In a recent study published in Molecular Psychiatryresearchers performed structural magnetic resonance imaging (sMRI) to develop a machine learning classifier and distinguish neuroanatomical patterns between healthy controls (HCs) and those developing psychotic illness (CHR-PS+).
Study: Using structural brain neuroimaging measures to predict psychosis onset for clinically high-risk individuals. Image credit: Nomad_Soul/Shutterstock.com
Record
Structural MRI is used to diagnose disease, although its ability to identify psychosis is unclear. The clinical high-risk (CHR) paradigm helps in early diagnosis and prevention of psychotic disorder.
Clinically high-risk individuals are more likely to develop psychosis than healthy controls. However, the majority do not transition or have reduced symptoms.
CHR status correlates with changes in brain anatomy, including gray matter volume, cortical surface area, and cortical thickness. Cross-sectional MRI results reveal that those in CHR had lower CT.
About the study
In the present study, the researchers constructed a machine learning model to differentiate CHR-PS+ individuals from HCs.
They also investigated whether the model could differentiate CHR-PS+ patients from those who did not show signs of psychosis (CHR-PS-) or from those with an unknown status at follow-up (CHR-UNK).
Researchers collected T1-weighted sMRI brain images from 1,029 HC and 1,165 CHR subjects at 21 sites of the ENIGMA CHR for Psychosis Working Group. They used the Desikan-Killiany atlas and the ENIGMA quality assessment pipeline to extract structural data from 153 sites of interest.
They identified clinically high-risk status using the Structured Interview for Prodromal Syndromes (SIPS) and the Comprehensive Assessment of At-Risk Mental States (CAARMS).
The team used ComBat tools to standardize measurements of cortical thickness, surface area and subcortical volume.
They used measurements of cortical surface area, cortical thickness, intracranial volume, and subcortical volume to predict potential psychosis conversion. They included age, sex, procedure and side effects as variables.
The researchers applied generalized additive models (GAMs) to HC data, generated non-linear sMRI features corrected for age and biological sex, and regressed intracranial volume effects.
They created an XGBoost classifier that uses CHR-PS+ and HC data to detect aberrations in neuroanatomical developmental patterns. They assessed the model’s predictive ability using the remaining site data.
The researchers evaluated the model’s performance in two steps, dividing the information into training type, test, independent group, and independent-validation information.
They conducted external validation using the trial type and independent confirmatory data sets, while the independent panel data set identified individuals with CHR-UNK and CHR-PS- status at all sites.
The team trained the final classifier model using the optimal hyperparameters and training data and evaluated the predictive ability of the machine learning classifier against independent group data.
They performed four decision curve comparisons and evaluated the classifier on four different feature sets: cortical thickness, surface area, subcortical volumes only, and all features.
They used the model showing the best predictive performance using the independent confirmatory data for further analysis.
Results
The team found that peripheral cortical surface significantly influenced the categorization of CHR-PS+ individuals by HC. Individuals in the CHR-UNK and CHR-PS- categories were more likely to be identified as HC.
A nonlinearly fitted SA feature classifier outperformed the CHR-PS+ and HC clusters. The model achieved 85% accuracy using the training data. The team achieved the best estimate using the test data (68%) and independent confirmatory data (73%).
They determined the best ten feature weights for separating the HC from CHR-PS+ groups, which included the insula, superior frontal, superior temporal, superior parietal, isthmus gyrus, fusiform, metacentral gyrus, and parahippocampus.
Individuals with more clinical symptoms had lower cortical areas in the anterior cingulate, lateral prefrontal and medial prefrontal regions, and parahippocampal gyrus.
Machine learning-based classifiers trained on 152 MRI brain structural features performed worse in confirmatory analysis than sex- and age-matched classifiers.
The researchers also attempted to distinguish clinically high-risk individuals from healthy controls and individuals in the CHR-PS+ category from those in the CHR-PS- category, but achieved only 50% accuracy.
Statistically significant differences in categorized labels were noted, with healthy controls showing an increased likelihood of being classified as controls compared to CHR-PS+ subjects (73% vs. 30%). Independent group data showed no differences between the CHR-UNK and CHR-PS- groups.
The study revealed significant differences in categorized labels and predicted probabilities in four groups of CHR-PS+ patients.
CHR-PS+ subjects differed from those of the other groups, while CHR-PS- subjects fell between these healthy control and CHR-PS+ groups.
Although the predicted probability varied by age and group, the team did not observe statistically significant interactions between age groups. The decision curves showed that obtaining a prediction from the current classifier resulted in a higher net benefit for the CHR Discoverer transition.
conclusion
Overall, the study findings indicated that sMRI scans could help determine the prognosis of individuals with CHR and discriminate between CHR-PS+ individuals and healthy controls.
The model achieved 85% accuracy in two-class categorization with non-linear fitting of cortical surface variables for gender and age.
Neuroanatomical changes helped identify individuals in the CHR-PS+ group. Superior temporal, insular, and frontal regions contributed most to distinguishing CHR-PS+ from HCs.