Machine Learning Model Predicts Incident Afib in CKD Patients


Using data from the CRIC study, a team of investigators has designed a machine learning model they suggest could predict AF incidence in patients with chronic kidney disease.

Leila Zelnick, PhD

Leila Zelnick, PhD

New data from a study presented at ASN Kidney Week 2020 provides clinicians suggests a pair of biomarkers used to identify patients at increased risk of myocardial injury could help predict which chronic kidney disease (CKD) patients were most likely to develop atrial fibrillation (AF).

The most common form of arrhythmia, results of the study indicate high sensitivity troponin T (hsTnT) and N-terminal brain natriuretic peptide (NT-proBNP) could help improve prediction of atrial fibrillation in patients with CKD, especially when paired in machine learning algorithms.

“The application of such a model could be used to identify patients with CKD who may benefit from enhanced cardiovascular care and also to identify selection of patients for clinical trials of AF therapies,” said lead author Leila Zelnick, PhD, assistant professor in the Division of Nephrology at the University of Washington, in a statement.

With previous studies demonstrating the prognostic value of hsTnT and NT-proBNP for predicting AF in community-based populations, Zelnick and a team from institutions across the US sought to examine the value of these biomarkers for predicting AF using machine learning methods. With this in mind, investigators identified the Chronic Renal Insufficiency Cohort (CRIC) as the data source for their analysis.

Beginning in 2001, CRIC contains data related to nearly 4000 CKD in the US. For the purpose of the analysis, investigators only included those without prior AF–yield a population of 2690 participants. Investigators noted participants were also excluded if they did not have complete cardiac biomarker, demographics, medical history, lifestyle information, medication history, physical characteristics, and laboratory data available for analysis.

Investigators performed their analyses using Cox regression, LASSO, ridge regression, elastic net, boosting methods, and a previously-validated prediction model to predict incident AF. Of note, the clinical prediction model used in the analysis was validated in the CHARGE-AF study.

The mean age of the 2690-patient population was 57 (SD, 11) years, 55% were men, 38% were black, and the group had a mean eGFR of 45 (SD, 15) mL/min/1.73m2. During a follow-up period lasting a mean of 7.3 (SD, 2.8) years, 251 participants experienced an incident AF event.

In their analyses, the CHARGE-AF predication equations using original and re-estimated coefficients both had a cross-validated C-index of 0.69. In a LASSO model analysis using only clinical data, results indicated the model achieved a C-index of 0.69. When adding NT-proBNP, hsTnT, or both biomarkers, the LASSO model achieved a C-index of 0.75, 0.73, and 0.76, respectively (P for all <.0001).

This study, "Prediction of Atrial Fibrillation Using Clinical and Cardiac Biomarker Data: The CRIC Study,” was presented at the American Society of Nephrology Kidney Week 2020.

Related Videos
© 2024 MJH Life Sciences

All rights reserved.