基于灰色关联分析及BP神经网络的下肢骨折术前深静脉血栓形成风险研究

Study on the risk of preoperative deep vein thrombosis after lower limb fracture based on grey relational analysis and BP neural network

目的 探究运用下肢骨折患者术前血液指标构建人工智能算法模型在下肢深静脉血栓形成(deep vein thrombosis, DVT)中的预测效能。方法 回顾性选择2018年1月—2022年12月于德阳市人民医院骨科治疗的下肢骨折患者,收集患者年龄、性别、身高、体重等基础和临床资料,以及入院时实验室检查指标,计算中性粒细胞与淋巴细胞比值(neutrophi to lymphocyte ratio, NLR)、单核细胞与淋巴细胞比值(monocyte to lymphocyte ratio, MLR)和血小板与淋巴细胞比值(platelet to lymphocyte ratio, PLR),并根据入院时血管彩色多普勒超声提示是否有DVT将患者分为DVT组和非DVT。进行数据预处理后采用灰色关联分析(grey relational analysis, GRA)筛选DVT的重要预测特征组合模型,建立BP神经网络预测模型,最后评价模型的精确度,同时评估不同模型在DVT临床预测中的效能。结果 最终纳入下肢骨折患者4 033例,其中DVT组3 127例,非DVT组906例。GRA选择了7个重要预测特征,分别为淋巴细胞绝对值、NLR、MLR、PLR、血浆D-二聚体、直接胆红素、总胆红素;logistic回归分析、随机森林、决策树、单一BP神经网络以及GRA-BP神经网络组合模型精确度分别为74%、76%、75%、84%、87%,以GRA-BP神经网络组合模型的精确度最高。结论 该研究所选GRA-BP神经网络在下肢骨折患者术前DVT风险预测中的精确度最高,能够为DVT预防策略制订提供参考依据。

Objective To explore the efficiency of artificial intelligence algorithm model using preoperative blood indexes on the prediction of deep vein thrombosis(DVT)in patients with lower limb fracture before operation.Methods Patients with lower limb fracture treated in the Department of Orthopedics of Deyang People’s Hospital between January 2018 and December 2022 were retrospectively selected.Their basic and clinical data such as age,gender,height and weight,and laboratory examination indicators at admission were collected,then the neutrophi to lymphocyte ratio(NLR),monocyte to lymphocyte ratio(MLR),and platelet to lymphocyte ratio(PLR)were calculated.According to color Doppler ultrasound indication of DVT in lower extremities at admission,the patients were divided into DVT group and non-DVT group.After data preprocessing,grey relational analysis(GRA)was used to screen the combination model of important predictive features of DVT,and BP neural network prediction model was established using the selected features.Finally,the accuracy of BP neural network prediction model was evaluated,and was compared with those of different models in clinical prediction of DVT.Results A total of 4033 patients with lower limb fracture were enrolled,including 3127 cases in the DVT group and 906 cases in the non-DVT group.GRA selected seven important predictive features:absolute lymphocyte value,NLR,MLR,PLR,plasma D-dimer,direct bilirubin,and total bilirubin.The accuracies of logistic regression analysis,random forest,decision tree,BP neural network and GRA-BP neural network combination model were 74%,76%,75%,84%and 87%,respectively.The GRA-BP neural network combination model had the highest accuracy.Conclusion The GRA-BP neural network selected in this paper has the highest accuracy in preoperative DVT risk prediction in patients with lower limb fracture,which can provide a reference for the formulation of DVT prevention strategies.