| BackgroundSurgical complications are often defined as any deviation from the ideal postoperative course,which can significantly increase mortality,length of stay,and health care costs for surgical patients,resulting in a significant burden on the health care system.Studies on the grading and risk assessment of complications have been carried out to explore better management of complications.However,the existing grading system and risk assessment tools have some limitations in terms of the quantitative approach.(1)The most commonly used clinical grading system of complications is the ClavienDindo classification,which takes the therapy used to treat the complication as the principle to grade complications,has the advantage of being simple and easy to implement,and can accommodate all complications.However,the same complication can be graded in different grades(e.g.,an incisional infection is graded to different grades according to whether it needs antibiotic treatment),making the complications included in each grade are not fixed,and the same patient can have multiple complications in different grades at the same time,so it is difficult to quantitatively assess the relationship between grades.(2)Previous risk assessment tools of surgical complications based on classical regression models with patient risk factors require the assumption of linear independence and additivity,which is often contrary to clinical practice;while machine learning-based assessment tools,which have been in the research hotspot in recent years,can accommodate complex relationships among many risk factors and achieve high predictive performance,but their "black box" nature makes it difficult to obtain transparent explanations,limiting their wide clinical application.(3)Due to methodological limitations,both regression-based and machine learningbased assessment tools have adopted a compromise approach to simplify the outcome form,either by evaluating a single complication,by evaluating all complications together as a composite(i.e.,with or without complications,actually equivalent to a single complication),or by simply evaluating severe complications defined as Clavien-Dindo classification≥3(actually also equivalent to a single complication).In fact,there is a complex network relationship among complications,one complication itself can cause multiple secondary complications,and severe complications tend to occur in clusters.It will be a challenge to elucidate and quantify the network relationships among complications,and presenting the occurrence,progression,and eventual death of complications from a global perspective will undoubtedly greatly facilitate the systematic management of complications.Sepsis,as a complication with serious conditions,high mortality,and close relationship with other severe complications,has always been the focus of clinical attention.From the perspective of a global network,priority should be given to focusing on key targets and establishing local detailed network relations.While verifying the tension of a global network to an individual complication,a feasible data-guided framework will be proposed for clinical priority to solve critical and difficult problems.In addition,the reproduction and cross-corroboration of the relationships and key targets found by the network-based evaluation methods on other machine learning methods will further support the effectiveness of the network methods and the robustness of the results.ObjectiveTo construct a network grading system for surgical complications,develop a risk assessment tool for joint prediction of different grades of complications,describe complication occurrence and progression pathways and identify key targets on the pathway;Based on this,the aforementioned pathway and prediction framework are applied to sepsis to achieve prediction and pathway description of occurrence and progression for sepsis;In addition,the grading system was also applied to different machine learning methods to validate the effectiveness of the developed network system from different perspectives.The above exploration can deepen the global understanding of complications to promote systematic clinical management of complications.MethodsThis study was based on the Modern Surgery and Anesthesia Safety Management System Construction and Promotion(MSCP)project,and prospectively collected data on surgical patients from four regionally representative tertiary hospitals during January-June 2015 and January-June 2016.The included population were surgical inpatients aged≥14 years who had at least one anesthesiologist involved in the surgery.The main information collected included demographic characteristics,preoperative physical examination,preoperative laboratory tests,surgery-related information,and postoperative outcomes,with the main outcome of interest being postoperative death and 22 surgical complications.Part I:(1)Complication grading:Complication clusters were obtained by a datadriven approach(cluster analysis)and an initial grading was obtained based on medical perceptions;the initial grading was evaluated and adjusted by a team of 54 expert surgeons and anesthesiologists;relevance between the final grading and mortality were used to illustrate the validity of the grading.(2)Risk factor clustering:17 individual preoperative risk factors were classified into clusters of risk factors by cluster analysis.(3)Bayesian network construction:The risk factor clusters and complication grading groups were used as nodes,and the structure learning was performed by using the hill-climbing algorithm based on score search,and the Akaike information criterion was used as the scoring function;maximum likelihood estimation was used for parameter learning.(4)Model evaluation:the area under the ROC curve(AUC)was used to evaluate the model discrimination;the Brier score,Hosmer-Lemeshow test,and calibration graph are used to evaluate the calibration of the model.The validation was performed in the training set,tenfold crossover validation in the training set,and validation set,respectively.(5)Model application:Given different node information,evaluate the ratio of the change in the probability of complications at each grade relative to the baseline probability(i.e.,to get probability ratio)to determine node importance;develop an online complication risk calculator to enable the translational application of complication prediction.Part Ⅱ:Based on Part Ⅰ,the Bayesian network modeling approach was applied to patients undergoing abdominal surgery in general surgery,with the outcome of focus selected as sepsis to describe the entire process of sepsis onset,progression,and eventual death.The scalability of the global network assessment framework in Part Ⅰ was reproduced and corroborated in the form of a local focus of attention.Part Ⅲ:(1)Using separate preoperative risk factors and risk factor clusters from PartⅠ as predictor variables and complication grading as the outcome,respectively,three multilabel machine learning methods,binary relevance(BR),classifier chain(CC),and multilabel k-nearest neighbor(MLKNN),were used to construct multi-label prediction models for different grades of complications and validate the predictions in the validation set.(2)The multi-label prediction results were interpreted at the population and individual levels using SHAP(Shapley additive explanations)interpretable methods.ResultsPart Ⅰ:This part developed a versatile network system based on data from 51030 surgical inpatients and the experience of 54 clinical experts,enabling grading,prediction,and pathway visualization of complications.In this network system,a total of 11 nodes represent 5 clusters of risk factors and 6 grades of complications(Ⅰ,Ⅱ,Ⅲ,Ⅳa,Ⅳb,and Ⅴfrom mild to severe);32 directed arcs represent the direct association between two nodes(two nodes connected by a series of directed arcs are indirectly associated).Some key nodes were identified through the network:malnourished status had the most arcs(7/32)and directly influenced other risk factor clusters and Grade Ⅰ,Ⅲ,and Ⅳb complications,being the most basic but most widely influencing factor.ASA(American Society of Anesthesiologists)score≥3 is a core element condensed by other risk factor clusters,directly dependent on all other risk factor clusters,and directly influences all serious complications(≥Grade Ⅲ).Grade Ⅲ complications in the network are the bridge connecting risk factors and complications.The upstream directly depends on 4/5 risk factor clusters,and the downstream directly affects all other grades of complications.At the same time,in the network,the clustering pattern of severe complications is very obvious,the Grade Ⅲ-Ⅴ complications are directly connected,and the correlation coefficient is relatively large(0.31-0.60).In addition,regardless of the grade of complications,the risk of complications itself is more likely to cause other grades of complications than the risk factor cluster(the probability ratio relative to baseline is 4.0-69.8 vs.1.0-3.4).Part Ⅱ:This part constructs a network model based on the data of 8211 patients undergoing abdominal surgery and realizes the whole perioperative risk prediction of sepsis and death,and the AUCs are 0.861 and 0.882,respectively.The whole network pathway of sepsis is visualized,namely,"risk factors→infection→sepsis→(organ failure)→death".In this pathway,high heart rate and emergency surgery had a greater impact on sepsis,and the probability ratios relative to baseline were 2.7 and 3.4,respectively.The impact of infectious complications on sepsis was greater than risk factors.The probability ratios of pneumonia,abdominal infection or fistula,and incision infection were 11.1,7.4,and 4.5,respectively.In addition,sepsis co-occurring with other complications significantly increased the mortality rate.In the presence of sepsis,the probability of death in patients with or without pneumonia was 73.4%vs.56.7%,and the probability of death increased to more than 90%in patients with acute renal failure,MODS,and cardiac arrest requiring CPR,respectively.Part Ⅲ:This part uses multi-label machine learning to predict complications of different grades based on the data of 43847 surgical inpatients and achieves good prediction results.For example,when five risk factor clusters are used as predictors,the AUCs of BR,CC,and MLKNN for Grade Ⅴ complications are 0.808,0.823 and 0.814,respectively;When 17 preoperative risk factors were used as predictors,the AUC for GradeⅤ complications were 0.885,0.886,and 0.825,respectively.The trends of the AUC were very consistent in the two scenarios.ASA score≥3 and emergency status were the two most important factors in the global interpretation of risk factor clusters;In the global interpretation of individual risk factors,ASA score≥3,emergency surgery and high WBC count had a greater impact on severe complications(Grade Ⅲ,Ⅳb and Ⅴ),while low albumin and age≥65 years had a greater impact on mild complications(Grade Ⅰ andⅡ).ConclusionCombining multidisciplinary surgical big data and clinical expert experience,this study proposed a multifunctional,whole-process network system that clearly elucidated the occurrence and progression pathways of complications and identifies key targets for quality improvement on the pathways,integrating the management of patients at high risk of complications into the beginning and end of clinical diagnosis and treatment,and providing a clear and systematic perspective for the prevention and treatment of complications.The further application of this systematic network concept to postoperative sepsis after abdominal surgery also illustrates the complexity of the complication relationship and the need for the introduction of a network approach.The fact that similar findings to the network system were obtained using other different machine learning methods also validates and reinforces the robustness of the results of this study.Further studies need to be taken in the future to confirm whether the network system can truly prompt clinical outcomes in surgical patients. |
PDF Full Text Request