Blood donation adverse reaction risk prediction method and system based on stratified blood volume estimation

Abstract

This invention discloses a method and system for predicting the risk of adverse reactions to blood donation based on stratified blood volume estimation, belonging to the field of medical informatics and intelligent risk prediction technology. The prediction method includes the following steps: collecting basic information of blood donors; calculating body mass index (BMI) based on height and weight; selecting a blood volume estimation formula based on BMI stratification and calculating the estimated blood volume EBV; calculating the relative blood donation load (DR) based on the donated blood volume (V) and the estimated blood volume EBV; inputting the relative blood donation load (DR) into a trained risk prediction model to obtain a risk score for the blood donor to experience severe adverse reactions to blood donation; and outputting corresponding treatment suggestions based on the preset risk level interval to which the risk score belongs. This invention can reduce risk assessment errors caused by deviations in blood volume estimation and transform risk scores into actionable on-site tiered treatment suggestions, providing decision support for risk management of adverse reactions such as vasovagal reactions related to blood donation.

Description

Technical Field

[0001] This invention relates to the field of medical information technology and intelligent risk prediction technology, and in particular to a method and system for predicting adverse blood donation reaction risks based on stratified blood volume estimation. Background Technology

[0002] Blood donation is a vital public health service ensuring the safety and supply of blood for clinical use. During the donation process, some donors may experience adverse reactions of varying degrees, with severe cases manifesting as fainting, loss of consciousness, and significant hypotension. While the incidence of such events is relatively low, they significantly impact donor health and experience, potentially requiring on-site medical intervention or transportation, thereby affecting blood donation safety management and donors' willingness to donate again. Therefore, how to achieve risk identification, early warning, and tiered management of severe adverse reactions at the blood donation site is one of the core technical requirements in this field.

[0003] In blood donation risk assessment, estimating individual blood volume is a crucial physiological basis for establishing safe blood donation volume control and assessing blood donation load. For ease of field application, existing technologies often use estimated blood volume (EBV) formulas based on gender, height, and weight. In academic research, EBV is often further utilized to construct derivative indicators such as "relative blood donation load" (i.e., the ratio of donated blood volume to blood volume) to more accurately characterize the circulatory load level imposed on an individual by blood donation behavior from a physiological perspective. However, in standardized risk assessment practices within the industry, the systematic application of such refined indicators is not yet widespread.

[0004] While some machine learning-based blood donation response prediction solutions exist, these solutions typically do not incorporate the more precise blood volume stratification calculation rules mentioned above into their feature engineering systems. The "relative blood load" indicator in their model inputs may still be calculated based on biased blood volume values. Furthermore, existing risk prediction models or real-time monitoring systems generally fail to further translate risk assessment results into actionable, tiered treatment recommendations on-site, resulting in insufficient closed-loop management capabilities for "assessment-treatment-follow-up."

[0005] In the field of adverse reaction risk management in blood donation, existing technologies are mainly developing along two technical paths: early warning systems that focus on real-time monitoring, and machine learning models that focus on risk prediction.

[0006] 1) Real-time monitoring and early warning system

[0007] The core of this type of technical solution is real-time monitoring of physiological indicators and threshold alarms during blood donation. For example, patent CN117100234A, "A Blood Donation Reaction Early Warning Monitoring System and Method," discloses a typical solution. This system collects real-time physiological indicators of blood donors and compares them with preset safety thresholds; an immediate alarm is triggered if any indicator is abnormal. It is worth noting that when calculating the safe blood collection range, the system's built-in parameters rely on a single, fixed blood volume estimation formula (such as the Nadler formula). The advantage of this type of solution is its rapid response, but it mainly focuses on in-process monitoring or alarms after an anomaly occurs, lacking the ability to predict risks in advance and implement hierarchical management.

[0008] 2) Machine learning prediction models

[0009] This type of technical solution aims to use historical data to build predictive models and achieve early risk identification. It can be further divided into two types:

[0010] General prediction framework: For example, patent CN202511597860, "A method and system for predicting the risk of adverse reactions in whole blood donation," discloses a complete machine learning modeling process. This approach predicts serious adverse reactions and adverse reaction types through steps such as data preprocessing, hierarchical modeling (distinguishing between primary and secondary outcome variables), and interpretability analysis, and covers the entire process from data collection to model self-optimization.

[0011] Feature engineering-optimized prediction scheme: Patent CN202511262235A, "Method, Device, Electronic Equipment and Storage Medium for Predicting Blood Donation Reactions," goes a step further in feature mining and model integration, and can be considered the prior art most similar to the implementation idea of ​​this invention. The core of this scheme is as follows: First, a pre-constructed feature grouping strategy is used to divide the original features into explicit feature groups (based on medical prior combinations), implicit feature groups (using decision trees to mine higher-order interaction relationships), and original feature groups; then, multiple base models are used to perform initial predictions on each feature group; finally, based on the weights of each base model, an integrated prediction is performed through a meta-model to output the probability of blood donation reactions. This scheme aims to deeply mine data value to improve prediction accuracy through feature grouping and model integration, and mentions dynamic updates of base model weights and time-sensitivity verification.

[0012] Although the aforementioned existing technologies (especially CN202511262235A) have been optimized in terms of model complexity and feature engineering, the following key shortcomings still exist in achieving the goal of more accurate, physiologically sound, and easily applicable individualized risk assessment for on-site management:

[0013] 1) Inherent biases exist in the quantification of core physiological risk indicators: Existing solutions fail to fundamentally address the accuracy issue in calculating the core input feature—"relative blood donation load" (i.e., the proportion of donated blood to an individual's total blood volume)—when constructing risk prediction models. Whether it's the fixed formula built into real-time monitoring systems (such as CN117100234A) or the blood volume calculation method not explicitly specified in the prediction model, a single blood volume estimation formula (such as the Nadler formula) is generally used or assumed to be employed. However, research and industry monitoring data suggest that a single blood volume estimation formula may have systematic biases in high BMI populations (e.g., overestimation of blood volume), leading to an underestimation of the calculated relative blood donation load. This means that even if the model subsequently adopts a complex architecture, the foundational data upon which its decisions are based is already biased, particularly affecting the accuracy of risk identification for high-risk groups with high BMI.

[0014] 2) Feature engineering is not compatible with physiological principles and industry practices: Although existing machine learning solutions (such as CN202511262235A) incorporate medical priors in feature construction, they fail to incorporate key features such as using different blood volume estimation methods for different body size ranges to optimize relative blood donation load, resulting in an insufficient characterization of core physiological mechanisms.

[0015] 3) Disconnect between risk assessment results and on-site response: Existing technical solutions have limitations in their functional endpoints. Real-time monitoring systems (such as CN117100234A) stop at immediate alarms and lack the ability to generate tiered response suggestions based on individualized risk predictions. Predictive model solutions (such as CN202511597860 and CN202511262235A) focus primarily on outputting risk probabilities without further elaborating on how to translate these probabilities into specific, actionable on-site response instructions for different risk levels (such as differentiated observation duration, fluid resuscitation plans, educational priorities, and follow-up priorities). This makes it difficult for risk assessment results to directly guide frontline staff in taking precise preventative measures, and prevents the formation of an intelligent management closed loop of "assessment-early warning-response-feedback."

[0016] In summary, existing technologies have not yet provided a solution that can deeply embed a more accurate blood volume calculation method based on body size stratification rules into the feature engineering of machine learning prediction models, and on this basis, build an intelligent system that can output tiered treatment recommendations and achieve closed-loop management on-site. Summary of the Invention

[0017] To address the shortcomings of existing blood donation risk assessment technologies, such as imprecise quantification of core physiological indicators, weak integration with physiological principles and industry practices, and disconnect from closed-loop on-site management, this invention aims to provide a method and system for predicting adverse blood donation reactions based on stratified blood volume estimation. This reduces errors in core risk indicators caused by deviations in blood volume estimation, achieves closed-loop decision output from risk prediction to tiered on-site management, and provides a basis for on-site decision support in rare adverse blood donation reaction scenarios. These adverse blood donation reactions include, but are not limited to, blood donation-related vasovagal reactions.

[0018] The technical solution adopted by this invention to achieve its objective is a method for predicting the risk of adverse reactions to blood donation based on stratified blood volume estimation, comprising the following steps:

[0019] S1. Collect basic information about blood donors, including gender, height, weight, and blood donation volume V;

[0020] S2. Calculate the Body Mass Index (BMI) based on the blood donor's height and weight;

[0021] S3. Select the blood volume estimation formula based on the body mass index (BMI) and calculate the estimated blood volume EBV. When the body mass index (BMI) is less than the preset threshold, the first blood volume estimation formula is used. When the body mass index (BMI) is greater than or equal to the preset threshold, the second blood volume estimation formula is used.

[0022] S4. Calculate the relative blood donation load (DR) based on the blood donation volume (V) and the estimated blood volume (EBV). ;

[0023] S5. Input the relative blood donation load (DR) into the trained risk prediction model to obtain the risk score of blood donors experiencing severe adverse reactions to blood donation;

[0024] S6. Output corresponding treatment suggestions based on the preset risk level range to which the risk score belongs. The treatment suggestions include at least one of the following: observation duration, fluid resuscitation suggestions, health education tips, and follow-up priority. The treatment suggestions can be displayed or pushed to the staff terminals of the blood donation vehicle information system, blood donation center information system, or blood bank business system.

[0025] Furthermore, in step S1, covariate information of blood donors is also collected, including age, whether it is the first time donating blood, and the type of blood collection location; in step S5, the covariate information is also input into the trained risk prediction model.

[0026] Furthermore, the adverse reactions to blood donation include blood donation-related vasovagal reactions.

[0027] Furthermore, the first blood volume estimation formula is the Nadler formula, and the second blood volume estimation formula is the Holme formula.

[0028] Furthermore, the preset threshold is determined in the following way: a candidate threshold set is set, a hierarchical blood volume estimation rule is constructed for each candidate threshold and a relative blood donation load (DR) feature is generated, a training set / validation set is used for evaluation by cross-validation or repeated segmentation and extrapolation, and the preset threshold is obtained by screening with the discrimination index AUC and / or the prediction error index Brier score as the optimization target.

[0029] Furthermore, the preset threshold is located in the range of 25 to 28.

[0030] Furthermore, the preset threshold is located in the range of 26 to 27.5.

[0031] Furthermore, the training method for the risk prediction model is as follows:

[0032] S501. A training sample set is constructed using historical blood donation records. Each training sample corresponds to one blood donation event and includes at least: gender, height, weight, and blood donation volume V. The label is "Severe" for severe adverse blood donation reaction, where Severe = 1 indicates that a severe adverse blood donation reaction has occurred, and Severe = 0 indicates that no severe adverse blood donation reaction has occurred. The Severe can correspond to one or a combination of severe adverse blood donation reaction events such as syncope, loss of consciousness, need for medical treatment or transportation.

[0033] S502. Construct features for the training samples, specifically including: firstly, calculating the body mass index (BMI), and then selecting either the first blood volume estimation formula or the second blood volume estimation formula based on the relationship between BMI and a preset threshold to calculate the estimated blood volume (EBV), and finally calculating the relative blood donation load (DR) as the core input feature. ;

[0034] S503. Data Preprocessing and Sample Screening: Perform unit conversion and format standardization on training samples; remove, correct, or impute missing or outlier values.

[0035] S504. Data partitioning and stratified sampling: The training sample set is divided into a training set and a validation set according to a preset ratio, and stratified sampling is used to maintain the ratio of positive to negative samples.

[0036] S505. Train the risk prediction model on the training set to minimize the loss function between the severe adverse blood donation risk score output by the risk prediction model and the label. The loss function is logarithmic loss or cross-entropy loss.

[0037] S506. To reduce the randomness caused by a single partition, K-fold cross-validation or repeated partitioning extrapolation evaluation is used to validate the risk prediction model and obtain the evaluation results. In one embodiment, repeated partitioning evaluation can be performed multiple times (e.g., 200 times). Each time, the model is fitted on the training set and a risk score is output on the corresponding validation / test set. This score is used to statistically analyze the mean, standard deviation, and confidence interval of model performance indicators (AUC, area under the precision-recall curve, calibration error, net benefit of decision curve, etc.).

[0038] S507. Model Selection and Determination of Risk Level Ranges or Thresholds: Based on the evaluation results of the validation set and / or the evaluation results obtained by the K-fold cross-validation and repeated segmentation extrapolation evaluation, select the parameters or model structure of the risk prediction model and determine the risk level ranges or threshold sets for use in generating disposal recommendations and on-site hierarchical management.

[0039] S508. Model Stabilization: Save the trained risk prediction model parameters as a callable model file or parameter table.

[0040] Furthermore, the risk prediction model is one of the following: logistic regression model, generalized linear model, support vector machine, random forest, gradient boosting tree, and ensemble learning model.

[0041] Furthermore, the trained risk prediction model is implemented using an ensemble prediction structure, which may include three stages: feature grouping, base model training and initial prediction, and fusion output.

[0042] The feature grouping specifically involves dividing the input features into multiple feature groups, which include explicit feature groups, implicit feature groups, original feature groups, and basic transformation feature groups, wherein:

[0043] Explicit feature sets, constructed based on medical or business priors, include at least the relative blood donation burden index DR and at least one donor covariate.

[0044] Implicit feature groups are obtained by combining the original features through decision tree models or gradient boosting tree models, and are used to represent the high-order interaction relationships between the original features.

[0045] The original feature set consists of the original input variables without any transformation, including gender, height, weight, and blood donation volume.

[0046] The basic transformation feature set is a feature set obtained by performing at least one mathematical transformation on the original feature set. The mathematical transformation includes at least one of logarithmic transformation, square transformation, and binning discretization.

[0047] The base model training and initial prediction specifically involve: training at least one base model for different feature groups and outputting the initial risk score or probability prediction result corresponding to the base model. The base model can be any trainable classification or regression model, and cross-validation or repeated split extrapolation is used to estimate the generalization performance.

[0048] The fusion output specifically refers to: taking the initial risk score or probability prediction result of the base model as input, training the fusion model to output the final risk score; the fusion model is implemented by weighted fusion or stacked fusion, wherein the weights or fusion parameters are determined based on the cross-validation results or repeated segmentation extrapolation evaluation results.

[0049] Furthermore, after step S6, real-time monitoring and alarm linkage are also carried out. The specific method is to collect real-time physiological indicators, output alarm information when the real-time physiological indicators meet the preset abnormal conditions, and provide enhanced prompts or upgrade outputs for handling suggestions.

[0050] This invention also provides a blood donation adverse reaction risk prediction system based on stratified blood volume estimation. This system can be deployed in blood donation vehicle information systems, blood donation center information systems, or blood bank business systems, and can interact with on-site blood collection equipment and monitoring equipment. Its specific structure includes:

[0051] The data acquisition module is used to collect basic information of blood donors, including gender, height, weight, and blood donation volume V.

[0052] The body mass index calculation module is used to calculate the body mass index (BMI) based on the blood donor's height and weight.

[0053] The blood volume stratification calculation module is used to select a blood volume estimation formula based on body mass index (BMI) and calculate the estimated blood volume EBV. When the body mass index (BMI) is less than a preset threshold, the first blood volume estimation formula is used, and when the body mass index (BMI) is greater than or equal to the preset threshold, the second blood volume estimation formula is used.

[0054] The relative blood donation load calculation module is used to calculate the relative blood donation load DR based on the blood donation volume V and the estimated blood volume EBV. ;

[0055] The risk prediction module is used to input the relative blood donation load (DR) into the trained risk prediction model and output a risk score for blood donors to experience severe adverse reactions to blood donation.

[0056] The decision output module is used to output corresponding treatment suggestions based on the preset risk level range to which the risk score belongs. The treatment suggestions include at least one of the following: observation duration, fluid resuscitation suggestions, health education tips, and follow-up priority.

[0057] Furthermore, the risk prediction system also includes an alarm module, which outputs alarm information when real-time physiological indicators meet preset abnormal conditions, and provides enhanced prompts or upgraded outputs for handling suggestions.

[0058] Furthermore, the data acquisition module is also used to collect blood donor covariate information, which includes age, whether it is the first time donating blood, and the type of blood collection location; the risk prediction module is also used to input the blood donor covariate information into the trained risk prediction model.

[0059] The beneficial effects of this invention are as follows:

[0060] 1) Reduce the systematic underestimation of core workload indicators in high BMI populations and improve the inclusion capacity of high-load individuals.

[0061] This invention employs a BMI-stratified blood volume estimation method to construct the relative blood donation load (DR). In the BMI < 26 population, the DR difference between the stratification rule and the traditional formula is 0; in the BMI ≥ 26 population, the DR difference is generally positive, with an average upward adjustment of approximately 1.25 percentage points (median approximately 1.21, P90 approximately 2.07, P95 approximately 2.15), indicating that the stratification rule has a systematic corrective effect on the high BMI population. It should be noted that the "traditional model" here is defined as a model that uniformly uses the Nadler formula to calculate EBV and DR for the entire population. The stratification rule model of this invention is equivalent to the "traditional model" in the BMI < 26 population, hence the DR difference is 0; while in the BMI ≥ 26 population, by switching to the Holme formula, the aforementioned positive correction is generated. Furthermore, when using the 80th percentile of traditional DR in the BMI≥26 population as the high load threshold, this invention reclassified 33 / 144 (22.9%) blood donors from "non-high load" to "high load". The incidence of severe adverse blood donation reactions in this group was approximately 15.2% (5 / 33), which was higher than the approximately 8.54% (7 / 82) in the consistently low load group. This provides a basis for the hierarchical allocation of resources such as on-site observation, fluid replacement and follow-up.

[0062] 2) Maintain a certain comprehensive predictive advantage in low-occurrence event scenarios.

[0063] For rare events like severe adverse reactions to blood donation, which have an extremely low incidence (approximately 11% in the examples), the AUC metric is not sensitive to improvements in model performance on the minority class. However, the area under the precision-recall curve more accurately reflects the model's ability to identify positive samples (those experiencing severe reactions). As shown in Table 5, the proposed solution achieves superior performance on the key subgroup (BMI ≥ 26). To verify the advantage of stratification rules in identifying positive individuals in low-incidence event scenarios, 200 repeated paired segmentation extrapolations were performed on the BMI ≥ 26 subgroup. Paired t-test results showed that the stratification rule model... The difference was significantly higher than the traditional model (mean difference +0.014, t=4.264, p<0.001), with a Cohen's d effect size of 0.302, suggesting a small to moderate improvement in rare event scenarios. The Brier scores of the two models were similar (0.116 vs 0.115), indicating that while the stratification rule significantly improved the ability to identify positive samples, the overall calibration performance of probability prediction was comparable to that of the traditional model.

[0064] 3) Support closed-loop disposal and optimized resource allocation

[0065] In individuals with a BMI ≥ 26, within a threshold probability range of approximately 0.05–0.30, the overall net benefit of the proposed solution is no less than that of the traditional model and is superior to the negative net benefit scenario that occurs under a higher threshold with the "full intervention strategy," thereby supporting the hierarchical allocation and closed-loop management of resources such as observation, fluid replacement, and follow-up.

[0066] 4) It forms a feasible closed-loop disposal mechanism, possessing practical deployment and transformation value.

[0067] This invention can correlate risk scores with risk level ranges, outputting actionable treatment recommendations (including observation duration, oral rehydration / fluid replacement, enhanced education and risk warnings, follow-up priority, etc.), and can further link with on-site real-time physiological indicator monitoring and alarm modules to achieve automatic calculation, automatic prompts, and standardized output. Therefore, without significantly increasing hardware investment, this invention helps improve donor safety and experience, reduces the risks of medical treatment and transportation due to severe adverse reactions to blood donation, and reduces the risk of complaints and disputes arising from such reactions, demonstrating clear value for on-site application and commercial deployment. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the risk prediction method in Embodiment 1 of the present invention;

[0069] Figure 2 This is a scatter plot showing the consistency between the Holme formula and the Nadler formula in estimating blood volume EBV in Embodiment 1 of the present invention.

[0070] Figure 3 This is a scatter plot of DR consistency in Embodiment 1 of the present invention;

[0071] Figure 4 This is a distribution map of the corrected DR magnitude for the population with a BMI ≥ 26 in Embodiment 1 of the present invention;

[0072] Figure 5 This is a box plot showing the relationship between the relative blood donation load difference and severe adverse blood donation outcomes in a population with a BMI ≥ 26 in Embodiment 1 of the present invention.

[0073] Figure 6 This is a stability evaluation chart of the BMI stratification threshold in Embodiment 1 of the present invention;

[0074] Figure 7 This is a calibration curve diagram of Embodiment 1 of the present invention;

[0075] Figure 8 This is a decision curve analysis diagram of Embodiment 1 of the present invention in a population with a BMI ≥ 26.

[0076] Figure 9 In Embodiment 1 of the present invention Histogram of difference distribution;

[0077] Figure 10 In Embodiment 1 of the present invention Performance distribution box plot;

[0078] Figure 11 This is a box plot of the Brier Score performance distribution in Embodiment 1 of the present invention;

[0079] Figure 12 This is a structural diagram of the risk prediction system in Embodiment 2 of the present invention. Detailed Implementation

[0080] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0081] Example 1

[0082] Figure 1 This invention illustrates a specific implementation of the method for predicting the risk of adverse reactions to blood donation based on stratified blood volume estimation. This embodiment takes blood donation-related vasovagal nerve reactions as an example and includes the following steps:

[0083] S1. Collect basic information and covariate information of blood donors. The basic information includes gender, height, weight, and blood donation volume V. The covariate information includes age, whether it is the first time donating blood, and the type of blood collection location.

[0084] S2. Calculate the Body Mass Index (BMI) based on the blood donor's height and weight; the calculation formula is as follows: Where W is weight (kg) and H is height (m);

[0085] S3. Select the blood volume estimation formula based on body mass index (BMI) and calculate the estimated blood volume EBV. The preset threshold for BMI is 26, and the unit of EBV is ml.

[0086] When BMI < 26, the first blood volume estimation formula, namely Nadler's formula, is used, specifically:

[0087] male:

[0088] female:

[0089] When BMI ≥ 26, the second blood volume estimation formula, namely the Holme formula, is used, specifically:

[0090] male:

[0091] female:

[0092] Where height h = H × 100, that is, the unit of h is cm;

[0093] S4. Calculate the relative blood donation load (DR) based on the blood donation volume (V) and the estimated blood volume (EBV). The blood donation volume (V) is measured in ml.

[0094] S5. Input the relative blood donation load (DR) and covariate information into the trained risk prediction model to obtain the risk score of blood donors experiencing severe adverse reactions to blood donation;

[0095] S6. Output corresponding treatment suggestions based on the preset risk level range to which the risk score belongs. The treatment suggestions include at least one of the following: observation duration, fluid resuscitation suggestions, health education tips, and follow-up priority.

[0096] S7. Real-time monitoring and alarm linkage: This involves collecting real-time physiological indicators, which can be acquired through a blood pressure and heart rate monitoring module integrated with the blood donation chair or through wearable devices, and transmitted to this system via wired or wireless means. When the real-time physiological indicators meet preset abnormal conditions, an alarm message is output, and enhanced or upgraded handling suggestions are provided.

[0097] Special Note: In another optional implementation, the preset threshold is a configurable threshold, which can be determined based on the characteristics of the target population and historical blood donation records. It can be set in the range of 25 to 28 and adjusted according to the data. Specifically, a candidate threshold set can be set, and for each candidate threshold, a stratified blood volume estimation rule can be constructed and a relative blood donation load (DR) feature can be generated. The training set / validation set is divided and K-fold cross-validation or repeated segmentation extrapolation is used for evaluation to calculate the discrimination index AUC and / or the prediction error index Brier score. The model performance corresponding to different candidate thresholds is compared using the training set / validation set division combined with K-fold cross-validation or repeated segmentation extrapolation, and the optimal threshold is determined accordingly.

[0098] In this embodiment, through repeated sampling evaluation, it was found that when the threshold is in the range of 26 to 27.5, the mean of the extrapolated evaluation AUC is relatively high and stable. Finally, the preset threshold in this embodiment is set to 26.

[0099] In this embodiment, the training method for the risk prediction model is as follows:

[0100] S501. A training sample set is constructed using historical blood donation records. Each training sample corresponds to one blood donation event and includes basic information and covariate information. The basic information includes gender, height, weight, and blood donation volume V. The covariate information includes age, whether it is the first time donating blood, and the type of blood collection location. The label is "Severe" for severe adverse blood donation outcome, where Severe=1 indicates that a severe adverse blood donation outcome occurred, and Severe=0 indicates that no severe adverse blood donation outcome occurred.

[0101] The severe adverse reaction outcome (Severe) of blood donation is a binary categorical variable, and its determination criteria are as follows:

[0102] A severe adverse reaction to blood donation is defined as Severe = 1 if a blood donor experiences any of the following conditions during or after the donation process; otherwise, it is defined as Severe = 0:

[0103] (1) Loss of consciousness / syncope occurs;

[0104] (2) Emergency medical intervention or hospitalization is required;

[0105] (3) External injury occurs (e.g., physical injury caused by falling due to fainting).

[0106] It should be noted that the above criteria can be adjusted or refined according to industry standards or monitoring standards; the method of this invention is also applicable without changing the modeling premise of "severe adverse blood donation outcome as a binary label".

[0107] S502. Construct features for the training samples, specifically including: firstly, calculating the body mass index (BMI), and then selecting either the first blood volume estimation formula or the second blood volume estimation formula based on the relationship between BMI and a preset threshold to calculate the estimated blood volume (EBV), and finally calculating the relative blood donation load (DR) as the core input feature. ;

[0108] S503. Data Preprocessing and Sample Screening: Perform unit conversion and format standardization on training samples; remove, correct, or impute missing or outlier values.

[0109] S504. Data Partitioning and Stratified Sampling: The training sample set is divided into a training set and a validation set according to a preset ratio, and stratified sampling is used to maintain the positive and negative sample ratio. In this embodiment, the training sample set is constructed using historical whole blood donation records collected by blood banks. After quality control and sample screening of the original records (including valid outcome labels, non-missing key variables, reasonable range of height / weight / blood donation volume values, and removal of records with a blood donation volume of 0), a total of 702 samples were finally included in the analysis. Among them, 78 cases of severe adverse blood donation reactions (Severe=1) and 624 cases of no severe adverse blood donation reactions (Severe=0) occurred, with a positive and negative sample ratio of approximately 1:8.0 (positive samples accounted for approximately 11.1%). To maintain the consistency of this ratio during the training and evaluation process, the training sample set is divided into a training set and a validation set at a ratio of 7:3, and stratified sampling is used to ensure that the positive and negative sample ratios in the training set and validation set are consistent with the original sample set.

[0110] S505, Model Fitting and Objective Function: Train the risk prediction model on the training set to minimize the loss function between the risk score of severe adverse blood donation reactions output by the risk prediction model and the label. The loss function is logarithmic loss or cross-entropy loss. The risk prediction model is one of the following: logistic regression model, generalized linear model, support vector machine, random forest, gradient boosting tree, and ensemble learning model.

[0111] S506. To reduce the randomness caused by a single partition, K-fold cross-validation or repeated partitioning extrapolation evaluation is used to validate the risk prediction model. Multiple repeated partitioning evaluations (200 times) can be performed. Each time, the model is fitted on the training set and a risk score is output on the corresponding validation set. This score is used to calculate the mean, standard deviation, and confidence interval of the model performance indicators (AUC, area under the precision and recall curves, calibration error, net benefit of the decision curve, etc.).

[0112] S507. Model selection and determination of risk level ranges or thresholds: Based on the evaluation results of the validation set, select the parameters or structure of the risk prediction model and determine the risk level ranges or threshold sets for subsequent disposal recommendations and on-site hierarchical management.

[0113] S508. Model Stabilization: Save the trained risk prediction model parameters as a callable model file or parameter table.

[0114] Model Update Mechanism: Considering that the characteristics of blood donors and blood donation scenarios may change over time (e.g., changes in age structure, seasonal fluctuations, etc.), the method of this invention supports dynamic updates to the risk prediction model. New blood donation data can be included according to a preset time period or triggering conditions, and the above training and validation process can be re-executed to maintain the model's timeliness and predictive accuracy. For example, the time period can be every quarter or every six months; or, for example, the triggering conditions can be that the cumulative number of new samples reaches a preset threshold, or the model's performance indicators on recent data are lower than a preset threshold. When updating the model, it can be released and replaced only after successful validation, to achieve iterative model maintenance and traceable management.

[0115] The trained risk prediction model described in this embodiment is implemented using an ensemble prediction structure. This ensemble prediction structure may include three stages: feature grouping, base model training and initial prediction, and fusion output.

[0116] The feature grouping specifically involves dividing the input features into multiple feature groups, which include explicit feature groups, implicit feature groups, original feature groups, and basic transformation feature groups, wherein:

[0117] Explicit feature sets, constructed based on medical or business priors, include at least the relative blood donation burden index DR and at least one donor covariate.

[0118] Implicit feature groups are obtained by combining the original features through decision tree models or gradient boosting tree models, and are used to represent the high-order interaction relationships between the original features.

[0119] The original feature set consists of the original input variables without any transformation, including gender, height, weight, and blood donation volume.

[0120] The basic transformation feature set is a feature set obtained by performing at least one mathematical transformation on the original feature set. The mathematical transformation includes at least one of logarithmic transformation, square transformation, and binning discretization. The base model training and initial prediction specifically involve: training at least one base model for different feature groups and outputting the initial risk score or probability prediction result corresponding to the base model. The base model can be any trainable classification or regression model, and cross-validation or repeated split extrapolation is used to estimate the generalization performance.

[0121] The fusion output specifically refers to: taking the initial risk score or probability prediction result of the base model as input, training the fusion model to output the final risk score; the fusion model is implemented by weighted fusion or stacked fusion, wherein the weights or fusion parameters are determined based on the cross-validation results or repeated segmentation extrapolation evaluation results.

[0122] During system operation, the base model and / or fusion model can be retrained or updated based on newly added blood donation data.

[0123] The quantitative impact analysis of the stratification rules used in this invention on core indicators:

[0124] It should be noted that the following analysis for this embodiment is used to illustrate the influence of different blood volume estimation methods on the construction of core features, and does not take any one estimation method as the gold standard for true blood volume.

[0125] To intuitively verify the necessity of selecting the blood volume estimation formula based on BMI stratification from a data perspective, and to demonstrate how this rule affects core risk prediction indicators, the following comparative analysis was conducted on the overall sample:

[0126] (a) EBV Consistency and Systematic Deviation

[0127] Figure 2 This is a scatter plot showing the consistency between the Holme formula and the Nadler formula in estimating blood volume EBV in Embodiment 1 of the present invention, used to illustrate the estimation relationship between the two blood volume estimation formulas under different body mass index strata. Figure 2 As shown, the EBV estimated by the two formulas are generally positively correlated, and most of the scatter points are below the diagonal, indicating that the EBV estimated by the Holme formula is generally lower than that estimated by the Nadler formula. The deviation from the diagonal is more obvious in the BMI ≥ 26 population, suggesting that the use of a single blood volume estimation formula in the high BMI population may lead to systematic bias in EBV estimation, thereby affecting the calculation of core derivative indicators such as relative blood load (DR).

[0128] (II) Distribution of DR correction magnitude in the BMI≥26 population

[0129] Figure 4 This is a distribution map of the DR correction magnitude for the population with BMI ≥ 26 in Embodiment 1 of the present invention, used to illustrate the DR difference. The distribution characteristics are shown, where the dashed line represents the median of the DR difference; further, represents the magnitude of the correction by the stratification rule in the high BMI population, such as Figure 4 As shown, in the BMI≥26 population, the DR difference The overall positive values ​​and the somewhat discrete distribution indicate that the stratification rule generally improved the estimate of the relative blood donation load in this population.

[0130] (III) Quantitative impact of stratification rules on core risk indicators (DR)

[0131] The core risk indicator, "relative blood donation load (DR)," is calculated from the donated blood volume and estimated blood volume (DR = V / EBV × 100). In this embodiment, the stratification rule uses the first blood volume estimation formula to calculate EBV for the BMI < 26 population and the second blood volume estimation formula to calculate EBV for the BMI ≥ 26 population. Therefore, the difference between the DR calculated by the stratification rule and the traditional formula is 0 in the BMI < 26 population, while a systematic positive correction occurs in the BMI ≥ 26 population. Table 1 quantifies this impact:

[0132] Table 1. Statistics of DR differences between stratification rules and traditional formulas under different BMI stratifications.

[0133]

[0134] Table 2. Differences in EBV and DR estimated by the Holme formula and the Nadler formula under different BMI strata.

[0135]

[0136] It should be noted that Table 2 is used to show the estimation differences between the Holme formula and the Nadler formula under different BMI strata. This difference reflects the systematic differences between the two blood volume estimation formulas themselves and is not equivalent to the difference between the stratification rule and the traditional formula. Specifically, the stratification rule of this invention does not switch to the Holme formula in the BMI < 26 population. Therefore, the difference in DR between the stratification rule and the traditional formula in the BMI < 26 population is 0 (see Table 1), while Table 2 shows the difference in DR between the stratification rule and the traditional formula in the BMI < 26 population. The DR difference is only used to illustrate the difference between the two formulas. Data shows (see Table 1) that in the BMI≥26 subgroup, the stratification rule increased the relative blood donation load index by an average of approximately 1.25 percentage points compared to the traditional formula. This change is consistent with the calculation relationship DR=V / EBV×100 (a decrease in the denominator EBV will lead to an increase in DR), suggesting that using a more body-type-adaptive EBV estimation method in high BMI populations will have a systematic impact on core characteristics.

[0137] (iv) Association between DR difference and severe adverse reaction outcomes of blood donation

[0138] Most importantly, this change in DR value caused by the formula switch is associated with the predicted target, namely severe adverse reactions to blood donation. Figure 5 This is a box plot illustrating the relationship between the relative blood donation load difference and severe adverse blood donation outcomes in a population with a BMI ≥ 26 according to Embodiment 1 of the present invention. The relative blood donation load difference is... It should be noted that the stratification rule for the BMI ≥ 26 population uses the Holme formula to estimate EBV; therefore, this difference is related to... The DR difference is equivalent. For example... Figure 5As shown, box plots comparing donors who experienced severe adverse blood donation reactions (DRs) revealed differences in the distribution of relative blood donation burden (DRs). In the group experiencing severe DRs, the DR distribution was more centrally located and exhibited a more pronounced right skew. These results suggest that using the Holme formula to correct DR values ​​for high BMI individuals can help include individuals who may be underestimated by traditional methods.

[0139] Comparative Examples: Using fixed formulas or alternative features, the covariates and model structure remain unchanged in the comparative examples; only the construction method of the core indicators is replaced. Specifically, the following three models are used:

[0140] Traditional model: Estimate EBV using Nadler's formula for all blood donors and calculate DR;

[0141] Holme model was used for all blood donors: EBV was estimated using the Holme formula, and DR was calculated.

[0142] Body Mass Index (BMI) Only Model: Outputs a risk score using BMI (and the same covariates).

[0143] Severe adverse reactions to blood donation were assessed using the Severe criterion.

[0144] It should be noted that in the BMI≥26 subgroup, the stratification rule uses the Holme formula to estimate EBV according to the definition. Therefore, the DR composition of the stratification rule model in this subgroup is consistent with the DR composition of "all adopt the Holme model" in this subgroup.

[0145] To verify the contribution of the key technical feature of “body mass index stratification for EBV formula and construction of relative blood donation load DR index” to risk identification performance, repeated segmentation extrapolation was used for evaluation (e.g., 200 stratifications, test set ratio, e.g., 0.3), and the discrimination was measured by the area under the receiver operating characteristic curve (AUC).

[0146] (a) Overall population

[0147] The overall evaluation results show that, as shown in Table 3, the mean AUC of the stratified rule model is approximately 0.682, and the mean AUC of the traditional model is approximately 0.681, with both being close overall; and both outperform the body mass index-only model (mean AUC of approximately 0.637).

[0148] Table 3 Comparison of AUC performance of each model for the overall population

[0149]

[0150] Note: B_valid represents the number of valid repetitions; auc_mean represents the AUC mean; auc_sd represents the AUC standard deviation; auc_median represents the AUC median; auc_lo represents the lower bound of the AUC 95% confidence interval; auc_hi represents the upper bound of the AUC 95% confidence interval.

[0151] (ii) Subgroup with BMI ≥ 26

[0152] In the BMI≥26 subgroup (see Table 4), the mean AUC of the stratified rule model and the traditional model were generally similar, suggesting that under the current sample size and assessment criteria, the difference in discrimination between the two is not significant; however, considering the aforementioned DR difference distribution (Table 1, ... Figure 3 and Figure 4 Based on the results of subsequent reclassification analysis, the stratification rules of this invention can systematically correct the core load indicators of high BMI populations and provide input features that are more in line with physiological mechanisms for risk stratification and treatment recommendations. Figure 3 This is a scatter plot of DR consistency in Embodiment 1 of the present invention, used to illustrate the relationship between the relative blood donation burden DR (CSBT) calculated by the stratified rule and the DR (Nadler) calculated by the traditional formula. The results show that the two are generally positively correlated; the scatter plot for the BMI < 26 group is basically close to the diagonal, while the scatter plot for the BMI ≥ 26 group is generally above the diagonal and deviates more significantly, suggesting that the DR calculated by the stratified rule is generally higher than the DR calculated by the traditional formula in the high BMI group.

[0153] Table 4 Comparison of AUC performance of various models for the BMI≥26 population

[0154]

[0155] Note: B_valid represents the number of valid repetitions; auc_mean represents the AUC mean; auc_sd represents the AUC standard deviation; auc_median represents the AUC median; auc_lo represents the lower bound of the AUC 95% confidence interval; auc_hi represents the upper bound of the AUC 95% confidence interval.

[0156] Real-world reliability verification (BMI≥26): Threshold sensitivity, precision-recall curves and probability error

[0157] To verify the robustness of the proposed solution in the target population (BMI≥26), its ability to identify rare events, and the accuracy of probability prediction, the following three evaluations were conducted:

[0158] (a) Threshold sensitivity (robustness)

[0159] To verify the robustness and configurability of the stratification threshold, a candidate BMI threshold set (22–32, step size 0.5) was set, and the model's discriminative power under different thresholds was evaluated using training / test splits and repeated sampling (100 times). The results show that the mean AUC on the test set is high and relatively stable within the range of approximately 26.5–27.5 (mean approximately 0.648, empirical range approximately 0.55–0.75), indicating that the effectiveness of this invention remains stable when the threshold varies within this range. Figure 6 This is a BMI stratification threshold stability evaluation chart from Embodiment 1 of the present invention, used to display the repeated sampling mean and fluctuation range of the model discrimination index (AUC) under different BMI thresholds. Figure 6 As shown, the mean AUC values ​​differ little under different thresholds, indicating that the present invention is robust near the threshold. Furthermore, it performs best in the 26.5–27.5 range, which can be considered a preferred threshold range. Therefore, the preset threshold can be set to 26 as a preferred implementation, or it can be determined data-driven within the range of 25–28 as a configurable threshold.

[0160] (ii) Area under the precision-recall curve (imbalanced event availability)

[0161] Severe adverse reactions to blood donation are low-incidence events. In individuals with a BMI ≥ 26, repeated segmentation assessments showed that the stratified rule model was superior in identifying positive individuals (i.e., those experiencing severe adverse reactions to blood donation). See Table 5 for a detailed performance comparison. It should be noted that for low-incidence events like severe adverse reactions to blood donation (with a positive rate of only about 12.5% ​​in this example), the area under the receiver operating characteristic (AUC) curve, because it focuses on overall ranking ability, is less sensitive to performance improvements in the minority class. In contrast, the area under the precision-recall curve... Focusing more on the model's ability to identify positive samples is a more suitable indicator for evaluating the performance of rare event prediction models. In the BMI≥26 subgroup, using 200 repeated segmentation extrapolation and paired t-tests, the results showed:

[0162] Hierarchical rule model The results were significantly higher than those of the traditional model (mean 0.296 vs 0.282, mean difference = 0.014, t = 4.264, df = 199, p < 0.001), and contained no zero values, indicating that the method of this invention has a statistically significant improvement in the ability to identify positive samples from high BMI populations (see Table 6 and ...). Figure 10 ).

[0163] Table 5 Comparison of repeated segmentation PR-AUC and Brill score in the BMI≥26 population

[0164]

[0165] Figure 9 In Embodiment 1 of the present invention The difference distribution histogram is used to illustrate the differences between the hierarchical rule model and the traditional model in the extrapolation evaluation of 200 repeated paired segmentation. Distribution characteristics of the differences. As shown in the figure, the overall distribution of the differences lies to the right of zero, indicating that the hierarchical rule model outperforms the traditional model in most segmentations, demonstrating its robustness. The dashed line in the figure represents the zero reference line, and the solid line represents the mean difference.

[0166] The Brier scores of the two models are close (0.116 vs 0.115), indicating that while the stratification rule improves the ability to identify positive samples, its calibration performance in probability prediction is comparable to that of the traditional model. Furthermore, this invention provides a reclassification analysis: in the BMI ≥ 26 population, the stratification rule reclassified 33 / 144 (22.9%) blood donors from 'non-high load' to 'high load'. The incidence of severe adverse blood donation reactions in this group was as high as 15.2%, far exceeding the 8.54% in the consistently low load group. This result further validates the practical value of the stratification rule in identifying high-risk individuals and provides an operational basis for on-site triage and treatment. Figure 11 ).

[0167] Table 6 Statistical Test Results

[0168] PR-AUC mean difference 0.014 PR-AUC t-statistic 4.264 PR-AUC p-value <0.001 PR-AUC Cohen's d 0.302 Brier mean difference 0.00023 Brier t statistic 7.011 Brier p-value <0.001 Brier Cohen's d 0.496

[0169] (iii) Brill Score (Probability Prediction Error)

[0170] The risk score output by the model, as a probability, is crucial for accurate on-site triage and treatment. In individuals with a BMI ≥ 26, repeated segmentation and extrapolation assessment results show that the stratified rule model... The mean score is approximately 0.296, higher than the traditional model's approximately 0.282. The Brier scores of the two models are close (0.116 vs 0.115), suggesting that while the hierarchical rule improves recognition ability, its probability prediction calibration performance is comparable to the traditional model. The consistency between probability and actual occurrence rate, such as... Figure 7 As shown, the calibration curves indicate that the predicted probabilities of the hierarchical rule model are generally consistent with the actual occurrence rates, suggesting that its probability output has good calibration performance.

[0171] Decision Value and Potential Business Value: Risk Stratification—Resolution Recommendations (BMI≥26)

[0172] The core value of this invention lies in transforming risk prediction into actionable on-site management decisions. Figure 8This is a decision curve analysis graph of Embodiment 1 of the present invention in a population with BMI≥26, used to show the changes in net benefit of various prediction models and control strategies (full intervention strategy and no intervention strategy) under different threshold probabilities, including stratified rule model, traditional model, body mass index only model and Holme model for all.

[0173] In people with a BMI ≥ 26, such as Figure 8 The decision curve analysis shown indicates that, within the threshold probability range of approximately 0.05 to 0.30, the overall net benefit of the stratified rule model is no less than that of the traditional model and is superior to the negative net benefit of the "full intervention strategy" at higher thresholds, thus providing a basis for decision-making regarding the tiered allocation of limited resources such as observation, fluid replacement, and follow-up.

[0174] In the BMI ≥ 26 group, using the 80th percentile of the DR calculated by the traditional formula as the "high load threshold" for comparison, the stratification rule of this invention reclassified 33 / 144 (22.9%) blood donors from "non-high load" to "high load". Among the blood donors included in the up-classification, the incidence of severe adverse blood donation reactions was approximately 15.2% (5 / 33), which was higher than approximately 8.54% (7 / 82) in the consistently low load group. This result suggests that this invention can include high-risk individuals who may be underestimated by traditional methods within the scope of attention, providing a basis for subsequent management strategies such as observation, fluid replacement, education, and follow-up prioritization.

[0175] Based on the preset threshold range to which the risk score belongs, a tiered treatment strategy can be formed. Below are a set of example thresholds (0.10, 0.22, 0.25) for treatment suggestions, which can be directly used as system output or operational guidelines for staff (these thresholds can be flexibly configured according to the resource conditions or risk tolerance of different blood donation sites):

[0176] Risk score <0.10: Routine observation; routine education; routine follow-up recommended;

[0177] Risk score 0.10–<0.22: Extend observation time by 10 minutes; administer oral rehydration solution; enhance patient education and risk awareness; prioritize follow-up.

[0178] Risk score 0.22–<0.25: Extend observation time by 20 minutes; oral rehydration or fluid replacement; enhanced education and risk warnings; mandatory follow-up.

[0179] Risk score ≥ 0.25: Extend observation time by 20 minutes; administer intravenous fluids; strengthen health education and risk warnings; follow-up is mandatory and family members are advised to accompany the patient upon departure.

[0180] The aforementioned risk-based tiered approach can prioritize the allocation of limited on-site care and follow-up medical resources to high-risk individuals without significantly increasing hardware investment. This helps reduce the risk of requiring emergency medical treatment or transfer due to severe adverse reactions to blood donation, minimizes unnecessary intervention for low-risk donors, and thus improves the overall donor experience and the likelihood of future blood donations. Therefore, this invention has clear value for on-site applications and commercial deployment potential.

[0181] Example 2

[0182] Figure 12 This invention illustrates a specific implementation of a blood donation adverse reaction risk prediction system based on stratified blood volume estimation, comprising:

[0183] The data acquisition module is used to collect basic information and covariate information of blood donors. The basic information includes gender, height, weight, and blood donation volume V. The covariate information includes age, whether it is the first time donating blood, and the type of blood collection location.

[0184] The body mass index calculation module is used to calculate the body mass index (BMI) based on the blood donor's height and weight.

[0185] The blood volume stratification calculation module is used to select a blood volume estimation formula based on body mass index (BMI) and calculate the estimated blood volume EBV. When the body mass index (BMI) is less than a preset threshold, the first blood volume estimation formula is used, and when the body mass index (BMI) is greater than or equal to the preset threshold, the second blood volume estimation formula is used.

[0186] The relative blood donation load calculation module is used to calculate the relative blood donation load DR based on the blood donation volume V and the estimated blood volume EBV. ;

[0187] The risk prediction module is used to input relative blood donation load (DR) and covariate information into the trained risk prediction model and output a risk score for blood donors to experience severe adverse reactions to blood donation.

[0188] The decision output module is used to output corresponding treatment suggestions based on the preset risk level range to which the risk score belongs. The treatment suggestions include at least one of the following: observation duration, fluid resuscitation suggestions, health education tips, and follow-up priority.

[0189] The alarm module is used to output alarm information when real-time physiological indicators meet preset abnormal conditions, and to provide enhanced prompts or upgrades for handling suggestions.

[0190] Using the risk prediction system of this embodiment, the method for predicting adverse reactions to blood donation in blood donors is as shown in the steps of Embodiment 1, wherein the training process of the risk prediction model is also as described in Embodiment 1.

[0191] This system can be embedded into existing blood donation vehicle, blood donation center information systems, or blood bank central business systems to automatically calculate risk scores, automatically prompt and generate treatment suggestions, and further link with real-time physiological indicator monitoring and alarm modules to form a closed-loop management process of "assessment-early warning-treatment-feedback" to improve the efficiency of safety management at blood donation sites.

Claims

1. A method for predicting the risk of adverse reactions to blood donation based on stratified blood volume estimation, characterized in that, Includes the following steps: S1. Collect basic information about blood donors, including gender, height, weight, and blood donation volume V; S2. Calculate the Body Mass Index (BMI) based on the blood donor's height and weight; S3. Select the blood volume estimation formula based on the body mass index (BMI) and calculate the estimated blood volume EBV. When the body mass index (BMI) is less than the preset threshold, the first blood volume estimation formula is used. When the body mass index (BMI) is greater than or equal to the preset threshold, the second blood volume estimation formula is used. S4. Calculate the relative blood donation load (DR) based on the blood donation volume (V) and the estimated blood volume (EBV). ; S5. Input the relative blood donation load (DR) into the trained risk prediction model to obtain the risk score of blood donors experiencing severe adverse reactions to blood donation; S6. Output corresponding treatment suggestions based on the preset risk level range to which the risk score belongs. The treatment suggestions include at least one of the following: observation duration, fluid resuscitation suggestions, health education tips, and follow-up priority.

2. The method for predicting adverse blood donation reaction risk based on stratified blood volume estimation according to claim 1, characterized in that: In step S1, covariate information of blood donors is also collected, including age, whether it is the first time donating blood, and type of blood collection location; In step S5, covariate information is also input into the trained risk prediction model.

3. The method according to claim 1, characterized in that: The adverse reactions to blood donation include blood donation-related vasovagal reactions.

4. The method for predicting adverse blood donation reaction risk based on stratified blood volume estimation according to claim 1, characterized in that: The first blood volume estimation formula is the Nadler formula, and the second blood volume estimation formula is the Holme formula.

5. The method for predicting adverse blood donation reaction risk based on stratified blood volume estimation according to claim 1, characterized in that: The preset threshold is determined by setting a candidate threshold set, constructing a hierarchical blood volume estimation rule for each candidate threshold and generating a relative blood donation load (DR) feature, using a training set / validation set partition and combining cross-validation or repeated segmentation extrapolation for evaluation, and using the discrimination index AUC and / or the prediction error index Brier score as optimization targets to screen and obtain the preset threshold.

6. The method for predicting adverse blood donation reaction risk based on stratified blood volume estimation according to claim 5, characterized in that: The preset threshold is in the range of 25 to 28.

7. The method for predicting adverse blood donation reaction risk based on stratified blood volume estimation according to claim 6, characterized in that: The preset threshold is in the range of 26 to 27.

5.

8. The method for predicting adverse blood donation reaction risk based on stratified blood volume estimation according to claim 1, characterized in that, The training method for the risk prediction model is as follows: S501. A training sample set is constructed using historical blood donation records. Each training sample corresponds to one blood donation event and includes at least: gender, height, weight, and blood donation volume V. The label is "Severe" for severe adverse blood donation reaction, where Severe = 1 indicates that a severe adverse blood donation reaction occurred, and Severe = 0 indicates that no severe adverse blood donation reaction occurred. S502. Construct features for the training samples, specifically including: first, calculating the body mass index (BMI), and then selecting either the first blood volume estimation formula or the second blood volume estimation formula based on the relationship between BMI and a preset threshold to calculate the estimated blood volume (EBV), and finally calculating the relative blood donation load (DR) as the core input feature. ; S503. Data Preprocessing and Sample Screening: Perform unit conversion and format standardization on training samples; remove, correct, or impute missing or outlier values. S504. Data partitioning and stratified sampling: The training sample set is divided into a training set and a validation set according to a preset ratio, and stratified sampling is used to maintain the ratio of positive to negative samples. S505, Model Fitting and Objective Function: Train the risk prediction model on the training set to minimize the loss function between the risk score of severe adverse blood donation reaction output by the risk prediction model and the label. The loss function is log loss or cross-entropy loss. S506. Validate the risk prediction model using K-fold cross-validation or repeated splitting extrapolation evaluation method, and obtain the evaluation results; S507. Based on the assessment results, select risk prediction model parameters or risk prediction model structure, and determine the risk level range or threshold set. S508. Model Stabilization: Save the trained risk prediction model parameters as a callable model file or parameter table.

9. The method for predicting the risk of adverse reactions to blood donation based on stratified blood volume estimation according to claim 8, characterized in that: The risk prediction model is one of the following: logistic regression model, generalized linear model, support vector machine, random forest, gradient boosting tree, and ensemble learning model.

10. The method for predicting the risk of adverse reactions to blood donation based on stratified blood volume estimation according to claim 8, characterized in that: The trained risk prediction model is implemented using an ensemble prediction structure, which includes three stages: feature grouping, base model training and initial prediction, and fusion output. The feature grouping specifically involves dividing the input features into multiple feature groups, which include explicit feature groups, implicit feature groups, original feature groups, and basic transformation feature groups, wherein: Explicit feature sets, constructed based on medical or business priors, include at least the relative blood donation burden index DR and at least one donor covariate. Implicit feature groups are obtained by combining the original features through decision tree models or gradient boosting tree models, and are used to represent the high-order interaction relationships between the original features. The original feature set consists of the original input variables without any transformation, including gender, height, weight, and blood donation volume. The basic transformation feature set is a feature set obtained by performing at least one mathematical transformation on the original feature set. The mathematical transformation includes at least one of logarithmic transformation, square transformation, and binning discretization. The base model training and initial prediction specifically involve: training at least one base model for different feature groups and outputting the initial risk score or probability prediction result corresponding to the base model. The base model can be any trainable classification or regression model, and cross-validation or repeated split extrapolation is used to estimate the generalization performance. The fusion output specifically refers to: taking the initial risk score or probability prediction result of the base model as input, training the fusion model to output the final risk score; the fusion model is implemented by weighted fusion or stacked fusion, wherein the weights or fusion parameters are determined based on the cross-validation results or repeated segmentation extrapolation evaluation results.

11. The method for predicting the risk of adverse reactions to blood donation based on stratified blood volume estimation according to claim 1, characterized in that: After step S6, real-time monitoring and alarm linkage are also carried out. The specific method is to collect real-time physiological indicators, output alarm information when the real-time physiological indicators meet the preset abnormal conditions, and provide enhanced prompts or upgrade outputs for handling suggestions.

12. A blood donation adverse reaction risk prediction system based on stratified blood volume estimation, characterized in that, include: The data acquisition module is used to collect basic information of blood donors, including gender, height, weight, and blood donation volume V. The body mass index calculation module is used to calculate the body mass index (BMI) based on the blood donor's height and weight. The blood volume stratification calculation module is used to select a blood volume estimation formula based on body mass index (BMI) and calculate the estimated blood volume EBV. When the body mass index (BMI) is less than a preset threshold, the first blood volume estimation formula is used, and when the body mass index (BMI) is greater than or equal to the preset threshold, the second blood volume estimation formula is used. The relative blood donation load calculation module is used to calculate the relative blood donation load DR based on the blood donation volume V and the estimated blood volume EBV. ; The risk prediction module is used to input the relative blood donation load (DR) into the trained risk prediction model and output a risk score for blood donors to experience severe adverse reactions to blood donation. The decision output module is used to output corresponding treatment suggestions based on the preset risk level range to which the risk score belongs. The treatment suggestions include at least one of the following: observation duration, fluid resuscitation suggestions, health education tips, and follow-up priority.

13. The adverse blood donation reaction risk prediction system based on stratified blood volume estimation according to claim 12, characterized in that, Also includes: The alarm module is used to output alarm information when real-time physiological indicators meet preset abnormal conditions, and to provide enhanced prompts or upgrades for handling suggestions.

14. The adverse blood donation reaction risk prediction system based on stratified blood volume estimation according to claim 12, characterized in that: The data acquisition module is also used to collect blood donor covariate information, which includes age, whether it is the first time donating blood, and the type of blood collection location; the risk prediction module is also used to input the blood donor covariate information into the trained risk prediction model.

Images