[Development and validation of a machine learning-based prediction model for fatty liver in Wilson disease].

内容摘要

To construct a machine learning-based predictive model for fatty liver in patients with Wilson disease (WD). Clinical data retrospectively collected from 1862 WD patients at the First Affiliated Hospital of Anhui University of Chinese Medicine were divided into a training set (70%) and a validation set (30%). The least absolute shrinkage and selection operator (LASSO) was employed to screen the key predictive variables. Seven algorithms, namely logistic regression (LR), decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), support vector machine (SVM), and artificial neural network (ANN), were compared for their performance using the area under the receiver-operating characteristic (ROC) curve (AUC), precision-recall (PR) curve, calibration curves, and decision curve analysis (DCA). The contribution of each feature to model prediction was assessed using SHAP analysis. Among the 1862 WD patients, 1296 (69.60%) were complicated with fatty liver. LASSO regression identified platelet count (PLT), red cell distribution width (RDW), alanine aminotransferase (ALT), total bile acids (TBA), type IV collagen (CIV), and indirect bilirubin (IBIL) as the key predictive variables. The LightGBM model demonstrated optimal overall performance, with a training set AUC of 0.826 (95% CI: 0.801-0.849) and good calibration (Brier score 0.143); its validation set AUC was 0.815 (95% CI: 0.776-0.852), and the PR curve showed a high average precision (AP=0.903) with good calibration (Brier score 0.138) and significant clinical net benefit across all the diagnostic thresholds as confirmed by DCA. SHAP analysis indicated that ALT, IBIL, TBA, and CIV all had significant positive effects on model outputs, while RDW and PLT contributed minimally to the cumulative predictive outcomes. Among the 7 predictive models for fatty liver in WD patients, the LightGBM model demonstrates superior performance to potentially facilitate early screening and risk stratification of WD patients at high risk of fatty liver. 目的: 构建基于机器学习的Wilson病（WD）患者合并脂肪肝的预测模型，并评估其效能。方法: 回顾性收集安徽中医药大学第一附属医院1862例WD患者的临床资料，按7∶3分为训练集（n=1296）和验证集（n=566），采用最小绝对收缩和选择算子（LASSO）筛选关键预测变量，运用逻辑回归（LR）、决策树（DT）、随机森林（RF）、极端梯度提升（XGBoost）、轻量梯度提升机（LightGBM）、支持向量机（SVM）和人工神经网络（ANN）7种算法构建预测模型，结合受试者工作特征曲线下面积（AUC）、精准-召回曲线（PR）、校准曲线、决策曲线分析（DCA）评估模型性能，通过SHAP分析每个特征对模型预测的贡献。结果: 本研究纳入的1862例WD患者中1296 例（69.60%）合并脂肪肝。LASSO回归筛选出血小板计数（PLT）、红细胞分布宽度（RDW）、丙氨酸氨基转移酶（ALT）、总胆汁酸（TBA）、IV型胶原（CIV）、间接胆红素（IBIL）6项关键预测变量。在7种模型中，LightGBM模型综合效能最优:训练集AUC为0.826（95% CI:0.801~0.849），校准度良好（Brier评分0.143）;验证集AUC为 0.815（95% CI:0.776~0.852），PR平均精确率高（AP=0.903），校准度良好（Brier评分0.138），DCA证实其具有临床净获益。SHAP分析表明ALT、IBIL、TBA、CIV对模型输出有显著的正向影响，RDW、PLT对预测结果的累积影响极小。结论: 本研究构建并验证了7种WD脂肪肝的预测模型，其中LightGBM模型更具优势，可为WD患者脂肪肝的早期筛查与风险分层提供一种可靠的评估工具。.