目的:肝细胞癌(HCC)在中国的发病率上升是一个紧迫的问题,需要早期诊断和治疗。本研究旨在通过将机器学习(ML)技术与人口统计、病史,和非侵入性生物标志物数据。这些模型可以增强医生对乙型肝炎病毒(HBV)相关性肝硬化患者的决策能力,这些患者的血清甲胎蛋白(AFP)水平较低。
方法:纳入2012年1月至2018年12月期间接受治疗的6,980例患者。获得了治疗前的实验室检查和临床数据。确定了HCC的重要危险因素,并使用ML和单变量回归分析计算影响其诊断的每个变量的相对风险。然后将数据集随机划分为验证集(20%)和训练集(80%)以开发ML模型。
结果:使用高斯朴素贝叶斯确定了肝癌的12个独立危险因素,极端梯度提升(XGBoost),随机森林,和最小绝对收缩和选择操作回归模型。多变量分析显示,男性,年龄>60岁,碱性磷酸盐>150U/L,AFP>25ng/mL,癌胚抗原>5ng/mL,纤维蛋白原>4g/L是危险因素,而高血压,钙<2.25mmol/L,钾≤3.5mmol/L,直接胆红素>6.8μmol/L,血红蛋白<110g/L,谷丙转氨酶>40U/L是肝癌患者的保护因素。基于这些因素,构造了一个列线图,显示0.746的曲线下面积(AUC)(灵敏度=0.710,特异性=0.646),显著高于AFPAUC0.658(灵敏度=0.462,特异度=0.766)。与几种ML算法相比,XGBoost模型的AUC为0.832(敏感性=0.745,特异性=0.766),独立验证AUC为0.829(敏感性=0.766,特异性=0.737),使其成为两组中表现最好的模型。外部验证结果证明了XGBoost模型的准确性。
结论:提出的XGBoost显示了在低水平AFP的HBV相关肝硬化患者中个体化预测HCC的有希望的能力。
OBJECTIVE: The increasing incidence of hepatocellular carcinoma (HCC) in
China is an urgent issue, necessitating early diagnosis and treatment. This study aimed to develop personalized predictive models by combining machine learning (ML) technology with a demographic, medical history, and noninvasive biomarker data. These models can enhance the decision-making capabilities of physicians for HCC in hepatitis B virus (HBV)-related cirrhosis patients with low serum alpha-fetoprotein (AFP) levels.
METHODS: A total of 6,980 patients treated between January 2012 and December 2018 were included. Pre-treatment laboratory tests and clinical data were obtained. The significant risk factors for HCC were identified, and the relative risk of each variable affecting its diagnosis was calculated using ML and univariate regression analysis. The data set was then randomly partitioned into validation (20 %) and training sets (80 %) to develop the ML models.
RESULTS: Twelve independent risk factors for HCC were identified using Gaussian naïve Bayes, extreme gradient boosting (XGBoost), random forest, and least absolute shrinkage and selection operation regression models. Multivariate analysis revealed that male sex, age >60 years, alkaline phosphate >150 U/L, AFP >25 ng/mL, carcinoembryonic antigen >5 ng/mL, and fibrinogen >4 g/L were the risk factors, whereas hypertension, calcium <2.25 mmol/L, potassium ≤3.5 mmol/L, direct bilirubin >6.8 μmol/L, hemoglobin <110 g/L, and glutamic-pyruvic transaminase >40 U/L were the protective factors in HCC patients. Based on these factors, a nomogram was constructed, showing an area under the curve (AUC) of 0.746 (sensitivity=0.710, specificity=0.646), which was significantly higher than AFP AUC of 0.658 (sensitivity=0.462, specificity=0.766). Compared with several ML algorithms, the XGBoost model had an AUC of 0.832 (sensitivity=0.745, specificity=0.766) and an independent validation AUC of 0.829 (sensitivity=0.766, specificity=0.737), making it the top-performing model in both sets. The external validation results have proven the accuracy of the XGBoost model.
CONCLUSIONS: The proposed XGBoost demonstrated a promising ability for individualized prediction of HCC in HBV-related cirrhosis patients with low-level AFP.