基于LASSO-机器学习结合CT-FFR构建冠状动脉支架内再狭窄的诺模图预测模型

Construction of a nomogram prediction model for coronary in-stent restenosis based on LASSO-machine learning combined with CT-FFR

摘要:

目的]基于冠状动脉CT-血流储备分数(CT-FFR)联合机器学习(ML)方法,构建冠状动脉支架内再狭窄(ISR)的诺模图预测模型,以评估ISR发生风险。 [方法]回顾性分析2022年1月—2025年1月期间于本院行经皮冠状动脉介入治疗(PCI)术后复查的患者。依据排除标准共纳入210例患者,其中ISR患者100例,非ISR患者110例。按7∶3的比例随机将其分为训练集和测试集,经单变量分析筛选潜在预测因子后,通过LASSO回归获取非零系数特征变量,再经随机森林(RF)、支持向量机(SVM)、极限梯度提升算法(XGB)三种ML算法对显著因素进行重要性排序,取各算法前10的变量交集作为输入,再经双向逐步多因素Logistic回归筛选后构建ISR风险评分,并以诺模图可视化。 [结果]经LASSO回归筛选出14个预测因子,分别为舒张压、C反应蛋白、甘油三酯(TG)、N末端脑钠肽前体(NT-proBNP)、低密度脂蛋白胆固醇(LDLC)、最小支架直径＜3 mm、收缩压、ΔCT-FFR、CT-FFR、白细胞介素6(IL-6)、体质指数、糖化血红蛋白(HbA1c)、高血压史以及高密度脂蛋白胆固醇(HDLC)。进一步通过3种ML算法及Logistic回归逐步筛选,最终确定6个ISR危险因素:高ΔCT-FFR、IL-6、NT-proBNP、TG和CT-FFR值以及最小支架直径＜3 mm；训练集和测试集曲线下面积分别为0.995(95%CI:0.989～1.000)和0.965(95%CI:0.927～1.000),决策曲线显示其分别在0～1.00及0～0.92范围内净收益高,整合6个预测因子的诺模图准确度高,临床实用性强。 [结论]基于LASSO-ML结合CT-FFR技术的ISR诺模图预测模型对ISR具有较高的准确性和临床实用性。

Abstract:

Aim Based on coronary CT-fractional flow reserve (CT-FFR) combined with machine learning methods, a nomogram prediction model for coronary in-stent restenosis (ISR) was developed to assess the risk of ISR. Methods Retrospective analysis was performed on patients who underwent re-examination after PCI at our hospital from January 2022 to January 2025. According to the exclusion criteria, a total of 210 patients were enrolled, including 100 cases of ISR and 110 cases of non-ISR. The dataset was randomly divided into training and test sets at a 7∶3 ratio. After univariate analysis to screen potential predictors, LASSO regression was applied to identify feature variables with non-zero coefficients. Subsequently, three machine learning (ML) algorithms including random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGB) were used to rank the importance of the significant factors. The intersection of the top 10 variables from each algorithm was used as input for bidirectional stepwise multivariate Logistic regression. An ISR risk score was then constructed and visualized using a nomogram. Results A total of 14 predictive factors were identified through LASSO regression, including diastolic blood pressure, C-reactive protein, triglycerides (TG), N-terminal pro-brain natriuretic peptide (NT-proBNP), low density lipoprotein cholesterol (LDLC), minimum stent diameter＜3 mm, systolic blood pressure, ΔCT-FFR, CT-FFR, interleukin-6 (IL-6), body mass index, glycosylated hemoglobin (HbA1c), history of hypertension, and high density lipoprotein cholesterol (HDLC). Following stepwise screening using three ML algorithms and Logistic regression, six independent risk factors for ISR were identified:elevated ΔCT-FFR, IL-6, NT-proBNP, TG and CT-FFR values, and minimum stent diameter＜3 mm. The area under the curve for the training set and test set were 0.995 (95%CI:0.989～1.000) and 0.965 (95%CI:0.927～1.000), respectively.Decision curve analysis demonstrated high net benefit across threshold probabilities of 0～1.00 in the training set and 0～0.92 in the test set. The nomogram integrating these six predictors exhibited high accuracy and clinical utility. Conclusion The ISR nomogram prediction model based on LASSO-ML combined with CT-FFR technology has high accuracy and clinical utility for ISR.