Development and validation of an interpretable prognostic model for in-hospital GCS-based neurological recovery after intracerebral hemorrhage in young and middle-aged adults: a multicenter retrospective cohort study

Prognostic tools for intracerebral hemorrhage (ICH) are largely derived from older or mixed-age cohorts and rarely incorporate lifestyle factors, limiting their relevance to young and middle-aged adults. We aimed to develop and externally evaluate an interpretable model for unfavorable in-hospital GCS-based neurological recovery in adults aged 18–59 years with primary spontaneous ICH. This multicenter retrospective cohort study (2021–2025) included a development cohort from Tianjin Medical University General Hospital and Shanxi Provincial People's Hospital and an independent external test cohort from Binzhou Medical University Hospital. The endpoint was a study-specific normalized GCS recovery index; in-hospital death was classified as unfavorable recovery. Exploratory univariate and multivariable screening used the development cohort. Imputation, standardization, model fitting, and hyperparameter tuning were restricted to the training subset; an internal-validation subset was used for model selection, and the external cohort was reserved for final evaluation. Seven algorithms were compared. SHapley Additive exPlanations (SHAP) and decision curve analysis assessed interpretability and decision-analytic net benefit. Among 2,172 patients, 1,658 formed the development cohort and 514 formed the external test cohort. An 8-feature random forest model was selected, including consciousness disturbance, hematoma volume, glucose, lymphocyte percentage, creatine kinase, monocyte percentage, alkaline phosphatase, and tea consumption frequency. In the external test cohort, the model achieved an area under the receiver-operating characteristic curve of 0.854 and an area under the precision–recall curve of 0.811, showed acceptable calibration (Hosmer–Lemeshow P = 0.412), and provided net benefit across the evaluated thresholds. SHAP identified consciousness disturbance and hematoma volume as the leading contributors. Tea-consumption categories showed differing SHAP contributions that require cautious, non-causal interpretation. A Streamlit research prototype was created to demonstrate individualized predictions and SHAP explanations. We developed and externally tested an interpretable admission-based model for estimating unfavorable in-hospital GCS-based neurological recovery in young and middle-aged adults with primary spontaneous ICH. The model showed good discrimination and acceptable calibration in the independent external test cohort. Because the endpoint is study-specific and the lifestyle association is observational, prospective validation against standardized functional outcomes is required before clinical use.