Machine-learning models to predict iron recovery after blood donation: a model development and external validation
Wanjin Li1, Chen-Yang Su2, Amber Meulenbeld3
1Department of Epidemiology, Biostatistics and Occupational Health, McGill School of Population and Global Health, Montréal, Canada.
Background:
Machine-learning models directly predicting iron biomarkers after blood donation could help to manage donation-associated iron deficiency and avoid low haemoglobin deferrals. No such models have been externally validated internationally. Our aim was to develop and externally validate machine-learning models predicting returning blood donors' haemoglobin and ferritin.
Methods:
We developed machine-learning models using retrospective blood donation data. The training cohort included 2425 repeat donors (2007-09 US-based RISE study); external validation used 2014-23 cohorts from the USA, South Africa, and the Netherlands. Models predicted donors' ferritin and haemoglobin at return donations by use of variables that are commonly measured by blood collectors (time until donors return, donation history, demographics, and baseline iron biomarkers). Models were selected via cross-validation and externally validated in donors aged at least 15 years in two contexts: those with baseline ferritin and haemoglobin measured (haemoglobin and ferritin) and those with only baseline haemoglobin measured (haemoglobin only). Model performance was assessed by use of root-mean-square percentage error (RMSPE).
Findings:
When predicting return haemoglobin in the RISE cohort, model performance was similar in the haemoglobin and ferritin dataset (n=2625 donation visits, RMSPE=6·78) and haemoglobin only dataset (n=3488 donation visits, RMSPE=6·78). In the external datasets, containing 11 000 to 514 000 donations, RMSPE never increased more than 8%. When predicting return ferritin in RISE, performance was better in the haemoglobin and ferritin dataset (RMSPE=14·9) than in the haemoglobin only dataset (RMSPE=27·4). In external validation, RMSPE never increased more than 0·4% and 28% in the haemoglobin-only datasets and the haemoglobin and ferritin datasets, respectively.
Interpretation:
Machine-learning models predicting haemoglobin and ferritin at return donations generalised well across diverse settings and could enable individualised approaches to manage iron deficiency while maintaining a sufficient blood supply.
Funding:
The Association for the Advancement of Blood and Biotherapies.
Translation:
For the Dutch translation of the abstract see Supplementary Materials section.
More Related Videos
07:15Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
07:13Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
