Automated Pipeline for Prediction of Spontaneous Paediatric Intracerebral Haemorrhage Outcome Based on CT Images

Abstract Purpose This study aimed to develop a machine learning model for predicting the outcome of spontaneous paediatric intracerebral haemorrhage (sICH) by leveraging automatically segmented volumes of total brain (TB), intraparenchymal haemorrhage (IPH), intraventricular haemorrhage (IVH), and perihaematomal oedema (PHE) from cranial CT images in conjunction with clinical data. Methods A retrospective, bi-center study of sICH patients was conducted. The Fudan University Children's Hospital (FUCH) cohort was split 7:3 for training/internal testing, while the Zhejiang University Children’s Hospital (ZUCH) cohort served for external validation. Automated and manual segmentation extracted regions of interest (TB, IPH, IVH, PHE), from which quantitative hematoma volume parameters (e.g., IPH/total brain volume ratio) were derived. Three models were developed: a hematoma location/complication (HLC) model, a hematoma quantification (HQ) model, and a combined model integrating both feature sets. Classifiers (logistic regression, support vector machine, decision tree) were trained with 10-fold cross-validation. Performance was evaluated via receiver operating characteristic (ROC) and decision curve analysis (DCA) on both test sets; SHapley Additive exPlanations (SHAP) interpreted the combined model. Results Among 147 patients (median age 6.6 years ± 4.28 [SD]; 84 (57.1%) males), 49 (33.3%) had poor outcomes (mRS ≥ 3). Automatic segmentation showed high consistency with radiologists (Dice: IPH 0.837, IVH 0.745; Lin coefficient > 0.8). The combined model outperformed the HLC and HQ models, achieving AUCs of 0.798 (internal) and 0.822 (external). DCA indicated greater clinical net benefit for the combined model across thresholds of 0.10–0.59. The pipeline completed automated segmentation and volume analysis per patient in under 17 seconds. Conclusion The automated pipeline for spontaneous intracerebral haemorrhage based on CT images can objectively and accurately predict the outcome of intracerebral haemorrhage in children.