Skip to main content
2024 AABB Annual Meeting Main banner
Icon Legend
This session is not in your favorites.
This session is in your favorites. Click again to remove it.
Presentation Icons
Live Stream Sessions
Live Stream Sessions

Abstract

Blood Center/Blood Hospital-Based Donor Center

Oral Abstract - Blood Donor Issues

OA2-AM24-ST-17 - External Validation of Machine Learning Models Predicting Iron Recovery After Blood Donation

Saturday, October 19, 2024
11:15 AM - 12:15 PM  
Location: 372
CE: Zero
Background/Case Studies: Models predicting ferritin and hemoglobin when blood donors return could help manage donation-associated iron deficiency and avoid wasteful low hemoglobin deferrals. Such biomarker prediction models have not been externally validated internationally.

Study

Design/Methods: We use a cohort of 2,454 repeat blood donors from the 2009-2011 REDS-II Donor Iron Status Evaluation (RISE) study in the United States (US) to develop machine learning models predicting hemoglobin and log10 ferritin. In a ‘hemoglobin only’ dataset, we used donation history, demographics, and baseline hemoglobin at index donations to predict biomarkers at a return visit at which hemoglobin and ferritin were both measured. In a smaller ‘hemoglobin and ferritin’ dataset, we also used baseline ferritin to predict return biomarkers. We assessed >1,000 model configurations using three repeats of five-fold cross validation, selecting the configuration that minimized root mean square percent error (RMSPE). We retrained the selected model configuration on the entire RISE dataset and externally validated on 2014-2023 operational data from the US, South Africa (SA), and the Netherlands (NL).

Results/Findings:

Selected models used Catboost and gradient boosted machine algorithms. In the ‘hemoglobin only’ RISE dataset (n=3,488 donations), the selected models’ cross-validation RMSPE was 6.5 for predicting return hemoglobin and 24.2 for predicting return log10 ferritin. When externally validating to the US (n=60,403), SA (n=253,537), and the NL (n=514,117), RMSPE increased by < 15% whether predicting hemoglobin (RMSPE=7.2, 7.3, and 5.9) or log10 ferritin (RMSPE=22.9, 27.6, and 19.1; Figure A). In the ‘hemoglobin and ferritin’ RISE dataset (n=2,625 donations), the selected models’ cross-validation RMSPE was 6.6 for predicting return hemoglobin and 14 for predicting return log10 ferritin. When externally validating to the US (n=11,025), SA (n=12,564) and NL (n=179,423), RMSPE increased by < 2% when predicting hemoglobin (RMSPE=4.6, 6.7, 6.3), and increased by < 33% when predicting log10 ferritin (RMSPE=14.9, 18.9 [+32%], 15).

Conclusions: Machine learning models generalized well across diverse settings, particularly when predicting return hemoglobin. Measuring baseline ferritin had a small impact on ability to predict return hemoglobin but greatly improved prediction of return log10 ferritin. Limitations include that our cross validation RMSPE may be optimistic due to overfitting and that each cohort’s policy on which donations are tested for ferritin could bias results.