PS-BPP03-1: PREDICTIVE ABILITY FOR NEW-ONSET CHRONIC KIDNEY DISEASES BY MACHINE LEARNING ALGORITHM USING HEALTH CHECKUP DATA

Abstract

Background: In Japan, many data on relatively healthy individuals exist because specific health checkups are conducted annually for individuals aged 40 and older for the purpose of early detection and intervention of lifestyle-related diseases. Efforts to effectively utilize these data to predict the future development of chronic kidney disease (CKD) are important from the perspective of improving prognosis and quality of life, as well as healthcare economics. Aim: We aimed to evaluate the predictive ability of machine learning algorithms for newly developed CKD using health checkup data. Methods: We included individuals aged 30–69 years who had a health checkups between April 2008 and March 2018 with baseline (year X) and 5-year (year X + 5) data. CKD was defined as < 60 ml/min/1.73m2, and those with hypertension in year X were excluded. Three different machine learning algorithms, including Random Forest, Logistic Regression, and Neural Network, were evaluated to predict the development of new-onset CKD at year X + 5 from baseline (year X) data, and the AUCs of the ROC curve were used to assess the predictive ability of the models. Explanatory variables included physical measurements, blood pressure and pulse rate, laboratory values, lifestyle-related indicators, and medication history. Results: A total of 58,418 participants (age 53.8 ± 10.2 years, male 49.9%) were included in this study. Among them, 3,997 individuals (6.8%) had newly developed hypertension in year X + 5. The indices of predictive ability were as follows: Random Forrest, AUC 0.88, accuracy 0.93, recall 0.99, precision 0.93; Logistic Regression, AUC 0.89, accuracy 0.93, recall 0.99, precision 0.94; Neural Network, AUC 0.89, accuracy 0.93, recall 0.99, precision 0.93. The variables with the highest importance features were baseline creatinine level (0.49), age (0.24), systolic blood pressure (0.07), and anti-hypertensive medication (0.06). Conclusions: The AUCs of the model predicting the development of CKD after 5 years ranged from 0.88 to 0.89, which were generally superior to previously reported risk models using statistical methods. Baseline creatinine levels, age, and blood pressure were found to contribute significantly to new-onset CKD.