Kidney Week

Abstract: TH-PO389

Development of the Deep Neural Network for Estimating Glomerular Filtration Rate

Session Information

• CKD: Risk Scores and Translational Epidemiology
  November 07, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: CKD (Non-Dialysis)

• 2101 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention

Authors

• Yang, Jae Won, Wonju Christian Severance Hospital, Wonju, Kangwon do, Korea (the Republic of)

Jae Won Yang,

• Lee, Jun Young, Wonju Christian Severance Hospital, Wonju, Kangwon do, Korea (the Republic of)

Jun Young Lee,

• Kim, Jae seok, Wonju Christian Severance Hospital, Wonju, Kangwon do, Korea (the Republic of)

Jae seok Kim,

• Eom, Minseob, Yonsei Univ. Wonju College of Medicine, Wonju, Korea (the Republic of)

Minseob Eom,

• Chai, Moonhee, Wonju Christian Severance Hospital, Wonju, Kangwon do, Korea (the Republic of)

Moonhee Chai,

• Choi, Seung-Ok, Wonju Christian Severance Hospital, Wonju, Kangwon do, Korea (the Republic of)

Seung-Ok Choi,

• Jeong, Jin-Jae, Wonju Christian Severance Hospital, Wonju, Kangwon do, Korea (the Republic of)

Jin-Jae Jeong,

Background

A variety of calculation formulas to estimate glomerular filtration rate (GFR) have been developed for decades. Recently, modern clinical medicine has been trying to use a deep neural network (DNN) in various clinical fields. Thus, we aimed to use DNN model for estimating GFR in the study.

Methods

A total of 241 patients with chronic kidney disease were enrolled in the study. All participants had technetium-99m diethylenetriaminepentaacetic acid (^99mTc-DTPA) renogram to obtain the standard value of GFR. We measured serum creatinine levels from all participants and calculated GFRs using various formulas such as MDRD and CKD-EPI. Furthermore, we developed a DNN model with three hidden layers. The first, second, and third hidden layers included 40, 20, and 10 nodes respectively. We compared GFR values of MDRD, CKD-EPI, and DNN model against standard GFR values from renogram in various statistical ways.

Results

The mean differences of GFR value from MDRD, CKD-EPI, and DNN methods with standard GFR were 2.35, 2.86, and 1.87 mL/min respectively. The mean root-mean-square-error values of MDRD, CKD-EPI, and DNN methods against standard GFR were 19.45, 18.9, and 16.78 (mL/min)² respectively suggesting that the GFR values of DNN model are closest to standard GFR values. When estimating the accuracy in classifying CKD stages, the degree of accuracy of MDRD, CKD-EPI, and DNN methods were 83.0, 84.3, and 85.1% respectively suggesting that the DNN model is the most accurate method in classifying CKD stages.

Conclusion

GFR measurement using DNN model is believed to be useful and accurate.