Machine learning integrating high-resolution T1-weighted MR radiomics of subcortical structures with blood biomarkers to identify depressed patients with a history of suicide attempts.

This study aimed to develop a machine learning (ML) model combining radiomic signatures from subcortical structures on T1-weighted MR images with peripheral blood biomarkers to distinguish depressed patients with versus without a history of suicide attempts. We enrolled 164 depressed patients, including 60 patients with a history of suicide attempts (DP-SAs) and 104 patients without a history of suicide attempts (DP-NAs), of which 41 DP-SAs and 47 DP-NAs were measured for peripheral inflammatory factors and kynurenine (KYN) pathway metabolites. Radiomic features from subcortical structures were extracted, and a combination of four feature selection methods and four ML algorithms was used to identify the most robust radiomic model for distinguishing the DP-SAs from the DP-NAs. The radiomic scores obtained from the optimal model and statistically significant clinical characteristics in patients with blood data were incorporated into the nomogram. The performance of the nomogram was assessed by a calibration curve and decision curve analysis (DCA). Compared with the other subcortical regions, the right caudate nucleus and bilateral thalamus achieved higher areas under the curve (AUCs) (0.793, 0.762 and 0.693, respectively) for distinguishing DP-SAs from DP-NAs. The model combining multiple subcortical regions demonstrated improved performance, with AUCs of 0.803 and 0.821 in the cross-validation and test cohorts, respectively. Peripheral blood kynurenic acid (KYNA) and KYNA/quinolinic acid (QA) were identified as independent clinical predictors. Owing to the collinearity between KYNA and KYNA/QA, we constructed clinical-radiomic nomograms that integrate radiomic scores with KYNA or KYNA/QA, with C-indexes of 0.919 and 0.914, respectively. Calibration curve analysis and DCA revealed that the nomogram had good reliability and net benefit. The nomogram integrating subcortical radiomic features with blood biomarkers exhibits excellent discriminative performance, providing an exploratory biomarker and preliminary proof-of-concept modelling to distinguish depressed patients with versus without a history of suicide attempts.