Machine Learning Predicts Adverse Events Following Transforaminal Lumbar Interbody Fusion

Part 1 Machine learning analyzes risk factors of high hospitalization cost following transforaminal lumbar interbody fusionObjective: Based on cost and length of stay,unsupervised cluster method analyzes risk factors of high hospitalization cost following TLIF.Method: These patients are enrolled into research,who undergo TLIF for lumbar degenerative diseases between January 2018 and December 2020 at spine surgery center,Zhongda Hospital affiliated with Southeast University.Based on cost and length of stay,KMeans models are built and evaluated by silhouette coefficient values and scatter plot.The univariate analysis is performed for features between subgroups,like: demographic characteristics,lumbar disease characteristics,operation characteristics,preoperative laboratory characteristics,postoperative disposition,data of hospital stay and AEs.Features with significant differences in univariate analysis are put into the multivariate analysis.Results: As the best manner,samples are divided into three clusters by KMeans model.The third cluster is identified as outliers and excluded.In total 1276 patients,the cost of second cluster is significantly higher than the first one as: 53739.53±10384.57(yuan)and 34079.31±5226.61(yuan).In the multivariate analysis,significant factors of high hospitalization cost included following: smoking status,AEs,Braden value,number of lumbar diseases,length of stay,operation quarter,operation time,allogenic blood transfusion,jugular vein catheterization time,drainage time and time of using antibiotics.Conclusion: Those patients could be divided into two clusters with different average cost by unsupervised algorithm.AEs may be the most important factor associated with high hospitalization cost.Preoperative identification of potential AEs patients and adjustment of individualized treatment may help improve outcomes and save medical resources.Part 2 Machine learning predicts adverse events following transforaminal lumbar interbody fusionObjective: ML algorithms predict AEs after TLIF,and efficacy of different feature selection methods for models are compared.Method: The ML predictive models are developed using preoperative and surgical features.The following feature selection methods are used: 1.features with significant differences in univariate analysis(P < 0.05);2.demographic and lumbar disease characteristics;3.surgical features;4.preoperative laboratory characteristics.Seven models are built,like: XGBoost,KNN,Decision Tree,Random Forest,SVC,Logistic Regression and ANN.Evaluated measures include following: ROC curve,AUC,ACC,Recall,F1 score and Precision.Results: A total of 1278 patients are enrolled and the incidence of AEs is 25.27%.It is different between predictive performances of models through different feature selections.Those models achieve fine performance which use significant features in univariate analysis,such as: XGBoost,KNN,Random Forest and SVC.Conclusion: Using preoperative and surgical characteristics,ML model could predict AEs after TLIF,which may contribute to surgical selection,optimization of patient management and saving medical resources.Those models obtained superior predictive performance,that using significant features in univariate analysis.Part 3 Interpretable machine learning model predicts major complications following transforaminal lumbar interbody fusionObjective: ML models predict major complications after TLIF,and interpretable model is used for global explanation and local explanation.Method: Features with significant differences in univariate analysis(P < 0.05)are inputted into model.Interpretable model is developed for global explanation and local explanation.Evaluated measures include following: ROC curve,AUC,ACC,Recall,F1 score and Precision.Results: Among 1278 patients,42 patients(3.29%)have postoperative major complications.Based on XGBoost model,global explanation is carried out,and important factors are following: operation time,glucose,Charlson comorbidity index,neutrophils count,albumin,age,number of performed pedicle screws,monocytes to lymphocyte ratio,lymphocyte ratio and neutrophils ratio.Local explanation of single sample could reveal the relationship between patient characteristics and prediction,through SHAP(Shapley additive explanations).Conclusion: ML model can predict major complications after TLIF.Interpretable model makes ML more convincing and reliable for medical decision makers and patients.Based on the XGBoost model,global explanation can clarify the feature importance.Local explanation could reveal the relationship between patient characteristics and prediction.