Dynamic Bayesian networks to predict loss of kidney function: a cross-institution use case in a large cohort with or...

Substantial loss of kidney function, measured as ≥40% decline in estimated glomerular filtration rate (eGFR) within a 2-year period, is associated with a tenfold increase in the risk for kidney failure.

We developed and externally validated dynamic Bayesian networks (DBNs) to predict ≥ 40% eGFR decline using electronic health record (EHR) data from Providence and UCLA Health.

Of 2.25 million patients, with and at-risk for chronic kidney disease (CKD), 6.49% (146,043 individuals) experienced at least one occurrence of ≥40% decline from baseline eGFR over six years of follow-up. The DBNs demonstrated strong predictive performance, as measured by the area under the receiver operating characteristic curve (AUCROC) and average precision (AP), with the highest values observed in the final year (Year 6), ranging from 0.83-0.89 and 0.28-0.37, respectively. A comparison of existing gold-standard CKD-Prognosis Consortium risk equations in real-world clinical settings with missing data demonstrated that DBNs' performance remained unaffected, while risk equations performed close to random due to their inability to handle missing data. The temporal structure of the DBNs captured longitudinal features changes and their interactions, with the most important observations over time including the urine albumin-creatinine ratio (UACR), the urine protein-creatinine ratio, and hemoglobin A1c. Notably, comparing DBN performance across institutions revealed that training on larger datasets generalized better.

This study offers valuable insights into the development of DBNs using real-world population data from EHRs. The DBNs successfully identified patients at-risk for ≥40% eGFR decline and, despite missing data, offer opportunities for timely intervention and risk mitigation to preserve kidney function.