One of the most important transitions in the life of a patient with chronic kidney disease (CKD) is the development of disease progression to kidney failure that necessitates treatment with hemodialysis, peritoneal dialysis, or kidney transplantation. While over 100,000 Americans with CKD progress to end stage renal disease (ESRD) annually, clinicians are often unable to predict which of their CKD patients will progress to kidney failure over time. Existing risk prediction models are limited because they rely on demographic and laboratory values from a single time point. In its report Best Care at Lower Cost, the Institute of Medicine notes that just as the information revolution has transformed many other fields, growing stores of data hold the same promise for improving clinical research, clinical practice, and clinical decision making. This proposal seeks to tackle this goal for patients with CKD by deriving and validating a prediction model for kidney failure using concepts derived from notes using natural language processing. Because each patient's electronic health record contains thousands of concepts, building a prediction model requires the use of big data analytic methods and significant computational power that have only become available recently due to advances in biostatistics and computing. In the first aim, two competing models (lasso Cox regression and random survival forest) will be developed to predict the onset of end stage renal disease in patients referred to the Brigham and Women's Hospital nephrology clinic. One of the two models will be chosen based on performance on measures of discrimination and calibration. In the second aim, the chosen high-dimensional natural language processing-derived model will be externally validated in patients seen in the Massachusetts General Hospital nephrology clinic and compared against a contemporary prediction model. If successful, this proposal will change the paradigm of risk prediction in nephrology and better equip nephrologists to counsel CKD patients on their kidney failure risk. Future research stemming from this work will focus on testing whether implementation of the prediction rule in a randomized trial can improve the care of patients with CKD, as assessed by timely AV fistula referral and patient reported outcome measures.
One of the most difficult questions that patients with chronic kidney disease ask their kidney doctors is 'when will my kidneys fail?' Clinical notes in the electronic health record contain a wealth of information that may enable physicians to make more accurate predictions about the onset of kidney failure. This proposal uses novel methods including natural language processing and big data analytics in order to improve the prediction of kidney failure among patients referred to a kidney disease clinic.