Prediction of blood donor return behaviour using machine learning.

Area of Science:

• Biomedical Informatics
• Public Health
• Data Science

Background:

• Traditional methods for assessing blood donor return behavior rely on questionnaires.
• There is a need for enhanced donor invitation models that predict donation dynamics using behavioral variables.
• Machine learning (ML) offers a promising approach to process blood donor datasets for improved return rates.

Purpose of the Study:

• To incorporate machine learning into processing blood donor datasets.
• To develop an enhanced donor invitation model for predicting donation dynamics.
• To increase blood donor return rates through precision marketing.

Main Methods:

• A retrospective study was conducted using blood donor records from the Kaohsiung Blood Center, Taiwan (2022-2023).
• Three machine learning models (MLMs) were employed: eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost).
• Models were cross-validated and evaluated using accuracy, average precision, balanced accuracy, F1 score, and log loss. The best model generated predictive donor return lists for testing.

Main Results:

• All three MLMs demonstrated strong predictive power, with CatBoost performing best.
• Key influential factors identified were return donation interval and donation location.
• A significant improvement in donor return rates was observed (57.99% vs. 15.54%, p < 0.0001).

Conclusions:

• Machine learning models, utilizing behavioral and demographic variables, can effectively predict and identify high-return-rate blood donors.
• Precision marketing strategies based on these predictions significantly enhance donor return rates.
• This approach enables blood banks to more rapidly meet transfusion demands, especially during emergencies.