A machine learning model for predicting worsening renal function using one‐year time series data in patients with type 2 diabetes

ABSTRACT Background and Aims To prevent end‐stage renal disease caused by diabetic kidney disease, we created a predictive model for high‐risk patients using machine learning. Methods and Results The reference point was the time at which each patient's estimated glomerular filtration rate (eGFR) first fell below 60 mL/min/1.73 m2. The input period spanned the reference point to 1 year prior. The primary endpoint was a 50% decrease in eGFR from the mean of the input period over the 3 year evaluation period. We created predictive models for patients’ primary endpoints using time series data of various variables over the input period. Among 2,533 total patients, 1,409 had reference points, 31 had records for their input and evaluation periods and had reached their primary endpoints, and 317 patients had not. The area under the curve (AUC) of the predictive model peaked (0.81) when the minimum eGFR, the difference between maximum and minimum eGFR, and both maximum and minimum urinary protein values were included in the features. Conclusion The accuracy of prognosis prediction can be improved by considering the variable components of urinary protein and eGFR levels. This model will allow us to identify patients whose renal functions are relatively preserved with eGFR of more than 60 mL/min/1.73 m2 and are likely to benefit clinically from immediate treatment intensification. Diabetic kidney disease is the leading cause of end‐stage renal disease in developing and developed countries. It would be very meaningful to predict worsening renal function at a time before the eGFR decreases when therapeutic intervention is useful. We created a model focusing on the variability of time‐series data.