A donor-recipient matching prediction system based on multi-source heterogeneous data fusion

Abstract

This invention belongs to the field of medical informatics, specifically relating to a donor-recipient matching prediction system based on multi-source heterogeneous data fusion. The system includes: a data access unit for receiving metagenomic, host phenotype, and historical efficacy data of recipients and candidate donors; a multi-dimensional feature engineering processing unit for converting heterogeneous data into feature vectors of uniform dimensions and constructing a multi-scale feature space including niche compatibility, functional complementarity, and host adaptation features; and an ensemble learning prediction engine unit for performing nonlinear inference using a pre-trained optimized random forest model, calculating the probability of reaching the clinical endpoint during the follow-up period, and outputting the feature weight contribution. This invention improves the accuracy and interpretability of donor-recipient matching through multi-scale heterogeneous data fusion and ensemble learning inference, achieving a shift from experience-based matching to vectorized, precise matching, and reducing the waste of medical resources.

Description

Technical Field

[0001] This invention belongs to the field of medical informatics, specifically relating to a donor-recipient matching prediction system based on multi-source heterogeneous data fusion. Background Technology

[0002] With the rapid rise of microbiome medicine, fecal microbiota transplantation (FMT), as a key intervention for restoring gut microbiota balance, has demonstrated clinical value in the treatment of diseases such as recurrent Clostridium difficile infection and inflammatory bowel disease. By transplanting the entire gut microbiota of a healthy donor into the recipient's intestine, the aim is to correct gut microbiota dysbiosis and restore the host's immune balance. In the current clinical translation process, extracting key biomarkers from complex biological information and establishing precise donor screening systems have become core aspects of improving the safety and efficacy of biotherapy.

[0003] Compatibility matching between donors and recipients is a key technological direction determining transplant success rates and clinical outcomes. This direction aims to establish scientific matching models by analyzing multivariate data, including differences in microbial composition, metabolic complementarity, and host physiological phenotypes between donors and recipients. Its fundamental principle lies in identifying the feasibility of donor microbes colonizing the recipient's gut microenvironment and assessing their ability to replenish deficient functions in the recipient.

[0004] Current technologies primarily rely on a single microbiome structure matching paradigm, typically using species abundance similarity or community distance as core metrics. This neglects the ecological interaction mechanisms and functional metabolic compensation effects among microorganisms, limiting the accuracy of clinical efficacy prediction. Most current matching strategies are based on linear logistic analysis or traditional statistical methods, which struggle to analyze the high-dimensional nonlinear interactions between donor-specific microorganisms and recipient-specific ecological niches. Furthermore, consideration of host phenotypic characteristics is limited to coarse-grained indicators such as age, height, and weight, failing to capture the impact of sex, hormone, and metabolic differences on colonization outcomes. The lack of dynamic fusion capabilities for multi-source heterogeneous data means that screening decisions often rely on clinical experience rather than quantitative predictive evidence, easily leading to incompatibility and metabolic disorders in the recipient's gut environment, thus hindering improvements in clinical response rates. Summary of the Invention

[0005] The purpose of this invention is to provide a donor-recipient matching prediction system based on multi-source heterogeneous data fusion, which can solve the problems of lack of scientific quantitative basis for donor screening, unpredictable clinical efficacy, and resource inefficiency caused by incompatibility of donor and recipient niches in the above-mentioned background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A donor-recipient matching prediction system based on multi-source heterogeneous data fusion includes a data access unit, a multi-dimensional feature engineering processing unit, an ensemble learning prediction engine unit, and a matching decision and report generation unit, as follows: The data access unit is used to receive multi-source heterogeneous biomedical data from recipients and candidate donors. The data includes standardized metagenomic analysis results, host basic phenotypic data, and historical transplant efficacy records. The metagenomic analysis results include diversity indices reflecting community diversity, overlap indices characterizing niche overlap, pathway abundance matrices reflecting metabolic potential, and microbial interaction network parameters. The host basic phenotypic data includes sex information, body mass index information, and age information. The multidimensional feature engineering processing unit is connected to the data access unit and is used to perform domain knowledge-driven transformation and normalization processing on the accessed raw data, transform multi-source heterogeneous data into feature vectors with unified dimensions, and construct a multi-scale feature space including niche compatibility features, functional complementarity features and host adaptation features. The integrated learning prediction engine unit is connected to the multidimensional feature engineering processing unit. It is loaded with pre-trained optimized random forest model weights to perform nonlinear inference on the feature vector, calculate the probability prediction value of the donor and recipient reaching the primary clinical endpoint within the predetermined follow-up period, and simultaneously output the influence weight contribution of each feature on the prediction result. The matching decision and report generation unit is connected to the integrated learning prediction engine unit and is used to sort all candidate donors in descending order according to the probability prediction value, filter out the donor set that is higher than the preset probability threshold, and generate a structured matching report containing matching scores and failure attribution analysis.

[0007] Preferably, the niche compatibility features are extracted by the multidimensional feature engineering processing unit, including a microbial community richness and evenness difference index, an overall community structure difference index, and a niche overlap index. The microbial community richness and evenness difference index is used to characterize the ecological stability of the gut microbiota. The overall community structure difference index is obtained by calculating the ratio of the absolute value of the abundance difference between each taxonomic unit of the donor and recipient to the total abundance, and is used to measure basic compatibility. The niche overlap index is calculated by summing the products of the resource utilization probabilities of the donor and recipient at each metabolic site, and then standardizing the results by combining the geometric mean of the sum of the squares of their respective resource utilization probabilities, in order to reveal the colonization feasibility of the donor strain in the recipient's gut microenvironment.

[0008] Furthermore, the functional complementarity feature is constructed by the multidimensional feature engineering processing unit based on information from the metabolic pathway database, including a metabolic pathway complementarity index, a microbial network compatibility index, and a probiotic targeting index. The metabolic pathway complementarity index is obtained by quantifying the donor's ability to supplement key metabolic functions missing in the recipient, which cover short-chain fatty acid synthesis and bile acid conversion pathways. The microbial network compatibility index is obtained by calculating the crossover and union ratio of synergistic and antagonistic relationships between the donor and recipient microbiota, reflecting the degree to which the donor microbiota is accepted by the recipient's original network. The probiotic targeting index is obtained by statistically analyzing the number of predetermined core bacterial genera that are specific to the donor and missing in the recipient.

[0009] Furthermore, the host adaptation features include sex matching status, absolute value of body mass index difference, and absolute value of age difference; the sex matching status is encoded as a binary variable to reflect the influence of sex hormone environment on microbial colonization; the body mass index difference and age difference are introduced after standardization to reflect differences in hosts at the physiological phenotype level.

[0010] Preferably, the normalization processing performed by the multidimensional feature engineering processing unit includes a differentiated transformation strategy: for continuous features, a linear normalization method is used, which subtracts the minimum observation value in the predetermined training set from the original value, and then divides it by the difference between the maximum and minimum observation values ​​to map the value to a predetermined interval of 0 to 1; for count features reflecting the number of probiotics, a logarithmic transformation method with a predetermined constant as the base is used, which adds 1 to the original value and then takes the logarithm to alleviate the long-tail effect of the data distribution; for binary features such as gender, the original logical value is maintained.

[0011] Furthermore, the integrated learning prediction engine unit employs an optimized random forest model that includes a preset number of decision trees as base learners, and sets a preset maximum depth and a preset minimum number of leaf node samples. During the model training phase, this unit adopts a feature importance-guided oversampling strategy, that is, based on the weight of the feature's contribution to classification, it performs weighted synthesis on minority class samples, i.e., clinical responders, to eliminate the impact of imbalanced sample class distribution on prediction accuracy.

[0012] Furthermore, the integrated learning prediction engine unit incorporates an early stopping mechanism based on Shapley sum and explanatory value during the inference process. That is, during the model iteration process, the consistency of the Shapley sum and explanatory value output by each decision tree is calculated in real time. When the change in explanatory value within a predetermined number of consecutive rounds is less than a preset small threshold, the inference is automatically stopped to prevent the model from overfitting to specific sample noise.

[0013] Preferably, the integrated learning prediction engine unit is also equipped with an uncertainty calibration function, which integrates the prediction results of multiple decision trees and calculates the standard deviation based on the out-of-bag sample variance to output the corresponding confidence interval for each probability prediction value, providing a measure of the reliability of the prediction results.

[0014] Furthermore, when generating a report, the matching decision and report generation unit performs a matching failure attribution analysis, that is, by tracing back the features with negative Shapley values ​​and large absolute values, it identifies the core factors that cause the donor and recipient to mismatch. The core factors include the absence of key bacterial genera, metabolic function conflicts, or host phenotype incompatibility.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The system provided by this invention overcomes the limitations of single-microbiome structure matching. By integrating heterogeneous data at multiple scales, including niche compatibility layers, functional complementarity layers, and host adaptation layers, a comprehensive donor-recipient compatibility evaluation system is constructed. Compared to traditional models that rely solely on community similarity, this invention can accurately capture the complex biological logic of donor microbiota colonization and function in the recipient's gut, improving the accuracy of clinical efficacy prediction and achieving a high pre-set area under the curve (AUC) index.

[0016] 2. This invention employs an optimized random forest algorithm and its accompanying feature importance-guided oversampling strategy to address the common sample imbalance problem in clinical data. By introducing an early stopping mechanism based on Shapley sum and explained values, along with uncertainty calibration output, the system not only ensures high robustness of predictions but also provides interpretability to the prediction results. It can quantify the contribution of each feature to the matching score, providing clinicians with understandable and reliable decision support.

[0017] 3. This invention represents a paradigm shift in fecal microbiota transplantation (FMT) from empirical matching to precise vector-based matching. By automatically traversing a candidate donor database and generating a Top-K recommendation list, the system can customize the selection of donors with a high probability of therapeutic efficacy based on the recipient's gut microbiota characteristics and host physiological background, reducing treatment failures and resource waste caused by incompatibility. The system's failure attribution analysis function provides in-depth insights for clinical research, helping to reveal key ecological mechanisms affecting transplant prognosis, and possesses both clinical application value and research reference value. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the present invention for performing clinical endpoint probability prediction based on an ensemble learning prediction engine; Figure 3This is a flowchart illustrating the logical flow of the multi-dimensional feature engineering processing unit in this invention, which constructs a multi-scale feature space. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of multi-source heterogeneous biomedical data between the recipient and the candidate donor in this invention; Figure 5 This is a flowchart illustrating the logical flow of matching decision screening, donor ranking, and matching failure attribution analysis in this invention. Detailed Implementation

[0019] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 A donor-recipient matching prediction system based on multi-source heterogeneous data fusion includes a data access unit, a multi-dimensional feature engineering processing unit, an integrated learning prediction engine unit, and a matching decision and report generation unit. The data access unit, deployed on a data center server cluster with high-speed network access capabilities, is used to receive multi-source heterogeneous biomedical data from recipients and multiple candidate donors through a standardized application programming interface. In specific engineering implementations, the data access unit is configured as an asynchronous message queue structure with high concurrency processing capabilities, capable of processing raw metagenomic sequencing data and host phenotype data packets sent from multiple clinical terminals.

[0020] The data access unit includes a data format verification subunit, a data cleaning and denoising subunit, a missing value multiple imputation subunit, and a data privacy desensitization subunit.

[0021] The asynchronous message queue structure of the data access unit is constructed using an advanced message queue protocol or a message queue telemetry transport protocol. Each data packet sent by the clinical terminal is encapsulated as a structured message containing a message header and a message body. The message header contains a unique identifier for the recipient or donor, the data type (metagenomics, phenotype, efficacy), and a timestamp; the message body contains the compressed raw data payload. The differential privacy algorithm executed by the data privacy desensitization subunit is specifically a Laplace mechanism. For continuous sensitive fields such as host age and body mass index, the desensitization process is as follows: first, the sensitivity of the query function is determined. Then generate a distribution that follows a Laplace distribution. The random noise is added to the original value to obtain the desensitized value.

[0022] For continuous sensitive field values Its desensitized value The calculation is as follows: ; in, This indicates that the position parameter is 0 and the scale parameter is... The Laplace distribution has a probability density function of: ; The global sensitivity of the query function is defined as the maximum difference between query results on adjacent datasets. Norm; The preset privacy budget typically ranges from [value range missing]. The smaller the value, the stronger the privacy protection.

[0023] The data format verification subunit is configured to perform strict protocol consistency checks on the incoming raw biomedical data, ensuring that metagenomic analysis results conform to predetermined standard naming conventions, such as using unified taxonomic unit level labeling. The data cleaning and denoising subunit filters out unqualified samples with insufficient sequencing depth or excessive contamination rates, improving the signal-to-noise ratio of the input data. The missing value multiple imputation subunit fills in missing minor host phenotypic indicators that may be missing during clinical data collection using a chain equation-based multivariate imputation method to ensure the integrity of the feature space. The data privacy desensitization subunit uses a differential privacy algorithm to desensitize fields involving host-sensitive information before the data enters subsequent processing stages, ensuring compliance with medical data requirements.

[0024] The multidimensional feature engineering processing unit establishes a data communication connection with the data access unit through a high-speed local area bus. Its core function is to perform domain knowledge-driven deep transformation on the received heterogeneous data.

[0025] When constructing the niche overlap index, the multidimensional feature engineering processing unit defines the "resource utilization probability" as the relative metabolic activity of microorganisms at specific metabolic resource sites. For the first... A metabolic site, receptor resource utilization probability The calculation is the weighted sum of the relative abundances of all metabolic enzyme genes in the recipient metagenomics corresponding to that locus. Similarly, the donor abundance is calculated... resource utilization probability Subsequently, the niche overlap index The calculation is performed through the following steps: First, calculate the molecular term, that is, the donor and acceptor in all... The sum of the product of the resource utilization probabilities at each metabolic site is used; next, the denominator is calculated, which is the square root of the product of the sum of squares of the recipient resource utilization probabilities and the sum of squares of the donor resource utilization probabilities; finally, the numerator is divided by the denominator to obtain the standardized overlap value.

[0026] The multidimensional feature engineering processing unit is configured to map highly fragmented microbiome and host physiological data to a high-dimensional and physically consistent feature vector space. This unit is configured to construct three nested feature sub-layers: a niche-compatible feature layer, a functionally complementary feature layer, and a host-adaptive feature layer.

[0027] In the niche compatibility feature layer, the multidimensional feature engineering processing unit is configured to extract microbial community richness and evenness difference indicators reflecting the stability of the gut microbiota. The microbial community richness is obtained by calculating the Shannon index of taxa, which is calculated by summing the products of the relative abundance of each taxa and their natural logarithms, and then taking the negative of the sum. The system assesses the recipient gut's capacity to accommodate external microbiota by comparing the Shannon index deviations between the recipient and candidate donors. This unit also calculates an overall community structure difference indicator, i.e., by summing the absolute values ​​of the abundance differences between the recipient and candidate donors at each taxonomic level, and using the ratio of this sum to the total abundance as a basic compatibility measure. Crucially, this unit is also configured to calculate a niche overlap indicator, which is calculated by summing the products of the resource utilization probabilities of the donor and recipient at each specific metabolic site, and then dividing by the geometric mean of the squared sums of their respective resource utilization probabilities. Through the standardization of this indicator, the system can quantitatively predict the colonization feasibility of donor strains in the complex microenvironment of the recipient.

[0028] In the functional complementary feature layer, the multidimensional feature engineering processing unit is configured to associate with a preset metabolic pathway database in real time to extract metabolic pathway complementarity indicators, microbial network compatibility indicators, and probiotic targeting indicators.

[0029] Before constructing the microbial community network compatibility index, the multidimensional feature engineering processing unit first constructs a microbial co-occurrence network. Specifically, it calculates the pairwise correlation between the relative abundance of all species based on the Spearman correlation coefficient and performs a significance test on the correlation coefficient. ), and define species pairs with significant positive correlation as synergistic edges, assigning a weight of 1; define species pairs with significant negative correlation as antagonistic edges, assigning a weight of -1. Consider a receptor network. The set of edges is donor network The set of edges is The microbial community network compatibility index The calculation is as follows: ; Let e ​​be the weight of edge e; This value ranges from 0 to 1, with a larger value indicating a more compatible network structure.

[0030] The generation logic of the metabolic pathway complementarity index is as follows: key metabolic functions missing in the recipient, such as short-chain fatty acid synthesis pathways or secondary bile acid conversion pathways, are retrieved; the donor's compensatory capacity for these missing pathways is quantified, specifically manifested as the degree of enhancement in the abundance of corresponding metabolic enzymes in the donor. The microbial network compatibility index is obtained by calculating the cross-union ratio (CUNR) of synergistic and antagonistic relationships between the donor microbial network and the recipient's in-situ microbial network; a value approaching 1 indicates that the donor microbial community is highly readily accepted by the recipient's original ecological network. The probiotic targeting index specifically counts the number of strains from 12 core genera present in the donor that are completely missing in the recipient and have been clinically validated to have colonization capabilities.

[0031] In summary, the microbial community network compatibility index of this invention The calculation formula is: ; in Indicates the connection node and The weights of the edges are as follows: edges with cooperative relationships have a weight of 1, and edges with antagonistic relationships have a weight of -1. Indicates taking the absolute value; and These represent the intersection and union operations of sets, respectively.

[0032] In the host adaptation feature layer, the multidimensional feature engineering processing unit is configured to process non-microbiome host features, including sex matching status, absolute difference in body mass index (BMI), and absolute difference in age. The sex matching status is logically quantified into a binary state quantity: 0 if the recipient and donor are of the same sex, and 1 if they are different, reflecting the varying impacts of different sex hormone backgrounds on microbial community colonization. The BMI and age differences are characterized by calculating the absolute differences between the donor and recipient values ​​to represent the degree of adaptation at the physiological phenotypic level.

[0033] The multidimensional feature engineering processing unit is also configured to execute differentiated feature normalization strategies to eliminate the influence of different dimensions on model training. For continuous features such as the Shannon index and niche overlap, the unit uses linear transformation logic to subtract the preset minimum observation value from the original observation value, and then divide by the difference between the maximum and minimum observation values, linearly mapping the data to a closed interval of 0 to 1. For count features reflecting the number of probiotics, the unit uses logarithmic transformation logic with base 10, adding 1 to the original number and then taking the logarithm to alleviate the long-tailed distribution bias caused by high-abundance strains. For binary features such as sex, the system transmits the original logical state value.

[0034] The integrated learning prediction engine unit is connected to the multidimensional feature engineering processing unit and is loaded with pre-trained optimized random forest model weight parameters.

[0035] The ensemble learning prediction engine unit employs a feature importance-guided oversampling strategy during the training phase, specifically an improved algorithm based on Synthetic Minority Oversampling Technique (SMOTE) (FISMOTE). First, the average Gini importance weight for each feature is calculated for the minority class samples (clinical responders). Then, for each minority class sample... Searching for its in the feature space nearest neighbor samples ( Randomly select a neighbor from among them. The interpolation process is guided by the weights of feature importance. New samples are then synthesized. The formula is: ; in, and These represent the new sample and the original sample at the th... Values ​​on each feature for Random numbers within an interval. It is the first The normalized importance weight of each feature is calculated as the proportion of the decrease in average Gini impurity for that feature in the random forest model. It is a minority class sample The One eigenvalue; It is a sample A randomly selected nearest neighbor sample The 1 eigenvalue, This allows the features that the model considers important to have a larger range of variation when synthesizing samples, thus more effectively expanding the sample space near the decision boundary.

[0036] This unit is configured to take the normalized multidimensional feature vectors mentioned above as input and perform a complex nonlinear inference process. The optimized random forest model consists of 500 classification and regression decision trees as base learners. The maximum depth of each decision tree is preset to 12, and the minimum number of leaf node samples is set to 5 to ensure that the model has good generalization performance while capturing high-order interaction effects.

[0037] The loss of the entire forest is the average loss of all decision trees on that sample. Training uses a bootstrap sampling method to build a training set for each tree. The model's parameters are updated through the splitting growth of the decision trees. At each node, the algorithm traverses all candidate features and all possible split points, selecting the feature and split point that maximizes the reduction of the weighted Gini impurity (or maximizes the information gain) of the samples at that node as the splitting rule, until the stopping condition is met (reaching the maximum depth or the number of leaf node samples is less than a threshold).

[0038] The binary cross-entropy loss function $L$ used to optimize the training of the random forest model is used for a single sample. Its expression is: ; in, The input feature vector; For the true label of the sample, Indicates clinical response.

[0039] This indicates no response; It is the model on the sample The probability of a positive prediction (response) is given by the average of the predictions from all decision trees in the random forest.

[0040] In the integrated learning prediction engine unit, to address the imbalance between clinical responders and non-responders in the clinical samples, the system is configured with feature importance-guided oversampling logic during the training phase. This logic calculates the Gini coefficient contribution of each feature to the classification decision, and performs linear interpolation in the feature space according to the weighted proportions for the minority class samples, i.e., clinical responders, to synthesize biologically plausible virtual samples, thereby adjusting the class distribution to a balanced state.

[0041] Furthermore, the ensemble learning prediction engine unit embeds early stopping logic based on Shapley sum and explanatory values ​​when performing inference prediction. Specifically, when the system performs multi-decision tree ensemble calculation on a donor-recipient pair, it monitors the consistency of the Shapley sum and explanatory values ​​of each feature generated in each iteration in real time. If, in five consecutive iterations, the average variation of the Shapley sum and explanatory values ​​of all features is less than a preset small threshold of 0.001, the prediction engine automatically terminates the subsequent tree calculation for that pair and outputs the current ensemble result, thus improving computational efficiency and avoiding overfitting to specific noisy data.

[0042] The integrated learning prediction engine unit is also configured to have an uncertainty calibration function. The system calculates the variance of the outputs of all 500 decision trees and, in conjunction with the error distribution based on out-of-bag samples, adds a corresponding standard deviation to each generated prediction probability value, outputting the confidence interval of the prediction result.

[0043] Probability prediction value The formula for calculating the confidence interval is: ; in, It is the standard deviation of the out-of-bag prediction probability, calculated using the following formula: ; in, This is the total number of decision trees that treat this sample as an out-of-bag sample; It is the first The probability of each decision tree predicting the sample; It is all The mean of the predicted probabilities.

[0044] This mechanism ensures that the prediction results include not only simple probability values, but also a measure of the reliability of those values, providing a deeper level of confidence reference for clinical decision-making.

[0045] The matching decision and report generation unit, connected to the integrated learning prediction engine unit, is used to perform descending sorting logic on all donors to be matched based on their predicted probability values. This unit is configured with a preset probability threshold; only donors with predicted probabilities higher than this threshold are included in the recommendation list. Typically, the system is configured to output the three optimal donors with the highest probabilities of therapeutic efficacy and generate a structured matching report.

[0046] The matching decision and report generation unit is also configured to perform in-depth attribution analysis logic for matching failures during the report generation process. For candidate donors with low predicted probabilities, this unit automatically identifies the core biological factors leading to the mismatch by backtracking features with negative Shapley values ​​and large absolute values. For example, if the Shapley value corresponding to a certain feature is negative, the system will explicitly indicate in the report that the mismatch is caused by specific reasons such as the absence of key bacterial genera, metabolic pathway conflicts, or a large age difference between hosts. This attribution analysis function allows clinicians to clearly understand the logic behind the matching results, enhancing the interpretability and clinical applicability of the system.

[0047] In this embodiment, the system operates as follows: the data access unit obtains metagenomic and phenotypic data of 169 pairs of real clinical samples from the clinical database and completes standardized preprocessing; the multidimensional feature engineering processing unit constructs a feature space containing 9 core indicators and performs refined normalization transformation; then, the integrated learning prediction engine unit loads the trained weights and performs probabilistic inference on new pairs to be matched, and the predicted area under the curve index reaches a certain level; the matching decision and report generation unit generates a personalized matching report containing specific scores and the contribution of driving features.

[0048] Example 2: Based on the donor-recipient matching prediction system based on multi-source heterogeneous data fusion described in Example 1, this example provides an alternative system implementation based on edge computing and cloud federated learning architecture to meet the data privacy protection and distributed processing requirements in large-scale medical consortia.

[0049] In Embodiment 2, the physical architecture of the system is divided into edge computing nodes deployed locally in the medical institution and a cloud coordination engine deployed in a centralized data center.

[0050] In the distributed processing module deployed on local edge nodes of the medical institution, the federated learning process specifically involves each edge node performing one or more rounds of gradient descent training on its local model copy using newly added clinical samples, thereby obtaining the changes in model parameters (gradient vectors). To protect privacy, each node adjusts the gradient vector. Gaussian noise is added or encryption is applied to generate an update package, which is then uploaded to the cloud coordination engine. The cloud coordination engine uses a federated averaging algorithm to aggregate the update packages uploaded by each node: calculating the average gradient vector in all valid update packages. Then update the global model parameters. The updated global model parameters are distributed back to each edge node, completing one round of co-evolution.

[0051] In Embodiment 2, the data access unit is deployed within local edge computing nodes of various medical institutions. This distributed deployment allows for preliminary parsing of raw metagenomic sequencing data from recipients and donors locally, eliminating the need to upload massive amounts of raw sequence files to the cloud, saving backbone network bandwidth overhead, and reducing the risk of data leakage. The local version of the data access unit integrates a high-performance local storage submodule for caching historical matching records to enable fast offline queries.

[0052] The multidimensional feature engineering processing unit, also deployed on a local edge node in Example 2, is configured to interface with a local bioinformatics analysis workstation to acquire species abundance tables and pathway abundance matrices in real time. When calculating niche overlap and metabolic pathway complementarity indices, this unit utilizes local computing resources to perform large-scale parallel matrix operations.

[0053] Niche overlap index The calculation formula is: ; in, The total number of preset metabolic resource sites; The recipient and donor were respectively in the 1st month. The probability of resource utilization at each metabolic site is calculated using the following formula: ; in, For catalysis A collection of enzyme genes at a metabolic site; The first in the receptor metagenomics The relative abundance of each enzyme gene; The catalytic efficiency weights for this enzyme are provided by a pre-defined metabolic pathway database. Donor The same calculation applies.

[0054] To ensure comparability of features between different medical institutions, the multidimensional feature engineering processing unit is configured to periodically download globally dynamically updated normalized parameters from the cloud coordination engine, namely the maximum and minimum values ​​in the training set.

[0055] The global model parameter aggregation formula executed by the cloud coordination engine is: ; in, These are the parameter tensors of the global model before and after aggregation, respectively; The number of edge nodes participating in this round of aggregation; For the first The amount of model parameter updates uploaded by each edge node is typically calculated as the parameter changes after that node performs one or more gradient descent operations on its local dataset. .

[0056] For the first The initial model parameters for each edge node before the start of this round of local training. No. The edge nodes update the model parameters after completing local training.

[0057] This mechanism ensures that even when there are offsets in the local data distribution, the system can still accurately map features to a globally consistent feature space.

[0058] In Embodiment 2, the ensemble learning prediction engine unit employs a model splitting and federated inference architecture. Its internal optimized random forest model weights are encrypted and stored locally, activating inference only upon receiving a matching request.

[0059] This unit is configured with incremental model update capabilities, enabling collaborative optimization of model parameters by uploading gradient information or model update amounts to the cloud coordination engine, based on locally added and follow-up-confirmed FMT efficacy labels from various medical institutions, without sharing the original data. During probabilistic inference, this engine unit utilizes locally configured hardware accelerators, such as graphics processors or dedicated tensor processing units, to handle the high-concurrency computation task of 500 decision trees.

[0060] In Example 2, the interpretability of the computational logic of the integrated learning prediction engine unit is further enhanced. In addition to calculating the Shapley sum and the interpreted value, the unit is configured to establish a local surrogate model for linearly approximating the prediction decision logic within the local neighborhood of the recipient data. This approach provides physicians with more intuitive feature sensitivity analysis reports, such as displaying a dynamic curve showing the current donor matching success rate as the recipient's Shannon index increases by 10%.

[0061] In Embodiment 2, the matching decision and report generation unit is configured to have multi-dimensional risk warning capabilities. When generating a structured matching report, this unit not only outputs a Top-K recommendation list but also automatically identifies potential matching pairs with high-risk, high-reward characteristics based on the uncertainty range of the prediction results. For example, for donors with predicted probabilities bordering on the threshold but possessing functional supplementation capabilities, the system will mark them as exploratory matching suggestions and prompt clinicians to strengthen postoperative monitoring. This unit is also configured to automatically interface with the electronic medical record system of medical institutions, pushing the matching report to the recipient's attending physician's workstation in encrypted PDF format.

[0062] At the hardware implementation level of Embodiment 2, the interaction between the various units of the system employs an encrypted communication protocol based on a trusted hardware execution environment. All feature vector transmissions and weight updates are performed within an isolated, secure memory area, preventing side-channel attacks targeting model parameters and sensitive features. This enhanced security design enables the system to be widely applied in multi-center clinical research and cross-regional donor bank sharing platforms where data security requirements are extremely high.

[0063] This embodiment, through improvements to the distributed architecture described above, maintains the high-precision prediction performance of Embodiment 1 while achieving consistent performance of the area under the prediction curve (AUC) on the external validation set, with its value remaining stable above 0.85. By distributing computational load to edge nodes, the system reduces the instantaneous load on the cloud server, increasing the number of concurrent matching requests that a single cluster can support. Through the introduction of a federated learning mechanism, the system can self-evolve as it continuously absorbs new clinical data, and its model prediction accuracy shows a steady upward trend with the accumulation of sample size.

[0064] In the normalization process of Example 2, the system introduced quantile-based preprocessing logic to address the differences in sequencing depth caused by different devices. This logic further eliminates the batch effect introduced by the differences in sequencing technology platforms by adjusting the abundance distribution of receptor samples to the same empirical distribution as the reference set, ensuring that the feature vectors output by the feature engineering unit are robust across different medical institutions.

[0065] The matching decision and report generation unit in Example 2 also integrates a dynamic donor bank management submodule. This submodule can dynamically adjust the matching score weights based on the remaining amount of donor biological samples and the frequency of collection. For example, for donor samples that have been stored for a long time and whose cell viability may have decreased, the system will automatically multiply their original predicted probability by a decay factor and prioritize displaying higher-quality donors with higher freshness and activity in the recommendation list, further ensuring clinical efficacy.

[0066] In summary, Example 2 provides a more scalable and secure system implementation scheme. By deeply integrating edge computing and federated learning, it solves the data silos and privacy ethics problems that are common in medical big data applications, laying a solid technical foundation for the large-scale clinical application of this invention.

[0067] Example 3: Building upon Examples 1 and 2, this example provides a highly available donor-recipient matching prediction system based on a containerized microservice architecture. This system is specifically designed to support ultra-large-scale fecal microbiota matching tasks and can handle tens of thousands of candidate donor screening requests.

[0068] In the third embodiment, each functional module of the system is decoupled into an independent microservice container and dynamically scaled under the scheduling of the cluster operating system.

[0069] In Embodiment 3, the data access unit is transformed into a high-performance API gateway service. This API gateway service is configured to use a non-blocking input / output model, capable of maintaining tens of thousands of concurrent long-connection requests. The unit integrates flow control logic based on a token bucket algorithm to prevent system crashes caused by sudden surges in traffic. The unit also features an automated metadata extraction sub-service, capable of extracting auxiliary dimension information such as sampling date, geographic coordinates, and sample type from user-uploaded raw compressed files, providing rich contextual background for subsequent multidimensional feature analysis.

[0070] The multidimensional feature engineering processing unit in Example 3 employs a distributed computing architecture. This unit is configured to utilize a distributed in-memory computing framework to break down computational tasks into multiple subtasks for parallel execution. For example, when calculating the niche overlap index of 169 pairs of samples, the system distributes the sample pairs across different computing nodes, simultaneously performing the summation of the numerator and denominator of the Piaka index. When processing the metabolic pathway complementarity index, this unit is configured to call a pre-built pathway abundance index database, quickly retrieving receptor-deficient metabolic pathways through high-speed key-value pair query logic. This unit also introduces a feature version management mechanism, which records the algorithm version and parameter configuration used for each feature extraction, ensuring the traceability of experimental results.

[0071] In Example 3, the ensemble learning prediction engine unit is constructed as a distributed inference service cluster. A load balancer is deployed within this cluster to distribute inference tasks to the least busy nodes based on their current resource utilization. The optimized random forest model in Example 3 is stored in a binary serialized format, supporting fast loading across platforms and languages. To further improve prediction speed, this unit is configured to have a resident feature cache pool in memory, storing the most frequently accessed donor feature vectors to avoid redundant feature transformation calculations.

[0072] In the FISMOTE oversampling logic of the integrated learning prediction engine unit, Example 3 introduces a neighborhood selection strategy based on manifold learning. Instead of blindly performing linear interpolation in the full feature space, the system uses a manifold learning algorithm to project high-dimensional features into a low-dimensional manifold space, identifying the manifold topology of minority class samples and generating new samples within this structure. This method ensures that the generated virtual samples are biologically more consistent with the actual distribution patterns of bacterial communities, improving the stability and reliability of the model when dealing with highly imbalanced data.

[0073] Furthermore, the SHAP value early stopping mechanism of the integrated learning prediction engine unit was distributed and optimized in Embodiment 3.

[0074] The early stopping mechanism employs the TreeSHAP algorithm to quickly calculate the Shapley sum and explained value (SHAP value) of each feature. TreeSHAP utilizes the structure of a decision tree, recursively traversing it to calculate the expected marginal contribution of each feature to the prediction result. Specifically, after each iteration (i.e., traversing a portion of the decision tree), the average SHAP value output by all traversed decision trees for each feature is calculated. The mechanism is triggered when the variation in the average SHAP value of all features is less than a threshold over five consecutive iterations. When the inference stops, the uncertainty calibration function assumes that the out-of-bag (OOB) prediction error follows a normal distribution. For each sample, its prediction probability on all decision trees that treat it as an OOB sample is calculated, resulting in a probability sequence. The standard deviation of this sequence is... As a measure of prediction uncertainty, the final output is the probability prediction value. The 95% confidence interval is calculated as follows: .

[0075] The system is configured to compute SHAP consensus for different subsets of decision trees in parallel across multiple servers and aggregate the final results through a distributed consensus protocol. This architecture enables the system to complete early termination determination in a short time, even for extremely deep and complex models, thus shortening the response time for clinical matching.

[0076] In Embodiment 3, the matching decision and report generation unit is equipped with multi-level ranking logic. In addition to a primary ranking based on predicted probabilities, the system is configured to perform secondary ranking based on dimensions such as donor biosafety scores, historical success rates, and logistics transportation distance. This unit integrates a rule engine, allowing clinical administrators to dynamically adjust the weighting factors of different dimensions according to actual needs. For example, in emergency transplant scenarios, the system is configured to automatically increase the weighting of logistics distance features to ensure that donor samples can reach the recipient's location in the shortest possible time.

[0077] In Example 3, the matching report generated by the matching decision and report generation unit is presented in a dynamic and visual format. The system utilizes a dedicated graphics rendering submodule to present the feature matching degree between the donor and recipient in the form of radar charts and correlation network diagrams. In the normalized feature display, the system can compare the recipient's scores in the three dimensions of niche compatibility, functional complementarity, and host adaptation with the mean of the entire population, helping doctors quickly identify the recipient's individualized shortcomings. In Example 3, the attribution analysis logic is configured to support deep drill-down queries; doctors can click on negative contribution features in the report to view the corresponding original species abundance distribution.

[0078] In the operation and maintenance support aspect of Example 3, the system is configured with end-to-end health monitoring and automatic fault recovery logic. When a microservice instance experiences an anomaly, the monitoring system can detect it in real time and automatically launch a new instance to replace it. The system is also equipped with comprehensive log auditing functions, recording in detail the input, intermediate feature values, and final output results of each matching request, providing detailed evidence for handling medical disputes and secondary development of scientific research data.

[0079] This embodiment, through the introduction of a microservice architecture, enables the system to control the end-to-end matching time to within minutes when processing a massive donor library (e.g., containing more than 10,000 candidate donors). Under stress testing, the system exhibits excellent linear scalability, with the area under the prediction curve remaining at a high level under different load conditions, demonstrating that the architecture, while ensuring high accuracy, possesses engineering usability and commercial potential.

[0080] Example 4: Based on the above examples, this example provides a customized parameter configuration and algorithm optimization scheme for a specific clinical application scenario of inflammatory bowel disease (IBD), aiming to further improve the matching accuracy and clinical benefits of the system in a specific disease area.

[0081] In Embodiment 4, the multidimensional feature engineering processing unit is configured to introduce a specialized feature layer targeting IBD pathological features. This layer constructs a disease-sensitive feature vector by extracting clinical indicators such as fecal calprotectin levels, endoscopic disease activity scores, and the degree of intestinal mucosal damage. The system captures the interaction between pathogenic microorganisms and the receptor inflammatory environment by cross-producting these disease-specific indicators with microbial features.

[0082] The multidimensional feature engineering processing unit is configured to pay particular attention to the tryptophan metabolism pathway, butyrate synthesis pathway, and bile acid hydrolysis conjugated pathway when calculating metabolic pathway complementarity. Since these pathways play a central role in the pathogenesis of IBD, the system is configured to assign higher feature weights to the abundance differences of these key pathways. Specifically, after normalization, the feature scores of the aforementioned pathways are multiplied by an enhancement coefficient between 1.5 and 2.0.

[0083] In Example 4, the integrated learning prediction engine unit introduces time-series prediction logic to address the high recurrence rate of IBD. This unit not only outputs the probability of a single transplant success but is also configured to predict the duration of therapeutic efficacy after transplantation by loading dynamic data from multiple previous follow-up visits of the recipient. Under this logic, the random forest base learner is configured to process extended feature vectors with timestamp information, identifying which donor characteristics are conducive to long-term stable colonization of the recipient's gut microbiota.

[0084] In the algorithm optimization of Example 4, the FISMOTE oversampling logic of the integrated learning prediction engine unit is replaced with cost-sensitive weighted sampling. The system dynamically adjusts the sampling ratio of responders to non-responders based on the severity of the consequences of misjudgment. Since recommending unqualified donors to recipients with severe IBD in clinical practice may pose safety risks, the system is configured to impose extremely high penalty weights on false positive cases misjudged as responses, forcing the model to be more cautious and robust in prediction, and further narrowing the confidence interval of its output.

[0085] In Example 4, the matching decision and report generation unit adds a postoperative management suggestion module. This module automatically generates corresponding dietary recommendations or probiotic support programs based on recipient functional deficiencies identified during the matching prediction process. For example, if attribution analysis shows that the mismatch between the recipient and donor is mainly concentrated in insufficient short-chain fatty acid synthesis capacity, the system will suggest appropriately increasing the intake of fermentable fiber after transplantation to synergistically improve the colonization efficiency of the donor microbiota.

[0086] At the hardware deployment level, Example 4 employs a highly reliable dual-machine hot standby architecture and is configured with an encrypted storage area for the IBD-specific disease database. By interfacing with the hospital's intranet PACS system, the system can automatically acquire the recipient's endoscopic image reports, extract disease severity keywords using a built-in natural language processing engine, and convert them into structured data input into the feature engineering unit.

[0087] This embodiment, through in-depth optimization for the IBD field, achieves a high level of predictive AUC for the matching success rate of the system in recurrent Crohn's disease and ulcerative colitis. In the correlation analysis with clinical gold standards (such as Mayo score improvement), the matching score output by the system showed a positive correlation, improving the clinical remission rate and quality of life of IBD patients after receiving FMT treatment.

[0088] Example 5: Based on all the foregoing examples, this example describes a closed-loop matching prediction system with self-learning and continuous evolution capabilities, focusing on its data feedback and automated model iteration mechanism.

[0089] In embodiment 5, the system adds a data feedback and annotation subsystem. This subsystem establishes a two-way communication channel with the clinical terminal and is configured to automatically collect the actual clinical performance of recipients after receiving transplantation from a system-recommended donor. Clinicians provide feedback on the recipient's clinical response status (e.g., complete remission, partial remission, ineffectiveness) at 4, 12, and 24 weeks post-transplantation through a preset interface.

[0090] In Embodiment 5, the data access unit adds a result feedback receiving interface, which has timeliness verification and multi-source mutual verification logic. The system is configured to only label a sample as a high-quality tagged sample and store it in the incremental learning library when the conclusions from the doctor's feedback, endoscopy report, and biomarker detection (such as calprotectin) are consistent.

[0091] In Embodiment 5, the multidimensional feature engineering processing unit is configured to have dynamic feature selection capabilities. As the amount of refluxed data increases, this unit dynamically evaluates the effectiveness of the feature set by calculating the mutual information value between each feature and the actual efficacy label. For features that exhibit low discriminative power in long-term clinical practice, the system automatically reduces their weight in subsequent calculations or removes them in the next round of model iteration, thus achieving automatic pruning of the feature space.

[0092] The integrated learning prediction engine unit in Embodiment 5 integrates an automated machine learning pipeline. Whenever the number of new labeled samples in the incremental learning library reaches a preset number (e.g., 50), the system automatically starts a background training task. This task is configured to use a hyperparameter optimization algorithm to automatically find the optimal number of decision trees, maximum depth, and splitting criterion within a preset range. The newly trained model is evaluated on an independent shadow validation set. Only when the new model's AUC, accuracy, and F1 score all surpass the currently running online model will the system perform an online hot update of the model weights through a smooth release mechanism.

[0093] The integrated learning prediction engine unit is also configured with a knowledge forgetting protection mechanism when processing incremental data. While learning the features of new samples, the system retains the model's memory of the basic biological laws contained in the early core samples (i.e., the initial 169 pairs of real clinical samples) by introducing a regularization term into the loss function, preventing drastic fluctuations in model performance caused by local data distribution fluctuations.

[0094] In Example 5, the matching decision and report generation unit adds a model evolution transparency display function. In each matching report, the system indicates the currently used model version, the training sample size, and the clinical validation accuracy of that version in the most recent quarter. This transparent design enhances medical institutions' trust in the AI ​​decision-making system. The unit is also equipped with a research analysis engine, allowing researchers to access historical system data to perform multi-dimensional correlation mining, such as analyzing the performance differences of donors from different seasons and regions in specific recipient populations, providing scientific evidence for clinical research.

[0095] In terms of security and robustness, Example 5 introduces anomaly detection logic. The system monitors the distribution characteristics of the returned data in real time. If it detects a drift between the newly received data and the previous distribution (e.g., a situation where the clinical manifestations of a group do not match the predictions), the system will immediately issue an alert, suspend the automatic model iteration logic, lock the current stable version, and wait for review and intervention by human experts.

[0096] Through this continuously evolving closed-loop mechanism, the system described in Example 5, after one year of operation, possesses self-healing and self-reinforcing properties that enable it to adapt to the ever-changing progress in microbial research and the evolution of clinical diagnostic and treatment standards, thus becoming a vibrant medical auxiliary decision-making platform.

[0097] Example 6: This example provides a containerized, highly scalable hardware implementation architecture to support the physical operation of all the above units.

[0098] A donor-recipient matching prediction system based on multi-source heterogeneous data fusion, the hardware infrastructure of which includes: High-performance computing node cluster: Each computing node is configured with at least two multi-core central processing units with a clock speed of no less than 3.0 GHz and no less than 256 GB of synchronous dynamic random access memory. This cluster is connected via a high-speed interconnect network that supports remote memory access technology and is used to perform massively parallel computing tasks of multi-dimensional feature engineering processing units and random forest inference tasks of integrated learning prediction engine units.

[0099] Dedicated hardware acceleration cards: At least two GPU-based computing acceleration cards are inserted into each compute node, each with at least 16 GB of video memory. These acceleration cards are configured to accelerate the parallel traversal logic of decision trees and the multiplication operations of high-dimensional matrices through the CUDA or OpenCL framework, thereby reducing prediction response latency.

[0100] Distributed storage system: A distributed file storage system built with an all-flash array, with a total capacity of no less than 500 TB and supporting multiple redundant backups. This system is used to store massive amounts of raw metagenomic data, preprocessed intermediate feature vectors, and encrypted recipient-donor matching history reports.

[0101] Load balancing device: Located at the system entry point, it uses a high-speed packet processing chip at the hardware level to distribute tasks to each microservice instance of the data access unit, ensuring that the system can maintain a low-latency response when faced with high concurrency access.

[0102] Security Encryption Module: The system is equipped with a hardware security module used to perform encryption and decryption operations on recipient privacy data, digital signature verification, and identity authentication. All data flows entering and leaving the system must pass through this module in real time for auditing.

[0103] Backup and Disaster Recovery Center: A backup data center with a synchronized mirror is established in a remote location, achieving real-time data synchronization via a dedicated fiber optic link. In the event of an unforeseen failure in the primary data center, the backup center can take over all business requests within minutes, ensuring system business continuity.

[0104] Supported by the hardware architecture of this embodiment, the system can process over 10,000 matching prediction requests per day. The average time for each matching task, from raw data input to the final generation of a structured report, is controlled within 30 seconds, and the system maintains high availability throughout the year. This industrial-grade hardware implementation provides a solid material guarantee for the large-scale commercial deployment of this invention.

[0105] Those skilled in the art should understand that although the system structure and processing logic of the present invention have been described in detail in the above embodiments, in practical applications, the functional limits, parameter configurations, and implementation methods of each module can be reasonably adjusted and equivalently replaced according to different clinical needs and hardware environments. For example, the random forest algorithm can be replaced with an ensemble learning framework with similar performance, such as gradient boosting trees or extreme gradient boosting trees. As long as it follows the core logic of multi-scale feature fusion and interpretable reasoning proposed in this invention, it should be included within the protection scope of this invention.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A donor-recipient matching prediction system based on multi-source heterogeneous data fusion, characterized in that, include: The data access unit is used to receive multi-source heterogeneous biomedical data from recipients and candidate donors. The multi-source heterogeneous biomedical data includes standardized metagenomic analysis results, host basic phenotypic data, and historical transplant efficacy records. A multi-dimensional feature engineering processing unit, connected to the data access unit, is used to perform domain knowledge-driven transformation and normalization processing on the accessed raw data, transforming the multi-source heterogeneous biomedical data into feature vectors with unified dimensions, and constructing a multi-scale feature space that includes niche compatibility features, functional complementarity features, and host adaptation features. An integrated learning prediction engine unit is connected to the multi-dimensional feature engineering processing unit. It is loaded with pre-trained optimized random forest model weights to perform nonlinear inference on the feature vector, calculate the probability prediction value of the donor and recipient reaching the primary clinical endpoint within the predetermined follow-up period, and simultaneously output the influence weight contribution of each feature on the prediction result. The matching decision and report generation unit, connected to the integrated learning prediction engine unit, is used to sort all candidate donors in descending order according to the probability prediction value, filter out the donor set that is higher than the preset probability threshold, and generate a structured matching report containing matching scores and failure attribution analysis.

2. The donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The data access unit includes a data format verification subunit, which performs protocol consistency checks on the accessed raw biomedical data to ensure that the metagenomic analysis results conform to the predetermined taxonomic unit level labeling specifications. The data cleaning and noise reduction subunit is used to filter out sample data with a sequencing depth lower than a preset depth threshold or a contamination rate higher than a preset contamination threshold. The missing value multi-imputation subunit is used to fill in missing host phenotypic indicators with numerical values ​​through multivariate imputation logic based on chain equations, so as to maintain the integrity of the feature space. The data privacy desensitization subunit is used to perform desensitization processing on fields involving host-sensitive information using differential privacy algorithms; The data access unit communicates with the clinical terminal through an asynchronous message queue structure to concurrently receive diversity indices, overlap indicators, pathway abundance matrices, and microbial interaction network parameters contained in the metagenomic analysis results.

3. The donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The niche compatibility features constructed by the multidimensional feature engineering processing unit include: microbial community richness and evenness difference index, which is used to characterize the ecological stability of gut microbiota. The microbial community richness is obtained by calculating the Shannon index of the taxonomic unit. The calculation logic is as follows: the relative abundance of each taxonomic unit is accumulated by multiplying its natural logarithm, and the negative of the sum is taken. The overall community structure difference index is used to measure basic compatibility. Its calculation logic is as follows: the sum of the absolute values ​​of the abundance differences between the receptor and the candidate donor at each taxonomic level is calculated, and the sum of the absolute values ​​is divided by the total abundance to obtain the corresponding ratio. The niche overlap index is used to quantify the colonization feasibility of donor strains in the recipient's gut microenvironment. Its generation logic is as follows: calculate the sum of the product of the resource utilization probabilities of the donor and recipient at each metabolic site, and divide the sum of the product of the resource utilization probabilities by the geometric mean of the sum of the squares of the resource utilization probabilities of the donor and recipient.

4. The donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The complementary features constructed by the multi-dimensional feature engineering processing unit include: Metabolic pathway complementarity index is used to quantify the donor’s ability to supplement key metabolic functions missing in the receptor; the key metabolic functions include short-chain fatty acid synthesis pathway and bile acid conversion pathway. Microbial network compatibility index is used to reflect the degree to which the donor microbial community is accepted by the recipient's original network by calculating the crossover and union ratio of synergistic and antagonistic relationships between the donor microbial community network and the recipient's in-situ microbial community network. Probiotic targeting indicators are used to count the number of predetermined core bacterial genera present in the donor and absent in the recipient. The multidimensional feature engineering processing unit, by real-time association with a preset metabolic pathway database, retrieves the metabolic functions lacking in the receptor and extracts the enhancement degree of the corresponding metabolic enzyme abundance in the donor, thereby generating the metabolic pathway complementarity index.

5. The donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 4, characterized in that, The host adaptation features constructed by the multidimensional feature engineering processing unit include: gender matching status, which is encoded as a binary variable. If the recipient and the donor are the same gender, it is encoded as a first logical value; if they are different, it is encoded as a second logical value. The absolute value of the difference in body mass index is used to reflect the physical differences between hosts at the level of nutritional metabolism; the absolute value of the difference in age is used to reflect the degree of adaptation between hosts at the level of physiological cycle. The gender matching status, the absolute value of the difference in body mass index, and the absolute value of the difference in age together constitute the host adaptation dimension component in the feature vector.

6. The donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 5, characterized in that, The normalization process performed by the multi-dimensional feature engineering processing unit adopts the following differential transformation strategy: For continuous features, linear normalization logic is used to subtract the minimum observation value in the predetermined training set from the original observation value, and then divide by the difference between the maximum and minimum observation values, thereby mapping the values ​​to a closed interval between 0 and 1. For count-type features reflecting the number of probiotics, logarithmic transformation logic is used. The logarithm to base 10 is calculated by adding 1 to the original number to mitigate the long-tail effect caused by high abundance data. For binary features including gender, the original logic value is kept unchanged.

7. The donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 6, characterized in that, The optimized random forest model in the integrated learning prediction engine unit consists of a preset number of decision trees as base learners; The number of decision trees is set to 500, the maximum depth of each decision tree is limited to 12, and the minimum number of leaf node samples is set to 5. During the model training phase, the ensemble learning prediction engine unit adopts a feature importance-guided oversampling strategy. The logic is as follows: first, the contribution of each feature to the classification decision is calculated; then, based on the weight ratio determined by the contribution of the Gini coefficient, linear interpolation in the feature space is performed on the clinical responders who are minority class samples to synthesize virtual samples to balance the sample class distribution.

8. The donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 7, characterized in that, The integrated learning prediction engine unit incorporates an early stopping mechanism and uncertainty calibration function during the inference process, specifically including: The early stopping mechanism is used to calculate the Shapley sum and explanatory value of each feature output by each decision tree in real time during the model iteration process; When the average variation of the Shapley sum and explanatory value of all features is less than a preset threshold in five consecutive iterations, the inference calculation will automatically stop and the current ensemble result will be output. The uncertainty calibration function is used to calculate the variance of all decision tree outputs and, in conjunction with the error distribution pattern based on out-of-bag samples, to calculate the corresponding confidence interval for the probability prediction value.

9. A donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 8, characterized in that, The matching decision and report generation unit performs the following logic when performing matching failure attribution analysis: By tracing the Shapley sum and explanation values ​​output by the integrated learning prediction engine unit, key features with negative values ​​and absolute values ​​greater than a preset impact threshold are identified. Based on the feature sub-layer to which the key features belong, determine the core biological factors that led to the matching failure; The core biological factors include the absence of key bacterial genera, metabolic function conflicts, and host phenotype incompatibility.

10. A donor-recipient matching prediction system based on multi-source heterogeneous data fusion according to claim 9, characterized in that, The system also includes a distributed processing module deployed on local edge nodes of medical institutions, and a coordination engine deployed on a cloud central server; The data access unit and the multidimensional feature engineering processing unit are deployed on the edge node and are used to perform parsing and feature extraction of raw metagenomic sequencing data locally. The multidimensional feature engineering processing unit is used to download globally dynamically updated normalized parameters from the coordination engine of the cloud central server to ensure that the feature vector has consistent dimensions across different medical institutions. The integrated learning prediction engine unit is used to adopt a federated learning architecture to achieve the co-evolution of optimizing the weights of the random forest model by uploading the updated model parameters to the coordination engine without disclosing the local raw data.

Images