Abstract
There is a growing interest in automatic classification of mental disorders based on neuroimaging data. Small training data sets (subjects) and very large amount of high dimensional data make it a challenging task to design robust and accurate classifiers for heterogeneous disorders such as schizophrenia. Most previous studies considered structural MRI, diffusion tensor imaging and task-based fMRI for this purpose. However, resting-state data has been rarely used in discrimination of schizophrenia patients from healthy controls. Resting data are of great interest, since they are relatively easy to collect, and not confounded by behavioral performance on a task. Several linear and non-linear classification methods were trained using a training dataset and evaluate with a separate testing dataset. Results show that classification with high accuracy is achievable using simple non-linear discriminative methods such as k-nearest neighbors (KNNs) which is very promising. We compare and report detailed results of each classifier as well as statistical analysis and evaluation of each single feature. To our knowledge our effects represent the first use of resting-state functional network connectivity (FNC) features to classify schizophrenia.
Introduction
Schizophrenia is a complex, heterogeneous mental disorder that lacks standard biologically-based clinical tests, making differential diagnosis a significant challenge. While neuroimaging has long been used to study the disorder, most classification research has focused on structural MRI or task-based fMRI. Task-based fMRI can be problematic because results are often confounded by a patient's ability to perform the specific task. Resting-state fMRI (rs-fMRI) is a promising alternative—it’s easier to collect and reflects the brain's intrinsic functional architecture. However, the "curse of dimensionality" (having a massive amount of high-dimensional data but very few subjects) makes designing robust classifiers difficult. This paper tackles these issues by using Functional Network Connectivity (FNC) to reduce data dimensionality while retaining highly discriminative information.
Methods
The researchers developed a pipeline that moves from raw imaging data to a final diagnostic classification. The study utilized a dataset of 56 participants: 28 healthy controls (HC) and 28 patients diagnosed with schizophrenia (SZ). The data was divided into a training set (16 HC + 16 SZ) and a testing set (12 HC + 12 SZ).
The methodology involved three primary stages:
- Group ICA: Independent Component Analysis (ICA) was used to decompose the fMRI data into 20 spatially independent components. From these, 9 non-artifactual "brain networks" were selected, including auditory, visual, motor, and default-mode networks.
- FNC Extraction: Instead of looking at individual voxels, the team calculated the maximum lagged correlation between the time-courses of these 9 networks. This resulted in 36 unique FNC features per subject. Allowing for "lag" accounts for variations in hemodynamic response shapes between different brain regions and individuals.
- Classification: 12 different linear and non-linear classifiers were compared, including Linear Discriminant Classifier (LDC), Support Vector Machines (SVM), k-Nearest Neighbors (KNN), and Artificial Neural Networks (ANN).
Results
The study found that the interaction between brain networks over time (FNC) is more informative for classification than the static spatial maps of the networks themselves. Non-linear classifiers significantly outperformed linear ones, indicating that the boundary between healthy and schizophrenic brain connectivity is complex and cannot be defined by a simple linear hyperplane.
KNN (k=1), RBF-SVM, and Polynomial SVM all achieved a peak overall accuracy of 96% on the testing dataset. These top-performing models showed 100% sensitivity (correctly identifying all patients) and 92% specificity. In contrast, the best linear method, Linear SVM, reached only 83% accuracy.
Statistical analysis revealed that schizophrenia patients showed significantly lower connectivity between nodes of the Default Mode Network (DMN) compared to healthy controls. Furthermore, decision tree analysis highlighted that specific thresholds in temporal-visual and temporal-motor connectivity were highly predictive of a schizophrenia diagnosis. To ensure results weren't just picking up on medication effects, the researchers re-ran the analysis excluding motor-related features; the high accuracy persisted, suggesting the FNC abnormalities are intrinsic to the disorder.
Discussion
This study represents the first use of resting-state FNC features to automatically classify schizophrenia patients. The primary limitation is the small sample size (56 subjects). While the results are statistically significant, machine learning models generally require much larger cohorts to ensure generalizability. Future work should involve applying this FNC-based approach to larger, multi-site datasets and exploring longitudinal data to see if these connectivity biomarkers change as the disease progresses or responds to treatment. Additionally, a study on "first-episode, drug-naive" patients would be the gold standard to completely rule out pharmacological confounds.
Summary updated