Radiomic texture features for classifying individuals at clinical high risk for psychosis

Background: Structural alterations in the brain have been associated with the risk of psychosis, nevertheless, the clinical criteria for identifying patients at clinical high risk (CHR) are insufficiently specific[1]. Radiomic texture features can identify microstructural variations in MRI data by assessing spatial intensity patterns and can serve as valuable biomarkers[2],[3].

Objective: This study assessed the potential of texture-based radiomic features to differentiate CHR from healthy controls (HC) through machine learning (ML) using the data from Personalised Prognostic Tools for Early Psychosis Management (PRONIA) study.

Methods: Texture features were extracted from MRI scans using the Gray-Level Co-occurrence Matrix (GLCM), encompassing measures such as difference entropy, and contrast. Difference entropy quantifies the randomness in intensity disparities between neighboring voxels, whereas contrast captures local intensity shifts, highlighting differences between adjacent voxel values[2],[3]. The analysis included a sample of n = 406 participants, consisting of 271 HC and 135 CHR.

First, a sensitivity analysis was undertaken to determine whether radiomic features could predict potential confounding variables such as age, sex, total intracranial volume (TIV), and scanner site.

Several ML algorithms were assessed to discriminate CHR from HC, including support vector machines, random forests, and XGBoost. XGBoost demonstrated the most robust performance among these options. Model training and evaluation were performed using a repeated nested cross-validation (20 outer folds with 10 repetitions, 10-fold stratified inner loops), guaranteeing unbiased feature selection and generalization.

Feature selection, a critical step before model training, seeks to eliminate non-informative voxels and reduce dimensionality. This process involved benchmarking various strategies: comparing variance thresholds to eliminate low-variance voxels among subjects, implementing ANOVA filtering, and examining conventional dimensionality reduction techniques like Principal Component Analysis and Singular Value Decomposition.

Results: Texture features shown significant predictive capability for scanner site (e.g., difference of entropy: balanced accuracy = 0.942) and sex (e.g., contrast: 0.858). Total intracranial volume (TIV) exhibited moderate predictability (contrast: R² = 0.241), whereas age shown weak predictability (Contrast R² = 0.169). Consequent to these findings, scanner site, sex, and TIV were incorporated as covariates in the CHR and HC classification models.

Of the evaluated features, contrast demonstrated the highest classification performance, attaining a accuracy of 65.25%. A comparable outcome was achieved using the difference of entropy (accuracy = 65.21%). However, both models demonstrated an imbalanced performance profile, characterized by low sensitivity and high specificity. This suggests that the classifiers were more effective in accurately identifying HC (true negatives) while often misclassifying CHR as healthy (false negatives).

Conclusions: The average classification performance, particularly the low sensitivity in identifying CHR individuals, underscores the ongoing challenge of CHR detection via MRI modalities. Previous research indicates that the classification of CHR against HC seldom surpasses 70% balanced accuracy, with several models performing near chance in external validation[4]. This can be explained by the heterogeneity of CHR group, encompassing varied symptom severity, outcomes and trajectories[4],[3]. Future research should focus on identifying more homogeneous CHR subgroups through grouping according to symptom severity or clinical outcomes, rather than a binary classification[3]. In addition, mapping radiomic features to brain region can help with biomarker identification.