SP347RISK AND CLINICAL OUTCOMES ASSOCIATED WITH CKD ANEMIA IN THE USA: A PREDICTIVE MACHINE LEARNING MODELING APPROACH USING NHANES

Abstract

INTRODUCTION: Anemia is a common complication among non-dialysis dependent (NDD) chronic kidney disease (CKD) patients. CKD patients with anemia generally have lower quality of life and higher risk of negative clinical outcomes. Early identification of NDD CKD patients in need of more frequent monitoring for and treatment of anemia may benefit this population. Objectively developed predictive models that assess both risk for anemia and negative clinical outcomes associated with anemia among NDD CKD patients can support timely treatment. We present a systematic study based on machine learning towards this goal. METHOD(S): We analyzed lab values, demography and prior comorbidity data for 54,811 subjects from the National Health and Nutrition Examination Survey (NHANES, 1999-2016) and National Death Index (NDI, 1999-2011). 4,306 NDDCKD patients were identified among these subjects via eGFR calculated using the CKDEPI creatinine equation, which was also used for CKD staging. KDIGO definition of anemia was used in this analysis. Standard classification algorithms were applied to assess how different risk factors are associated with the discriminatory patterns within the NDD population, as well as to estimate a statistical model using a fraction of the data for risk of anemia, and clinical outcomes such as hospitalization, major adverse cardiovascular events (MACE) and all-cause mortality among CKD patients. The heldout data was used to determine model accuracy, which was enumerated via area under the curve (AUC). Additionally, Cox regression was applied to estimate survival functions, supported by log-rank tests. RESULT(S): A total of 983 (23%) NDD-CKD patients were found to be anemic in this cohort, of whom 785 (80%) were CKD stage 3, 150 (15%) stage 4 and 48 (5%) stage 5 patients. Our analysis revealed serum iron level, age and hemoglobin level as most important determinants of risk for developing anemia and corresponding negative clinical outcomes in NDD-CKD patients. The predictive accuracy of the models, as measured by AUC, were as follows-risk of anemia: 0.82, hospitalization: 0.68, MACE: 0.73, mortality: 0.70. Importantly, our analysis reconfirmed that presence of CKD, in conjunction with anemia, is associated with higher risk of all-cause mortality than CKD or anemia alone. Results of pairwise log-rank tests were highly significant: CKD and Anemia vs CKD (mortality HR = 1.62, 95% confidence interval: 1.36-1.92, p< 0.001), and were consistent across different comorbidities, such as diabetes, hypertension and congestive heart failure. CONCLUSION(S): Analysis of recent US NHANES data clearly demonstrates that anemia is a potential multiplier of risk of mortality in NDD CKD. The machine learning models described may be used to assess the risk of anemia in CKD patients, thus allowing an opportunity for early identification and intervention if necessary, thereby supporting healthcare decision making.