Risk assessment for patients with sickle cell disease (SCD) remains challenging as it depends on an individual physician's experience and ability to integrate a variety of test results. We aimed to provide a new risk score that combines clinical, laboratory, and imaging data. In a prospective cohort of 600 adult patients with SCD, we assessed the relationship of 70 baseline covariates to all-cause mortality. Random survival forest and regularised Cox regression machine learning (ML) methods were used to select top predictors. Multivariable models and a risk score were developed and internally validated. Over a median follow-up of 4·3 years, 131 deaths were recorded. Multivariable models were developed using nine independent predictors of mortality: tricuspid regurgitant velocity, estimated right atrial pressure, mitral E velocity, left ventricular septal thickness, body mass index, blood urea nitrogen, alkaline phosphatase, heart rate and age. Our prognostic risk score had superior performance with a bias-corrected C-statistic of 0·763. Our model stratified patients into four groups with significantly different 4-year mortality rates (3%, 11%, 35% and 75% respectively). Using readily available variables from patients with SCD, we applied ML techniques to develop and validate a mortality risk scoring method that reflects the summation of cardiopulmonary, renal and liver end-organ damage. Trial Registration: ClinicalTrials.gov Identifier: NCT#00011648.
Trial registration: ClinicalTrials.gov NCT00011648.
Keywords: machine learning; risk assessment; sickle cell anaemia.
Published 2021. This article is a U.S. Government work and is in the public domain in the USA.