Development and validation of a machine learning-based model for 90-day prognosis outcome in spontaneous intracerebral hemorrhage patients based on non-contrast computed tomography: a multicenter retrospective observational study.

Background: Spontaneous Intracerebral hemorrhage (sICH) is a disease with high mortality and disability. Non-contrast computed tomography (NCCT) is the most commonly used imaging method in the diagnosis and treatment of sICH. This study aimed to develop a clinically useful prediction model for the short-term prognosis of sICH patients based on NCCT features using a machine learning model. Methods: We retrospectively collected data from sICH patients from four centers in China between January 2021 and June 2024, used data from three centers as training cohort to build the model, and another single center data for external validation. The NCCT imaging features were combined with the basic clinical characteristics of sICH patients as training features for machine learning. We developed and verified the effectiveness of five models: support vector machine (SVM), logistic regression (LR), random forest (RF), eXtreme Gradient Boosting (XGboost) and Light Gradient Boosting Machine (LightGBM). The clinical feature set, NCCT imaging feature set and fusion feature set were modeled separately and externally validated. The performance of machine learning models with different features was comprehensively evaluated using ROC curves, accuracy and other related indicators. The SHapley Additive exPlanations (SHAP) diagram was used to illustrate the importance of variables in the model, and the Sequential Forward Selection (SFS) was used to screen out the core features. Finally, a convenient and practical prognosis prediction platform was built based on the core features. This study is registered with ClinicalTrials.gov (NCT06535438). Findings: A total of 1091 sICH patients from three centers were included as the training cohort, and 102 patients from a single center were included as the external validation cohort. The LightGBM model showed the best performance in predicting the short-term prognosis of sICH patients, with an area under the receiver operating characteristic curve (AUROC) of 0.813 ± 0.012. The clinical feature cohort model (AUC: 0.822, 95% CI (0.763-0.881)), the NCCT imaging feature model (AUC: 0.770, 95% CI (0.704-0.835)) and the fusion model (AUC: 0.852, 95% CI (0.797-0.906)) were developed respectively. The external validation cohort were the clinical feature model (AUC: 0.792, 95% CI (0.689-0.894)), the NCCT imaging feature model (AUC: 0.746, 95% CI (0.637-0.855)), and the fusion feature (AUC: 0.796, 95% CI (0.694-0.897). Finally, the core factors obtained through screening, including Glasgow Coma Scale (GCS) score at admission, intraventricular hemorrhage (IVH), National Institutes of Health Stroke Scale (NIHSS) score at admission, hematoma volume, mean CT value, and black hole sign were incorporated into the model to generate a publicly accessible online platform (https://surge-ustc.shinyapps.io/multi_para_sih_prognosis/). Interpretation: The prediction model based on NCCT features established by the LightGBM model has a reliable predictive effect on the short-term prognosis of sICH patients and is of great clinical convenience and practicality. Funding: Funding provided by National Natural Science Foundation of China (82427808, 82171382, U23A20424, 82090034 and 82371201), the Anhui Province Clinical Medical Research Transformation Special Project (202204295107020006 and 202204295107020028) and Research Fund of Anhui Institute of translational medicine (2022zhyx-B11).