The Validation Of Glomerular Filtration Rate Equation In Tumor Population And New Equation Derived By Machine Learning Method

PartⅠ:Validation of Glomerular Filtration Rate Equations for Tumor PopulationObjective:Accurate evaluation of glomerular filtration rate（GFR）is crucial for tumor patients receiving anti-tumor drug treatment.This part aims to explore the clinical characteristics of tumor patients who need to monitor renal function,and apply the patients’data to commonly used equations for estimating glomerular filtration rate（eGFR）to verify their performance in tumor patients.Methods:Patients performed renal dynamic imaging ⁹⁹mTc-DTPA in our hospital to obtain measured glomerular filtration rate（mGFR）from December 2012 to December 2021 with histologically confirmed tumors were collected.And then analyzing the characteristics of tumor patients and exploring the clinical factors related to different stages of mGFR were performed.The ability of commonly used eGFR equations in tumor patients to distinguish between different stages of mGFR,as well as their bias,accuracy,and consistency with mGFR were analyzed.The mean percentage error（MPE）,mean absolute percentage error（MAPE）,root mean square error（RMSE）,quartile range of residuals,and the percentage of patients with absolute percentage errors within 30%,20%,and 10%（P30,P20,P10）mGFR were used as indicators to evaluate and compare the performance of the eGFR equations.Results:The tumor population in this study is mainly elderly,with a median age of 58 years.The proportion of elderly people with low mGFR is relatively high,while the proportion of young people with high mGFR is relatively high.An increase in age was positively associated with a risk of GFR<60 ml/min（OR:1.06;95%confidence interval:1.05-1.07）.In addition to indicators related to renal function,serum albumin and hemoglobin levels in tumor population in this study showed strong correlation with CKD stage.There were statistically significant differences in serum albumin levels among the three populations:patients with<60ml/min,60-89ml/min,and≥90ml/min.Currently,all published equations tended to overestimate mGFR with poor consistency and accuracy,with P30 ranging from 28.1%to 70.7%.The best performing equation is the Jelliffe equation,with a minimum MPE of 9.9%and a maximum accuracy P30 of 70.7%.However,the average and 95%confidence interval of the difference between the eGFR of Jelliffe equation and mGFR are 4.92（-30.49,40.34）ml/min,still showing poor consistency.Compared to higher hemoglobin and albumin levels,the eGFR equations in lower hemoglobin and albumin levels had a higher coincidence rate and lower misclassification rate for classifying different mGFR levels,and had higher precision and accuracy.Conclusions:In addition to the indicators related to renal function,the serum albumin level and hemoglobin level in the tumor population in this study showed a strong correlation with CKD stage.High albumin level can reflect better renal function,while low hemoglobin level can reflect poorer renal function.The commonly used eGFR equations showed poor consistency,precision and accuracy in tumor population.The best performing equation is the Jelliffe,which has the smallest bias,but still exhibits poor consistency compared to mGFRPart Ⅱ: New Equation for Glomerular Filtration Rate in Tumor Population Derived by Machine Learning MethodObjective: Currently,neither the estimated glomerular filtration rate（eGFR）equations commonly recommended for the general population nor the eGFR equations developed abroad for tumor patients is accurate when applied to Chinese tumor population.Therefore,this section aims to develop more accurate new eGFR equations for Chinese tumor patients through symbolic regression algorithms that provide visual calculation equations.Methods: Patients performed renal dynamic imaging 99 mTc-DTPA in our hospital to obtain measured glomerular filtration rate（mGFR）from December 2012 to December 2021 with histologically confirmed tumors were collected.1001 tumor patients and 485 tumor patients with cystatin C data in our hospital were randomly divided into glomerular filtration rate formula derivation dataset and internal validation dataset at 7:3.Through symbolic regression of machine learning methods,two interpretable eGFR equations based on serum creatinine versus serum creatinine and cystatin C were established using the two derivation datasets,namely,the new equation_Scr and equation_Scr-cysC.Results: From the 1.9 * 109 candidate equations,the new equation_Scr and equation_Scr-cysC were selected,taking into account the requirements of low complexity and high precision.The new equations included the variables gender,age,body surface area,serum creatinine,serum urea nitrogen,serum albumin,and hemoglobin（and cystatin C）Conclusions: New equations for eGFR in tumor population has been developed using symbolic regression algorithms.Part Ⅲ: Internal and External Validation of the New Equation for Glomerular Filtration Rate in Tumor PatientsObjective: The purpose of this part of the study is to evaluate and validate the new equations for estimating glomerular filtration rate（eGFR）in tumor patients derived in Part Ⅱ in both internal and external populations.Methods: The study population consisted of 296 internal validation population for the new equation_Scr,145 internal validation population for the new equation_Scr-cysC in Part Ⅱ,and 195 adult tumor patients from external hospital.The differences between the external hospital population and the internal validation population of the hospital were analyzed.Validating and comparing the new equations were in internal and external validation populations.The agreement between the eGFR of equations and the measured glomerular filtration rate（mGFR）,the ability to distinguish different stages of CKD,and its precision and accuracy were analyzed when the equations were validated.The mean percentage error（MPE）,mean absolute percentage error（MAPE）,root mean square error（RMSE）,quartile range of residuals,and the percentage of patients with absolute percentage errors within 30%,20%,and 10%（P30,P20,P10）were used as indicators to evaluate and compare the performance of the eGFR equations.Results: Compared with the published equations,the new equation_Scr had a higher consistency between eGFR and mGFR,higher classification ability at different mGFR stages,and smaller bias in estimating GFR（MPE,4.4%;95% CI,1.1% to 7.6%）and is more accurate（P30,74.3%,95% CI,69.3% to 79.3%）.In addition,the accuracy of the new equation_Scr-cysC had also been improved.In the external validation dataset,it was confirmed that the new equations had better performance than the published equations.In addition,the accuracy of the two new equations is similar(P30,the new equation_Scr is 79.0% vs the new equation_Scr-cysC is 79.5%).But the bias in the new equation_Scr-cysC was significantly reduced compared with the new equation_Scr(MPE,the new equation_Scr-cysC is 0.3% vs the new equation_Scr is 3.6%);The new equations were more suitable for patients with lower body mass index and patients with mGFR<90 ml/min.Conclusions: The new eGFR equations in this study had greater consistency,accuracy and classification ability at different mGFR stage,and lower bias than commonly published equations in both internal and external populations.