Kidney Week

Abstract: SA-PO550

Development of an Exhaustive-Risk-Prediction System Using Deep Learning and Different Patterns of Diabetic Kidney Disease Progression Based on Patient Characteristics

Session Information

• Diabetic Kidney Disease: Pathology, Epidemiology
  November 09, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Diabetic Kidney Disease

• 602 Diabetic Kidney Disease: Clinical

Authors

• Kanda, Eiichiro, Kawasaki Medical School, Kurashiki, OKAYAMA, Japan

Eiichiro Kanda,

• Tokuyama, Atsuyuki, Kawasaki Medical School, Kurashiki, OKAYAMA, Japan

Atsuyuki Tokuyama,

• Itano, Seiji, Kawasaki Medical School, Kurashiki, OKAYAMA, Japan

Seiji Itano,

• Nagasu, Hajime, Kawasaki Medical School, Kurashiki, OKAYAMA, Japan

Hajime Nagasu,

• Epureanu, Bogdan I., University of Michigan, Ann Arbor, Michigan, United States

Bogdan I. Epureanu,

• Kashihara, Naoki, Kawasaki Medical School, Kurashiki, OKAYAMA, Japan

Naoki Kashihara,

Background

Diabetic kidney disease (DKD) is a risk factor for end-stage kidney disease (ESKD) and death. An accurate prediction of these risks at an individual level is required to improve DKD patients’ prognosis. In this study, we developed a new system for the prediction of chronic kidney disease (CKD) progression using deep learning (DL) and the CKD big database in Japan, and investigated DKD progression patterns based on their characteristics.

Methods

The associations among patients’ characteristics, laboratory data and an outcome (ESKD or death) in five years were evaluated by DL. The kidney disease progression risk for a virtual patient was simulated using the trained DL model.

Results

Among the patients (n=3877, 83701 measured data), 53.0% were male; average age, 59.8±17.9 years; estimated glomerular filtration rate (eGFR), 51.1±29.0 ml/min/1.73 m²; and diabetes mellitus, 18.8%. In the test dataset, the accuracy of DL was 0.95, which was higher than those of multivariate logistic regression models (0.84) and support vector machine models (0.84). Then, various patterns of characteristics of a 60-year-male patient were evaluated. The predicted risk of the outcome is shown in heat maps (Figure 1), and various patterns are observed. A basic pattern (e.g., CKD severity category) shows high risks in categories of low eGFR and high urinary protein level. And, the nephrosclerosis-like pattern shows higher risks at low urinary protein levels than at high urinary protein levels. Moreover, DKD shows a mix pattern, which suggests that DKD patients are a heterogenous population.

Conclusion

We developed a new exhaustive risk-prediction system using DL and found different patterns of kidney disease progression based on patient characteristics. This system may be useful for identifying patients at an increased risk of DKD progression for early treatment.