External Validation Of Predictive Model Of Myelosuppression By Paclitaxel Plus Platinum Chemotherapy In Ovarian Cancer

Objectives:Our preliminary research have established a model to predict myelosuppression by TC regimen chemotherapy in ovarian cancer.The aims are to validate the model,and to visualize the model by providing nomograms,web calculator and scoring system to facilitate clinical assessment of the risk of myelosuppression.Methods:Preliminary study:The preliminary study collected the ovarian cancer patients who underwent paclitaxel plus platinum chemotherapy from May 2013 to May 2018 in the Third Affiliated Hospital of Kunming Medical University to establish myelosuppression prediction model,and the model formula was as follows:The probability of myelosuppression P=ex/(1+ex)Model 1(Myelosuppression):x=4.885-0.539×(prealbumin)-0.712×(HB count before chemotherapy)-0.775×(G-CSF)-0.966x(recurrence)+0.608×(previously myelosuppression index)-0.745×(platinum RDI)-0.739×(chemotherapy cycle)AUC=0.788(95%CI 0.755-0.820).Model 2(Leukopenia):x=2.636+0.591×(BSA)-0.512×(HB count before chemotherapy)-0.977×(G-CSF)-0.889×(recurrence)+0.629×(previous occurrence of myelosuppression grade)-0.677×(chemotherapy cycle)-0.984×(prealbumin)AUC=0.803(95%CI 0.769-0.837).Model 3(Neutropenia):x=2.368-0.890×(G-CSF)-0.993×(recurrence)+0.623×(previous myelosuppression grade)-0.694×(chemotherapy cycle)-0.693×(cholesterol)AUC=0.771(95%CI:0.734-0.808).Model 4(Hemoglobin reduction):x=2.625+1.010x(age)-1.626x(pre-chemotheray RBC count)-0.941×(pre-chemotherapy PLT count)-1.312x(pre-chemotherapy HB count)-0.741×(chemotherapy cycle)AUC=0.864(95%CI 0.810-0.918).Model 5(Thrombocytopenia):x=-2.130-0.73 3x(pre-chemotherapy PLT count)+0.359x(previously myelosuppression grade)AUC=0.703(95%CI 0.624-0.782).Methods of this study:We reviewed the medical records(from June 2018 to May 2019)of ovarian cancer patients who received paclitaxel plus platinum chemotherapy in the Third Affiliated Hospital of Kunming Medical University.The index of the discrimnation and the calibration were calculated by R Studio software,the discrimnation index was AUC,and the calibration index included Brier score,discrimination slope,calibration slope,calibration-in-the-large and calibration plot.In order to visualize the model,nomograms and web calculators were build by R Studio software respectively,and a scoring system was made by Excel.Results:A total of 2061 cases were reviewed,1842 cases were qualified to enter the validation corhort.1.Discrimination indexIn external validation,AUC of the model are 0.738,0.740,0.706,0.843,0.711,respectively,and all the AUC are greater than 0.7,showing a good discrimination.2.Calibration indexIn external validation,the Brier scores are 0.158,0.098,0.134,0.014,0.007,respectively.All of them are less than 0.25 and close to 0,indicating that the model has predictive ability and good calibration.the calibration-in-the-large of models are-1.923,-1.884,-2.031,-1.739,-1.672,respectively.The intercepts are all less than 0,indicating that the actual occurrences of events in the validation cohort is lower in that of the derivation cohort.In terms of calibration slopes,the slopes are 0.663,0.670,0.618,0.757,and 1.263,separately.The slope in model 1,2,3,and 4 are less than 1,indicating that there is an over-prediction problem in each model.The discrimination slopes of models are 0.204,0.189,0.159,0.062,0.023,respectively.The calibration curves in models are acceptable.In model 1,2,3,the curves are well fitted in the low probability interval and poorly fitted in the high probability interval,whereas are well fitted in model 4 and 5.3.Model visualizationNomograms,web calculators and scoring systems were build,according to the final predictive model formula.Conclusions:1.In external validation,all the AUCs shows good discrimination.2.The overall calibration of models is acceptable,but there is a problem of over-prediction.3.We recommend that the myelosuppression,leukocytes,neutrophils model can be used clinically.However,you should pay attention to the problem of over-prediction when using the model.To solve this problem,we can group the population into low-risk group and high-risk group according to the threshold(clinicians can define the threshold by themselves),which can guide clinicians to screen low-risk groups.4.The hemoglobin and the platelet model are well calibrated,but due to the low incidence and insufficient sample size,the model for reference only.5.Five models are only applicable to ovarian cancer patients receiving paclitaxel plus platinum chemotherapy.