P0844PREDICTION MODEL FOR RENAL FUNCTION DECLINE WITH SUGGESTING HOW TO CONTROL CLINICAL PARAMETERS TO MAINTAIN RENAL FUNCTION BY AN INTERPRETABLE MACHINE LEARNING IN JAPANESE OUTPATIENTS WITH CHRONIC KIDNEY DISEASE. (6th June 2020)

• Record Type: Journal Article
• Authors: Ogata, Soshiro; Akashi, Yumi; Sakusabe, Takaya; Maeda, Kenji; Nakai, Shigeru
• Is Part Of: Nephrology dialysis transplantation. Volume 35(2020)Supplement 3
• Publication Date: 2020-06-06
• Subjects: Nephrology -- Periodicals; Hemodialysis -- Periodicals; Kidneys -- Transplantation -- Periodicals; Hemodialysis; Kidneys -- Transplantation; Nephrology; Periodicals; 616.61
• DOI: 10.1093/ndt/gfaa142.P0844
• Languages: English
• ISSNs: 0931-0509

Abstract

Abstract: Background and Aims: Prediction models for renal function decline has already been developed. However, few prediction models can suggest how to control clinical parameters to maintain renal function. The present study aimed to develop a prediction model for renal function decline that can suggest how to control clinical parameters to maintain renal function by an interpretable machine learning in Japanese outpatients with chronic kidney disease (CKD). Method: The present study used a retrospective cohort design. The eligible criteria of the present study were outpatients with CKD stage 3, 4, and 5, treated in Daiko-Sunadabashi Clinic in Japan between 2005 and 2019. Among the patients, we analyzed patients whose 24-hour urine samples were collected 3 times and over between 6 and 36 months. The primary outcome was renal function decline defied by glomerular filtration rate (GFR) decreased by >30% from the baseline. As predictors, we used baseline CKD stage and baseline values and change rate from the baseline to the next-to-last observations of the following clinical parameters: urine volume, GFR, quantity of protein in urine, intake of protein and salt, and body mass index (BMI). We used a random forest model with actionable feature tweaking to develop the prediction model with suggesting how to control those clinical parameters. This method has been developed recently and to improve interpretability of the random forest method. This can suggest how to alter predictor values in order to change prediction results from bad to favorable outcome (e.g., changing from predicted decline to predicted maintenance of renal function). Predictability of the prediction model was assessed by accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) in training and testing datasets. Results: We analyzed 355 participants whose mean age (SD) was 69.2 (13.0) years old. Of the participants, 65.4% were men, 37.5%, 37.2% and 25.3% were CKD stage 3, 4, and 5, respectively. Of the participants, the 80% were included in the training dataset, and the 20% in the testing dataset. Predictability of the developed prediction model was summarized in Table 1, showing the prediction model had good predictability. The developed prediction model suggested that decreasing protein in urine and salt intake, and increasing GFR could be important to change from predicted decline to predicted maintenance of renal function. The prediction model could also suggest how to control those clinical parameters at individual-level (e.g., #53 and #67 patients in Figure 1). Conclusion: The present study developed the prediction model for renal function decline with high predictability. The prediction model suggested that decreasing protein in urine and salt intake, and increasing GFR could be important to maintain renal function.