机器学习算法筛选髋部骨折延长术后住院时间特征变量及预测模型构建  

The machine learning algorithm screened the characteristic variables of prolonged hospital stay after hip fracture and constructed the prediction model

作　　者：苏长英 黄丽莎 钟洁斯 彭蓝梅 王静茹 高辉 杨俊 Su Changying;Huang Lisha;Zhong Jiesi;Peng Lanmei;Wang Jingru;Gao Hui;Yang Jun(Department of Joint Sports Medicine,Yingtan 184th Hospital,Yingtan 335000,China;Nursing Department,Yingtan 184th Hospital,Yingtan 335000,China)

机构地区：[1]鹰潭一八四医院关节运动医学科,鹰潭335000 [2]鹰潭一八四医院护理部,鹰潭335000

出　　处：《中国实用护理杂志》2024年第19期1454-1461,共8页Chinese Journal of Practical Nursing

摘　　要：目的通过机器学习算法识别髋部骨折患者延长术后住院时间(PPOLOS)风险变量并构建Nomogram模型。方法采用回顾性病例对照研究,以方便抽样法选取鹰潭一八四医院2019年6月至2023年6月诊治的248例髋部骨折患者为研究对象。使用2种机器学习算法最小绝对收缩选择算子(LASSO)和支持向量机-递归特征消除(SVM-RFE)筛选PPOLOS风险变量。基于交集风险变量构建预测髋部骨折患者PPOLOS风险的Nomogram模型。采用内部数据集进行模型的验证。结果248例患者男79例,女169例,年龄(64.49±8.02)岁。平均术后住院时间为(7.98±5.68)d,中位值7 d。LASSO算法识别出7个风险变量;SVM-RFE算法识别出8个风险变量。交集风险变量为:年龄、体质量指数(BMI)、Charlson共病指数(CCI)、手术类型和中性粒细胞与淋巴细胞比值(NLR)。多因素Logistic回归(交集风险变量)分析结果显示年龄[OR(95%CI)为1.649(1.235~2.202),P<0.05]、BMI[OR(95%CI)为1.603(1.204~2.134),P<0.05]、CCI[OR(95%CI)为1.670(1.236~2.258),P<0.05]、手术类型[OR(95%CI)为1.620(1.209~2.170)、1.699(1.243~2.321),均P<0.05]和NLR[OR(95%CI)为3.258(2.299~4.617),P<0.05]与PPOLOS风险独立相关。内部数据集验证结果显示:Nomogram模型一致性指数为0.865(95%CI 0.768~0.945);曲线下面积为0.852(95%CI 0.748~0.962);当风险阈值>0.08时,能提供显著临床净收益;临床影响曲线显示在高风险群体中能有效识别出PPOLOS患者。结论Nomogram模型能指导医护人员尽早做出临床诊疗决策以规避风险并实现合理分配医疗资源目的,提高护理质量。Objective To identify risk variables for prolonged postoperative length of stay(PPOLOS)in hip fracture patients by machine learning algorithms and construct Nomogram models.Methods A retrospective case-control study was conducted to select 248 patients with hip fracture diagnosed and treated in Yingtan 184th Hospital from June 2019 to June 2023 by convenient sampling method.Two machine learning algorithms were used(least absolute shrinkage and selection operator,LASSO and support vector machine-Recursive Feature Elimination,SVM-RFE)to screen PPOLOS risk variables.Construct a Nomogram model to predict the risk of PPOLOS in patients with hip fracture based on intersection risk variables.The model was validated using internal data sets.Results Among the 248 patients with hip fracture,there were 79 males and 169 females,aged(64.49±8.02).The mean postoperative length of hospital stay of 248 patients was(7.98±5.68)d,and the median was 7 d.LASSO algorithm identifies 7 risk variables,the SVM-RFE algorithm identified 8 risk variables.Intersectional risk variables were age,body mass index(BMI),Charlson comorbidity index(CCI),type of surgery,and central granulocytocyte ratio(NLR).Multivariate Logistic regression analysis(intersection risk variables)showed that age[OR=1.649(1.235-2.202)],BMI[1.603(1.204-2.134)],CCI[OR=1.670(1.236-2.258)],type of surgery[OR=1.620(1.209-2.170),1.699(1.243-2.321)],and NLR[OR=3.258(2.299-4.617)]were independently associated with the risk of PPOLOS(all P<0.05).The results showed that the conformity index of the Nomogram model was 0.865(95%CI=0.768-0.945).The area under the curve was 0.852(95%CI=0.748-0.962).When the risk threshold was>0.08,it could provide significant clinical net benefit.The clinical impact curve showed effective identification of PPOLOS patients in high-risk groups.Conclusions This Nomogram model can guide medical staff to make clinical diagnosis and treatment decisions as soon as possible to avoid risks,allocate medical resources rationally,and improve nursing quality.

关 键 词：机器学习 髋部骨折 术后 住院时间 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] R687.3[自动化与计算机技术—控制科学与工程]