Skip to main content
Journal of Southern Medical University logoLink to Journal of Southern Medical University
. 2026 Aug 20;46(8):1861–1869. [Article in Chinese] doi: 10.12122/j.issn.1673-4254.2026.08.13

基于机器学习的Wilson病脂肪肝预测模型的开发与验证

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

SHI Shiheng 1,2, HUA Daiping 1, XIANG Shang 1, SUN Lanting 1, XUAN Qiaoyu 1, YANG Wenming 1,2, WANG Han 1,2,✉
PMCID: PMC13458563  PMID: 42576494

Abstract

Objective

To construct a machine learning-based predictive model for fatty liver in patients with Wilson disease (WD).

Methods

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.

Results

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.

Conclusion

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.

Keywords: Wilson disease, fatty liver disease, machine learning, predictive model, LightGBM


Wilson病(WD)是一种由ATP7B基因突变引发的常染色体隐性遗传铜代谢障碍性疾病,其核心病理机制为铜离子在肝脏、脑等多器官异常沉积[1]。铜离子首先在肝内大量蓄积,引起肝脏脂质代谢紊乱和脂肪变性[2]。研究表明,在WD患者中,脂肪肝可能是肝脏最早且唯一的病理改变,常发生在出现明显症状(如黄疸、腹水)或肝硬化之前[3]。然而,脂肪肝作为无症状早期病变易被临床忽视,导致干预延迟,进而进展为肝纤维化、肝硬化。目前,肝活检虽是脂肪肝的诊断金标准,但其有创性等缺陷限制了应用;腹部超声虽为脂肪肝常规初筛手段,但WD患者需长期规律监测,反复超声检查成本效益较低,加之医疗资源分布不均,导致WD合并脂肪肝的检出率低、漏诊风险高。因此,构建基于常规实验室指标的预测模型,确保高危患者获得评估,有助于早期识别高危人群并优化WD患者的全程管理策略。

机器学习(ML)能够处理高维数据并识别复杂关系,在疾病预测模型中优势显著[4, 5]。研究表明,相较于传统线性模型,机器学习更具灵活性且可能具备更优的预测能力[6, 7]。先前研究多集中在WD相关肝病的终末期表现,如肝硬化及肝衰竭等[8, 9],对于WD病程中更早出现的脂肪肝病变,仍缺乏有效的早期识别工具。虽有非酒精性脂肪性肝病的机器学习预测模型[10],但WD脂肪肝的铜-脂代谢交互特殊病理机制,现有模型难以直接适用。

鉴于此,本研究收集WD患者的临床基线资料、实验室检测指标及影像学数据,构建逻辑回归(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种机器学习算法在WD合并脂肪肝人群中的预测性能。采用SHAP算法对最优模型进行可解释性分析,以量化分析各特征对模型预测的贡献度及其方向性。

1. 资料和方法

1.1. 研究对象

本研究为回顾性队列研究,通过收集2020年7月~2025年7月首次在安徽中医药大学第一附属医院接受治疗的WD患者实验室参数及影像学数据,共纳入1862例。WD患者纳入标准:参照《肝豆状核变性临床实践指南》[11],年龄≥18岁,且病历资料完整;并结合莱比锡量表[12]评估。该量表由2位具有副高及以上职称的神经科医师独立完成,总分≥4分者方可入选。脂肪肝诊断标准:依据《代谢相关(非酒精性)脂肪性肝病防治指南(2024年版)》[13]。若符合以下任一条件即可确诊:肝脏近场回声弥漫性增强(肝脏显影较亮),且肝脏回声强度高于肾脏;肝脏内部管状结构无法清晰显示;肝脏远场回声逐渐衰减。排除标准:存在全身性内分泌疾病的患者;由病毒性肝炎、血吸虫病感染及酒精中毒等引发的肝脏疾病。本研究方案经安徽中医药大学第一附属医院医学伦理委员会批准(批准号:2025AH-68),严格遵循2013年修订的《赫尔辛基宣言》。鉴于研究采用回顾性设计且使用匿名患者信息,故免除患者知情同意。结局指标为WD合并脂肪肝的情况,根据患者结局分为脂肪肝组与非脂肪肝组,研究流程见图1。

图1.

图1

WD患者脂肪肝预测模型的开发与验证流程图

Fig.1 Flowchart of the development and validation of the fatty liver prediction model in WD patients.

1.2. 变量选取

收集患者的一般资料:年龄、性别、身高、体质量;实验室指标:白细胞(WBC)、红细胞(RBC)、血红蛋白(Hb)、血小板(PLT)、红细胞分布宽度(RDW)、丙氨酸氨基转移酶(ALT)、天门冬氨酸氨基转移酶(AST)、总蛋白(TP)、球蛋白(GLB)、谷氨酰基转移酶(GGT)、碱性磷酸酶(ALP)、乳酸脱氢酶(LDH)、血糖(GLU)、甘油三酯(TG)、总胆固醇(TC)、高密度脂蛋白胆固醇(HDL-C)、低密度脂蛋白胆固醇(LDL-C)、胆汁酸(TBA)、脂蛋白(a) [Lp(a)]、同型半胱氨酸(Hcy)、尿酸(UA)、尿素氮(BUN)、血肌酐(Cr)、Ⅲ型前胶原N端肽(PⅢNP)、层粘连蛋白(LN)、透明质酸(HA)、Ⅳ型胶原(CIV)、国际标准化比值(INR)、纤维蛋白原(FIB)、凝血酶原时间(PT)、活化部分凝血活酶时间(APTT)、总胆红素(TBIL)、直接胆红素(DBIL)、间接胆红素(IBIL)、24 h尿铜(24 hUCu);计算体质量指数(BMI)、甘油三酯和高密度脂蛋白胆固醇比值(TG/HDL-C)、白蛋白和球蛋白比值(A/G)、肝胆胰脾彩超。

1.3. 研究设计

将数据集按7∶3的比例随机分为训练集(n=1296)和验证集(n=566),使用训练集对模型进行构建与超参数优化,验证集评估模型的最终预测性能与泛化能力。使用多重插补法对缺失值进行处理,剔除缺失率>20%变量。在训练集上使用最小绝对收缩和选择算子(LASSO)结合10折交叉验证筛选变量以识别潜在预测因子并控制过拟合[14],同时通过计算方差膨胀因子(VIF)评估多重共线性问题,当VIF<5时,说明各变量之间均相互独立,不存在显著的共线性问题[15]。构建并比较LR、DT、RF、XGBoost、LightGBM、SVM和ANN 7种机器学习算法模型。最后选取最优模型进行SHAP解释。

1.4. 模型评价

为全面评估预测模型的性能,采用以下指标与方法:区分能力:使用受试者工作特征(ROC)曲线进行评估,主要评价指标为受试者工作特征曲线下面积(AUC)。AUC值的95%置信区间(CI)通过Bootstrap法(有放回随机抽样1000次)进行估计。精准-召回曲线:采用精准-召回曲线(PR)比较不同模型的预测效能,计算平均精确率(AP)量化曲线整体效能,AP综合反映了模型在类别不平衡数据中的分类性能,其值越高,表示模型在精确率与召回率之间的权衡越优[16]。校准度:通过Brier评分和校准曲线评估模型预测概率与实际观测风险之间的一致性,Brier评分越低表示模型的预测准确性越高[17]。临床实用性:利用决策曲线分析(DCA)量化模型在不同决策阈值下的临床净收益[18]。分类性能:采用Youden指数确定模型的最佳诊断阈值,并据此计算准确度、精确率、灵敏度、特异度、F1分数、阳性预测值与阴性预测值[19]。

1.5. SHAP解释

为解释最佳预测模型的决策依据,采用SHAP值量化各预测变量对模型输出的贡献[20, 21]。通过SHAP摘要图展示变量的全局重要性。SHAP瀑布图展示单个样本的预测结果如何由各特征值的SHAP贡献累积而成,以解释特定预测的生成过程及各特征对偏离基线预测的具体贡献度。

1.6. 统计学分析

使用R(版本4.5.1)与Python(版本3.10.4)进行数据统计分析,相关代码在Visual Studio Code集成环境中开发执行。连续变量由于均呈非正态分布,以中位数(四分位数间距)表示,组间比较采用Mann-Whitney U检验。分类变量以频数(百分比)表示,组间比较采用χ²检验。所有假设检验均采用双侧检验,P<0.05表示差异具有统计学意义。

2. 结果

2.1. 基线资料比较

1862例WD患者中,合并脂肪肝者1296例(69.60%),非脂肪肝者566例(30.40%);男性1057例(56.77%),年龄中位数30岁。WD脂肪肝组与非脂肪肝组相比,WBC、RBC、PLT、RDW、ALT、AST、GGT、ALP、LDH、TBA、TP、A/G、PⅢNP、LN、HA、CIV、TG、HDL-C、INR、PT、APTT、TBIL、DBIL、IBIL差异具有统计学意义(P<0.05);性别、年龄、BMI、Hb、GLB、GLU、TC、LDL-C、TG/HDL-C、Lp(a)、Hcy、UA、BUN、Cr、24-h UCu差异无统计学意义(P>0.05,表1)。

表1.

WD患者脂肪肝组与非脂肪肝组的基线特征

Tab.1 Baseline characteristics of WD patients with and without fatty liver [Median (IQR)]

Variable Overall (n=1862) Non-fatty liver group (n=566) Fatty liver group (n=1296) P
Male[n (%)] 1057 (56.77) 312 (55.12) 745 (57.48) 0.371
Age (years) 30.00 (23.00, 37.00) 29.00 (23.00, 36.00) 30.00 (23.00, 37.00) 0.170
BMI (kg/m2) 21.26 (19.05, 23.79) 21.22 (19.16, 23.61) 21.30 (19.05, 23.92) 0.600
WBC (×10⁹/L) 4.60 (3.61, 5.83) 4.72 (3.84, 5.96) 4.52 (3.51, 5.73) 0.003
RBC (×10¹²/L) 4.37 (3.99, 4.79) 4.39 (4.07, 4.78) 4.36 (3.95, 4.79) 0.038
Hb (g/L) 130.00 (118.00, 142.00) 130.00 (119.00, 142.00) 130.00 (117.00, 142.00) 0.377
PLT (×10⁹/L) 144.00 (94.00, 203.00) 159.50(111.00, 217.00) 138.00 (87.75, 196.25) <0.001
RDW (%) 13.30 (12.70, 14.30) 13.10 (12.50, 13.90) 13.40 (12.70, 14.50) <0.001
ALT (U/L) 30.40 (20.02, 53.10) 24.10 (17.00, 40.30) 34.00 (21.98, 59.12) <0.001
AST (U/L) 29.00 (22.00, 43.60) 26.00 (20.40, 36.00) 30.65(23.00, 46.00) <0.001
TP (g/L) 63.60 (60.00, 67.60) 64.10 (60.60, 67.77) 63.40 (59.80, 67.43) 0.048
GLB (g/L) 24.40 (21.90, 27.70) 24.20 (21.90, 27.08) 24.50 (21.90, 27.90) 0.141
A/G 1.61 (1.36, 1.86) 1.64 (1.41, 1.87) 1.59 (1.34, 1.86) 0.014
GGT(U/L) 31.00 (19.00, 58.00) 28.00 (18.00, 52.00) 32.00 (20.00, 61.25) <0.001
ALP (U/L) 96.00 (76.00, 124.00) 93.00 (75.00, 118.00) 98.00 (78.00, 127.00) 0.002
LDH (U/L) 168.00 (146.00, 198.00) 166.00 (145.25, 192.00) 170.00 (146.75, 201.25) 0.010
GLU (mmol/L) 4.61 (4.32, 4.93) 4.62 (4.34, 4.95) 4.60 (4.31, 4.92) 0.454
TG (mmol/L) 0.92 (0.71, 1.22) 0.97 (0.75, 1.29) 0.88 (0.69, 1.19) <0.001
TC (mmol/L) 4.05 (3.51, 4.66) 4.08 (3.61, 4.66) 4.04 (3.47, 4.65) 0.131
HDL-C (mmol/L) 1.26 (1.07, 1.48) 1.27 (1.08, 1.52) 1.26 (1.06, 1.47) 0.008
LDL-C (mmol/L) 2.38 (2.01, 2.82) 2.38 (2.01, 2.79) 2.38 (2.00, 2.83) 0.748
TG/HDL-C 0.72 (0.52, 1.02) 0.74 (0.53, 1.02) 0.70 (0.51, 1.02) 0.349
TBA (μmol/L) 7.80 (4.50, 16.20) 5.80 (3.70, 9.70) 9.15 (5.00, 21.92) <0.001
Lp(a) (mg/L) 53.75 (26.70, 115.50) 58.45 (27.50, 129.10) 51.65 (26.17, 110.93) 0.079
Hcy (μmol/L) 11.30 (8.30, 15.80) 11.10 (8.20, 15.88) 11.30 (8.30, 15.72) 0.718
UA (μmol/L) 222.50 (173.25, 285.75) 221.50 (174.00, 282.00) 223.00 (172.75, 286.00) 0.860
BUN (mmol/) 4.88 (3.99, 5.90) 4.93 (4.11, 5.98) 4.87 (3.95, 5.85) 0.076
Cr (μmol/L) 65.20 (54.00, 79.18) 65.55 (53.12, 82.07) 64.95 (54.68, 78.23) 0.541
PⅢNP (ng/mL) 14.24 (9.92, 23.26) 12.37 (9.18, 19.11) 15.21 (10.20, 24.79) <0.001
LN (ng/mL) 104.36 (80.90, 140.61) 98.98 (76.04, 126.21) 108.03 (82.45, 151.14) <0.001
HA (ng/mL) 123.04 (71.61, 226.08) 102.69 (61.83, 176.16) 134.66 (77.42, 248.96) <0.001
CIV (ng/mL) 66.87 (49.04, 101.60) 57.42 (40.35, 79.25) 72.22 (51.76, 109.73) <0.001
INR 1.01 (0.95, 1.09) 1.00 (0.94, 1.07) 1.02 (0.96, 1.11) <0.001
FIB (g/L) 2.08 (1.78, 2.42) 2.17 (1.86, 2.54) 2.04 (1.74, 2.38) <0.001
PT (s) 31.00 (29.10, 33.10) 30.70 (28.92, 32.68) 31.00 (29.10, 33.40) 0.002
APTT (s) 11.50 (10.80, 12.40) 11.30 (10.60, 12.00) 11.50 (10.90, 12.60) <0.001
TBIL (μmol/L) 15.20 (11.30, 20.58) 13.60 (9.10, 17.98) 15.91 (12.00, 21.70) <0.001
DBIL (μmol/L) 3.20 (2.40, 4.50) 2.80 (1.90, 3.90) 3.40 (2.50, 4.73) <0.001
IBIL (μmol/L) 11.80 (8.80, 16.00) 10.30 (7.00, 13.90) 12.40 (9.47, 17.00) <0.001
24-h UCu (μg/24 h) 748.37 (423.13, 1151.57) 703.16 (420.82, 1108.14) 762.30 (427.12, 1162.81) 0.294

BMI: Body mass index; WBC: White blood cell count; RBC: Red blood cell count; Hb: Hemoglobin; PLT: Platelet count; RDW: Red blood cell distribution width; ALT: Alanine aminotransferase; AST: Aspartate transaminase; TP: Total protein; GLB: Globulin; A/G: Albumin/globulin ratio; GGT: Gamma-glutamyl transpeptidase; ALP: Alkaline phosphatase; LDH: Lactate dehydrogenase; GLU: Glucose; TG: Triglyceride; TC: Total cholesterol; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; TG/HDL-C:Triglyceride/high-density lipoprotein cholesterol ratio; TBA: Total bile acid; Lp(a): Lipoprotein(a); Hcy: Homocysteine; UA: Uric acid; BUN: Blood urea nitrogen; Cr: Creatinine; PⅢNP: Type III procollagen N-terminal peptide; LN: Laminin; HA: Hyaluronic acid; CIV: Type IV collagen; INR: International normalized ratio; FIB: Fibrinogen; PT: Prothrombin time; APTT: Activated partial thromboplastin time; TBIL: Total bilirubin; DBIL: Direct bilirubin; IBIL: Indirect bilirubin; 24-h UCu: 24-hour urinary copper.

2.2. 变量筛选

使用LASSO回归进行变量筛选,当采用λ的1倍标准误准则(λ=0.015)时模型表现最佳,识别出6个具有非零系数的变量(PLT、RDW、ALT、TBA、CIV、IBIL),各变量间不存在多重共线性(VIF范围为1.01~1.21,图2)。

图2.

图2

LASSO回归筛选WD患者脂肪肝预测变量的特征选择图

Fig.2 Feature selection plot of LASSO regression for screening predictors of fatty liver in WD patients. A: Distribution of LASSO regression coefficients. B: Determination of the regularization parameter (λ) for the LASSO model by 10-fold cross-validation. The left dashed line represents the λmin, and the right dashed line represents λ1se.

2.3. 模型评价

LR、DT、RF、XGBoost、LightGBM、SVM、ANN模型,在训练集中AUC分别为0.738、0.785、0.840、0.811、0.826、0.735、0.771,验证集分别为0.729、0.785、0.809、0.804、0.815、0.722、0.734,LightGBM在训练集和验证集中AUC均较高。训练集与验证集中各模型的准确度、精确率、灵敏度、特异性、F1分数、Youden指数、阳性预测值、阴性预测值(表2)。

表2.

不同机器学习模型预测WD患者脂肪肝的性能表现对比分析

Tab.2 Performance comparison of different machine learning models for predicting fatty liver in WD patients

Model AUC (95% CI) Accuracy Precision Sensitivity Specificity F1 Score Youden's J PPV NPV
Training set
LR 0.738 (0.710-0.765) 0.728 0.752 0.910 0.312 0.823 0.222 0.752 0.602
DT 0.785 (0.759-0.811) 0.800 0.777 0.999 0.345 0.874 0.344 0.777 0.993
RF 0.840 (0.816-0.863) 0.802 0.780 0.994 0.360 0.875 0.355 0.780 0.966
XGBoost 0.811 (0.785-0.837) 0.785 0.767 0.992 0.310 0.865 0.302 0.767 0.946
LightGBM 0.826 (0.801-0.849) 0.797 0.778 0.991 0.353 0.872 0.344 0.778 0.946
SVM 0.735 (0.706-0.762) 0.741 0.739 0.970 0.217 0.839 0.187 0.739 0.761
ANN 0.771 (0.744-0.795) 0.759 0.778 0.915 0.403 0.841 0.318 0.778 0.675
Validation set
LR 0.729 (0.684-0.772) 0.713 0.751 0.879 0.331 0.810 0.210 0.751 0.544
DT 0.785 (0.744-0.824) 0.808 0.786 0.995 0.379 0.878 0.374 0.786 0.970
RF 0.809 (0.770-0.848) 0.813 0.793 0.990 0.408 0.881 0.398 0.793 0.945
XGBoost 0.804 (0.764-0.843) 0.799 0.780 0.990 0.361 0.873 0.351 0.780 0.938
LightGBM 0.815 (0.776-0.852) 0.820 0.798 0.995 0.420 0.885 0.415 0.798 0.973
SVM 0.722 (0.677-0.767) 0.731 0.743 0.938 0.254 0.829 0.193 0.743 0.642
ANN 0.734 (0.691-0.780) 0.754 0.777 0.907 0.402 0.837 0.310 0.777 0.654

LR: Logistic regression; DT: Decision tree; RF: Random forest; XGBoost: Extreme gradient boosting; LightGBM: Light gradient boosting machine; SVM: Support vector machine; ANN: Artificial neural network; AUC: Area under the curve.

校准曲线:校准曲线与理想对角线走形基本一致,各模型均具有较好的预测准确度,LightGBM在校准曲线中表现良好。Brier评分在训练集中由低到高排序为RF(0.140)、LightGBM(0.143)、DT(0.144)、XGBoost(0.154)、ANN(0.169)、LR(0.180)、SVM(0.181);验证集中排序为LightGBM(0.138)、RF(0.141)、DT(0.142)、XGBoost(0.149)、ANN(0.179)、SVM(0.181)、LR(0.182)。

决策曲线:训练集显示,当概率阈值处于0~0.75时,各模型的净获益均高于“干预所有”与“不干预”基准线,具有临床实用价值;验证集决策曲线趋势与训练集基本一致,且在0~0.6阈值内,LightGBM模型的净获益曲线整体更靠前,净获益表现最优(图3)。

图3.

图3

WD患者脂肪肝预测模型在训练集与验证集的AUC、校准曲线及决策曲线

Fig.3 AUC, calibration curves and decision curves of fatty liver prediction models for WD patients in the training and validation sets. A: Training-set AUC. B: Validation-set AUC. C: Training-set calibration curves. D: Validation-set calibration curves. E: Training-set decision curves. F: Validation-set decision curves.

AP:在验证集中由高到低排序为LightGBM(0.903)、RF(0.899)、XGBoost(0.895)、LR(0.870)、SVM(0.864)、ANN(0.864)、DT(0.860)(图4)。

图4.

图4

WD患者脂肪肝预测模型在验证集的PR曲线图

Fig.4 PR curve of the prediction model for fatty liver in WD patients in the validation set. X-axis: Recall; Y-axis: Precision. Colored curves: machine learning models (annotated with AP values, higher = better performance). Red dashed line: sample prevalence (69.7%, random guessing baseline).

2.4. LightGBM的SHAP解释

对LightGBM进行可解释分析,SHAP蜂群图显示:红色点为高风险样本、蓝色点为低风险样本,IBIL、CIV升高时,SHAP值正向偏移;PLT、RDW升高时,SHAP值负向偏移;TBA、ALT的SHAP值随样本特征差异双向波动(图5A)。SHAP特征重要性排序显示:IBIL与CIV的平均绝对SHAP值最高,对预测贡献最大;TBA、ALT次之;RDW与PLT的影响相对较弱(图5B)。

图5.

图5

LightGBM模型预测WD患者脂肪肝的SHAP可解释性分析

Fig.5 SHAP interpretability analysis of the LightGBM model for predicting fatty liver in WD patients. A: Beeswarm plot of SHAP values for each feature. Each point represents a sample. Greater dispersion of points indicates a more significant impact on the model output. Colors represent feature values (red=high, blue=low). B: Bar plot of mean absolute SHAP values. Each row represents a feature, with higher values indicating greater overall contribution to the model prediction. C: Waterfall plot of SHAP values for a single sample. Baseline E[f(X)] denotes the expected model output, and final f(x)denotes the sample-specific prediction. Red bars represent positive SHAP contributions (pushing the prediction upward), while blue bars represent negative contributions (pushing the prediction downward). Labels show each feature and its corresponding SHAP value.

单样本瀑布图显示:目标事件阳性样本(脂肪肝组)中,ALT、IBIL、TBA、CIV的SHAP值均为正,累积推动模型输出概率达0.757;阴性样本(非脂肪肝组)中,上述指标SHAP值均为负,共同使输出概率降至0.243;RDW、PLT在两类样本中SHAP值均接近0,贡献度相对较小但效应方向稳定(图5C)。

3. 讨论

WD是由ATP7B基因突变所致的铜代谢紊乱性疾病,呈常染色体隐性遗传,全球范围内发病率约为2/10万[22],我国发病率约为5.87/10万[1],是少数可以治疗的神经遗传病之一。肝脏是铜离子沉积最早且最主要的靶器官,铜代谢失常引起肝脏脂质的分解与合成异常,导致脂质在肝脏累积,进而发展成脂肪变性[23]。目前,针对肝脂肪变性的现有检测手段,包括超声、CT、MRI及肝活检,均不适用于大规模筛查。因此,研发具有高灵敏度的WD脂肪肝筛查工具,具有重要的临床现实意义。

本研究纳入1862例WD患者,通过LASSO筛选出PLT、RDW、ALT、TBA、CIV及IBIL 6项关键预测特征,构建并验证了LR、DT、RF、XGBoost、LightGBM、SVM和ANN 7种机器学习模型。综合比较,LightGBM模型展现出最优综合性能,在验证集中AUC为0.815(95% CI:0.776~0.852)、准确性0.820、敏感性0.995、特异性0.420、F1分数0.885、PR曲线平均精确率高(AP=0.903),校准度良好(Brier评分0.138),DCA证实其在各诊断阈值下均具有显著临床净获益。本研究构建的模型呈高敏感性(0.995)、低特异性(0.420),适用于WD患者合并脂肪肝的初级筛查。高敏感性可有效降低漏诊风险,但低特异性会导致假阳性增加,可能造成不必要的医疗资源消耗。因此,在临床应用中,建议对阳性结果联合肝脏瞬时弹性成像进行二次验证[14],以在确保筛查敏感度的同时,减少不必要的医疗资源浪费。

LightGBM模型的树结构优化算法在处理高维数据时能够更好地减少过拟合,高效稳健的处理数据[24],该模型在训练集和验证集上的性能差异较小,表明其具有较好的泛化能力。SHAP可解释性分析进一步明确了各特征的重要性依次为IBIL、CIV、TBA、ALT、RDW、PLT。

IBIL平均绝对SHAP值最高,是预测模型中贡献度最大的核心驱动因素。IBIL升高是肝细胞功能障碍的敏感标志,在WD中,其本质反映了铜离子蓄积所导致的肝细胞线粒体损伤[25]。线粒体功能异常不仅直接阻碍胆红素的结合与排泄,导致IBIL升高;同时也是肝细胞内脂质β-氧化受阻、引发脂质代谢紊乱和脂肪变性的关键环节[26]。表明在WD患者中IBIL不仅提示肝脏损伤,还可能是连接“铜毒性-线粒体功能障碍-脂肪肝”这一特异性通路的关键指标。

CIV作为肝纤维化的经典血清学标志物,其预测价值同样得到既往研究的支持。既往在NAFLD患者中的研究发现,CIV水平升高不仅与肝纤维化程度相关,还与脂肪性肝炎的发生风险显著关联[27],肝星状细胞激活是肝纤维化与脂肪变性的共同病理环节,脂肪变性引发的肝细胞损伤可激活肝星状细胞,促进包括CIV在内的胶原成分合成,而胶原沉积又会进一步加重肝细胞脂质代谢紊乱[28]。WD患者因铜蓄积直接激活肝星状细胞,叠加脂肪肝的协同作用,导致CIV水平升高。

ALT是WD脂肪肝发生的重要预测因子,这与既往研究“ALT是NAFLD独立危险因素”结论一致[29]。既往研究也证实,ALT水平与肝细胞脂肪变性程度呈正相关[30],脂肪变性导致肝细胞细胞膜完整性受损,使ALT释放入血。但NAFLD相关的ALT升高,核心诱因是代谢紊乱引发的脂肪毒性损伤,过量脂肪在肝细胞内堆积形成脂肪变性,而WD 患者的ALT升高是铜毒性损伤与脂肪变性损伤的协同作用结果,且铜毒性为核心驱动因素[31]。

胆汁酸代谢异常是脂肪肝的重要诱因之一[32],胆汁酸不仅参与脂质消化吸收,还可通过调控核受体影响脂质合成与分解。当胆汁酸排泄障碍导致其血中水平蓄积时,对脂质代谢的调控失衡会促进肝细胞脂肪变性[33]。WD患者因铜沉积损伤胆小管上皮细胞,导致胆汁分泌与排泄障碍,TBA在脂肪肝患者中升高,SHAP分析也证实其对预测的重要性。若WD患者出现TBA升高,尤其是伴随腹胀、消化不良等症状时,需要关注胆汁淤积与脂肪变性的协同损伤。

RDW与PLT可作为WD患者脂肪肝预测的次要辅助指标,其关联符合慢性肝病的潜在病理逻辑[34]。既往研究表明,RDW升高与慢性肝病的炎症、氧化应激相关,PLT轻度降低则与脂肪肝早期轻度门静脉高压、脾功能亢进有关[35, 36],二者均为慢性肝损伤的轻度辅助标志。在WD患者中,铜蓄积引发的氧化应激会加剧炎症,且早期肝细胞损伤可诱发轻度门静脉高压,使二者与脂肪肝的关联更具疾病特异性。本研究SHAP分析显示,二者对模型预测贡献稳定但不明显,平均SHAP值在6项特征中最低,提示其异常仅为风险补充信号。

本研究构建的LightGBM模型较既往同类模型更具优势,其聚焦WD早期脂肪肝这一易被忽视的阶段,为疾病“早识别、早干预”提供实用支撑。LightGBM模型借助梯度提升框架与直方图优化算法,能高效捕捉临床指标间的非线性关联,降低过拟合风险、提升泛化稳定性[37],且以CIV、IBIL等WD疾病特异性指标构建,避免了普通人群NAFLD预测模型依赖代谢指标可能导致的漏判问题。本研究筛选的6项预测特征均为血常规、肝功能、肝纤4项等临床常规检测项目,无需额外增加检测成本、获取便捷,适用于基层医疗机构开展WD患者脂肪肝早期筛查与风险分层,助力优化临床管理流程。

本研究仍存在一定局限性:首先,本研究为单中心回顾性设计,未开展多中心外部验证,模型普适性有待进一步验证;其次,脂肪肝诊断基于超声而非肝活检金标准,可能存在误分类偏倚;此外,驱铜治疗方案及依从性未纳入分析,可能存在混杂偏倚;最后,饮食、运动等生活方式变量尚未收集,未来将进一步开展多模态数据研究。

基金资助

国家自然科学基金区域创新发展联合基金项目(U22A20366); 安徽省中医药科技攻关专项项目(202303a07020004);全国名老中医药专家传承工作室建设项目(国中医药人教函〔2022〕75号);2023年安徽省临床医学研究转化专项项目(202304295107020109);安徽省自然科学基金(2208085MH266);安徽省卫生健康科研项目(AHWJ2022b036)

Supported by Regional Innovation and Development Joint Fund of National Natural Science Foundation of China (U22A20366).

参考文献

  • 1. Członkowska A, Litwin T, Dusek P, et al. Wilson disease[J]. Nat Rev Dis Primers, 2018, 4: 21. doi: 10.1038/s41572-018-0018-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Mazi TA, Shibata NM, Medici V. Lipid and energy metabolism in Wilson disease[J]. Liver Res, 2020, 4(1): 5-14. doi: 10.1016/j.livres.2020.02.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Medici V, Shibata NM, Kharbanda KK, et al. Wilson's disease: changes in methionine metabolism and inflammation affect global DNA methylation in early liver disease[J]. Hepatology, 2013, 57(2): 555-65. doi: 10.1002/hep.26047 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Shamout F, Zhu TT, Clifton DA. Machine learning for clinical outcome prediction[J]. IEEE Rev Biomed Eng, 2021, 14: 116-26. doi: 10.1109/rbme.2020.3007816 [DOI] [PubMed] [Google Scholar]
  • 5. Greener JG, Kandathil SM, Moffat L, et al. A guide to machine learning for biologists[J]. Nat Rev Mol Cell Biol, 2022, 23(1): 40-55. doi: 10.1038/s41580-021-00407-0 [DOI] [PubMed] [Google Scholar]
  • 6. Khene ZE, Bigot P, Doumerc N, et al. Application of machine learning models to predict recurrence after surgical resection of nonmetastatic renal cell carcinoma[J]. Eur Urol Oncol, 2023, 6(3): 323-30. doi: 10.1016/j.euo.2022.07.007 [DOI] [PubMed] [Google Scholar]
  • 7. Yu YD, Lee KS, Man Kim J, et al. Artificial intelligence for predicting survival following deceased donor liver transplantation: Retrospective multi-center study[J]. Int J Surg, 2022, 105: 106838. doi: 10.1016/j.ijsu.2022.106838 [DOI] [PubMed] [Google Scholar]
  • 8. Rao ZH, Yang WM, Yang YL, et al. Revolutionizing Wilson disease prognosis: a machine learning approach to predict acute-on-chronic liver failure[J]. J Transl Med, 2025, 23(1): 999. doi: 10.1186/s12967-025-06987-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Chen K, Wan Y, Mao J, et al. Liver cirrhosis prediction for patients with Wilson disease based on machine learning: a case-control study from southwest China[J]. Eur J Gastroenterol Hepatol, 2022, 34(10): 1067-73. doi: 10.1097/meg.0000000000002424 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. 程琳傑, 袁 晴, 李宛凇, 等. 基于机器学习的非酒精性脂肪性肝病预测模型的构建及验证[J]. 中华疾病控制杂志, 2025, 29(6): 682-7, 696. [Google Scholar]
  • 11. European Association for the Study of the Liver . EASL-ERN Clinical Practice Guidelines on Wilson' s disease[J]. J Hepatol, 2025, 8278(24): 2706-15. [DOI] [PubMed] [Google Scholar]
  • 12. Ferenci P, Caca K, Loudianos G, et al. Diagnosis and phenotypic classification of Wilson disease 1[J]. Liver Int, 2003, 23(3): 139-42. doi: 10.1034/j.1600-0676.2003.00824.x [DOI] [PubMed] [Google Scholar]
  • 13. 中华医学会肝病学分会 . 代谢相关(非酒精性)脂肪性肝病防治指南(2024年版)[J]. 中华肝脏病杂志, 2024, 32(5): 418-34. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Sanchez-Pinto LN, Venable LR, Fahrenbach J, et al. Comparison of variable selection methods for clinical predictive modeling[J]. Int J Med Inform, 2018, 116: 10-7. doi: 10.1016/j.ijmedinf.2018.05.006 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Kim JH. Multicollinearity and misleading statistical results[J]. Korean J Anesthesiol, 2019, 72(6): 558-69. doi: 10.4097/kja.19087 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Huang CX, Li SX, Caraballo C, et al. Performance metrics for the comparative analysis of clinical risk prediction models employing machine learning[J]. Circ Cardiovasc Qual Outcomes, 2021, 14(10): 7526-38. doi: 10.1161/circoutcomes.120.007526 [DOI] [PubMed] [Google Scholar]
  • 17. Ralston A, Liu P, Warrener K, et al. Small field diode correction factors derived using an air core fibre optic scintillation dosimeter and EBT2 film[J]. Phys Med Biol, 2012, 57(9): 2587. doi: 10.1088/0031-9155/57/9/2587 [DOI] [PubMed] [Google Scholar]
  • 18. Van Calster B, Wynants L, Verbeek JFM, et al. Reporting and interpreting decision curve analysis: a guide for investigators[J]. Eur Urol, 2018, 74(6): 796-804. doi: 10.1016/j.eururo.2018.08.038 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Collins GS, Dhiman P, Andaur Navarro CL, et al. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence[J]. BMJ Open, 2021, 11(7): e048008. doi: 10.1136/bmjopen-2020-048008 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Islam MM, Rahman MJ, Rabby MS, et al. Predicting the risk of diabetic retinopathy using explainable machine learning algorithms[J]. Diabetes Metab Syndr, 2023, 17(12): 102919. doi: 10.1016/j.dsx.2023.102919 [DOI] [PubMed] [Google Scholar]
  • 21. Petch J, Di S, Nelson W. Opening the black box: the promise and limitations of explainable machine learning in cardiology[J]. Can J Cardiol, 2022, 38(2): 204-13. doi: 10.1016/j.cjca.2021.09.004 [DOI] [PubMed] [Google Scholar]
  • 22. Poujois A, Woimant F. Wilson’s disease: a 2017 update[J]. Clin Res Hepatol Gastroenterol, 2018, 42(6): 512-20. doi: 10.1016/j.clinre.2018.03.007 [DOI] [PubMed] [Google Scholar]
  • 23. 张梦影,赵晨玲,田丽伟,等. 肝豆扶木汤通过GPX4/ACSL4/ALOX15通路抑制铁死亡改善Wilson病小鼠的肝脏脂肪变性[J]. 南方医科大学学报, 2025, 45(7): 1471-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Zou LX, Wang X, Hou ZL, et al. Machine learning algorithms for diabetic kidney disease risk predictive model of Chinese patients with type 2 diabetes mellitus. Ren Fail. 2025, 47(1): 2486558. doi: 10.1080/0886022x.2025.2486558 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25. Tang S, Ren F, Hou W, et al. Uncovering the critical role of cuproptosis in Wilson disease: insights into potential therapeutic targets[J]. J Cell Mol Med, 2025, 29(21): e70946. doi: 10.1111/jcmm.70946 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Zischka H, Einer C. Mitochondrial copper homeostasis and its derailment in Wilson disease[J]. Int J Biochem Cell Biol, 2018, 102: 71-5. doi: 10.1016/j.biocel.2018.07.001 [DOI] [PubMed] [Google Scholar]
  • 27. Stefano JT, Guedes LV, de Souza AAA, et al. Usefulness of collagen type IV in the detection of significant liver fibrosis in nonalcoholic fatty liver disease[J]. Ann Hepatol, 2021, 20: 100253. doi: 10.1016/j.aohep.2020.08.070 [DOI] [PubMed] [Google Scholar]
  • 28. Ge HM, Ge H, Tian M, et al. Extracellular matrix stiffness: new areas affecting cell metabolism[J]. Front Oncol, 2021, 11: 631991. doi: 10.3389/fonc.2021.631991 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. 李 楠, 王雪莹, 郭佳桐, 等. 中青年人群非酒精性脂肪肝发生风险预测模型的建立[J]. 中国慢性病预防与控制, 2021, 29(3): 167-71. doi: 10.16386/j.cjpccd.issn.1004-6194.2021.03.002 [DOI] [Google Scholar]
  • 30. Tutunchi H, Saghafi-Asl M, Asghari-Jafarabadi M, et al. The relationship between severity of liver steatosis and metabolic parameters in a sample of Iranian adults[J]. BMC Res Notes, 2020, 13(1): 218. doi: 10.1186/s13104-020-05059-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Gao DJ, Zeng T, Chong YT, et al. Copper and hepatic lipid dysregulation: Mechanisms and implications[J]. World J Hepatol, 2025, 17(8): 107803-12. doi: 10.4254/wjh.v17.i8.107803 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. 田 盟, 严 俊, 李 汛. 胆汁酸受体在非酒精性脂肪性肝病中的作用[J]. 中国生物化学与分子生物学报, 2022, 38(5): 587-94. [Google Scholar]
  • 33. Gillard J, Clerbaux LA, Nachit M, et al. Bile acids contribute to the development of non-alcoholic steatohepatitis in mice[J]. JHEP Rep, 2022, 4(1): 100387. doi: 10.1016/j.jhepr.2021.100387 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Michalak A, Guz M, Kozicka J, et al. Red blood cell distribution width derivatives in alcohol-related liver cirrhosis and metabolic-associated fatty liver disease[J]. World J Gastroenterol, 2022, 28(38): 5636-47. doi: 10.3748/wjg.v28.i38.5636 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Aslam H, Oza F, Ahmed K, et al. The role of red cell distribution width as a prognostic marker in chronic liver disease: a literature review[J]. Int J Mol Sci, 2023, 24(4): 3487. doi: 10.3390/ijms24043487 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36. Li L, Yu JX, Zhou ZW. Association between platelet indices and non-alcoholic fatty liver disease: a systematic review and meta-analysis[J]. Rev Esp Enferm Dig, 2024, 116(5): : 264-73. [DOI] [PubMed] [Google Scholar]
  • 37. Deng L, Lu KJ, Hu HH. An interpretable LightGBM model for predicting coronary heart disease: Enhancing clinical decision-making with machine learning[J]. PLoS One, 2025, 20(9): e0330377. doi: 10.1371/journal.pone.0330377 [DOI] [PMC free article] [PubMed] [Google Scholar]

Articles from Journal of Southern Medical University are provided here courtesy of Editorial Department of Journal of Southern Medical University

RESOURCES