Prediction of blood donor return behaviour using machine learning

Blood donor return behaviour is frequently examined through questionnaires assessing donors' motivations, attitudes, awareness and willingness to donate. An enhanced donor invitation model that better predicts donation dynamics using specific behavioural variables is needed. This study aims to incorporate machine learning into processing blood donor datasets, thereby increasing donor return rates through precision marketing.

This retrospective study was conducted between 1 January 2022 and 31 December 2023, focusing on blood donors whose records were maintained by the Kaohsiung Blood Center in Taiwan. Three machine learning models (MLMs), namely eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and Categorical Boosting (CatBoost), were employed. Each model underwent cross-validation, and their performance was evaluated using accuracy, average precision, balanced accuracy, F1 score and log loss. The best performing model was selected to generate predictive donor return lists, which were subsequently tested.

All three MLMs demonstrated strong predictive power in generating donor return prediction lists. The performance metrics for the three models were as follows: accuracy: 0.8093, 0.8096 and 0.8106; specificity: 0.9227, 0.9256 and 0.9229. CatBoost was found to be the best performing model. The most influential predictive factors were return donation interval and donation location. Results showed a significant improvement in donor return rates (57.99% vs. 15.54%, p < 0.0001).

MLMs leveraging both behavioural and demographic variables can predict and identify high-return-rate donors by prioritizing return thresholds. Precision marketing significantly improves donor return rates, enabling blood banks to rapidly meet transfusion demands during emergencies.