Kidney Week

Abstract: TH-PO026

Predictive Model of the Time to Renal Replacement Therapy Using Machine Learning

Session Information

• AI, Digital Health, Data Science - I
  November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Augmented Intelligence, Digital Health, and Data Science

• 300 Augmented Intelligence, Digital Health, and Data Science

Authors

• Okita, Jun, Department of Endocrinology, Metabolism, Rheumatology and Nephrology, Faculty of Medicine, Oita University, Yufu, Oita, Japan

Jun Okita,

• Nakata, Takeshi, Department of Endocrinology, Metabolism, Rheumatology and Nephrology, Faculty of Medicine, Oita University, Yufu, Oita, Japan

Takeshi Nakata,

• Uchida, Hiroki, Department of Endocrinology, Metabolism, Rheumatology and Nephrology, Faculty of Medicine, Oita University, Yufu, Oita, Japan

Hiroki Uchida,

• Kudo, Akiko, Department of Endocrinology, Metabolism, Rheumatology and Nephrology, Faculty of Medicine, Oita University, Yufu, Oita, Japan

Akiko Kudo,

• Fukuda, Akihiro, Department of Endocrinology, Metabolism, Rheumatology and Nephrology, Faculty of Medicine, Oita University, Yufu, Oita, Japan

Akihiro Fukuda,

• Shibata, Hirotaka, Department of Endocrinology, Metabolism, Rheumatology and Nephrology, Faculty of Medicine, Oita University, Yufu, Oita, Japan

Hirotaka Shibata,

Background

In the treatment of chronic kidney disease, nephrologists are expected to prolong the renal prognosis for those at high risk for end-stage renal disease (ESRD). It is also important to estimate the time to renal replacement therapy (RRT) in patients at high risk of ESRD. In Japan, the “time-series data of estimated glomerular filtration rate (eGFR)” is used to estimate the time to RRT based on the annual decline rate. In this study, we used machine learning to predict the time to RRT from the "data obtained at a single time point.”

Methods

Patients who underwent hemodialysis at our hospital from April 2016 to March 2021 were included, and a data set, including 30 laboratory data items (BUN, creatinine [Cr], etc.), six patients' background demographic background, and medications were extracted retrospectively from the electronic medical records. 75% of the data were randomly split for training and 25% for testing, and the predictive models were created with several algorithms: linear regression, ridge regression, least absolute shrinkage selection operator ( LASSO ) regression, elastic nets, random forests, and gradient boosting decision trees. We also predicted the time to RRT using ”time-series data of eGFR” and compared the accuracy by the coefficient of determination (R²) and mean absolute error (MAE).

Results

A total of 13,323 data were extracted from 147 patients of which 99 were males, resulting in a final total of 1,801 data groups with no missing data. The mean age at dialysis induction was 60.8 years, the most common etiology for ESRD was diabetic nephropathy (44 %), the mean Cr was 7.6 ± 2.0 mg/dL, and the mean eGFR was 6.1 ± 1.7 ml/min/1.73 m². The prediction model based on LASSO was moderately accurate with R² 0.62 and MAE 416, while the prediction based on the "time-series data of eGFR" was highly inaccurate with R² of -17.1 and of MAE 2466, indicating that the machine learning is superior in predicting the onset of dialysis.

Conclusion

The machine learning method was used to predict the time to RRT using "data obtained at a single time point,” and a moderately accurate prediction model was obtained. The ability to specify the time to RRT, even approximately, is useful not only for medical staff to make treatment decisions, but also for patients to motivate themselves to receive treatment and making long-term plans.