Intelligent evaluation system for endometrial receptivity based on biosensor data processing

Abstract

This invention relates to the field of intelligent information processing technology and discloses an intelligent assessment system for endometrial receptivity based on biosensor data processing. The system acquires signals of multiple types of biomarkers in bodily fluid samples through multi-channel biosensor detection and performs structured preprocessing and quantitative transformation. Based on this, it establishes object-level associations between biosensor data and multi-source information to construct a unified feature data system. Furthermore, it constructs a weighted heterogeneous assessment graph and introduces a manifold stochastic feature modeling mechanism to generate manifold feature data. The manifold features and original features are jointly input into a shared-parameter cyclic Transformer model. Through cyclic inference and consistency constraints, multi-round feature fusion is achieved, outputting the endometrial receptivity assessment result. This invention improves the multi-source data modeling capability and the stability and consistency of the assessment process.

Description

Technical Field

[0001] This invention relates to the field of intelligent information processing technology, and in particular to an intelligent assessment system for endometrial receptivity based on biosensor data processing. Background Technology

[0002] Endometrial receptivity is an important indicator reflecting the state of the endometrium during a specific physiological cycle, and its changes are closely related to the expression of various biomarkers, fluctuations in endocrine levels, and the state of the local microenvironment. Current technologies for obtaining information related to endometrial receptivity mainly rely on ultrasound imaging parameters, hormone level detection, and some molecular marker detection methods. These methods analyze indicators such as endometrial thickness, morphology, and blood flow signals to reflect the endometrial condition. However, these technologies are mostly based on single-source or limited-dimensional data, lacking the ability to uniformly express and systematically integrate multi-source biological information, making it difficult to characterize complex multi-factor coupling relationships. With the development of biosensing technology, multi-channel detection methods based on microfluidic structures can achieve parallel detection of protein biomarkers, nucleic acid molecules, and metabolites, thus providing richer multimodal data sources. However, existing systems lack a unified coding and correlation modeling mechanism between biosensor data and clinically relevant data, and the structural relationships between the data are not effectively utilized. Meanwhile, existing data processing methods mostly employ rule-driven or simple statistical analysis approaches, which have limited ability to model nonlinear relationships and time-series dependent features in high-dimensional, multimodal data, making it difficult to achieve refined characterization of complex state features. Therefore, there is an urgent need for a technical solution that integrates biosensing detection and computer information processing to achieve structured expression and efficient modeling of multi-source data on endometrial receptivity. Summary of the Invention

[0003] This invention addresses the problems in existing technologies, such as the difficulty in unified modeling of multi-source biological information, insufficient ability to express complex feature relationships, and lack of stable reasoning mechanisms in the evaluation process. It proposes an intelligent assessment system for endometrial receptivity based on biosensor data processing. By constructing a multimodal data collaborative modeling framework, it achieves a high-dimensional structured expression of endometrial receptivity. This technology uses multi-channel biosensor detection data as the core input and introduces a manifold stochastic feature modeling mechanism. Heterogeneous data is approximated to a continuous manifold space through random walk propagation and kernel function, thereby characterizing multi-order relationships between nodes. Based on this, a Transformer model with a shared parameter cyclic structure is constructed, and cyclic reasoning is used to... The trajectory modeling and trajectory condition modulation mechanism enables round-by-round refinement and dynamic control of feature representation. At the same time, a multi-trajectory consistency constraint optimization strategy is introduced to improve the representation stability and inference consistency of the model at different computational depths. During execution, the multi-channel biosensor signals are first structured and the object-level alignment of multi-source data is completed. Then, a weighted heterogeneous evaluation graph is constructed and a manifold random feature representation is generated. The manifold features and the original features are then jointly input into a recurrent Transformer model for multi-round iterative inference and cross-modal fusion. Finally, a structured endometrial receptivity evaluation result is output, thereby improving the multimodal data fusion capability, complex relationship modeling capability, and evaluation stability.

[0004] This invention provides an intelligent assessment system for endometrial receptivity based on biosensor data processing, deployed in an environment including outpatient sampling terminals, laboratory point-of-care testing terminals, in-hospital edge computing terminals, in-hospital server terminals, and regional collaborative cloud environments. The system includes:

[0005] The biosensing detection module is deployed at the outpatient sampling terminal and the laboratory point-of-care testing terminal to acquire multi-channel raw sensor signals;

[0006] The sensor preprocessing module, deployed at the edge computing terminal within the hospital, is used to amplify, filter, perform baseline correction, drift compensation, time alignment, noise suppression, anomaly removal, and preliminary quantitative transformation on multi-channel raw sensor signals to generate structured biosensor data.

[0007] The information fusion module, deployed on the hospital's server, connects the hospital information system, laboratory information system, electronic medical record system, ultrasound examination terminal, and endocrine testing terminal. It is used to receive the age information, menstrual cycle information, previous assisted reproductive information, hormone level information, metabolic index information, inflammatory index information, and endometrial ultrasound phenotype information of the examinee, and perform unified indexing and object-level association with the structured biosensor data to construct endometrial receptivity characteristic data.

[0008] The manifold feature module, deployed on the hospital's server, constructs a weighted heterogeneous evaluation graph. A manifold stochastic feature modeling mechanism is introduced to perform manifold stochastic feature modeling processing based on random walk propagation and kernel function approximation on the weighted heterogeneous evaluation graph, generating manifold stochastic feature data to characterize the local manifold structure of endometrial receptivity and cross-object similarity relationships.

[0009] The endometrial receptivity assessment module is deployed on the hospital's server and in a regional collaborative cloud environment. A Transform model based on a multi-head self-attention mechanism and a feedforward network structure is established. A shared parameter loop mechanism is introduced to transform the fixed-level stacked structure of the Transform model into a reusable loop computation structure. A trajectory condition modulation mechanism based on normalized time-position parameters and step size parameters is introduced to dynamically scale and gate the normalized layers and residual connections in each iteration of the Transform model. A multi-trajectory consistency constraint optimization mechanism is also introduced to apply alignment constraints to the model output under different loop budget conditions, thereby optimizing the representation consistency and inference stability of the Transform model during multiple iterations and constructing a Looped-Transform model with elastic deep inference capabilities. Manifold random feature data and endometrial receptivity feature data are input into the Looped-Transform model to output the endometrial receptivity assessment results of the examined subjects.

[0010] The decision output module, deployed on the hospital's server, connects the clinical workstation and the assisted reproductive diagnosis and treatment terminal. It is used to generate suggestions on the timing of embryo transfer, re-examination, adjustment of endometrial preparation plan, and risk warnings based on the endometrial receptivity assessment results.

[0011] Furthermore, a manifold stochastic feature modeling mechanism is introduced to perform manifold stochastic feature modeling processing based on random walk propagation and kernel function approximation on the weighted heterogeneous evaluation graph, generating manifold stochastic feature data. The specific process includes the following:

[0012] Step S1: Perform object parsing and primitive processing on the endometrial receptivity characteristic data to construct a weighted heterogeneous evaluation map;

[0013] Step S2: Based on the weighted adjacency matrix of the weighted heterogeneous evaluation graph, construct a graph kernel expansion system to characterize the multi-order propagation relationship between node objects; unify the first-order direct association, second-order indirect propagation association and high-order cross-path diffusion association between nodes into a power series superposition form of the graph kernel matrix, and configure the corresponding kernel coefficient sequence according to the difference in the contribution of different propagation orders to node similarity.

[0014] Step S3: Introduce a modulation function to adjust the contribution of random walk path length and establish a mapping relationship between the modulation function and the kernel coefficient sequence; based on the mapping relationship, perform deconvolution of the modulation function to obtain a path length modulation function that matches the target graph kernel; at the same time, configure random walk start rules, neighborhood transfer rules, load update rules and termination determination rules for each node in the weighted heterogeneous evaluation graph.

[0015] Step S4: Based on the random walk initiation rule, neighborhood transition rule, load update rule, and termination determination rule, each node in the weighted heterogeneous evaluation graph is used as the starting node, and multiple random walk samplings are performed. During each random walk, when the random walk path visits the current node, the component of the discrete signature vector corresponding to the starting node at the current node's position is updated. The load value corresponding to the current random walk path is combined with the output value of the path length modulation function to form the current node's access contribution, and the current node's access contribution is accumulated and written into the discrete signature vector corresponding to the starting node. After completing the node jump, the load value is updated according to the current node's degree value, edge weight value, and preset normalization parameters, and the next walk continues until the termination condition is met. The node access contributions obtained from each random walk are aggregated and mean normalized to obtain a set of manifold random feature signature vectors.

[0016] Step S5: Based on the manifold random feature signature vector set, extract the corresponding signature vector for any node object, and perform pairwise matching calculation with the signature vectors of other node objects to obtain the signature response values ​​between each node pair; construct a supervised training sample set with the start node, reference node, and corresponding signature response value in the node pair as the basic unit; construct a continuous mapping function, such that the continuous mapping function takes the start node and reference node in the node pair as input and outputs the corresponding continuous response value; train the parameters of the continuous mapping function by constructing an error loss function between the continuous response value and the signature response value, and performing minimization optimization on the error loss function;

[0017] Step S6: After completing the training of the continuous mapping function, the node objects and their neighborhood extension positions in the weighted heterogeneous evaluation graph are used as the objects to be mapped. The trained continuous mapping function is called to calculate the continuous response value of the objects to be mapped relative to the reference node set. Based on the continuous response value, a target bivariate evaluation function is constructed to characterize the continuous response relationship between any two node objects in the manifold space. A continuous kernel field integral approximation model is introduced to decompose and approximate the target bivariate evaluation function to generate a continuous manifold response field. Based on the continuous manifold response field, manifold random feature data is generated.

[0018] Furthermore, the process of inputting manifold random feature data and endometrial receptivity feature data into the Looped-Transform model to output the endometrial receptivity assessment results of the tested subjects specifically includes the following:

[0019] Step B1: The manifold random feature data and endometrial receptivity feature data are aligned at the object level, registered at the temporal level, and uniformly encoded at the field level to construct a joint input feature sequence corresponding to the tested object; based on the joint input feature sequence, modality identifier, detection cycle identifier, sampling time sequence identifier, and clinical context identifier of each feature item are extracted, and the identifier information is superimposed with the feature embedding results to generate an initial hidden representation for iterative receptivity assessment;

[0020] Step B2: Input the initial hidden representation into the basic computational structure of the Looped-Transform model, which consists of multiple Transformer blocks, and perform parameter sharing processing on each Transformer block in the basic computational structure to construct a reusable shared parameter loop backbone; normalize the iterative evaluation process of the Looped-Transform model as a continuous inference trajectory from the start time to the end time, and regard the shared parameter loop backbone as an evolution operator on the continuous inference trajectory; perform discrete partitioning on the continuous inference trajectory according to the preset maximum number of iterations to obtain the normalized time position parameters and step size parameters for each iteration.

[0021] Step B3: For any given cycle, the normalized time position parameter and step size parameter corresponding to that cycle are input into the trajectory modulation coding unit. Frequency embedding mapping and nonlinear projection processing are performed to generate the trajectory condition representation corresponding to that cycle. Based on the trajectory condition representation, normalized scaling control parameters and residual gating control parameters are generated respectively. The normalized scaling control parameters are applied to the normalization layer in the shared parameter cyclic backbone, and the residual gating control parameters are applied to the residual connection positions of the multi-head self-attention branch and feedforward branch inside the shared parameter cyclic backbone. Structural modulation processing is performed on the shared parameter cyclic backbone to obtain the shared backbone after trajectory modulation.

[0022] Step B4: In each iteration, based on the shared backbone modulated by the trajectory, multi-head self-attention modeling, feedforward nonlinear mapping, and residual update processing are sequentially performed on the current hidden representation to realize cross-modal association modeling between manifold random feature data and endometrial receptivity assessment feature data; and the hidden representation is updated after each iteration, and the intermediate representation results corresponding to each iteration are recorded; when the number of iterations reaches the maximum number of iterations, the final representation result of the corresponding complete cyclic inference process is obtained, and the complete inference path formed from the initial hidden representation to the final representation result is defined as the longest cyclic trajectory;

[0023] Step B5: During the model training phase, perform the complete inference process based on the longest loop trajectory, randomly sample loop paths with fewer than the maximum number of loops as shortcut trajectories, and perform loop inference on the initial hidden representation under different computational budget conditions; calculate the main prediction loss under the longest loop trajectory and the quick prediction loss under the shortcut trajectory respectively, and construct a consistency constraint loss by combining the output results of the longest loop trajectory and the output results of the shortcut trajectory; perform joint optimization by combining the main prediction loss, quick prediction loss and consistency constraint loss to complete the optimized training of the Looped-Transform model parameters, and obtain the trained Looped-Transform model;

[0024] Step B6: Based on the trained Looped-Transform model, output the endometrial receptivity assessment results.

[0025] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0026] This invention constructs a multimodal collaborative modeling framework driven by multichannel biosensor data, which realizes the unified expression and structured fusion of multi-source information related to endometrial receptivity, improves the correlation modeling ability between data from different sources, and solves the problem of lack of unified coding and effective integration between biosensor data and clinically relevant data in the prior art. This enhances the overall characterization of endometrial receptivity and enables the assessment results to more comprehensively reflect the coupling characteristics of multiple factors.

[0027] This invention introduces a manifold stochastic feature modeling mechanism to achieve continuous expression and high-dimensional mapping of multi-order propagation relationships in weighted heterogeneous evaluation graphs. This enhances the characterization ability of complex nonlinear relationships and cross-modal structural dependencies, and solves the problem that traditional methods are unable to characterize the implicit manifold structure and deep associations between nodes in high-dimensional data. As a result, it enhances the ability to identify local structural features of endometrial receptivity and cross-object similarity relationships, making the evaluation process more precise and discriminative.

[0028] This invention constructs a Transformer model with a shared parameter loop structure and introduces a trajectory modulation mechanism and a multi-trajectory consistency constraint optimization strategy. This achieves stable inference and dynamic feature refinement under different computational budget conditions, improves the model's representation consistency and robustness in multiple iterations, and solves the problem of large fluctuations in results when the computational depth changes in existing evaluation methods. This enhances the stability and adaptability of endometrial receptivity evaluation results, making the system more reliable and scalable in practical applications. Attached Figure Description

[0029] Figure 1This is a schematic diagram of the structure of the biosensor detection chip proposed in Embodiment 1 of the present invention;

[0030] Figure 2 This is a flowchart illustrating the Looped-Transform model proposed in Example 5. Detailed Implementation

[0031] 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.

[0032] Example 1, according to Figure 1 This invention provides an intelligent assessment system for endometrial receptivity based on biosensor data processing. It is applied to scenarios such as pre-treatment assessment of assisted reproductive technology (ART) in elderly patients with polycystic ovary syndrome (PCOS), embryo transfer window determination, bedside point-of-care testing, and combined in-hospital and out-of-hospital reproductive medicine assessment. The system is deployed in an environment including outpatient sampling terminals, laboratory point-of-care testing terminals, in-hospital edge computing terminals, in-hospital server terminals, and regional collaborative cloud terminals. The system includes:

[0033] The biosensor detection module, deployed at the outpatient sampling end and the laboratory point-of-care testing end, connects a vaginal secretion sampler, a cervical mucus sampler, an endometrial cavity fluid micro-collection device, a microfluidic sample loading component, and a biosensor detection chip. The vaginal secretion sampler, cervical mucus sampler, and endometrial cavity fluid micro-collection device are used to collect bodily fluid samples from the subjects. The collected bodily fluid samples are quantitatively controlled and guided through the flow path by the microfluidic sample loading component before being loaded onto the biosensor detection chip. The biosensor detection chip is used to perform multi-channel parallel detection of endometrial receptivity-related biomarkers in bodily fluid samples before and after the implantation window corresponding to the subject's menstrual cycle. This includes sample quantitative allocation, detection channel loading, specific recognition response triggering, and signal transduction detection. The testing process generates multi-channel raw sensor signals corresponding to endometrial receptivity biomarkers. These biomarkers include integrin αvβ3, leukemia inhibitory factor LIF, glycodelin, miR-30d, miR-200 family molecules, long non-coding RNAs, inflammatory cytokines, endometrial fluid metabolites, and exosome payload molecules. The corresponding detection channels are selected based on the molecular type of each biomarker. Protein biomarkers are detected through the nucleic acid aptamer recognition channel, nucleic acid biomarkers through the nucleic acid hybridization channel, and metabolite and secretion biomarkers through enzyme reaction channels, colorimetric detection channels, or label-free impedance detection channels. This allows for the generation of combined multi-biomarker detection results within the same testing cycle.

[0034] The biosensor chip is a microfluidic integrated chip, which internally includes a sample introduction unit, a microfluidic distribution unit, a reaction chamber unit, a recognition molecule immobilization unit, a signal transduction unit, and a waste liquid recovery unit. The sample introduction unit introduces externally collected body fluid samples into the chip's internal flow system. The microfluidic distribution unit quantitatively distributes the sample to multiple reaction chamber units according to a preset fluid path, enabling parallel loading of multiple detection channels. The recognition molecule immobilization unit is located on the inner wall of the reaction chamber unit, and its surface is immobilized with antibodies, nucleic acid aptamers, nucleic acid probes, and enzymes, used to target protein markers, nucleic acid markers, metabolite markers, and exosome-related components, respectively. Specific recognition reaction; during the specific recognition process, the target molecule undergoes a binding reaction, hybridization reaction, enzymatic catalytic reaction, or conformational change with the corresponding recognition element, driving a change in the corresponding physical quantity; the signal transduction unit is used to convert the physical quantity change caused by the recognition reaction into a quantifiable output signal, including current change signal, potential change signal, impedance change signal, capacitance change signal, light intensity change signal, chemiluminescence signal, resonant wavelength drift signal, and mass load change signal; the waste liquid recovery unit is used to perform diversion and recovery treatment of the reaction residue after detection to avoid cross-contamination between different detection channels and maintain the stability of the detection environment;

[0035] In this embodiment, the overall structure of the biosensor chip, the connection relationships between its constituent units, and its collaborative working principle are as follows: Figure 1 As shown; Figure 1 This is a schematic diagram of the structure of the biosensor detection chip in an embodiment of the present invention. The biosensor detection chip is a microfluidic integrated structure, including a sample introduction unit, a microfluidic distribution unit, multiple reaction chamber units (reaction chamber unit A, reaction chamber unit B, and reaction chamber unit C), a recognition molecule fixation unit, a signal transduction unit, and a waste liquid recovery unit disposed within each reaction chamber unit. The sample introduction unit is connected to the microfluidic distribution unit and is used to introduce body fluid samples into the internal flow channel system of the chip. The microfluidic distribution unit is connected to each reaction chamber unit and is used to quantitatively distribute the sample to each reaction chamber unit according to a preset fluid path, achieving multi-channel parallel loading. Each reaction chamber unit contains a recognition molecule fixation unit, which is used to perform specific recognition reactions for different types of biomarkers. The output of each reaction chamber unit is connected to the signal transduction unit, which is used to convert the physical quantity changes caused by the recognition reaction into quantifiable signals. The signal transduction unit is further connected to the waste liquid recovery unit, which is used to centrally recover and process the reaction residue during the detection process. Finally, the chip outputs multi-channel raw sensing signals corresponding to each detection channel for subsequent endometrial receptivity assessment.

[0036] The sensor preprocessing module, deployed at the hospital's edge computing endpoint, is used to amplify, filter, perform baseline correction, drift compensation, temporal alignment, noise suppression, anomaly removal, and preliminary quantitative transformation on multi-channel raw sensor signals to generate structured biosensor data. Specifically, it sequentially performs the following actions on the multi-channel raw sensor signals: receiving continuous detection signals output from each sensor channel; performing zero-point correction and baseline normalization on the continuous detection signals based on the reference channel and blank control channel; and performing environmental processing on the continuous detection signals based on temperature, humidity, pH, ionic strength, and sample viscosity parameters. The data processing includes: compensation processing; timing synchronization alignment processing of signals from different sensor channels based on timestamp information; noise suppression and invalid segment removal processing of continuous detection signals based on preset noise thresholds, drift thresholds, and abnormal response modes; extraction of peak amplitude, integral area, phase shift, impedance spectrum parameters, response slope, recovery time constant, and spectral change characteristics of the processed signals based on different sensor channel types; and conversion of the extracted features into the concentration, expression, relative abundance, and response intensity values ​​of corresponding biomarkers based on mapping relationships, thereby forming structured biosensor data.

[0037] The information fusion module, deployed on the hospital's server, connects the hospital information system, laboratory information system, electronic medical record system, ultrasound examination terminal, and endocrine testing terminal. It receives data on the examinee's age, menstrual cycle, past assisted reproductive technology, hormone levels, metabolic indicators, inflammatory indicators, and endometrial ultrasound phenotype. This data is then uniformly indexed and object-level correlated with structured biosensor data to construct endometrial receptivity characteristic data. This endometrial receptivity characteristic data includes biosensor detection features, endocrine hormone characteristics, metabolic state characteristics, and inflammatory immune characteristics. The study considered three main characteristics: endocrine hormone characteristics, ultrasound phenotypic characteristics, and medical history contextual characteristics. Endocrine hormone characteristics included estradiol, progesterone, luteinizing hormone, follicle-stimulating hormone, testosterone, and insulin resistance-related indicators. Metabolic status characteristics included blood glucose, blood lipids, and body mass index-related indicators. Inflammatory and immune characteristics included TNF-α, IL-6, and immune cell ratio-related indicators. Ultrasound phenotypic characteristics included endometrial thickness, endometrial type, endometrial blood perfusion parameters, and uterine artery blood flow parameters. Medical history contextual characteristics included age stratification information, menstrual cycle information, ovulation induction protocol information, and previous embryo transfer outcome information.

[0038] The manifold feature module, deployed on the hospital server, constructs a weighted heterogeneous assessment graph. A manifold stochastic feature modeling mechanism is introduced to perform manifold stochastic feature modeling processing on the weighted heterogeneous assessment graph based on random walk propagation and kernel function approximation, generating manifold stochastic feature data to characterize the local manifold structure of endometrial receptivity and cross-object similarity relationships. The weighted heterogeneous assessment graph includes nodes for the examined object, testing cycle, sample type, biomarker, sensor channel, clinical indicator, and sampling points at different time sequences. The following connections are established between the nodes: a periodic relationship between the examined object node and the testing cycle node; a sampling source relationship between the testing cycle node and the sample type node; a detection correspondence relationship between the sample type node and the biomarker node; a response mapping relationship between the biomarker node and the sensor channel node; a statistical coupling relationship between the biomarker node and the clinical indicator node; and a sequential adjacency relationship between sampling points at different time sequences.

[0039] The tolerance assessment module is deployed on the hospital's server and in a regional collaborative cloud environment. A Transform model based on a multi-head self-attention mechanism and a feedforward network structure is established. A shared parameter loop mechanism is introduced to transform the fixed-level stacked structure of the Transform model into a reusable loop computation structure. A trajectory condition modulation mechanism based on normalized time-position parameters and step size parameters is introduced to perform dynamic scaling and gating control on the normalization layers and residual connections in each loop iteration of the Transform model. A multi-trajectory consistency constraint optimization mechanism is also introduced to perform alignment constraints on the model output under different loop budget conditions to optimize the representation of the Transform model during multiple iterations. To ensure consistency and inference stability, a Looped-Transform model with elastic deep inference capabilities was constructed. Manifold random feature data and endometrial receptivity feature data were jointly input into the Looped-Transform model. Based on a shared parameter cyclic structure, the Looped-Transform model iteratively performs position encoding, self-attention modeling, cross-modal feature interaction, and hierarchical fusion processing through multiple rounds, progressively refining and optimizing the feature representation under different computational budgets. The model outputs the endometrial receptivity assessment results for the examined subjects, including a comprehensive endometrial receptivity score, receptivity level, and implantation window fit status.

[0040] The decision output module, deployed on the hospital's server, connects the clinical workstation and the assisted reproductive diagnosis and treatment terminal. It is used to generate suggestions on the timing of embryo transfer, re-examination, adjustment of endometrial preparation plan, and risk warnings based on the endometrial receptivity assessment results.

[0041] Example 2 differs from Example 1 in that: the manifold feature module introduces a manifold stochastic feature modeling mechanism to perform manifold stochastic feature modeling processing based on random walk propagation and kernel function approximation on the weighted heterogeneous evaluation graph; the manifold feature module differs from Example 1, specifically including the following: deployed on the hospital server to construct the weighted heterogeneous evaluation graph; introducing a node representation learning mechanism based on graph neural networks, using a graph convolutional network (GCN) to perform node embedding calculation on the weighted heterogeneous evaluation graph, and performing node feature matrix and adjacency matrix processing... Multi-layer propagation operations are performed, and neighborhood feature aggregation and linear transformation are executed layer by layer. Combined with nonlinear activation functions, iterative updates of node representations are achieved, thereby completing feature fusion within the multi-layer neighborhood. During the propagation process at each layer, the relationship between nodes is modeled by sharing a parameterized mapping function, so that the node representation gradually integrates its neighborhood structure information and attribute information to obtain the embedding representation vector corresponding to each node. After completing the node embedding calculation, the final embedding representations of each node are collected and organized to generate manifold random feature data for characterizing the local manifold structure of endometrial receptivity and cross-object similarity relationships.

[0042] Example 3 differs from Example 1 in that: the acceptability assessment module constructs a Looped-Transform model with elastic deep reasoning capabilities; manifold random feature data and endometrial acceptability feature data are jointly input into the Looped-Transform model; the acceptability assessment module differs from Example 1, specifically including the following: constructing a Transform model; inputting manifold random feature data and endometrial acceptability feature data into the Transform model, the Transform model models global dependencies of each feature in the input sequence based on a multi-head self-attention mechanism, realizes cross-feature interaction through similarity calculation between query vectors, key vectors, and value vectors, and combines a feedforward nonlinear mapping layer to transform and enhance the feature representation layer by layer; in each layer of the model, residual connections and layer normalization operations maintain stable information propagation, improving the convergence and expressive power of the model during training; after completing multi-layer stacked calculations, the feature representation result fused with global context information is obtained; based on the feature representation result, classification calculation is performed through the output mapping layer to output the endometrial acceptability assessment result of the examined object.

[0043] Example 4, based on Example 1, introduces a manifold stochastic feature modeling mechanism to perform manifold stochastic feature modeling processing on the weighted heterogeneous evaluation graph based on random walk propagation and kernel function approximation, generating manifold stochastic feature data. Specifically, this includes the following:

[0044] Step S1: Perform object parsing and primitive organization on the endometrial receptivity characteristic data, mapping the tested objects, testing cycles, sample types, biomarkers, sensor channels, clinical indicators, and time-series sampling points to different types of nodes in the heterogeneous evaluation graph; establish graph edge connections based on the corresponding associations, time-series adjacency relationships, co-occurrence dependencies, numerical coupling relationships, and cross-modal mapping relationships between node objects, and assign edge weights to each edge based on the association strength, time interval, co-occurrence frequency, numerical correlation, and modal consistency of each edge, thereby constructing a weighted heterogeneous evaluation graph; regard the weighted heterogeneous evaluation graph as a discrete approximation of the distribution manifold of multimodal endometrial receptivity characteristics, with the node set as the discrete sampling points of the manifold, and the edge connections and edge weight structure as the local geometric neighborhood relationships;

[0045] Step S2: Based on the weighted adjacency matrix of the weighted heterogeneous evaluation graph, construct a graph kernel expansion system to characterize the multi-order propagation relationships between node objects; unify the first-order direct association, second-order indirect propagation association, and higher-order cross-path diffusion association between nodes into a power series superposition form of the graph kernel matrix, and configure the corresponding kernel coefficient sequence according to the difference in the contribution of different propagation orders to node similarity; on this basis, construct the graph diffusion propagation basis matrix, so that the graph diffusion propagation basis matrix can characterize the local connectivity effect, multi-step propagation effect, and cross-modal structural coupling effect between node objects in the heterogeneous evaluation graph, thereby forming the graph kernel target expression on which the modeling depends;

[0046] Step S3: Introduce a modulation function to adjust the contribution of random walk path length and establish a mapping relationship between the modulation function and the kernel coefficient sequence; based on the mapping relationship, perform deconvolution of the modulation function to obtain a path length modulation function that matches the target graph kernel; simultaneously, configure random walk start rules, neighborhood transfer rules, load update rules, and termination rules for each node in the weighted heterogeneous evaluation graph. The neighborhood transfer rule is used to select the next node from the set of neighboring nodes of the current node, the load update rule is used to gradually accumulate and normalize the path propagation contribution during the walk, and the termination rule is used to control the distribution of random walk length; through the above configuration, a consistent correspondence is established between the subsequent random walk process at the discrete graph sampling level and the multi-order propagation structure in the target graph kernel.

[0047] Step S4: Based on the random walk initiation rules, neighborhood transition rules, load update rules, and termination judgment rules, each node object in the weighted heterogeneous evaluation graph is used as the starting node, and multiple random walk samplings are performed. During each random walk, when the random walk path visits the current node object, the component of the discrete signature vector corresponding to the starting node at the current node object position is updated. The load value corresponding to the current random walk path is combined with the output value of the path length modulation function to form the current node access contribution, and the current node access contribution is accumulated and written into the discrete signature vector corresponding to the starting node. After completing the node jump, the load value is updated according to the current node degree value, edge weight value, and preset normalization parameters, and the next walk continues until the termination condition is met. The node access contributions obtained from each random walk are aggregated and mean normalized to obtain the manifold random feature signature vector set, thereby forming a graph random feature representation based on Graph Random Features supervision.

[0048] Step S5: Based on the manifold random feature signature vector set, extract the corresponding signature vector for any node object, and perform pairwise matching calculation with the signature vectors of other node objects to obtain the signature response values ​​between each node pair; construct a supervised training sample set with the starting node, reference node and corresponding signature response value in the node pair as the basic unit; construct a continuous mapping function, so that the continuous mapping function takes the starting node and reference node in the node pair as input and outputs the corresponding continuous response value; by constructing an error loss function between the continuous response value and the signature response value, and performing minimization optimization on the error loss function, train the parameters of the continuous mapping function so that the continuous mapping function learns the continuous distribution law of the discrete graph random features in the manifold space, thereby establishing the mapping relationship from discrete graph signature to continuous manifold feature field;

[0049] Step S6: After completing the training of the continuous mapping function, the node objects and their neighborhood extension positions in the weighted heterogeneous evaluation graph are used as the objects to be mapped. The trained continuous mapping function is called to calculate the continuous response value of the objects to be mapped relative to the reference node set. Based on the continuous response value, a target bivariate evaluation function is constructed to characterize the continuous response relationship between any two node objects in the manifold space. A continuous kernel field integral approximation model is introduced to decompose and approximate the target bivariate evaluation function, thereby realizing the continuous modeling and kernel field reconstruction of the multi-order association relationship between node objects and generating a continuous manifold response field. Based on the continuous manifold response field, the local response distribution features in the manifold space and the cross-object response relationship features relative to the reference node set are extracted for any node object. The local response distribution features and the cross-object response relationship features are fused and encoded to generate manifold random feature data to characterize the local manifold structure of endometrial receptivity and the cross-object similarity relationship.

[0050] The continuous kernel field integral approximation model is used to represent the bivariate response relationship between node objects as a kernel function integral form in manifold space. By decomposing the bivariate function through a shared reference position, the relationship between nodes is transformed from a discrete structure expression to a continuous kernel field coupled expression. The target bivariate evaluation function between any two node objects satisfies the following continuous response decomposition form:

[0051] ;

[0052] in, This represents the position of the first node object in the manifold space. This represents the position of the second node object in the manifold space. The objective bivariate evaluation function is used to reflect the local manifold proximity between two node objects; the local manifold proximity between two node objects; the multi-order propagation association strength between two node objects; and the coupling relationship between two node objects in the endometrial receptivity feature space. In manifold space Perform integration on the above; The parameter is Continuous mapping function, This indicates that the continuous mapping function is applied to the node object. Reference position The continuous response values ​​between This indicates that the continuous mapping function is applied to the node object. Reference position The continuous response values ​​between.

[0053] Example 5, according to Figure 2 This embodiment is based on Embodiment 4. In this embodiment, the process of inputting manifold random feature data and endometrial receptivity feature data into the Looped-Transform model and outputting the endometrial receptivity assessment results of the tested subject specifically includes the following:

[0054] Step B1: Feature Alignment and Embedding Construction: Manifold random feature data and endometrial receptivity feature data are aligned at the object level, registered at the temporal level, and uniformly encoded at the field level to construct a joint input feature sequence corresponding to the examined object; based on the joint input feature sequence, modality identifier, detection cycle identifier, sampling time sequence identifier, and clinical context identifier of each feature item are extracted, and the identifier information is superimposed with the feature embedding results to generate an initial hidden representation for iterative receptivity assessment;

[0055] Step B2: Cyclic Backbone Construction and Trajectory Partitioning: The initial hidden representation is input into the basic computational structure of the Looped-Transform model, which consists of multiple Transformer blocks. Parameter sharing processing is performed on each Transformer block in the basic computational structure to construct a reusable shared parameter cyclic backbone. The iterative evaluation process of the Looped-Transform model is normalized and represented as a continuous inference trajectory from the start time to the end time. The shared parameter cyclic backbone is regarded as an evolution operator on the continuous inference trajectory. The continuous inference trajectory is discretized according to the preset maximum number of iterations to obtain the normalized time position parameters and step size parameters of each iteration. This enables each iteration to have explicit trajectory position awareness and inference granularity awareness based on the shared backbone parameters.

[0056] Step B3: Trajectory Modulation and Parameter Control: For any given loop iteration, the normalized time position parameters and step size parameters corresponding to that loop iteration are input into the trajectory modulation coding unit. Frequency embedding mapping and nonlinear projection processing are performed to generate the trajectory condition representation corresponding to that loop iteration. Based on the trajectory condition representation, normalized scaling control parameters and residual gating control parameters are generated respectively. The normalized scaling control parameters are applied to the normalization layer in the shared parameter loop backbone, and the residual gating control parameters are applied to the residual connection positions of the multi-head self-attention branches and feedforward branches inside the shared parameter loop backbone. Structural modulation processing is performed on the shared parameter loop backbone to obtain the shared backbone after trajectory modulation. This allows the same set of shared parameters to form a dynamic computational behavior that matches the current inference stage under different loop iterations.

[0057] The trajectory condition representation satisfies:

[0058] ;

[0059] in, Indicates the first Trajectory condition representation corresponding to cycle. This represents a mapping function for the normalized time position parameter. Indicates the first Normalized time position parameters before the start of the cycle. This represents a mapping function for the step size parameter. Indicates the first The step size parameter corresponding to the cycle;

[0060] Step B4: Looping Inference and Feature Fusion: In each loop, based on the shared backbone modulated by the trajectory, multi-head self-attention modeling, feedforward nonlinear mapping, and residual update processing are sequentially performed on the current hidden representation to realize cross-modal association modeling between manifold random feature data and endometrial receptivity assessment feature data; and the hidden representation is updated after each loop, and the intermediate representation results corresponding to each loop are recorded; when the number of loops reaches the maximum number of loops, the final representation result of the corresponding complete looping inference process is obtained, and the complete inference path formed from the initial hidden representation to the final representation result is defined as the longest loop trajectory;

[0061] Step B5: Multi-trajectory joint optimization training: During the model training phase, the complete inference process is performed based on the longest loop trajectory. Loop paths with fewer than the maximum number of loops are randomly sampled as shortcut trajectories. Loop inference is performed on the initial hidden representation under different computational budget conditions. The master prediction loss under the longest loop trajectory and the quick prediction loss under the shortcut trajectory are calculated respectively. The output results of the longest loop trajectory and the output results of the shortcut trajectory are used to construct a consistency constraint loss. The master prediction loss, quick prediction loss and consistency constraint loss are combined to perform joint optimization to complete the optimized training of the Looped-Transform model parameters, resulting in the trained Looped-Transform model.

[0062] Step B6: Result Generation and Grade Determination: Based on the trained Looped-Transform model, hierarchical fusion and scoring mapping are performed on the hidden representation in the final loop corresponding to the maximum number of loops to obtain the comprehensive endometrial receptivity score for the examinee; then, the comprehensive score is graded according to the preset grading rules to generate the corresponding receptivity grade; and combined with the detection cycle position, implantation window time interval, and comprehensive score distribution results, the implantation window adaptation status of the examinee in the current cycle is determined, thereby outputting the endometrial receptivity assessment result for clinical applications in assisted reproduction.

[0063] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. An endometrial receptivity intelligent assessment system based on biosensor data processing, characterized in that, The system includes: The biosensing detection module acquires multi-channel raw sensing signals; The sensor preprocessing module processes the multi-channel raw sensor signals to generate structured biosensor data. The information fusion module acquires information about the tested object and correlates it with structured biosensor data to construct endometrial receptivity characteristic data. The manifold feature module constructs a weighted heterogeneous evaluation map based on endometrial receptivity feature data; it introduces a manifold stochastic feature modeling mechanism to perform manifold stochastic feature modeling processing based on random walk propagation and kernel function approximation on the weighted heterogeneous evaluation map, generating manifold stochastic feature data; The receptivity assessment module constructs a Looped-Transform model, inputting both manifold random feature data and endometrial receptivity feature data into the Looped-Transform model, and outputs the receptivity assessment results. The decision output module generates recommendations on the timing of embryo transfer, re-examination, adjustments to the endometrial preparation plan, and risk warnings based on the acceptability assessment results.

2. The intelligent assessment system for endometrial receptivity based on biosensor data processing according to claim 1, characterized in that: The weighted heterogeneous evaluation graph includes nodes for the tested object, testing cycle, sample type, biomarker, sensor channel, clinical indicator, and sampling points at different time sequences.

3. The intelligent assessment system for endometrial receptivity based on biosensor data processing according to claim 1, characterized in that: The Looped-Transform model is constructed as follows: a Transform model is established, a shared parameter loop mechanism is introduced, the fixed-level stacking structure of the Transform model is transformed into a loop computation structure that can be called repeatedly, the representation consistency and inference stability of the Transform model in the multi-round iteration process are optimized, and the Looped-Transform model is constructed.

4. The intelligent assessment system for endometrial receptivity based on biosensor data processing according to claim 2, characterized in that: The process of generating random feature data of a manifold includes the following: Step S1: Perform object parsing and primitive processing on the endometrial receptivity characteristic data to construct a weighted heterogeneous evaluation graph; the weighted heterogeneous evaluation graph uses the set of nodes as the discrete sampling points of the manifold; Step S2: Represent the first-order direct association, second-order indirect propagation association, and higher-order cross-path diffusion association between nodes in a unified form as a power series superposition of the graph kernel matrix, and configure the corresponding kernel coefficient sequence according to the difference in the contribution of different propagation orders to node similarity. Step S3: Introduce the modulation function and establish the mapping relationship between the modulation function and the kernel coefficient sequence; based on the mapping relationship, perform deconvolution of the modulation function to obtain the path length modulation function; configure random walk strategies for each node in the weighted heterogeneous evaluation graph. Step S4: Based on the random walk strategy, take each node object in the weighted heterogeneous evaluation graph as the starting node, perform multiple random walk samplings to obtain the set of manifold random feature signature vectors; Step S5: Based on the manifold random feature signature vector set, extract the corresponding signature vector for any node object in the weighted heterogeneous evaluation graph, and perform pairwise matching calculation with the signature vectors of other node objects to obtain the signature response value between each node pair; construct a supervised training sample set with the starting node, reference node and corresponding signature response value in the node pair as the basic unit; Construct a continuous mapping function and train its parameters using a supervised training sample set; Step S6: After completing the training of the continuous mapping function, take the node objects and their neighborhood extension positions in the weighted heterogeneous evaluation graph as the objects to be mapped, and call the trained continuous mapping function to calculate the continuous response values ​​of the objects to be mapped. A bivariate evaluation function for the target is constructed based on continuous response values; a continuous core field integral approximation model is introduced to decompose and approximate the bivariate evaluation function for the target, generating a continuous manifold response field; and manifold random feature data is generated based on the continuous manifold response field.

5. The intelligent assessment system for endometrial receptivity based on biosensor data processing according to claim 4, characterized in that: The configuration of the random walk strategy includes the random walk start rule, neighborhood transfer rule, load update rule, and termination determination rule.

6. The intelligent assessment system for endometrial receptivity based on biosensor data processing according to claim 3, characterized in that: The process of obtaining the acceptance assessment results specifically includes the following: Step B1: Process the manifold random feature data and endometrial receptivity assessment feature data to generate an initial hidden representation; Step B2: Input the initial hidden representation into the basic computational structure of the Looped-Transform model, and perform parameter sharing processing on each Transformer block in the basic computational structure to build a shared parameter loop backbone; The iterative evaluation process of the Looped-Transform model is normalized to a continuous inference trajectory, and the shared parameter loop backbone is regarded as an evolution operator on the continuous inference trajectory. The continuous inference trajectory is discretized according to the preset maximum number of iterations to obtain the normalized time position parameters and step size parameters for each iteration. Step B3: For any cycle, perform frequency embedding mapping and nonlinear projection processing on the normalized time position parameters and step size parameters corresponding to that cycle to generate the trajectory condition representation corresponding to that cycle. Based on the trajectory condition representation, normalized scaling control parameters and residual gating control parameters are generated respectively. The normalized scaling control parameters are applied to the normalization layer in the shared parameter cyclic backbone, and the residual gating control parameters are applied to the residual connection positions of the multi-head self-attention branches and feedforward branches inside the shared parameter cyclic backbone. The shared parameter cyclic backbone is subjected to structural modulation processing to obtain the shared backbone after trajectory modulation. Step B4: In each cycle, based on the shared backbone modulated by the trajectory, perform multi-head self-attention modeling, feedforward nonlinear mapping and residual update processing to update the hidden representation; when the cycle reaches the maximum number of cycles, the final representation result is obtained and defined as the longest cycle trajectory; Step B5: Randomly sample loop paths with fewer than the maximum number of loops as shortcut trajectories and perform loop inference on the initial hidden representation; calculate the main prediction loss under the longest loop trajectory and the quick prediction loss under the shortcut trajectory, and construct the consistency constraint loss; combine the main prediction loss, quick prediction loss and consistency constraint loss to perform joint optimization, complete the optimized training of the Looped-Transform model, and obtain the trained Looped-Transform model. Step B6: Based on the trained Looped-Transform model, output the tolerance evaluation results.

7. The intelligent assessment system for endometrial receptivity based on biosensor data processing according to claim 6, characterized in that: The longest loop trajectory is defined as the complete inference path formed from the initial hidden representation to the final representation result.

Images