基于Fisher判别构建CT引导下肺穿刺活检并发症的预测模型

The establishment of a model for predicting complications of CT-guided percutaneous transthoracic needle biopsy based on Fisher discriminant

目的探讨基于机器学习算法的Fisher判别,初步构建CT引导下经皮肺穿刺活检(PTNB)并发症的预测模型。方法回顾性分析227例CT引导下PTNB的肺部肿块或结节,用前187例筛选并发症危险因素,纳入有统计学意义的指标,构建Fisher判别式,然后采用交叉核实法和后40例评估预测模型。结果187例中出现并发症48例(25.7%),主要为气胸29例(15.5%)和肺出血26例(13.9%),其中包含有气胸合并肺出血7例(3.74%)。并发症的危险因素有病灶大小、合并肺气肿、病灶中心与膈面短径、穿刺深度、穿刺角度和穿刺次数,并设参数值:X1=病灶大小(0>2 cm;1≤2 cm)、X2=合并肺气肿等(0=是;1=否)、X3=病灶中心距离膈面短径(0>3 cm;1≤3 cm)、X4=穿刺深度(0≤5 cm;1>5 cm)、X5=穿刺胸膜角度(0≤50°;1>50°)、X6=穿刺时间(0≤20 min;1>20 min)、X7=穿刺次数(0=1次;1≥2次)。所得并发症的非标准化Fisher判别公式为Z=1.531X1+1.531X2+2.123X3+1.390X4+1.564X5+0.903X6+1.716X7-3.114,判别界值为0.514。预测模型的交叉核实法和40例实践测评的误判率分别是10.2%和7.5%,准确率为89.8%和92.5%,敏感度为85.4%和88.9%,特异度为91.4%和93.5%。结论Fisher判别模型可以用于辅助临床预测CT引导下PTNB并发症的发生概率。

Objective To discuss on Fisher’s discriminant based on machine learning algorithm,and to construct a model for predicting complications in CT-guided percutaneous transthoracic needle biopsy(PTNB).Methods The clinical data of 227 patients with pulmonary mass or nodule,who received CT-guided PTNB,were retrospectively analyzed.The first 187 patients were used to screen the risk factors of complications,and the statistically significant indicators were selected to construct the Fisher discriminant.Then the cross-validation method and the relevant data of the remaining 40 patients were used to evaluate the prediction model.Results Of the 187 patients,48 had complications(25.7%),including pneumothorax(n=29,15.5%)and pulmonary hemorrhage(n=26,13.9%),among them 7 patients(3.7%)had pneumothorax associated with pulmonary hemorrhage.The risk factors for complications included the size of the lesion,coexisting emphysema,the short distance between the center of the lesion and the diaphragmatic surface,the depth of puncturing,the puncture-pleural angle and the number of puncture times.The values of relevant parameters were set as follows:X1=lesion size(0>2 cm;1≤2 cm),X2=coexisting emphysema,etc.(0=yes;1=no),X3=short distance between the center of the lesion and the diaphragmatic surface(0>3 cm;1≤3 cm),X4=puncturing depth(0≤5 cm);1>5 cm),X5=puncture-pleural angle(0≤50°;1>50°),X6=time spent for puncturing(0≤20 min,1>20 min),and X7=the number of puncture times(0=1 time;1≥2 times).The non-standardized Fisher discriminant formula for complications thus obtained was Z=1.531X1+1.531X2+2.123X3+1.390X4+1.564X5+0.903X6+1.716X7-3.114,and the discriminant boundary value was 0.514.The misjudgment rates of cross-validation method with the prediction model and the actual evaluation results of 40 patients were 10.2%and 7.5%respectively,the accuracy rates were 89.8%and 92.5%respectively,the sensitivities were 85.4%and 88.9%respectively,and the rates of specificity were 91.4%and 93.5%respectively.Conclusion The Fisher discriminant model can be used to assist clinical prediction of the occurrence probability of complications in CT-guided PTNB.