Prediction of plasma volume and total hemoglobin mass with machine learning

Abstract

Hemoglobin concentration ([Hb]) is used for the clinical diagnosis of anemia, and in sports as a marker of blood doping. [Hb] is however subject to significant variations mainly due to shifts in plasma volume (PV). This study proposes a newly developed model able to accurately predict total hemoglobin mass (Hbmass) and PV from a single complete blood count (CBC) and anthropometric variables in healthy subject. Seven hundred and sixty-nine CBC coupled to measures of Hbmass and PV using a CO-rebreathing method were used with a machine learning tool to calculate an estimation model. The predictive model resulted in a root mean square error of 33.2 g and 35.6 g for Hbmass, and 179 mL and 244 mL for PV, in women and men, respectively. Measured and predicted data were significantly correlated (p < 0.001) with a coefficient of determination (R2) ranging from 0.76 to 0.90 for Hbmass and PV, in both women and men. The Bland–Altman bias was on average 0.23 for Hbmass and 4.15 for PV. We herewith present a model with a robust prediction potential for Hbmass and PV. Such model would be relevant in providing complementary data in contexts such as the epidemiology of anemia or the individual monitoring of [Hb] in anti-doping.