目的 探讨创伤性脑损伤（TBI）患者发生早期急性肾损伤（AKI）的独立相关因素，构建并验证列线图预测模型，并与随机森林及XGBoost模型进行判别效能比较。 方法 回顾性收集2023年1月至2025年12月徐州医科大学附属医院收治的TBI患者临床资料。依据入院后48 h内是否发生AKI分为AKI组和非AKI组。采用单因素及多因素logistic回归分析筛选早期AKI的独立相关因素，基于多因素模型构建列线图预测模型。进一步构建随机森林和XGBoost模型，通过5折交叉验证比较3种模型的判别效能。 结果 共纳入368例TBI患者，其中AKI组54例（14.7%），非AKI组314例（85.3%）。多因素logistic回归分析显示，年龄、输血、机械通气、全身性感染及凝血功能紊乱评分是TBI患者早期AKI的独立相关因素，而格拉斯哥昏迷评分（GCS）为保护因素（P < 0.05）。基于上述变量构建的列线图校准曲线平均绝对误差为0.034，校准度良好；决策曲线分析显示模型在较低至中等阈值概率范围内具有潜在的临床获益。5折交叉验证结果显示，logistic模型、随机森林模型、XGBoost模型曲线下面积（AUC）分别为0.862、0.870、0.852。DeLong检验显示3种模型间AUC差异均无统计学意义。SHAP分析揭示各变量在机器学习模型中的相对贡献及重要性排序。 结论 基于年龄、GCS评分、输血、机械通气、全身性感染及凝血功能紊乱评分构建的logistic列线图模型，具有良好的区分度、校准度及潜在临床参考价值，可为神经外科临床TBI患者早期AKI风险识别与个体化管理提供参考。鉴于本研究为单中心回顾性内部验证，模型仍需多中心、前瞻性队列进一步验证。

Objective To explore the independent factors associated with early acute kidney injury （AKI） in patients with traumatic brain injury （TBI）， construct and validate a nomogram prediction model， and compare its discriminative efficacy with that of random forest and XGBoost models. Methods Clinical data of TBI patients admitted to the Affiliated Hospital of Xuzhou Medical University from January 2023 to December 2025 were retrospectively collected. The patients were then divided into the AKI group and the non-AKI group according to whether AKI occurred within 48 hours after admission. Univariate and multivariate logistic regression analyses were employed to identify the independent factors associated with early AKI. Subsequently， a nomogram was constructed based on the multivariate model. In addition， random forest and XGBoost models were developed， and the discriminative efficacy of these three models was compared through 5-fold cross-validation. Results A total of 368 patients with TBI were included， among whom 54 （14.7%） developed AKI. Multivariate logistic regression analysis demonstrated that age， blood transfusion， mechanical ventilation， systemic infection， and coagulation dysfunction score were independently associated with early AKI， whereas the Glasgow Coma Scale （GCS） score served as a protective factor （all P < 0.05）. The nomogram presented excellent calibration， with a mean absolute error of 0.034. Decision curve analysis suggested potential clinical benefits across low-to-moderate threshold probabilities. The 5-fold cross-validation area under the curve （AUC） values were 0.862 for the logistic model， 0.870 for the random forest model， and 0.852 for the XGBoost model. According to DeLong tests， there were no statistically significant differences among the three models. SHAP analysis disclosed the relative contribution and importance ranking of each variable. Conclusions The logistic nomogram model， which is based on age， GCS score， blood transfusion， mechanical ventilation， systemic infection， and coagulation dysfunction score， demonstrated excellent discrimination， calibration， and significant potential clinical reference value. It can offer a reliable reference for the early identification of AKI risk and the individualized management of TBI patients in neurosurgical practice. Considering that this was a single-center retrospective study with internal validation， the model still needs to be further validated in multicenter prospective cohorts.