Answers are grounded in this paper and cite their sources.
Objective To develop and validate a risk prediction model for early hematoma progression (HP) in patients with bifrontal cerebral contusions, systematically compare the performance of multiple machine learning algorithms, identify an optimal model with robustness, generalizability, and clinical interpretability, and determine the key factors driving HP. Methods This dual-center retrospective study included 1,100 adult patients with radiologically confirmed bifrontal cerebral contusions. The modeling cohort comprised 454 patients from the General Hospital of Ningxia Medical University and was randomly divided into a training set of 318 patients and an internal validation set of 136 patients. An independent cohort of 646 patients from the Second Affiliated Hospital of the University of South China was used for external validation. Demographic characteristics, vital signs, laboratory indicators, and cranial computed tomography features were collected, with HP defined as the outcome. To minimize information leakage, preprocessing, feature selection, model fitting, and hyperparameter optimization were conducted exclusively within the training set. A two-stage feature-selection strategy combining random forest importance ranking and least absolute shrinkage and selection operator regression was used to identify the core predictors, which were subsequently applied unchanged to the internal and external validation cohorts. Thirteen machine learning models were developed, and their discrimination, classification performance, calibration, and clinical utility were evaluated. Shapley Additive Explanations were used to characterize global and patient-level contributions of the predictors to the selected model. Results The intersection of features selected by RF and Lasso yielded 10 stable core predictors: time to baseline CT, ventricular compression, sulcal effacement, hematoma density coefficient of variation (Hematoma.DCV), heart rate, white blood cell count, eosinophil count, albumin, monocyte-to-lymphocyte ratio (MLR), and prognostic nutritional index (PNI). Among the 13 models, LightGBM achieved the highest apparent training-set AUC of 0.872 but decreased to 0.652 in the internal validation set, a pattern consistent with limited generalizability and possible overfitting. In contrast, Lasso achieved the numerically highest AUC in the internal validation set, with an AUC of 0.764 (95% CI: 0.677–0.852). In the external cohort, its AUC decreased to 0.694 but remained numerically higher than those of the other models displayed in the external ROC analysis. Because formal pairwise statistical comparisons were not available, these findings should not be interpreted as demonstrating the statistical superiority of Lasso over the other models. External decision curve analysis indicated potential net benefit mainly at low threshold probabilities, whereas calibration was suboptimal (Hosmer–Lemeshow χ²=33.87, P < 0.001; Brier score = 0.2405), with a tendency to underestimate observed HP risk. SHAP analysis indicated that Hematoma.DCV, time to baseline CT, MLR, sulcal effacement, PNI, heart rate, and ventricular compression were the main drivers of model prediction. Conclusion In patients with bifrontal cerebral contusions, the Lasso model showed comparatively stable discrimination and greater structural interpretability than the higher-capacity ensemble models. However, external discrimination was modest, calibration was inadequate, and clinical net benefit was mainly confined to low threshold probabilities. Hematoma density heterogeneity, the timing of baseline CT, systemic inflammatory–immune–nutritional status, and local intracranial compression signs contributed importantly to model predictions. The model should currently be regarded as a supplementary risk-ranking approach rather than a stand-alone screening tool or a source of directly actionable absolute-risk estimates. External recalibration and larger multicenter prospective validation are required before clinical implementation.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: