Radiomics-based prediction of early antipsychotic treatment response for first-episode schizophrenia

Abstract

Background

This study aims to develop and validate cranial CT-based radiomics models for predicting early antipsychotic response in first-episode schizophrenia (SCZ).

Methods

We retrospectively enrolled 530 antipsychotic-naïve first-episode SCZ patients and split them 7:3 into training and independent test sets. Early response was defined as ≥25% Brief Psychiatric Rating Scale (BPRS) reduction at week four. CT images were resampled, registered to MNI space, and parcellated using the Harvard–Oxford atlas. Radiomic features (RFs) were extracted and selected via a multi-stage pipeline (ANOVA, mutual information, LASSO–SVM, and interaction screening). An XGBoost classifier was trained to generate radiomics-only predictions. A radiomics score (Rad-Score) was derived by logistic regression and combined with clinical predictors in a nomogram. Spearman's correlations between RFs and symptoms were tested with FDR control.

Results

Among the RFs of 530 patients (313 responders and 217 non-responders), we retained 11 radiomic features and one interaction term, predominantly from temporal regions and the thalamus. In the test set, AUCs were 0.72 for the clinical-only model and 0.86 for the radiomics-only model; the nomogram integrating Rad-Score with age, duration of untreated psychosis, marital status, and education achieved AUC = 0.92. Sensitivity analyses using ≥30% and ≥40% response thresholds showed consistent discrimination. Rad-Score correlated with symptom improvement (rs = 0.504; q < 0.001).

Conclusion

Baseline cranial CT radiomics, particularly when combined with routine clinical factors, enables accurate early prediction of antipsychotic response and yields interpretable biomarkers linked to symptom change in first-episode schizophrenia.