Machine learning-based post-transplant organ rejection monitoring system
Patent Information
- Application Number
- PCT/CN2026/089573
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-08-05
- Filing Date
- 2026-04-09
- Publication Date
- 2026-09-17
Smart Images
Figure CN2026089573_17092026_PF_FP_ABST
Abstract
Description
Machine Learning-Based Organ Transplant Rejection Monitoring System Technical Field
[0001] This invention relates to the field of post-organ transplant physiological data monitoring technology, specifically to a machine learning-based post-organ transplant rejection monitoring system. Background Technology
[0002] Post-organ transplant rejection monitoring utilizes advanced machine learning algorithms to analyze data such as biomarkers, clinical data, and imaging information of transplant patients to identify rejection reactions early and make accurate predictions. Through automated data processing and pattern recognition, it can more effectively monitor the occurrence of rejection reactions, thereby providing decision support for clinicians, optimizing immunosuppressive therapy regimens, and reducing the risk of rejection reactions.
[0003] However, traditional organ transplant rejection monitoring systems often rely on single-modality data, leading to the loss of key pathological signals, failure to reflect the systemic mechanisms of rejection, lack of temporal continuity in risk assessment, neglect of indicator evolution patterns, delay in identifying progressive rejection, and reliance on subjective threshold settings, resulting in poor interpretability and difficulty in generating stratified risk labels to support differentiated intervention strategies. Furthermore, traditional organ transplant rejection monitoring systems suffer from the following technical problems: the static architecture of traditional models cannot integrate dynamic biological signals such as physiological rhythms, leading to sluggish responses to rhythm-related rejection features; they lack dedicated structures designed for the bioelectrical and immune dual mechanisms of organ rejection, making it difficult to distinguish between rejection and ordinary postoperative complications; and traditional single-output frameworks only provide binary classification results, failing to simultaneously provide real-time warnings and risk evolution predictions. Summary of the Invention
[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a machine learning-based post-organ transplant rejection monitoring system. Traditional post-organ transplant rejection monitoring systems typically rely on single-modality data, leading to the loss of key pathological signals, failure to reflect the systemic mechanisms of rejection, lack of temporal continuity in risk assessment, neglect of indicator evolution patterns, delay in identifying progressive rejection, reliance on subjective threshold settings, poor interpretability, and difficulty in generating hierarchical risk labels to support differentiated intervention strategies. This solution creatively employs a multi-modal feature fusion module to enhance the dataset. It models the essential relationships between features across different dimensions through hierarchical embedding and utilizes graph networks to quantify the strength of interactions between layers, solving the problem of semantic fragmentation in multi-source data and avoiding misjudgments caused by isolated analysis of a single indicator. Simultaneously, it introduces time-constrained clustering to generate temporally continuous risk level labels, enhancing risk... The temporal continuity of the assessment provides an interpretable and intuitive basis for subsequent monitoring. Addressing the limitations of traditional organ transplant rejection monitoring systems—namely, the static architecture of conventional models fails to integrate dynamic biological signals such as physiological rhythms, resulting in sluggish responses to rhythm-related rejection features; the lack of dedicated structures for the bioelectrical and immune mechanisms of organ rejection makes it difficult to distinguish between rejection and ordinary postoperative complications; and the fact that traditional single-output frameworks only provide binary classification results, failing to simultaneously offer real-time warnings and risk evolution predictions—this solution creatively employs a physiological enhancement transformer model as the rejection monitoring model. It embeds physiological rhythms into the feature space through oscillatory coding and designs a dual-pathway network to simulate abnormal organ electrophysiological conduction and inflammatory cascade responses, respectively. Simultaneously, it uses dual output heads to synchronously predict immediate risks and short-term evolution trends, enhancing the model's ability to fit biomedical mechanisms and its predictive accuracy.
[0005] The technical solution adopted by the present invention is as follows: The postoperative rejection monitoring system based on machine learning provided by the present invention includes a postoperative data acquisition module, a postoperative data optimization module, a multimodal feature fusion module, a rejection monitoring model construction module, and a postoperative rejection monitoring module;
[0006] The postoperative data acquisition module collects data to obtain a raw postoperative monitoring dataset and sends the raw postoperative monitoring dataset to the postoperative data optimization module.
[0007] The postoperative data optimization module uses data cleaning and imputation, multi-source temporal alignment, feature standardization and dynamic feature enhancement to obtain past monitoring datasets and real-time monitoring datasets, and sends the past monitoring datasets and the real-time monitoring datasets to the multimodal feature fusion module.
[0008] The multimodal feature fusion module is used to fuse and obtain the comprehensive postoperative risk features of patients and perform preliminary classification. By integrating multi-source data and performing hierarchical interactive modeling, a real-time enhanced dataset, a monitoring training set, and a monitoring test set are obtained. The real-time enhanced dataset is sent to the postoperative rejection monitoring module, and the monitoring training set and the monitoring test set are sent to the rejection monitoring model construction module.
[0009] The rejection monitoring model construction module constructs a physiological enhancement transformer model as a rejection monitoring model and sends the rejection monitoring model to the postoperative rejection monitoring module.
[0010] The postoperative rejection monitoring module specifically uses the rejection monitoring model to monitor postoperative rejection after organ transplantation and obtain postoperative monitoring reference results.
[0011] Furthermore, in the postoperative data acquisition module, the postoperative monitoring raw dataset specifically includes the past postoperative monitoring raw dataset and the current postoperative monitoring raw dataset. Both the past postoperative monitoring raw dataset and the current postoperative monitoring raw dataset include basic physiological data, immune status data, drug metabolism data, and organ damage data.
[0012] Furthermore, in the postoperative data optimization module, the data cleaning and imputation are used to ensure data integrity and reliability, specifically by processing missing and outlier values through medical rationality verification and statistical imputation methods.
[0013] The multi-source time alignment is used to eliminate sampling time offset. Specifically, cubic spline interpolation is used to map data of different frequencies to a unified time grid with a time grid interval of 3 hours.
[0014] The feature standardization is used to eliminate dimensional differences, specifically by standardizing the original data using the Z-score standardization method.
[0015] The dynamic feature enhancement is used to capture the dynamic change pattern of indicators. Specifically, it calculates the 6-hour change rate and 24-hour change acceleration of each indicator in the original data and combines them with the original values.
[0016] By performing data cleaning and imputation, multi-source temporal alignment, feature standardization, and dynamic feature enhancement, the past original dataset and the current original dataset of postoperative monitoring are optimized to obtain the past monitoring dataset and the real-time monitoring dataset.
[0017] Furthermore, in the multimodal feature fusion module, the patient's postoperative comprehensive risk characteristics are fused and preliminarily classified. By integrating multi-source data, hierarchical interactive modeling is performed, and preliminary classification is carried out by combining time-constrained clustering to obtain risk level labels, risk change characteristics, and postoperative comprehensive risk characteristics of the patient with temporal continuity. This provides interpretable input for subsequent postoperative rejection monitoring after organ transplantation. The risk level labels specifically include low risk, medium risk, high risk, and emergency status.
[0018] The multimodal feature fusion module specifically includes hierarchical embedding interaction, hierarchical graph network design, time-constrained clustering design, and dataset augmentation processing.
[0019] The hierarchical embedding interaction is used to extract the essence of pathological features at each level. Specifically, it processes the features at each level through an adaptive activation function encoder and a multi-head attention mechanism to obtain high-level features at each level, including:
[0020] Feature stratification is used for semantic grouping of clinical features. Specifically, based on basic physiological data, immune status data, drug metabolism data, and organ damage data, it is divided into four levels: physiological level, immune level, drug level, and damage level, to obtain the basic features of each level.
[0021] Autoencoder dimensionality reduction is used for feature compression and dimensionality reduction. Specifically, it processes the basic features at each level through a two-layer encoding network activated by an adaptive ReLU activation function to obtain the dimensionality-reduced features at each level.
[0022] Multi-head attention interaction is used to capture the correlation of features within a hierarchy. Specifically, it calculates the correlation weights within each hierarchy through a multi-head attention mechanism to obtain the interaction features of each hierarchy.
[0023] Residual fusion is used to preserve the original feature information. Specifically, it fuses dimensionality-reduced features and interactive features through learnable weighted residual connections to obtain high-level features at various levels.
[0024] The hierarchical graph network design is used to model the pathological association mechanism between levels. Specifically, it models the clinical associations at each level by designing a gated graph neural network to obtain the patient's comprehensive postoperative risk characteristics, including:
[0025] Graph structure construction is used to define the interaction framework between levels. Specifically, high-level features of four levels are used as graph nodes, and a fully connected topology is established to obtain the clinical relationship graph structure.
[0026] Edge weights are obtained to quantify the strength of interactions between layers. Specifically, edge weights are dynamically calculated using feature cosine similarity and learnable parameters to obtain the interaction matrix between layers.
[0027] Gated node updates are used for cross-level information fusion. Specifically, the current level node status is updated by combining the status of neighboring nodes with the gated loop unit.
[0028] Cross-level feature fusion is used for feature integration output. Specifically, it involves splicing high-level features from each level along the feature axis and the updated final node state to obtain the patient's comprehensive postoperative risk characteristics.
[0029] The time-constrained clustering design is used to generate risk level labels. Specifically, it analyzes feature evolution patterns within a sliding window using time-constrained clustering to obtain risk level labels and risk change characteristics, including:
[0030] Sliding window construction is used to construct a time series analysis window, specifically by extracting the feature sequence of the patient's postoperative comprehensive risk characteristics for the previous 7 days at the current time step to obtain a local time series unit;
[0031] Time-constrained K-means clustering is used for risk pattern recognition. Specifically, it employs a K-means clustering algorithm with added time continuity constraints to cluster all local temporal units of each sample within a sliding window. The clustering label of the local temporal unit of the sample at the last time step within the sliding window is used as the risk level label of the sample. The time continuity constraint is used to constrain the smooth change of the clustering labels of local temporal units in adjacent time steps.
[0032] To obtain risk change characteristics, which are used to quantify the dynamic changes in risk, specifically, the risk difference between the current time step and previous time steps is calculated, and a vector is constructed to obtain the risk change characteristics.
[0033] The dataset augmentation process specifically involves taking the past monitoring dataset and the real-time monitoring dataset as inputs to the multimodal feature fusion module, and processing them through the hierarchical embedding interaction, the hierarchical inter-graph network design, and the time-constrained clustering design to obtain the past augmented dataset and the real-time augmented dataset. The past augmented dataset is then split into a monitoring training set and a monitoring test set.
[0034] Furthermore, in the rejection monitoring model construction module, a model is constructed to build the model required for post-organ transplant rejection monitoring, specifically a physiological enhancement transformer model, which serves as the rejection monitoring model.
[0035] The rejection monitoring model construction module specifically includes oscillation coding design, improved attention design, improved feedforward network design, risk prediction output, and model construction and training.
[0036] The oscillation coding design, used for feature space alignment, specifically involves fusing risk level labels, risk change features, and the patient's postoperative comprehensive risk features, followed by oscillation position coding to obtain oscillation coding features, the contents of which include:
[0037] Risk feature embedding involves transforming risk level labels through an embedding layer, linearly transforming and encoding risk change features, and then fusing them with the patient's postoperative comprehensive risk features to obtain comprehensive basic features.
[0038] Oscillatory position coding is used to add physiological time information. Specifically, it generates an oscillatory signal using a sine function through a vector composed of heart rate and respiratory rate, and obtains oscillatory coding features based on the oscillatory signal.
[0039] The improved attention design, used for adaptive feature enhancement, specifically adjusts the attention output by generating gating weights through a vector composed of heart rate and respiratory rate to obtain physiological enhancement features, including:
[0040] Rhythm gating generation is used for dynamic attention modulation. Specifically, it transforms the vector composed of heart rate and respiratory rate into gating values in the interval [0,1] through a fully connected layer to obtain rhythm gating weights.
[0041] Gated multi-head attention is used for risk feature focusing. Specifically, it uses a multi-head attention mechanism to process oscillatory coding features, and fuses the output features of the multi-head attention mechanism and the oscillatory coding features through rhythmic gating weights to obtain the gated multi-head attention output features.
[0042] Residual normalization specifically involves processing the gated multi-head attention output features through residual connections and layer normalization to maintain a stable feature distribution and obtain physiological enhancement features.
[0043] The improved feedforward network design is used to extract rejection-specific patterns. Specifically, it separates organ electrophysiological characteristics and inflammatory response characteristics through a dual-pathway structure and performs gated fusion to obtain risk response enhancement features, including:
[0044] Organ electrophysiological pathway design to simulate abnormal physiological electrical signal transduction, specifically by extracting features through a custom bioelectric activation function and depthwise separable convolution to obtain organ electrophysiological features;
[0045] Inflammatory response pathway design, used to characterize immune inflammatory responses, specifically by converting physiological enhancement features into inflammatory response features through a custom cytokine response activation function;
[0046] Gated fusion, used for feature fusion, specifically involves dynamically weighting organ electrophysiological features and inflammatory response features through a gating mechanism to obtain pathway fusion features;
[0047] Risk perception enhancement is used to strengthen risk perception capabilities. Specifically, risk level labels are embedded as bias injection pathway features to obtain risk response enhancement features.
[0048] The risk prediction output is used to output the model prediction results, specifically generating an immediate warning probability and a risk probability for the next three days using a dual-classification head. The content includes:
[0049] To obtain the instant warning probability for acute rejection diagnosis, a custom immune response activation function is used to process the risk response enhancement feature, and then combined with the sigmoid activation function to obtain the instant warning probability.
[0050] The probability of risk over the next three days is obtained to predict the evolution of risk. Specifically, the probability of risk over the next three days is predicted recursively through loop units.
[0051] The construction and training of the model specifically involves constructing a physiological enhancement transformer model through the oscillation coding design, the improved attention design, the improved feedforward network design, and the risk prediction output. The model is then trained and its performance is verified based on the monitoring training set and the monitoring test set to obtain the physiological enhancement transformer model, which serves as the rejection monitoring model.
[0052] Furthermore, in the postoperative rejection monitoring module, the real-time enhanced dataset is used as the input of the rejection monitoring model to monitor postoperative rejection after organ transplantation, obtain postoperative monitoring reference results, compare them with a preset risk warning threshold, select appropriate medical decisions based on the comparison results, and complete postoperative rejection monitoring after organ transplantation. The postoperative monitoring reference results specifically include the immediate warning probability output by the rejection monitoring model and the risk probability for the next three days.
[0053] The beneficial effects achieved by the present invention using the above solution are as follows:
[0054] (1) In view of the technical problems of traditional organ transplant post-transplant rejection monitoring systems, which usually rely on single-modality data, resulting in the loss of key pathological signals, failure to reflect the systemic mechanism of rejection, lack of temporal continuity in risk assessment, neglect of indicator evolution patterns, delay in the identification of progressive rejection, and reliance on subjective threshold setting, poor interpretability, and difficulty in generating hierarchical risk labels to support differentiated intervention strategies, this solution creatively adopts a multimodal feature fusion module to enhance the dataset. It models the essential correlation of features in each dimension through hierarchical embedding and uses graph networks to quantify the strength of interaction between hierarchical levels, solving the problem of semantic fragmentation of multi-source data and avoiding misjudgment caused by isolated analysis of a certain indicator. At the same time, it introduces time-constrained clustering to generate temporally continuous risk level labels, enhances the temporal continuity of risk assessment, and provides interpretable and intuitive evidence for subsequent monitoring.
[0055] (2) To address the technical problems of traditional organ transplant rejection monitoring systems, such as the inability of the static architecture of the traditional model to integrate dynamic biological signals such as physiological rhythms, resulting in a sluggish response to rhythm-related rejection features, the lack of a dedicated structure for the bioelectrical and immune dual mechanisms of organ rejection, making it difficult to distinguish between rejection and ordinary postoperative complications, and the fact that the traditional single-output framework only provides binary classification results and cannot simultaneously provide real-time warnings and risk evolution predictions, this solution creatively adopts a physiological enhancement transformer model as the rejection monitoring model. By embedding physiological rhythms into the feature space through oscillatory coding, a dual-pathway network is designed to simulate abnormal organ electrophysiological conduction and inflammatory cascade responses, respectively. At the same time, the dual output heads simultaneously predict real-time risks and short-term evolution trends, enhancing the model's ability to fit biomedical mechanisms and its prediction accuracy. Attached Figure Description
[0056] Figure 1 is a schematic diagram of the modules of the machine learning-based organ transplant post-transplant rejection monitoring system provided by the present invention;
[0057] Figure 2 is a flowchart of the postoperative data optimization module;
[0058] Figure 3 is a flowchart of the multimodal feature fusion module;
[0059] Figure 4 is a flowchart illustrating the construction module of the rejection monitoring model.
[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0062] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0063] Example 1, referring to Figure 1, the machine learning-based organ transplant post-transplant rejection monitoring system provided by the present invention includes a post-operative data acquisition module, a post-operative data optimization module, a multi-modal feature fusion module, a rejection monitoring model construction module, and a post-operative rejection monitoring module;
[0064] The postoperative data acquisition module collects data to obtain a raw postoperative monitoring dataset and sends the raw postoperative monitoring dataset to the postoperative data optimization module.
[0065] The postoperative data optimization module uses data cleaning and imputation, multi-source temporal alignment, feature standardization and dynamic feature enhancement to obtain past monitoring datasets and real-time monitoring datasets, and sends the past monitoring datasets and the real-time monitoring datasets to the multimodal feature fusion module.
[0066] The multimodal feature fusion module is used to fuse and obtain the comprehensive postoperative risk features of patients and perform preliminary classification. By integrating multi-source data and performing hierarchical interactive modeling, a real-time enhanced dataset, a monitoring training set, and a monitoring test set are obtained. The real-time enhanced dataset is sent to the postoperative rejection monitoring module, and the monitoring training set and the monitoring test set are sent to the rejection monitoring model construction module.
[0067] The rejection monitoring model construction module constructs a physiological enhancement transformer model as a rejection monitoring model and sends the rejection monitoring model to the postoperative rejection monitoring module.
[0068] The postoperative rejection monitoring module specifically uses the rejection monitoring model to monitor postoperative rejection after organ transplantation and obtain postoperative monitoring reference results.
[0069] Example 2, referring to Figure 1, is based on the above example. In the postoperative data acquisition module, the postoperative monitoring raw dataset specifically includes the past postoperative monitoring raw dataset and the current postoperative monitoring raw dataset. Both the past postoperative monitoring raw dataset and the current postoperative monitoring raw dataset include basic physiological data, immune status data, drug metabolism data, and organ damage data.
[0070] The basic physiological data specifically include postoperative body temperature data, postoperative blood pressure data, postoperative blood oxygen saturation data, postoperative respiratory rate waveform data, and postoperative electrocardiogram data.
[0071] The immune status data specifically includes postoperative T cell subset count and ratio data, postoperative cytokine concentration data, postoperative serum immunoglobulin detection data, and postoperative donor-specific antibody detection data.
[0072] The drug metabolism data specifically includes postoperative immunosuppressant blood concentration data, postoperative glucocorticoid dosage data, and postoperative metabolic enzyme genotype data.
[0073] The organ damage data specifically includes postoperative serum creatinine test data, postoperative transplant organ-related protease test data, postoperative transplant organ resistance index data, and tissue biopsy data.
[0074] Example 3, referring to Figures 1 and 2, is based on the above examples. In the postoperative data optimization module, the data cleaning and imputation are used to ensure data integrity and reliability. Specifically, missing values and outliers are processed through medical rationality verification and statistical imputation methods.
[0075] The multi-source time alignment is used to eliminate sampling time offset. Specifically, cubic spline interpolation is used to map data of different frequencies to a unified time grid with a time grid interval of 3 hours.
[0076] The feature standardization is used to eliminate dimensional differences, specifically by standardizing the original data using the Z-score standardization method.
[0077] The dynamic feature enhancement is used to capture the dynamic change pattern of indicators. Specifically, it calculates the 6-hour change rate and 24-hour change acceleration of each indicator in the original data and combines them with the original values.
[0078] By performing data cleaning and imputation, multi-source temporal alignment, feature standardization, and dynamic feature enhancement, the past original dataset and the current original dataset of postoperative monitoring are optimized to obtain the past monitoring dataset and the real-time monitoring dataset.
[0079] Example 4, referring to Figures 1 and 3, is based on the above examples. In the multimodal feature fusion module, it is used to fuse and obtain the comprehensive postoperative risk characteristics of patients and perform preliminary classification. By integrating multi-source data, hierarchical interactive modeling is performed, and preliminary classification is performed by combining time-constrained clustering to obtain risk level labels, risk change characteristics, and comprehensive postoperative risk characteristics of patients with temporal continuity. This provides interpretable input for subsequent postoperative rejection monitoring after organ transplantation. The risk level labels specifically include low risk, medium risk, high risk, and emergency status.
[0080] The multimodal feature fusion module specifically includes hierarchical embedding interaction, hierarchical graph network design, time-constrained clustering design, and dataset augmentation processing.
[0081] The hierarchical embedding interaction is used to extract the essence of pathological features at each level. Specifically, it processes the features at each level through an adaptive activation function encoder and a multi-head attention mechanism to obtain high-level features at each level, including:
[0082] Feature stratification is used for semantic grouping of clinical features. Specifically, based on basic physiological data, immune status data, drug metabolism data, and organ damage data, it is divided into four levels: physiological level, immune level, drug level, and damage level, to obtain the basic features of each level.
[0083] Autoencoder dimensionality reduction, used for feature compression and dimensionality reduction, specifically involves processing the basic features at each level through a two-layer encoding network activated by an adaptive ReLU activation function to obtain the dimensionality-reduced features at each level. The formula used is as follows:
[0084] In the formula, AReLU(·) represents the adaptive ReLU activation function, x represents the input independent variable, ad represents the learnable negative interval slope, and Fe a ′ represents the dimensionality reduction feature at the a-th level, Fe a This represents the basic characteristics of the a-th level. This represents the weights of the second layer of the coding network at level a. This represents the bias term of the second layer coding network of the a-th type. This represents the weight of the first layer of the encoding network in the a-th type. This represents the bias term of the first layer coding network of the a-th level;
[0085] Multi-head attention interaction is used to capture feature correlations within a hierarchy. Specifically, it calculates the correlation weights within each hierarchy using a multi-head attention mechanism to obtain the interaction features of each hierarchy. The formula used is as follows:
[0086] In the formula, Att h This represents the output of the h-th attention head, and softmax(·) represents the softmax activation function. This represents the query transformation weight of the h-th attention head at level a. This represents the key transformation weight of the h-th attention head at level a. This represents the weight transformation of the h-th attention head at level a. This represents the dimension of the key in the h-th attention head, and concat(·) represents the concatenation operation function. Let Att1 represent the interaction feature at level a, and Att2 represent the output of the first attention head. H This represents the output of the Hth attention head, where H represents the total number of attention heads, and T represents the transpose operation.
[0087] Residual fusion, used to preserve original feature information, specifically involves fusing dimensionality-reduced features and interactive features through learnable weighted residual connections to obtain high-level features at various levels. The formula used is as follows:
[0088] In the formula, Let represent the high-level features of the a-th level, and le represent the learnable residual fusion weights;
[0089] The hierarchical graph network design is used to model the pathological association mechanism between levels. Specifically, it models the clinical associations at each level by designing a gated graph neural network to obtain the patient's comprehensive postoperative risk characteristics, including:
[0090] Graph structure construction is used to define the interaction framework between levels. Specifically, high-level features of four levels are used as graph nodes, and a fully connected topology is established to obtain the clinical relationship graph structure.
[0091] Edge weights are obtained to quantify the strength of interactions between layers. Specifically, edge weights are dynamically calculated using feature cosine similarity and learnable parameters to obtain the inter-layer interaction matrix. The formula used is as follows:
[0092] In the formula, ew ab The elements of the inter-level interaction matrix represent the edge weights between level a and level b. `sigmoid(·)` represents the sigmoid activation function, and `Cos(·)` represents the cosine similarity calculation function. Let represent the mean of the high-level features at level a. W represents the mean of the high-level features at level b. ed pi represents the weight used to calculate the edge weights. ed This represents the bias term of the weight matrix used to calculate the edge weights;
[0093] Gated node updates are used for cross-level information fusion. Specifically, the current level node state is updated by combining the state of neighboring nodes with the state of the gated loop unit. The formula used is as follows:
[0094] In the formula, This represents the neighbor aggregation feature of the a-th node at the current time step. Let a represent the set of neighboring nodes of the a-th node. This represents the node state of the b-th node at the previous time step, and GRU(·) represents the gated cyclic unit function. This represents the node state of the a-th node at the current time step;
[0095] Cross-level feature fusion is used for feature integration output. Specifically, it involves concatenating high-level features from each level along the feature axis, along with the updated final node state, to obtain the patient's comprehensive postoperative risk characteristics. The formula used is as follows:
[0096] In the formula, Fe fu This indicates the patient's overall postoperative risk characteristics. Indicates higher-level physiological characteristics. Indicates high-level features of the immune system. Indicates a high-level characteristic of the drug. Indicates high-level features of damage. Represents the final node state at the physiological level. Indicates the final node state of the immune hierarchy. Indicates the final node state at the drug level. Indicates the final node state of the damage level;
[0097] The time-constrained clustering design is used to generate risk level labels. Specifically, it analyzes feature evolution patterns within a sliding window using time-constrained clustering to obtain risk level labels and risk change characteristics, including:
[0098] Sliding window construction is used to construct a time series analysis window, specifically by extracting the feature sequence of the patient's postoperative comprehensive risk characteristics for the previous 7 days at the current time step to obtain a local time series unit;
[0099] Temporally constrained K-means clustering is used for risk pattern recognition. Specifically, it employs a K-means clustering algorithm with added temporal continuity constraints. For each sample, all local temporal units within a sliding window are clustered. The cluster label of the local temporal unit at the last time step within the sliding window is used as the risk level label for that sample. The temporal continuity constraint is used to ensure the smooth change of cluster labels of local temporal units between adjacent time steps. The formula used is as follows:
[0100] In the formula, Ls time This represents the time-continuous constraint loss value, where tm represents the current time step, which is the last time step of the sample within the sliding window, and m and n represent the indices of different time steps within the sliding window, and fe m Ce represents the local temporal unit of a sample at time step m. a Let λ represent the a-th cluster center. time This represents the time constraint regularization parameter. If(lab) n ≠lab n-1() represents an indicator function whose value is 1 when the cluster label of a sample in a local time unit at time step n is the same as the cluster label of a sample in a local time unit at time step n-1; otherwise, its value is 0. n This represents the clustering label of a sample's local temporal unit at time step n. n-1 Let ||·|| represent the clustering label of the local temporal unit of the sample at time step n-1, and ||·|| represent the calculation of the L2 norm;
[0101] To obtain risk change characteristics for quantifying dynamic risk changes, specifically, the risk difference between the current time step and previous time steps is calculated, and a vector is constructed to obtain the risk change characteristics, which are represented as follows:
[0102] In the formula, lab tm The risk level label represents the clustering label of the local temporal unit of the sample at the current time step. tm-6 This represents the clustering label of a sample's local temporal unit at time step tm-6, lab tm-12 This represents the clustering label of a sample's local temporal unit at time step tm-12, lab tm-24 Fe represents the clustering label of the local temporal unit of the sample at time step tm-24. RC Indicates the characteristics of risk changes;
[0103] The dataset augmentation process specifically involves taking the past monitoring dataset and the real-time monitoring dataset as inputs to the multimodal feature fusion module, and processing them through the hierarchical embedding interaction, the hierarchical inter-graph network design, and the time-constrained clustering design to obtain the past augmented dataset and the real-time augmented dataset. The past augmented dataset is then split into a monitoring training set and a monitoring test set.
[0104] By performing the above operations, this solution addresses the technical problems of traditional organ transplant post-transplant rejection monitoring systems, which typically rely on single-modality data, leading to the loss of key pathological signals, failure to reflect the systemic mechanisms of rejection, lack of temporal continuity in risk assessment, neglect of indicator evolution patterns, delay in identifying progressive rejection, reliance on subjective threshold settings, poor interpretability, and difficulty in generating hierarchical risk labels to support differentiated intervention strategies. This solution creatively employs a multimodal feature fusion module to enhance the dataset. It models the essential relationships between features across different dimensions through hierarchical embedding and utilizes graph networks to quantify the strength of interactions between layers, solving the problem of semantic fragmentation in multi-source data and avoiding misjudgments caused by isolated analysis of a single indicator. Simultaneously, it introduces time-constrained clustering to generate temporally continuous risk level labels, enhancing the temporal continuity of risk assessment and providing interpretable and intuitive evidence for subsequent monitoring.
[0105] Example 5, referring to Figures 1 and 4, is based on the above examples. In the rejection monitoring model construction module, a model is constructed to build the model required for post-organ transplant rejection monitoring. Specifically, a physiological enhancement transformer model is constructed as the rejection monitoring model.
[0106] The rejection monitoring model construction module specifically includes oscillation coding design, improved attention design, improved feedforward network design, risk prediction output, and model construction and training.
[0107] The aforementioned oscillation coding design, used for feature space alignment, specifically involves fusing risk level labels, risk change features, and the patient's comprehensive postoperative risk features, followed by oscillation position coding to obtain oscillation coding features, the contents of which include:
[0108] Risk feature embedding specifically involves transforming risk level labels through an embedding layer, linearly transforming and encoding risk change features, and then fusing them with the patient's comprehensive postoperative risk features to obtain comprehensive basic features. The formula used is as follows: Fe base =concat(Fe fu EB(lab) tm ), W linear Fe RC );
[0109] In the formula, Fe base Represents the comprehensive basic features, EB(·) represents the embedding layer function, and W linear Represents the weights of the linear transformation;
[0110] Oscillatory position coding, used to add physiological timing information, specifically generates an oscillatory signal using a sine function from a vector composed of heart rate and respiratory rate, and obtains oscillatory coding features based on the oscillatory signal. The formula used is as follows:
[0111] In the formula, Pe represents the oscillation signal, fr represents the physiological frequency value, hr represents the vector composed of heart rate and respiratory rate, Ef represents the oscillation coding feature, and W Pe Indicates the weight of the oscillation encoding;
[0112] The improved attention design, used for adaptive feature attention enhancement, specifically adjusts the attention output by generating gating weights through a vector composed of heart rate and respiratory rate to obtain physiological enhancement features, including:
[0113] Rhythm gating generation, used for dynamic attention modulation, specifically involves transforming a vector composed of heart rate and respiratory rate into gating values within the interval [0,1] using a fully connected layer to obtain rhythm gating weights. The formula used is as follows: Bg=sigmoid(W bg ·hr);
[0114] In the formula, Bg represents the rhythm gating weight, and W bg This represents the weight matrix used to calculate the rhythm gating weights;
[0115] Gated multi-head attention is used for risk feature focusing. Specifically, it uses a multi-head attention mechanism to process oscillatory coding features. By fusing the output features of the multi-head attention mechanism and the oscillatory coding features through rhythmic gating weights, the gated multi-head attention output features are obtained. The formula used is as follows: GA=Bg⊙MHA(Ef)+(1-Bg)☉Ef;
[0116] In the formula, GA represents the gated multi-head attention output feature, MHA(·) represents the multi-head attention mechanism function, and ⊙ represents the element-wise multiplication operation;
[0117] Residual normalization, specifically, involves processing the gated multi-head attention output features through residual connections and layer normalization to maintain a stable feature distribution, resulting in physiological enhancement features. The formula used is as follows: Fe Bio =LN(GA+Ef);
[0118] In the formula, represents the physiological enhancement feature, and LN(·) represents the layer normalization function;
[0119] The improved feedforward network design is used to extract rejection-specific patterns. Specifically, it separates organ electrophysiological characteristics and inflammatory response characteristics through a dual-pathway structure and performs gated fusion to obtain risk response enhancement features, including:
[0120] Organ electrophysiological pathway design, used to simulate abnormal physiological electrical signal transduction, specifically involves extracting features using a custom bioelectrical activation function and depthwise separable convolution to obtain organ electrophysiological features. The formula used is as follows:
[0121] In the formula, BEA(·) represents the bioelectric activation function, Fe El Represents the organ's electrophysiological characteristics, DWCov(·) represents the depthwise separable convolution function, W El Indicates the weights used in organ electrophysiological calculations;
[0122] Inflammatory response pathways were designed to characterize the immune inflammatory response. Specifically, this involved converting physiological enhancement features using a custom cytokine response activation function to obtain inflammatory response characteristics, as shown in the following formula:
[0123] In the formula, CRA(·) represents the cytokine response activation function, and ρ represents the activation threshold. Fe represents the negative feedback adjustment coefficient. In Indicating inflammatory response characteristics, W In Indicates the weights used to calculate the inflammatory response;
[0124] Gated fusion, used for feature fusion, specifically involves dynamically weighting organ electrophysiological features and inflammatory response features through a gating mechanism to obtain pathway fusion features. The formula used is as follows:
[0125] In the formula, Gw represents the path fusion gating weight, and W gate Fe represents the weight matrix used to calculate the path fusion gating weights. Pfu Indicates pathway fusion characteristics;
[0126] Risk perception enhancement, used to strengthen risk perception capabilities, specifically involves embedding risk level labels as features in a bias injection pathway to obtain enhanced risk response features. The formula used is as follows: Fe risk =Fe Pfu +W risk ·EB(lab tm );
[0127] In the formula, Fe risk Indicating enhanced risk response characteristics, W risk This indicates the embedding weight of the risk level label;
[0128] The risk prediction output is used to output the model prediction results, specifically generating an immediate warning probability and a risk probability for the next three days using a dual-classification head. The content includes:
[0129] To obtain the immediate warning probability for acute rejection diagnosis, a custom immune response activation function is used to process the risk response enhancement features, combined with a sigmoid activation function, to obtain the immediate warning probability. The formula used is as follows:
[0130] In the formula, IA(·) represents the immune response activation function, and λ IA Let τ represent the immune response threshold regulation coefficient, τ represent the immune response intensity regulation coefficient, tanh(·) represent the hyperbolic tangent function, and Po represent the immune response threshold regulation coefficient. ins The probability of an immediate warning is represented by α, the weight of the immune response is represented by β, and the risk level weight is represented by W. IA Indicates the weights of the linear transformation of the immune response;
[0131] The probability of risk over the next three days is obtained to predict risk evolution. Specifically, the probability of risk over the next three days is predicted recursively using loop units, and the formula used is as follows:
[0132] In the formula, s0 represents the basic risk prediction, and W s s represents the basic prediction weights. q+1 W represents the risk forecast for day q+1. pre pi represents the weighting of future risk predictions.pre s represents the bias term for future risk prediction. q Po represents the risk forecast for day q. pre This represents the risk probability for the next three days, where s1 represents the risk prediction for the first day, s2 represents the risk prediction for the second day, and s3 represents the risk prediction for the third day.
[0133] The construction and training of the model specifically involves constructing a physiological enhancement transformer model through the oscillation coding design, the improved attention design, the improved feedforward network design, and the risk prediction output. The model is then trained and its performance is verified based on the monitoring training set and the monitoring test set to obtain the physiological enhancement transformer model, which serves as the rejection monitoring model.
[0134] By performing the above operations, this solution addresses the technical problems of traditional organ transplant rejection monitoring systems. Traditional models, with their static architectures, cannot integrate dynamic biological signals such as physiological rhythms, resulting in sluggish responses to rhythm-related rejection features. Furthermore, they lack dedicated structures designed for the bioelectrical and immune mechanisms of organ rejection, making it difficult to distinguish between rejection and ordinary postoperative complications. Additionally, traditional single-output frameworks only provide binary classification results, failing to simultaneously offer real-time warnings and risk evolution predictions. This solution creatively employs a physiological enhancement transformer model as the rejection monitoring model. It embeds physiological rhythms into the feature space through oscillatory coding and designs a dual-pathway network to simulate abnormal organ electrophysiological conduction and inflammatory cascade responses, respectively. Simultaneously, it uses dual output heads to synchronously predict immediate risks and short-term evolution trends, enhancing the model's ability to fit biomedical mechanisms and its predictive accuracy.
[0135] Example 6, referring to Figure 1, is based on the above examples. In the postoperative rejection monitoring module, the real-time enhanced dataset is used as the input of the rejection monitoring model to monitor postoperative rejection after organ transplantation. Postoperative monitoring reference results are obtained and compared with a preset risk warning threshold. Based on the comparison results, an appropriate medical decision is selected to complete the postoperative rejection monitoring of organ transplantation. The postoperative monitoring reference results specifically include the immediate warning probability and the risk probability for the next three days output by the rejection monitoring model.
[0136] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0137] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0138] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A machine learning-based post-organ transplant rejection monitoring system, characterized in that: The system includes a postoperative data acquisition module, a postoperative data optimization module, a multimodal feature fusion module, an rejection monitoring model construction module, and a postoperative rejection monitoring module; The postoperative data acquisition module obtains the postoperative monitoring raw dataset by collecting data. The postoperative monitoring raw dataset specifically includes the past postoperative monitoring raw dataset and the current postoperative monitoring raw dataset. The postoperative data optimization module uses data cleaning and imputation, multi-source temporal alignment, feature standardization and dynamic feature enhancement to obtain past monitoring datasets and real-time monitoring datasets. The multimodal feature fusion module is used to fuse and obtain the comprehensive postoperative risk features of patients and perform preliminary classification. By integrating multi-source data, hierarchical interactive modeling is performed, and preliminary classification is performed by combining time-constrained clustering to obtain risk level labels, risk change features, and comprehensive postoperative risk features of patients with temporal continuity. This provides interpretable input for subsequent postoperative rejection monitoring after organ transplantation. The risk level labels specifically include low risk, medium risk, high risk, and emergency status. The rejection monitoring model construction module is used to construct the model required for post-organ transplant rejection monitoring, specifically to construct a physiological enhancement transformer model as a rejection monitoring model. The postoperative rejection monitoring module specifically uses the real-time enhanced dataset as input to the rejection monitoring model to monitor postoperative rejection after organ transplantation, obtains postoperative monitoring reference results, compares them with a preset risk warning threshold, selects appropriate medical decisions based on the comparison results, and completes postoperative rejection monitoring after organ transplantation. The postoperative monitoring reference results specifically include the immediate warning probability and the risk probability for the next three days output by the rejection monitoring model.
2. The machine learning-based organ transplant rejection monitoring system according to claim 1, characterized in that: The multimodal feature fusion module specifically includes hierarchical embedding interaction, hierarchical graph network design, time-constrained clustering design, and dataset augmentation processing.
3. The machine learning-based organ transplant rejection monitoring system according to claim 2, characterized in that: The hierarchical embedding interaction is used to extract the essence of pathological features at each level. Specifically, it processes the features at each level through an adaptive activation function encoder and a multi-head attention mechanism to obtain high-level features at each level, including: Feature stratification is used for semantic grouping of clinical features. Specifically, based on basic physiological data, immune status data, drug metabolism data, and organ damage data, it is divided into four levels: physiological level, immune level, drug level, and damage level, to obtain the basic features of each level. Autoencoder dimensionality reduction is used for feature compression and dimensionality reduction. Specifically, it processes the basic features at each level through a two-layer encoding network activated by an adaptive ReLU activation function to obtain the dimensionality-reduced features at each level. Multi-head attention interaction is used to capture the correlation of features within a hierarchy. Specifically, it calculates the correlation weights within each hierarchy through a multi-head attention mechanism to obtain the interaction features of each hierarchy. Residual fusion is used to preserve the original feature information. Specifically, it fuses dimensionality-reduced features and interactive features through learnable weighted residual connections to obtain high-level features at various levels. The hierarchical graph network design is used to model the pathological association mechanism between levels. Specifically, it models the clinical associations at each level by designing a gated graph neural network to obtain the patient's comprehensive postoperative risk characteristics, including: Graph structure construction is used to define the interaction framework between levels. Specifically, high-level features of four levels are used as graph nodes, and a fully connected topology is established to obtain the clinical relationship graph structure. Edge weights are obtained to quantify the strength of interactions between layers. Specifically, edge weights are dynamically calculated using feature cosine similarity and learnable parameters to obtain the interaction matrix between layers. Gated node updates are used for cross-level information fusion. Specifically, the current level node status is updated by combining the status of neighboring nodes with the gated loop unit. Cross-level feature fusion is used for feature integration output. Specifically, it involves splicing high-level features from each level along the feature axis and the updated final node state to obtain the patient's comprehensive postoperative risk characteristics. The time-constrained clustering design is used to generate risk level labels. Specifically, it analyzes feature evolution patterns within a sliding window using time-constrained clustering to obtain risk level labels and risk change characteristics, including: Sliding window construction is used to construct a time series analysis window, specifically by extracting the feature sequence of the patient's postoperative comprehensive risk characteristics for the previous 7 days at the current time step to obtain a local time series unit; Time-constrained K-means clustering is used for risk pattern recognition. Specifically, it employs a K-means clustering algorithm with added time continuity constraints to cluster all local temporal units of each sample within a sliding window. The clustering label of the local temporal unit of the sample at the last time step within the sliding window is used as the risk level label of the sample. The time continuity constraint is used to constrain the smooth change of the clustering labels of local temporal units in adjacent time steps. To obtain risk change characteristics, which are used to quantify the dynamic changes in risk, specifically, the risk difference between the current time step and previous time steps is calculated, and a vector is constructed to obtain the risk change characteristics. The dataset augmentation process specifically involves taking the past monitoring dataset and the real-time monitoring dataset as inputs to the multimodal feature fusion module, and processing them through the hierarchical embedding interaction, the hierarchical inter-graph network design, and the time-constrained clustering design to obtain the past augmented dataset and the real-time augmented dataset. The past augmented dataset is then split into a monitoring training set and a monitoring test set.
4. The machine learning-based organ transplant rejection monitoring system according to claim 1, characterized in that: The rejection monitoring model construction module specifically includes oscillation coding design, improved attention design, improved feedforward network design, risk prediction output, and model construction and training.
5. The machine learning-based organ transplant rejection monitoring system according to claim 4, characterized in that: The aforementioned oscillation coding design, used for feature space alignment, specifically involves fusing risk level labels, risk change features, and the patient's comprehensive postoperative risk features, followed by oscillation position coding to obtain oscillation coding features, the contents of which include: Risk feature embedding involves transforming risk level labels through an embedding layer, linearly transforming and encoding risk change features, and then fusing them with the patient's postoperative comprehensive risk features to obtain comprehensive basic features. Oscillatory position coding is used to add physiological time information. Specifically, it generates an oscillatory signal using a sine function through a vector composed of heart rate and respiratory rate, and obtains oscillatory coding features based on the oscillatory signal. The improved attention design, used for adaptive feature attention enhancement, specifically adjusts the attention output by generating gating weights through a vector composed of heart rate and respiratory rate to obtain physiological enhancement features, including: Rhythm gating generation is used for dynamic attention modulation. Specifically, it transforms the vector composed of heart rate and respiratory rate into gating values in the interval [0,1] through a fully connected layer to obtain rhythm gating weights. Gated multi-head attention is used for risk feature focusing. Specifically, it uses a multi-head attention mechanism to process oscillatory coding features, and fuses the output features of the multi-head attention mechanism and oscillatory coding features through rhythmic gating weights to obtain the gated multi-head attention output features. Residual normalization specifically involves processing the gated multi-head attention output features through residual connections and layer normalization to maintain a stable feature distribution and obtain physiological enhancement features. The improved feedforward network design is used to extract rejection-specific patterns. Specifically, it separates organ electrophysiological characteristics and inflammatory response characteristics through a dual-pathway structure and performs gated fusion to obtain risk response enhancement features, including: Organ electrophysiological pathway design to simulate abnormal physiological electrical signal transduction, specifically by extracting features through a custom bioelectric activation function and depthwise separable convolution to obtain organ electrophysiological features; Inflammatory response pathway design, used to characterize immune inflammatory responses, specifically by converting physiological enhancement features into inflammatory response features through a custom cytokine response activation function; Gated fusion, used for feature fusion, specifically involves dynamically weighting organ electrophysiological features and inflammatory response features through a gating mechanism to obtain pathway fusion features; Risk perception enhancement is used to strengthen risk perception capabilities. Specifically, risk level labels are embedded as bias injection pathway features to obtain risk response enhancement features. The risk prediction output is used to output the model prediction results, specifically generating an immediate warning probability and a risk probability for the next three days using a dual-classification head. The content includes: To obtain the instant warning probability for acute rejection diagnosis, a custom immune response activation function is used to process the risk response enhancement feature, and then combined with the sigmoid activation function to obtain the instant warning probability. The probability of risk over the next three days is obtained to predict the evolution of risk. Specifically, the probability of risk over the next three days is predicted recursively through loop units. The construction and training of the model specifically involves constructing a physiological enhancement transformer model through the oscillation coding design, the improved attention design, the improved feedforward network design, and the risk prediction output. The model is then trained and its performance is verified based on the monitoring training set and the monitoring test set to obtain the physiological enhancement transformer model, which serves as the rejection monitoring model.
6. The machine learning-based organ transplant rejection monitoring system according to claim 1, characterized in that: Both the previous raw dataset for postoperative monitoring and the current raw dataset for postoperative monitoring include basic physiological data, immune status data, drug metabolism data, and organ damage data.
7. The machine learning-based organ transplant rejection monitoring system according to claim 1, characterized in that: In the postoperative data optimization module, the data cleaning and imputation are used to ensure data integrity and reliability. Specifically, missing and outlier values are processed through medical rationality verification and statistical imputation methods. The multi-source time alignment is used to eliminate sampling time offset. Specifically, cubic spline interpolation is used to map data of different frequencies to a unified time grid with a time grid interval of 3 hours. The feature standardization is used to eliminate dimensional differences, specifically by standardizing the original data using the Z-score standardization method. The dynamic feature enhancement is used to capture the dynamic change pattern of indicators. Specifically, it calculates the 6-hour change rate and 24-hour change acceleration of each indicator in the original data and combines them with the original values. By performing data cleaning and imputation, multi-source temporal alignment, feature standardization, and dynamic feature enhancement, the past original dataset and the current original dataset of postoperative monitoring are optimized to obtain the past monitoring dataset and the real-time monitoring dataset.