Blood donor recruitment currently faces significant challenges, as declining donor willingness continues to place increasing pressure on recruitment efforts. Developing precise strategies to improve both efficiency and quality has therefore become essential. With the ongoing digitalization of blood services and the emergence of large-scale databases, machine learning (ML) solutions are increasingly being adopted in transfusion medicine^1,2. ML techniques offer promising tools to address challenges across the transfusion supply chain—enhancing donor recruitment efficiency and optimizing blood resource management. Compared to traditional approaches, which randomly select a list of donors who meet the requirements for blood donation, ML models provide substantial advantages in processing large volumes of data and uncovering latent patterns. Notably, they can effectively capture complex nonlinear relationships without relying on linear assumptions, handle high-dimensional feature spaces, and reveal intricate interdependencies among variables. Furthermore, through incremental learning³, ML models can be continuously updated in real time, enabling rapid adaptation to new data. These capabilities confer greater efficiency, adaptability, and analytical power for identifying hidden trends in large-scale datasets. Although several ML-based approaches have been explored for blood and hematopoietic stem cell donor recruitment, most prior studies have focused primarily on evaluating model performance from a data modeling standpoint, without extending to real-world applications^4,5.

Our previous work⁶ developed a machine learning–based blood donor recruitment model using data from Yangzhou, where accuracy was employed as the primary evaluation metric. The results demonstrated that machine learning–based precise recruitment at a single center outperformed traditional methods. However, that study did not evaluate the generalizability of the recruitment framework. Moreover, because donor recruitment constitutes a highly imbalanced classification problem—where the number of non-donors greatly exceeds that of actual recruiters—the accuracy, which reflects the overall proportion of correctly predicted samples, may not be the most appropriate evaluation metric. In contrast, recall more effectively captures the proportion of true positive cases identified, making it a more suitable metric for this task. To better address this class imbalance, the present study adopts the principle of “recall as the primary metric, supported by other indicators” in model construction and performance evaluation. Specifically, models were initially trained on data from Nanjing and subsequently fine-tuned using 10% of the data from Yangzhou and Suzhou to assess cross-center transferability. Building on this foundation, optimized models were developed and applied to targeted SMS-based recruitment in all three cities—Nanjing, Suzhou, and Yangzhou—resulting in more effective and accurate donor recruitment outcomes.

Each donor is linked to their corresponding SMS message via a unique identification code, and their donation status is tracked within seven days of receiving the message. Using the donor ID, several variables are derived: the total number of donations, cumulative donation volume, donation interval (i.e., the time between the SMS and the most recent prior donation), and donation frequency—calculated as (date of last donation – date of first donation) / total number of donations. Additionally, donor age is computed as (SMS sending date – date of birth) / 365, and repeat donor status is defined as having at least two recorded donations in the historical data.

Recruitment models, with recall as the primary evaluation metric, were initially trained and fine-tuned using data from Nanjing. Donation records from 2017 to 2021 were analyzed in conjunction with SMS recruitment data to capture historical donor behavior and characterize donation patterns—including donation count, frequency, total volume, intervals, and age—thus generating comprehensive donor profiles. SMS data from 2022 were subsequently used to assess donor willingness by linking response outcomes to these profiles. This enabled the identification of features distinguishing high- and low-willingness donors and the construction of a labeled dataset comprising effective donors (those who donated within seven days of receiving the SMS) and ineffective donors (those who did not).

The SMS recruitment data collected from Nanjing in 2022 were divided into a training set and a test set using a 70:30 split. Several machine learning models—including eXtreme Gradient Boosting (XGBoost) ⁷, Support Vector Machine (SVM) ⁸, K-Nearest Neighbors (KNN), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF) ⁹, and Multi-Layer Perceptron (MLP) ¹⁰—were trained on the training set. To address class imbalance, sampling techniques such as Synthetic Minority Oversampling Technique (SMOTE) ¹¹ and under-sampling (US) were employed, in combination with cost-sensitive learning approaches, including MFE, MSFE, and weighted mean squared error loss functions ^12,13. Each model (XGBoost, SVM, KNN, LR, RF, DT, MLP) was evaluated under different sampling strategies (raw data, US, SMOTE) using a grid search over four performance metrics: accuracy, precision, recall, and F1-score. Model selection was primarily based on the recall value, with the remaining metrics used for reference. All models were implemented in Python programming language using the XGBoost, Scikit-learn ¹⁴, and PyTorch ¹⁵ libraries.

To evaluate the generalizability of our approach, MLP and RF models initially trained on data from Nanjing were fine-tuned using 10% of donor records from Suzhou and Yangzhou. An incremental learning paradigm was employed to gradually align the feature distributions of the new datasets with the original Nanjing framework. This strategy enabled the assessment of the models’ ability to generalize across different regional datasets while retaining learned knowledge from the source domain. The results provide valuable insights into the robustness and cross-center applicability of machine learning–based precision recruitment strategies.

Assuming that blood donation records and SMS recruitment data can be used to rank and predict donor willingness, the selected MLP and RF models were jointly employed to conduct prospective blood donor recruitment in Nanjing, Suzhou, and Yangzhou, China. The procedure was as follows: (1) Candidates were drawn from SMS recruitment lists provided by the local blood centers, and their key characteristics were input into the trained MLP and RF models to generate donation probability scores reflecting predicted willingness; (2) Donors with scores above 0·5 in both models were selected, prioritizing those appearing in the intersection of MLP and RF predictions, followed by those identified by either model individually; (3) To ensure comparability, the number of selected donors matched that of conventional recruitment efforts. These high-willingness donors were recruited for a controlled study, and donation outcomes within seven days of SMS delivery were recorded. Based on these trained models, an incremental learning paradigm 3 was further applied to fine-tune and iteratively update the MLP model using the most recent donation data. In total, 13 prospective recruitment studies were conducted across the three cities, leveraging the updated MLP and RF models and comparing their performance against traditional recruitment methods.

Blood donor recruitment in China currently faces several critical challenges: (1) a nationwide decline in donations—voluntary blood donations (15·821 million) and total blood volume collected (26·924 million units) decreased by 6·9% compared to 2023; (2) few institutions seeing growth in blood collection—only 4 out of 395 prefecture-level or higher blood collection institutions reported positive growth in 2024; and (3) an aging donor population—with a noticeable decline in younger donors, as the proportion of donors under age 35 dropped from 65% in 2019 to 52% in 2024· These developments raise serious concerns about the sustainability and security of the national blood supply, underscoring the urgent need for targeted recruitment strategies, data-driven interventions, and improved engagement of younger populations to ensure a safe, stable, and clinically sufficient blood supply.

In recent years, data-driven statistical and artificial intelligence (AI) methods have become increasingly important for predicting donor willingness and forecasting future blood supply needs. Hanapi WHWH et al. employed logistic regression to identify key factors influencing donation intention, including donation interval, frequency, and total volume ¹⁶. Salazar-Concha C et al. applied decision trees, achieving 84·17% accuracy in predicting donor re-engagement ¹⁷. Marade C et al. utilized SVM, decision trees (DT), and KNN to forecast monthly blood donations, aiding inventory management ¹⁸. Ding X et al. proposed a hybrid SARIMAX-LSTM neural network for accurate donation estimation ¹⁹, while El-Rashidy N et al. and Selvaraj P et al. reported that random forest models yielded the best performance in predicting donation and re-donation behaviors ^20,21. Other studies have leveraged SVM for donor classification and donation outcome prediction ^22,23, and Zulfikar W.B. et al. found decision trees to outperform naive Bayes classifiers ²⁴. Cloutier M. focused on predicting return rates among young donors ²⁵. Several reviews—including those by Kiarie et al. ²⁶, Li Y et al. ²⁷, Buturovic L et al. ²⁸, Sivasankaran A et al. ²⁹, and Gupta V et al. ³⁰—have highlighted the growing application of machine learning in donor retention, recruitment optimization, post-transplant outcomes, and hematopoietic stem cell transplantation (HSCT), while also acknowledging current limitations and challenges. Collectively, these studies underscore the value of AI and big data analytics in enhancing blood donor recruitment and ensuring a safe, sufficient, and sustainable clinical blood supply.

This study demonstrates the generalizability and practical effectiveness of our machine learning–based recruitment framework by adopting recall as the primary evaluation metric—contrasting with previous studies that largely evaluated performance from a data modeling perspective. In earlier work, we developed seven recruitment models using XGBoost and SVM on a dataset comprising 697,174 blood donation records (October 1998–2019) and 95,476 SMS recruitment records (April 2016–July 2019) from Yangzhou, incorporating 13 donor features. This laid the groundwork for the Nanjing recruitment model. In the present study, we extend this foundation into a prospective setting, employing multi-layer perceptron (MLP) and random forest (RF) models, with recall as the central metric. While MLP, as a deep learning algorithm, offers strong representational capacity, its performance is sensitive to data scale and resource demands. To address this, we complemented MLP with RF, a more lightweight and robust traditional machine learning method. Given the limited size of available blood donation data, we adopted an incremental learning paradigm: following each recruitment round, newly collected data were used to fine-tune the MLP model, improving its performance before deployment in subsequent rounds. As illustrated by Table S8 (Supplementary Material), we began with an initial dataset D₀. After performing the recruitments in some universities in Nanjing, China in our previous work, we obtained some most up-to-date blood donation data, denoted as D1· Therefore, we can get an augmented dataset D0∪D1 (denoted as D01). We randomly selected a training set and a test set in a 7:3 ratio on D01 and selected a new model trained on such a training set with a higher value recall than the old model on such a testing set. This new model was applied in the 1st recruitment in this work. After performing the 1st recruitment, we can obtain some new data D2 and use D01∪D2 (denoted as D11) to fine-tune the MLP via similar procedures as stated above. As shown in Tables S9 and S10, the updated models consistently outperformed their predecessors in terms of identifying high-willing donors and improving recruitment efficiency (measured by success rate per SMS). Some exceptions were observed during simultaneously conducted recruitments (e.g., 1st and 2nd, 4th and 5th, 6th and 7th rounds), where a single model could not optimize performance across both settings. Nonetheless, the overall trend confirmed the effectiveness of the incremental learning paradigm. Detailed fine-tuning results are provided in Tables S11–S14. Furthermore, models initially trained on Nanjing data were successfully fine-tuned with data from Suzhou and Yangzhou, validating the cross-center generalizability of our approach. Across 13 recruitment rounds, the machine learning framework improved the recruitment success rate for high-willing donors by an average of 151·57% compared with conventional methods, reduced SMS volume by 96·52%, and increased recruitment efficiency per SMS by 34·42%. These results highlight the framework’s ability to identify high-potential donors, reduce unnecessary messaging, and lower overall recruitment costs. Analysis of the Venn diagrams in Figures S1–S8 revealed minimal overlap between donors identified by both MLP and RF and those recruited via conventional methods—likely due to high redundancy in traditional approaches. By adjusting prediction thresholds to vary repetition ratios, we observed that lower overlap corresponded to greater performance improvements, as shown in Fig. 3. Model hyperparameters were as follows: the MLP consisted of three layers with 6, 8, and 2 neurons for the three cities, ReLU activation after the second layer, and a Softmax layer for classification. The RF model used the Gini impurity criterion, a maximum depth of 9, a minimum of 5 samples per leaf, and balanced class weights (1:1). As shown in Tables S13 and S14, MLP consistently achieved higher overall recruitment success than RF—both for high-willing donors and the general donor population—aligning with its superior dataset-level performance.

Blood donor recruitment presents a highly imbalanced classification problem, with substantially more non-responders than responders. In such scenarios, simultaneously achieving high accuracy and high recall is challenging, and the F1 score may not always offer meaningful insight. Because the primary goal is to identify as many willing donors as possible, recall is prioritized as the key evaluation metric. This section explains the rationale for adopting the principle of “recall as the primary metric, with other metrics serving as auxiliary criteria” in model selection. We begin by reviewing standard evaluation metrics in binary classification. A correctly predicted positive sample is termed a True Positive (TP), while a negative sample incorrectly predicted as positive is a False Positive (FP). A positive sample incorrectly predicted as negative is a False Negative (FN), and a correctly predicted negative sample is a True Negative (TN). Based on these definitions: Accuracy = (TP + TN) / (TP + TN + FP + FN) Precision = TP / (TP + FP) Recall = TP / (TP + FN) F1 score = 2 × (Precision × Recall) / (Precision + Recall) Accuracy reflects the overall proportion of correct predictions, precision measures the correctness among predicted positives, and recall quantifies the proportion of actual positives that are successfully identified. The F1 score serves as a harmonic mean of precision and recall, balancing the two. However, these metrics may lead to misleading interpretations in imbalanced settings. Consider a population of 100 individuals, where 20 are true positives and 80 are negatives. Suppose Model A predicts all 100 as positive—yielding a recall of 1·0 but an accuracy of only 0·20· In contrast, Model B correctly identifies 10 of the 20 positive cases and misclassifies only 5 of the 80 negatives as positives—resulting in a recall of 0·67 and an accuracy of 0·85· Although Model A achieves perfect recall, it produces many unnecessary messages; Model B, by contrast, strikes a more practical balance by recruiting half of the willing donors while minimizing redundant outreach. Therefore, while recall remains our primary metric—given the high stakes of missing willing donors—it must be interpreted alongside precision and accuracy to ensure practical and efficient recruitment.

In our previous work 6, model selection was based on accuracy. However, in the current datasets from Nanjing, Suzhou, and Yangzhou, the proportion of positive cases is extremely low—only 4·71%, 0·49%, and 3·67%, respectively. In such imbalanced settings, a model that predicts all samples as negative would still achieve high accuracy (approximately 95%, 99%, and 96% for the three cities), but would be of little practical value for identifying potential donors. Consequently, we adopt the principle of prioritizing recall as the primary evaluation metric, while treating accuracy, precision, and other metrics as auxiliary. In practical implementation, models with accuracy below 0·5 are discarded, and the optimal model is selected from the remaining candidates based primarily on recall. Additionally, recruitment contexts differ notably across cities. In Nanjing, recruitment efforts primarily target university students, resulting in a relatively small and concentrated donor pool. In contrast, recruitment in Suzhou and Yangzhou is community-based, encompassing a broader and more diverse population. These differences in target populations and recruitment volumes are likely to influence the extent of performance improvements achieved by machine learning models across regions.

In future work, the performance of machine learning models for donor recruitment can be further improved through several avenues. For example, model parameters can be optimized using additional evaluation metrics—such as accuracy and F1 score—to identify models that perform well across multiple criteria, beyond recall alone. Moreover, building on the insights from the threshold adjustment experiment, heuristic algorithms could be employed to dynamically determine the optimal prediction threshold for each recruitment round. This approach would aim to minimize overlap with conventional recommendation lists, thereby improving the overall effectiveness and efficiency of the recruitment strategy.

This study underscores and validates the effectiveness and generalizability of machine learning in enabling precise blood donor recruitment. The results demonstrate that machine learning–based recruitment models significantly outperform traditional methods by increasing recruitment success rates while substantially reducing the number of SMS messages required. These findings suggest that machine learning can greatly enhance the efficiency of SMS-based recruitment by minimizing outreach volume and maximizing donor response, thereby contributing to the development of more targeted and cost-effective donation strategies. Furthermore, the proposed models exhibit strong reproducibility and replicability, highlighting their potential for broad deployment across diverse settings. 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, and promoting the long-term sustainability of transfusion systems.