Development And Validation Of A Risk Prediction Model For Chronic Pain In Elderly Patients Undergoing Orthopedic Surgery

Objective Chronic Postsurgical Pain(CPSP)is a disease with high incidence and great harm.This study analyzed the occurrence status and risk factors of CPSP in elderly patients undergoing orthopedic surgery,and adopt a variety of Machine Learning(ML)and Synthetic Minority Oversampling Technique(SMOTE)algorithm,aim to explore the prevalence and severity of CPSP in elderly patients undergoing orthopedic surgery,and to develop a best prediction model of CPSP for elderly patients undergoing orthopedic surgery with clinical universality and extension.Methods This study is divided into three parts.Part I: Elderly patients(≥60 years)who underwent elective orthopedic surgery in our hospital from January 1,2020 to October19,2020 were selected.Based on evidence-based medical evidence,a structured CPSP questionnaire was designed,and the patients were followed up by telephone 3 months after surgery,including:(1)According to the 11 th Revision of International Classification of Diseases(ICD-11)of International Association for the Study of Pain(IASP)in 2019,diagnose whether patients have CPSP;(2)Numerical Rating Scale(NRS)was used to evaluate the pain degree of CPSP.The heaviest pain degree reported by patients(including resting,sports and cough,etc.)was taken as the pain degree of CPSP.According to the NRS score,they were divided into three grades: mild pain(NRS 1-3),moderate pain(NRS 4-6)and severe pain(NRS 7-10).(3)The nature of CPSP was evaluated with the Douleur Neuropathique 4questions(DN4).The scale contained 4 questions and a total of 10 options.If each option answered "yes",it was recorded as 1 point.Neuropathic Pain(NP)was diagnosed when respondents scored≥4.(4)The effects of CPSP on body function were evaluated by using Brief Pain Inventory(BPI),including daily activities,mood,mobility,interpersonal relationship,sleep and enjoyment of life.Part II: Bidirectional cohort study was used to collect model training set samples(January 1,2020 to October 19,2020),and prospective cohort study was used to collect model validation set samples(October 20,2020 to January31,2021).Data collected demographic and clinical characteristics as independent variables,including: age,sex,spouse or not,body mass index(BMI),education,smoking history,drinking history,hypertension,diabetes,preoperative pain history(at both the surgical site and non-surgical site),preoperative inflammatory state,American Society of Anesthesiologist(ASA)classification,anesthesia method,type of surgery,operation duration,intraoperative blood loss,postoperative analgesia,Acute Postsurgical Pain(APSP)or not,and postoperative hospital stay.Patients were divided into two groups according to whether they were diagnosed with CPSP 3 months after surgery,to analyze the risk factors for CPSP by univariate analysis and multivariate Logistic regression analysis.At the same time,construct and evaluate six CPSP prediction models for elderly orthopedic surgery patients preliminarily based on ML algorithm,including Logistic Regression(LR),Decision Tree(DT),Random Forests(RF),Extra Trees(ET),Adaptive Boosting(Ada Boost)and e Xtreme Gradient Boosting(XGBoost).After cross-validation in the training set,the accuracy,precision,recall,F1-score and Area Under Curve(AUC)of the model were evaluated.The appropriate model was determined according to the recall,and its generalization ability was tested in the validation set.Part III: Analyze the defects and deficiencies of the model,and further update the model with SMOTE algorithm,use the training data to build LR,DT,RF,ET,Ada Boost and XGBoost models for elderly orthopaedic surgery patients.Using cross-validation to do hyperparameter tuning after data balance by SMOTE,get the best model by cross validation evaluation with recall as the optimal standard.After the generalization capability is verified by external data,the model is visualized by a web.Compare the change of the evaluation indicators by cross validation of all models before and after SMOTE algorithm,judge whether SMOTE algorithm is effective for model update by the increase rate of recall.Results A total of 1227 elderly patients undergoing orthopedic surgery were enrolled in this study.In chronological order,815 in the training set and 412 in the validation set.Part I:The incidence of CPSP in elderly orthopedic patients was 30.8% and there was no difference in the incidence of CPSP among different types of orthopedic surgery(P>0.05).Moderate and severe CPSP pain accounted for 41%,in the same type of surgery,there was statistically significant difference between mild,moderate and severe pain(P<0.001).12% of patients had neuropathic pain(NP),and there was no difference between different types of orthopedic surgery(P>0.05),the degree of pain was positively correlated with the occurrence of NP(P<0.001).The occurrence of CPSP in elderly patients after orthopedic surgery has a significant impact on body function(P<0.001),in which daily activities and actions are most affected,CPSP had no difference in body function among different surgical types(P>0.05).Moderate and severe pain degree had more significant effects on body function(P<0.001).Part II: The independent risk factors for CPSP in elderly orthopedic surgery patients,in descending order of correlation,are: APSP(OR 8.151),preoperative non-operative area pain(OR 3.987),other limb surgery(OR 3.260),preoperative operative area pain(OR 2.393),prolonged hospitalization(OR 1.053),and no spouse(OR 0.409).The six CPSP prediction models for elderly orthopedic surgery patients constructed based on ML algorithm all showed good prediction effects.The cross-validation results of each evaluation index showed that: in terms of accuracy,DT,ET and Ada Boost models performed similarly(0.78),and were better than the other three models in predicting correct classification(CPSP and non-CPSP).In terms of precision,ET and Ada Boost models performed best(0.69).In terms of recall,DT model was the best(0.51),which correctly detected the highest proportion of CPSP occurrence;ET and XGBoost models had the highest F1-Score(0.56),while LR,ET and Ada Boost models had higher AUC(0.80).With recall as the selection criterion,DT model performed better in identifying CPSP patients(0.51),and was an appropriate model for predicting the occurrence of CPSP in elderly orthopaedic surgery patients.The DT model performed well in the external validation concentration as a whole(accuracy =0.76,accuracy=0.71,recall =0.44,F1-score=0.54,AUC=0.76),but the recall rate was low,which required further analysis and processing.Part III: Further model updated with SMOTE algorithm,comparison of the indicators after cross-validation shows that ET,RF and XGBoost models have a better accuracy(0.78).In terms of precision,XGBoost model showed the best performance(0.67).In terms of recall rate,LR model performed best(0.64),which correctly detected the highest proportion of CPSP occurrence;ET model had the highest F1-Score(0.60),while AUC of all the other models reached 0.80 except DT model and ET model.All of the six predictors show good predicting ability after SMOTE,LR model is the best(recall=0.64),this model has a good performance in external validation also(accuracy =0.70,accuracy =0.52,recall =0.71,F1-score=0.60,AUC=0.79).Compare the evaluation indicators of all models before and after SMOTE algorithm,the recall rate of the 6 models improves as a whole: LR model(0.46 vs 0.64),DT model(0.51 vs 0.56),RF model(0.48 vs 0.56),ET model(0.49 vs 0.58),Ada Boost model(0.46 vs 0.62),XGBoost model(0.49 vs 0.56),recall of the LR model increased 39%,proving that SMOTE algorithm is correct and effective for model update.A web calculator based on best model is developed:http://43.138.154.46:5000/.Conclusions Elderly orthopedic surgery patients have a high incidence of CPSP,which has a significant adverse effect on body function,and needs to be widely paid attention to.In this study,the best model built based on SMOTE and ML algorithm has strong prediction ability and can effectively identify CPSP patients.It is convenient and quick to use the web computing tool to predict the risk of CPSP,which is helpful for clinical staff to identify and intervene in the early stage of medium and high risk patients.