Abdominal CT-based body composition radiomics model for predicting liver injury in acute paraquat poisoning

Acute paraquat (PQ) poisoning frequently results in fatal liver injury, yet efficient non-invasive modalities for early prediction remain lacking. To evaluate the utility of CT-based radiomics models in predicting liver injury following PQ poisoning. A retrospective analysis was conducted on 152 patients with poisoning. Based on diagnostic criteria and the clinical trajectory of liver injury, patients were classified into early-onset and late-onset liver injury groups. Radiomic features of muscle, subcutaneous adipose tissue (SAT), and visceral adipose tissue (VAT) at the L3 vertebral level were extracted using LIFEx software. Following dimensionality reduction, ten machine learning models were constructed to generate radiomics scores (Rad-scores). Independent risk factors identified via multivariable logistic regression were integrated with the Rad-score to develop a clinical-radiomics nomogram. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis. In the early liver injury group, the highest AUC for muscle features was 0.719. Multivariable logistic regression identified a semi-quantitative level 4 of urinary PQ as an independent risk factor (OR = 43.57, 95%CI:13.25–200.91, P < 0.001). The resulting nomogram, integrating the radiomic score and this clinical factor, achieved an AUC of 0.910 in the training set and 0.943 (95%CI:0.879–1.000) in the testing set, with good calibration (Brier = 0.127, H-L P = 0.731) and clinical net benefit. In the late liver injury group, the highest AUC for muscle features was 0.843, with no clear independent clinical indicators identified. Abdominal CT-based body composition radiomics models demonstrated substantial predictive value for liver injury in PQ poisoning, with muscle features yielding the highest efficacy. The nomogram incorporating the muscle radiomics score and a semi-quantitative level 4 of urinary PQ effectively predicted early liver injury. However, validation in the late liver injury group is required due to limited sample size. Not applicable.