Research on the generalization of a blood donor recruitment framework based on machine learning.

BACKGROUND: Traditional short message service (SMS) recruitment, an essential method for recruiting blood donors, constitutes a highly imbalanced classification problem and still requires improvement in terms of efficiency and cost-effectiveness. METHODS: Using 513,523 donations and 11,616 SMS data from Nanjing, China, several machine learning models—including eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), and Multi-Layer Perceptron (MLP)—were initially trained, followed by an incremental learning strategy and a dynamic threshold adjustment mechanism, sampling techniques such as Synthetic Minority Oversampling Technique (SMOTE) and under-sampling (US) were employed, in combination with cost-sensitive learning approaches, including Minimum Feasibility Error (MFE), Minimum Feasibility Error (MSFE), and weighted mean squared error loss functions. Then, by prioritizing recall as the core evaluation metric (supplemented by other indicators), the 2 models with the best performance were selected and fine-tuned with 10% of data from Suzhou and Yangzhou, China. Finally, a prospective study using a machine learning-based recruitment framework was conducted in 3 cities and its performance was compared with traditional recruitment methods. RESULTS: 72 for MLP and 0.70 for RF. The fine-tuned models demonstrated good generalizability, achieving recall rates of 0.63 and 0.67 in Suzhou, and 0.58 and 0.63 in Yangzhou, respectively. Further optimization led to improved recall rates of 0 .70 and 0.77 in Suzhou, and 0.68 and 0.64 in Yangzhou. In prospective validations across 13 recruitment rounds in 3 cities, the ML-driven approach significantly outperformed conventional methods. It increased the recruitment success rate for high-willingness donors by an average of 151.57% (up to 408.28% in Nanjing) while recruitment efficiency increased by an average of 96.52% (up to 138.75% in Yangzhou). Consequently, recruitment efficiency per SMS improved by an average of 34.42% (up to 53.61% in Suzhou). CONCLUSIONS: This study underscores and validates the effectiveness and generalizability of machine learning in enabling precise blood donor recruitment. The ML–driven recruitment framework exhibits strong reproducibility and replicability, which highlight its potential for broad deployment across diverse settings and offers a practical, cost-effective strategy to enhance blood supply stability. This data-driven approach represents a meaningful advancement in the modernization of blood services in China and holds important implications for improving donor engagement, ensuring the safety and stability of the blood supply, thereby promoting a more efficient and sustainable approach to maintaining a stable blood supply. CLINICAL TRIAL NUMBER: Not applicable