糖尿病足溃疡复发风险预测模型的构建：基于 Logistic 回归、支持向量机和 BP 神经网络模型
onstruction of Recurrence Risk Prediction Model for Diabetic Foot Ulcer on the Basis of Logistic Regression，Support Vector Machine and BP Neural Network Model
背景 全球范围内糖尿病足溃疡（DFUs）首次复发与再次复发率逐年上升，且早期复发风险高于远期风险。导致 DFUs 复发的风险因素较多，目前缺乏系统的筛选，因此需要探索 DFUs 复发的危险因素，以便早期识别出复发高危人群。目的 探讨 Logistic 回归、支持向量机（SVM）和 BP 神经网络（BPNN）模型在糖尿病足溃疡（DFUs）复发风险中的预测价值。方法 选取 2020 年 1 月—2022 年 12 月在宁夏医科大学总医院烧伤整形美容科和伤口造口门诊就诊的 DFUs 患者 390 例作为开发模型的研究对象。根据患者出院后 1 年内 DFUs 是否复发分为复发组116 例（29.7%）和非复发组 274 例（70.3%）。收集两组患者的一般资料包括社会人口学特征、病史评估和临床病例资料并进行比较，采用糖尿病足部自我管理行为量表（DFSBS）评估患者糖尿病足部自我管理行为，采用慢性病风险感知问卷评估患者 DFUs 风险感知水平。采用多因素 Logistic 回归分析探讨 DFUs 患者出院后 1 年内 DFUs 复发的影响因素；将患者按照 70:30 划分为训练集和测试集，运用 Logistic 回归变量筛选策略，分别建立 Logistic 回归、SVM 和BPNN 模型；绘制各模型预测 DFUs 复发风险的受试者工作特征（ROC）曲线，采用查准率、召回率、正确率、F1 指数及 ROC 曲线下面积（AUC）比较各模型的预测效能。结果 两组 DFUs 患者年龄、BMI、发病后就诊时间、住院时间、吸烟史、足溃疡分级、受累足趾截肢史、溃疡位置在脚底、独居、足部存在行走障碍、溃疡因外伤引起、踝肱指数、糖尿病周围神经病变、下肢动脉粥样硬化、多重耐药菌感染、糖化血红蛋白、白蛋白、足部护理行为、风险感知水平比较，差异均有统计学意义（P<0.05）。多因素 Logistic 回归分析结果显示，住院时间〔OR=0.678，95%CI（0.557，0.826），P<0.001〕、 溃 疡 位 置 在 脚 底〔OR=0.078，95%CI（0.011，0.541），P=0.010〕、 独 居〔OR=5.689，95%CI（2.583，10.726），P<0.001〕、足部存在行走障碍〔OR=3.364，95%CI（2.742，5.638），P<0.001〕、糖尿病周围神经病变〔OR=3.089，95%CI（1.156，4.585），P=0.025〕、下肢动脉粥样硬化〔OR=4.033，95%CI（3.688，9.060），P<0.001〕、多重耐药菌感染〔OR=3.241，95%CI（1.728，7.361），P<0.001〕、糖化血红蛋白〔OR=0.209，95%CI（0.065，0.669），P=0.008〕、白蛋白〔OR=2.038，95%CI（1.515，2.741），P<0.001〕、足部护理行为〔OR=1.965，95%CI（0.874，3.208），P=0.014〕和风险感知水平〔OR=0.261，95%CI（0.197，0.825），P=0.002〕是 DFUs 患者 1 年内 DFUs 复发的影响因素（P<0.05）。Logistic 回归、SVM 和 BPNN 模型在测试集中预测 DFUs 复发风险的正确率分别 82.03%、94.87%、87.17%，AUC 分别为 0.843、0.930、0.800。Logistic 回归、SVM 和 BPNN 模型预测 DFUs 复发风险的 ROC 曲线 AUC 比较，差异有统计学意义（Z=8.826，P<0.05）；SVM 模型预测 DFUs 复发风险的 ROC 曲线 AUC 高于 Logistic 回归和 BPNN 模型（Z=5.672，P=0.014；Z=2.767，P<0.001）。结论 SVM 模型预测 DFUs 患者出院后 1 年内 DFUs 复发风险的正确率、灵敏度、特异度、AUC 等指标均较好，为相对最优的模型，建议进一步推广应用以验证预测模型的效能。
Background The rates of first and multiple recurrence of diabetic foot ulcersDFUsare increasing annually worldwideand the risk of early recurrence is higher than the distant recurrence. There are numerous risk factors for DFUs recurrenceand there is a lack of systematic screening. Thereforethere is a need to explore the risk factors for DFUs recurrence in order to identify high-risk population of recurrence at an early stage. Objective To explore the predictive value of Logistic regressionLRsupport vector machineSVMBP neural network modelBPNN in the recurrence risk of DFUs. Methods From January 2020 to December 2022a total of patients with DFUs attending the burn plastic and wound ostomy outpatient department in General Hospital of Ningxia Medical University were selected as the research objects and divided into the recurrence groupn=11629.7% and non-recurrence groupn=27470.3% according to the recurrence of DFUs within 1 year after discharge. General information was collected and compared between the two groups of patientsincluding sociodemographic characteristicsmedical history assessment and clinical case information. The Diabetes Foot Selfcare Behavior Scale DFSBS was used to assess the self-management behavior of diabetes foot in patients and chronic diseases risk perception questionnaire was used to assess the risk perception level of DFUs among patients. Multivariable Logistic regression analysis was used to explore the influencing factors of DFUs recurrence in patients within 1 year after discharge. The patients were divided into training and test sets according to the ratio of 70 to 30the LRSVM and BPNN recurrence risk prediction models were developed based on Logistic regression variable screening strategy. The receiver operating characteristicROC curves of each model to predict the recurrence risk of DFUs were plottedand the predictive efficacy of each model was compared using the precision raterecall rateaccuracy rateF1 index and area under curveAUC. Results There were significant differences in ageBMItime to visit after onsetlength of hospital stayhistory of smokingclassification of diabetic foot ulcershistory of involved toe amputationsole ulcerliving alonewalking impairmentulcers due to traumaankle-brachial indexdiabetic peripheral neuropathylower limb atherosclerosismultidrug-resistant bacteria infectionglycated hemoglobinalbuminfoot care behaviourlevel of risk perception in both groups of DFUs patientsP<0.05. Multivariable Logistic regression analysis showed that the length of hospital stayOR=0.67895%CI0.5570.826P<0.001sole ulcerOR=0.07895%CI0.0110.541P=0.010living aloneOR=5.68995%CI2.58310.726P<0.001walking impairmentOR=3.36495%CI2.7425.638P<0.001diabetic peripheral neuropathyOR=3.08995%CI1.1564.585P=0.025lower limb atherosclerosisOR=4.03395%CI3.6889.060P<0.001multidrug-resistant bacteria infectionOR=3.24195%CI1.7287.361P<0.001glycated haemoglobin OR=0.20995%CI0.0650.669P=0.008albuminOR=2.03895%CI1.5152.741P<0.001foot care behaviourOR=1.96595%CI0.8743.208P=0.014and level of risk perceptionOR=0.26195%CI0.1970.825P=0.002were influencing factors of the recurrence of DFUs within 1 yearP<0.05. The accuracy rates of LRSVM and BPNN models to predict the recurrence risk of DFUs in the test sets were 82.03%94.87% and 87.17%with AUCs of 0.8430.930 and 0.800respectively. There were significant differences in AUC of ROC curves of LRSVM and BPNN recurrence risk prediction models of DFUsZ=8.826P<0.05the AUC of ROC curve of SVM recurrence risk prediction model was higher than the LR and BPNN modelsZ=5.672P=0.014Z=2.767P<0.001. Conclusion SVM model can predict the recurrence risk of DFUs patients within 1 year after discharge with good accuracy ratesensitivityspecificityAUC and other indicatorswhich is the relative optimal model. It is recommended to further promote and apply the prediction model to verify its effectiveness.
张娟，李海芬，李小曼，姚苗，马惠珍，马强 *
10.12114/j.issn.1007-9572.2023.0175
临床医学内科学
糖尿病足溃疡糖尿病足复发Logistic 模型支持向量机模型BP 神经网络模型影响因素分析
张娟，李海芬，李小曼，姚苗，马惠珍，马强 *.糖尿病足溃疡复发风险预测模型的构建：基于 Logistic 回归、支持向量机和 BP 神经网络模型[EB/OL].(2023-07-25)[2025-08-02].https://chinaxiv.org/abs/202307.00695.点此复制
评论