Abstract: PO0529
A Machine Learning-Based Prediction Model for Trajectory of GFR in CKD Patients with Rapid Decline of GFR by Using a Big Database
Session Information
- CKD Health Services Research
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2101 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention
Authors
- Inaguma, Daijo, Fujita Health University School of Medicine, Department of Nephrology, Toyoake, Aichi, Japan
- Kitagawa, Akimitsu, Fujita Health University Bantane Hospital, Department of Nephrology, Nagoya, Japan
- Yanagiya, Ryosuke, Fujita Health University, Division of Medical Information System, Toyoake, Japan
- Koseki, Akira, IBM Research, Tokyo, Japan
- Iwamori, Toshiya, IBM Research, Tokyo, Japan
- Kudo, Michiharu, IBM Research, Tokyo, Japan
- Fukuma, Shingo, Kyoto University Graduate School of Medicine, Human Health Sciences, Kyoto, Japan
- Tsuboi, Naotake, Fujita Health University School of Medicine, Department of Nephrology, Toyoake, Aichi, Japan
- Yuzawa, Yukio, Fujita Health University School of Medicine, Department of Nephrology, Toyoake, Aichi, Japan
Background
There are various patterns of GFR trajectories in patients with chronic kidney disease (CKD), even among those with rapid declines in GFR. We sought to create a machine learning-based predictive model for extremely rapid decline of GFR in patients with CKD using a single hospital database.
Methods
We used a database, which included the electronic medical records of 286,494 patients. We selected patients with CKD and rapid decline in kidney function, which was defined as an estimated GFR (eGFR) decline of 30% or more within two years. We used longitudinal statistics using data extracted from baseline, 90-, 180-, and 360-day windows prior to baseline and exponentially weighted averages (ESAs) where the weight was calculated as 0.9*(days/decay parameter). The random forest algorithm and python code with the scikit-learn library (https://scikit-learn.org/) were used for model creation.
Results
Patients were automatically classified, using machine learning, into three groups according to eGFR at baseline (G1; high GFR, G2; intermediate GFR, G3; low GFR) and nine subgroups according to the slope of eGFR decline. The subgroup with the fastest GFR decline exhibited the steepest slope (Figure 1). The area under the curves for predicting the steepest (fastest) GFR decline by random forest model among the G1, G2, and G3 were 0.68, 0.72 and 0.81, respectively. Regarding feature importance, in the G1 group, hemoglobin of the 7-day ESAs and measures obtained 90 days prior to baseline ranked within the top five. Meanwhile, serum albumin and CRP at baseline ranked within the top seven in the G3 group.
Conclusion
The random forest model identified patients with extremely rapid GFR decline. Anemia in patients with higher eGFR, and nutritional status in patients with lower eGFR, emerged as strong risk factors.