Predicting haemoglobin deferral using machine learning models: Can we use the same prediction model across countries?

Area of Science:

• Transfusion medicine
• Predictive modeling in healthcare
• Blood donor management

Background:

• Personalized hemoglobin (Hb) prediction models can reduce donation deferrals and costs.
• Previous research indicated better model performance with high Hb deferral rates.
• The study explores the generalizability of Hb deferral prediction models across different blood collection agencies.

Purpose of the Study:

• To evaluate the performance of Hb deferral prediction models when shared between blood establishments.
• To determine if models trained in one setting perform well in others.
• To assess the impact of training data origin on model generalizability.

Main Methods:

• Random forest models were developed using 5 years of donation data from 10,000 donors across five countries.
• Trained models were exchanged between participating blood establishments.
• Model performance was quantified using the area under the precision-recall curve (AUPR); variable importance was assessed using SHAP values.

Main Results:

• The area under the precision-recall curve (AUPR) ranged from 0.05 to 0.43 across validation datasets and exchanged models.
• Exchanged models demonstrated similar performance irrespective of the training data's origin.
• Predictor variable importance was largely consistent across all trained models, with only minor variations.

Conclusions:

• Hb deferral prediction models exhibit similar performance when applied to validation datasets from different blood establishments.
• Model generalizability is not significantly affected by the deferral rate of the training data.
• Blood establishments appear to learn comparable associations relevant to Hb deferral prediction.