Kidney Week

Abstract: SA-PO934

Predicting Renal Function Decline From Readily Available Clinical Variables in Electronic Health Records Using Machine Learning

Session Information

• CKD: Observational Research and Patient-Oriented Interventions
  November 05, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
  Abstract Time: 10:00 AM - 12:00 PM

Category: CKD (Non-Dialysis)

• 2202 CKD (Non-Dialysis): Clinical‚ Outcomes‚ and Trials

Authors

• Mathew, Roy O., Loma Linda VA Health Care System, Loma Linda, California, United States

Roy O. Mathew,

• Odigwe, Brendan Elochukwu, University of South Carolina School of Engineering, Columbia, South Carolina, United States

Brendan Elochukwu Odigwe,

• Paudel, Sujay Dutta, Loma Linda University School of Medicine, Loma Linda, California, United States

Sujay Dutta Paudel,

• Odigwe, Celestine I., Thomas Hospital, Fairhope, Alabama, United States

Celestine I. Odigwe,

• Fung, Enrica, Loma Linda VA Health Care System, Loma Linda, California, United States

Enrica Fung,

• Norouzi, Sayna, Loma Linda University School of Medicine, Loma Linda, California, United States

Sayna Norouzi,

• Infante, Sergio, Loma Linda University School of Medicine, Loma Linda, California, United States

Sergio Infante,

• Abdi Pour, Amir, Loma Linda University School of Medicine, Loma Linda, California, United States

Amir Abdi Pour,

• Rangaswami, Janani, The George Washington University School of Medicine and Health Sciences, Washington, District of Columbia, United States

Janani Rangaswami,

• Valafar, Homayoun, University of South Carolina School of Engineering, Columbia, South Carolina, United States

Homayoun Valafar,

Background

This analysis sought to implement machine learning (ML) algorithms to incorporate readily available clinical variables from a nationally representative administrative dataset to predict future renal function decline.

Methods

All data retrieved from the Veterans Affairs (VA) corporate data warehouse. The outcome was the rate of estimated glomerular filtration rate (eGFR) decline over 3 years. An Artificial Neural Network (ANN) was developed as our machine learning technique of choice and was implemented utilizing the nerualnet package in R.
A total of 183,054 unique veterans with baseline eGFR between 15 and 60 ml/min/1.73m² with follow up serum creatinine annually for 3 years following the index creatinine value were included. Training cohort consisted of 75% of the total population, leaving 25% for ML model testing. A total of 101 variables were initially available. For the final training cohort, 74 variables were included (after excluding outcomes, dates, and identifiers). Loma Linda VA IRB provided expedited review approval.
For experimentation, we created a feed-forward Neural network with 2 hidden layers, having 32 and 16 neurons in the first and second hidden layers, respectively.

Results

Of the included patients, 81% had CKD G3a, 15% G3b, and 4% with G4. The optimal neural network architecture produced the prediction of the 3-year decline in eGFR in the testing cohort with mean square error of 0.04, and correlation coefficient of 0.998 (Figure).

Conclusion

ANNs accurately predicted the rate of progression. Such ML techniques can accurately identify high-risk patients for intensive risk reduction and/or targeted research efforts for novel interventions.

Predicted vs actual rate of eGFR decline using feed forward neural network from administrative health data.