AUTHOR=Wang Meng , Hu Ke , Fan Lingzhong , Yan Hao , Li Peng , Jiang Tianzi , Liu Bing 

TITLE=Predicting Treatment Response in Schizophrenia With Magnetic Resonance Imaging and Polygenic Risk Score

JOURNAL=Frontiers in Genetics

VOLUME=13

YEAR=2022

URL=https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2022.848205

DOI=10.3389/fgene.2022.848205

ISSN=1664-8021

ABSTRACT=<p><bold>Background:</bold> Prior studies have separately demonstrated that magnetic resonance imaging (MRI) and schizophrenia polygenic risk score (PRS) are predictive of antipsychotic medication treatment outcomes in schizophrenia. However, it remains unclear whether MRI combined with PRS can provide superior prognostic performance. Besides, the relative importance of these measures in predictions is not investigated.</p><p><bold>Methods:</bold> We collected 57 patients with schizophrenia, all of which had baseline MRI and genotype data. All these patients received approximately 6 weeks of antipsychotic medication treatment. Psychotic symptom severity was assessed using the Positive and Negative Syndrome Scale (PANSS) at baseline and follow-up. We divided these patients into responders (<italic>N</italic> = 20) or non-responders (<italic>N</italic> = 37) based on whether their percentages of PANSS total reduction were above or below 50%. Nine categories of MRI measures and PRSs with 145 different <italic>p</italic>-value thresholding ranges were calculated. We trained machine learning classifiers with these baseline predictors to identify whether a patient was a responder or non-responder.</p><p><bold>Results:</bold> The extreme gradient boosting (XGBoost) technique was applied to build binary classifiers. Using a leave-one-out cross-validation scheme, we achieved an accuracy of 86% with all MRI and PRS features. Other metrics were also estimated, including sensitivity (85%), specificity (86%), F1-score (81%), and area under the receiver operating characteristic curve (0.86). We found excluding a single feature category of gray matter volume (GMV), amplitude of low-frequency fluctuation (ALFF), and surface curvature could lead to a maximum accuracy drop of 10.5%. These three categories contributed more than half of the top 10 important features. Besides, removing PRS features caused a modest accuracy drop (8.8%), which was not the least decrease (1.8%) among all feature categories.</p><p><bold>Conclusions:</bold> Our classifier using both MRI and PRS features was stable and not biased to predicting either responder or non-responder. Combining with MRI measures, PRS could provide certain extra predictive power of antipsychotic medication treatment outcomes in schizophrenia. PRS exhibited medium importance in predictions, lower than GMV, ALFF, and surface curvature, but higher than measures of cortical thickness, cortical volume, and surface sulcal depth. Our findings inform the contributions of PRS in predictions of treatment outcomes in schizophrenia.</p>