistic regression, to classify schizophrenia from FC. METHODS: Resting-state fMRI data were collected from 100 first-episode psychosis (FEP) patients and 90 healthy controls. At the time of the MRI scan, the average duration of untreated psychosis in patients was 3.23 months (S.D. = 4.82), average duration of antipsychotic treatment 2.29 months (S.D. = 4.58). Controls were age- and sex-matched but had slightly higher education levels (p < 0.001). 3T MRI scanner (Siemens Magnetom Trio) at the Institute of Clinical and Experimental Medicine in Prague, Czech Republic was used. Functional images were obtained using T2-weighted echo-planar imaging (EPI) with blood oxygenation level-dependent (BOLD) contrast using SENSE imaging. GE-EPIs (TR/TE=2000/30 ms, flip angle=70°) with 35 axial slices acquired continuously in sequential decreasing order covering the entire cerebrum (voxel size=3×3×3 mm, slice dimensions 48x64 voxels). The next 400 functional volumes were used. Three preprocessing pipelines were compared: 1. Raw Data: motion correction and normalization. 2. Moderate Denoising: added regression of six motion parameters and mean signals from white matter and CSF, followed by bandpass filtering, 3. Stringent Denoising: regression of motion parameters, their derivatives, and five PCA components from white matter and CSF, plus linear detrending and filtering, corresponding to the default setting of the CompCor, i.e. component-based noise correction method (Behzadi, 2007) in the CONN toolbox (Whitfield-Gabrieli, 2012). FC was quantified in line with quantitative recommendations (Hlinka, 2011) via Pearson’s correlation; between 90 brain regions (AAL atlas). Dimensionality reduction using PCA was applied to the 4005 FC features, and logistic regression classifiers trained. Performance was evaluated using leave-one-out cross-validation, accuracy assessed across varying numbers of principal components. See (Hlinka, 2024) and (Tomecek, 2024) for details concerning the methods and group comparison results. RESULTS: Classification accuracy differed substantially across pipelines, see Fig. 1 for results visualization. For raw data, accuracy peaked at 74% using 44 components but declined with higher dimensions due to overfitting. In contrast, moderate denoising achieved the highest accuracy of 82% with just 24 components, and stringent denoising reached 81% with 45 components. Importantly, the results were robust across a range of components, with denoised data consistently outperforming raw data. Combining multiple denoising strategies did not further improve accuracy. CONCLUSIONS: Denoising significantly enhances the classification of schizophrenia from resting-state FC, with both moderate and stringent denoising strategies yielding comparable performance. Despite differences in preprocessing, accuracy improved by approximately 10% compared to raw data. Dimensionality reduction via PCA played a critical role. These findings underscore the importance of careful fMRI preprocessing in neuroimaging-based classification tasks. However, the lack of substantial improvement from combining strategies highlights persistent challenges in achieving higher classification accuracy. SUPPORTED BY: ERDF-Project Brain dynamics, No. CZ.02.01.01/00/22_008/0004643.