Acute kidney injury (AKI) is a common and serious complication of acute pancreatitis (AP) that substantially increases morbidity, mortality, and healthcare costs. Early, practical, and reliable risk stratification tools are needed to identify patients who may benefit from intensified monitoring and early interventions.

To develop and externally validate a liquid neural network (LNN)-based predictive model and a Streamlit-based bedside application for early detection of AKI in AP patients.

We included 586 AP patients treated between June 2020 and February 2025 at three hospitals in Changshu: Changshu No.1 People's Hospital (dataset 1, n=325), Changshu Shanghu Central Health Center (dataset 2, n=117), and Changshu Hospital of Traditional Chinese Medicine (dataset 3, n=144). Using Python, the combined development data (datasets 1 and 2) were randomly divided into training and validation cohorts at an 8∶2 ratio. Dataset 3 (prospective) served as an external test set for the bedside tool. For each patient in the development cohort, we extracted 31 candidate variables (demographics, medical history, laboratory indices). After data cleaning and missing-value handling, feature selection was performed using least absolute shrinkage and selection operator (LASSO). To address class imbalance, synthetic minority oversampling technique (SMOTE) was applied. We trained five models [Logistic regression (LR), decision tree (DCT), random forest (RF), extreme gradient boosting (XGBoost), and LNN] and compared performance by receiver operating characteristic (ROC) curve, sensitivity, specificity, and accuracy. SHapley Additive exPlanation (SHAP) analysis provided model interpretability. Finally, the best-performing model (LNN) was deployed as an interactive bedside risk calculator built with Streamlit and evaluated on the external test set.

Overall, 136 of 586 patients (23.21%) developed AKI. Baseline characteristics between training and validation cohorts were comparable (P>0.05). LASSO identified six predictors with the strongest association to AKI: albumin, blood urea nitrogen (BUN), alkaline phosphatase (ALP), serum potassium, serum creatinine, and neutrophil-to-lymphocyte ratio (NLR). In the validation cohort the LNN outperformed conventional and tree-based models, achieving an area under the curve (AUC) of 0.91 (95%CI=0.86-0.95). Feature-importance ranking identified six top predictors for AKI: NLR, serum creatinine, serum potassium, BUN, ALP, and albumin. SHAP force-plot visualizations clearly depicted how each of these features influenced individual patient risk. The best-performing LNN was implemented as an interactive bedside application using the Python-Streamlit framework. On the external test set, this AI bedside tool demonstrated an accuracy of 94.40%, sensitivity of 89.29%, and specificity of 95.69% for predicting AKI in patients with acute pancreatitis.

The LNN-based prediction model and its Streamlit bedside tool enable early identification of AP patients at high risk for AKI and have strong potential for clinical application.

急性胰腺炎（AP）常可并发急性肾损伤（AKI），显著增加死亡率和医疗负担。临床缺乏早期、简便且准确的预测工具，亟须构建有效模型用于识别高危患者。

构建基于液态神经网络（LNN）的预测模型及其床旁工具，实现对AP患者并发AKI的早期预测。

纳入2020年6月—2025年2月因AP就诊于常熟市第一人民医院（数据集1，n=325）、常熟市尚湖中心卫生院（数据集2，n=117）及常熟市中医院（数据集3，n=144）的患者586例。数据集1和数据集2采用回顾性收集，主要用于模型构建、内部验证及床旁工具开发；数据集3采用前瞻性收集，用于床旁工具的外部测试。使用Python按照8∶2的比例，随机将模型开发组数据（由数据集1和2组成）划分为训练队列和验证队列。收集训练队列和验证队列患者的临床资料，纳入人口学特征、既往病史及实验室检查等31个潜在特征。对数据进行缺失值处理、特征筛选，并采用合成少数类过采样技术（SMOTE）平衡样本后，分别构建逻辑回归（LR）、决策树（DCT）、随机森林（RF）、极端梯度提升（XGBoost）及LNN预测模型。通过受试者工作特征（ROC）曲线、灵敏度、特异度及准确度等指标比较模型性能，选出最佳模型，并利用SHapley加性解释（SHAP）方法进行可解释性分析。基于Streamlit框架开发网络应用，实现床旁风险预测。

586例AP患者中136例（23.21%）并发AKI。训练队列和验证队列的基线资料比较，差异均无统计学意义（P>0.05）。LASSO回归分析共筛选出6个关键预测特征，分别为白蛋白、血尿素氮、碱性磷酸酶、血钾、血肌酐、中性粒细胞与淋巴细胞比值（NLR）。在验证队列中，LNN模型表现最佳，优于LR和其他机器学习模型，AUC为0.91（95%CI=0.86~0.95）。特征重要性排序显示NLR、肌酐、血钾、尿素氮、碱性磷酸酶、白蛋白6项指标对AKI预测的贡献最大，相关力图可视化分析清晰呈现各特征对个体风险的影响。基于Python-Streamlit框架，将性能最优的LNN模型开发为具有可视化操作界面的AI床旁工具，在测试集中，AI床旁工具预测AP并发AKI的准确度、灵敏度和特异度分别为94.40%、89.29%和95.69%。

基于LNN的预测模型及床旁工具能够实现AP患者并发AKI风险的早期识别，具有重要的临床应用价值。