Method and device for predicting risk of ohss in assisted reproduction based on traditional chinese medicine syndrome differentiation

Abstract

The method and device for predicting the risk of OHSS in assisted reproduction based on traditional Chinese medicine syndrome differentiation can realize risk assessment and individualized traditional Chinese medicine syndrome differentiation intervention in the whole ovulation promotion cycle through data fusion of traditional Chinese and Western medicine, dynamic weighting of double models and support of prior knowledge base, while ensuring data privacy and real-time performance. The method includes: (1) collecting clinical data as a data set and preprocessing; (2) establishing a prior knowledge base; (3) based on the clinical data of the sub-data set, constructing an initial risk prediction model and a time series risk prediction model, and through a model confidence driven adaptive weight fusion mechanism, dynamically weighting the prediction results of the double models, and outputting the whole cycle fusion risk prediction results; (4) constructing a double model bias prediction early warning mechanism; (5) automatically collecting patient data and storing the collected patient data in block chain encryption, while real-time outputting the OHSS risk prediction results to provide quantitative support for traditional Chinese medicine syndrome differentiation intervention scheme.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a method for predicting the risk of OHSS in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation, and a device for predicting the risk of OHSS in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation. Background Technology

[0002] Against the backdrop of a global infertility rate rising by 15%-20%, assisted reproductive technology has become an important means of addressing fertility issues. Artificial ovulation induction, as a core component of the assisted reproductive cycle, stimulates the development of multiple follicles in the ovaries, and its safety directly impacts patient prognosis. Ovarian hyperstimulation syndrome (OHSS), a major complication of ovulation induction treatment, seriously threatens women's health.

[0003] Current OHSS prediction methods mainly rely on single-dimensional indicators from Western medicine, such as patient age, body mass index (BMI), antral follicle count (AFC), anti-Müllerian hormone (AMH) level, estradiol (E2) peak during ovulation induction, and number of oocytes retrieved. These methods have several problems: First, they focus solely on numerical changes in biological indicators, lacking a holistic assessment of the patient, leading to insufficient individualized risk identification. Second, they often use static models for initial risk screening before ovulation induction, failing to adapt to the dynamic changes in patient response to medication during treatment, resulting in insufficient risk warning capabilities in the mid-to-late cycles. Third, the lack of an efficient integration mechanism between Western and Traditional Chinese Medicine data hinders the realization of collaborative decision-making combining Western and Traditional Chinese Medicine approaches.

[0004] Traditional Chinese medicine (TCM) syndrome differentiation, as the core of TCM diagnosis and treatment, can systematically collect macroscopic syndrome information of patients through the four diagnostic methods of observation, auscultation and olfaction, inquiry and palpation, construct a dynamic assessment system for syndrome differentiation, systematically capture the dynamic evolution of syndromes in the body, and demonstrate unique advantages in individualized diagnosis and treatment.

[0005] Therefore, it is urgent to build an OHSS risk prediction system based on TCM syndrome differentiation and Western medicine data technology. By collecting multimodal data, modeling time series features, and integrating TCM and Western medicine knowledge, we can overcome the limitations of single-dimensional indicator prediction and promote the innovative application of TCM and Western medicine in the field of reproductive medicine. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a method for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation. This method can achieve risk assessment and personalized TCM syndrome differentiation intervention throughout the ovulation induction cycle by integrating TCM and Western medicine data, using dynamic weighting of dual models and prior knowledge base support, while ensuring data privacy and real-time performance.

[0007] The technical solution of this invention is: a method for predicting the risk of OHSS in assisted reproduction based on traditional Chinese medicine syndrome differentiation, which includes the following steps:

[0008] (1) Collect clinical data as a dataset and preprocess it. The static data before ovulation induction and the dynamic data of treatment time sequence during ovulation induction are both subsets of this dataset.

[0009] (2) Establish a prior knowledge base, including a diagnostic time series rule base based on the TCM diagnostic time series data of the subset dataset, and a TCM and Western medicine drug knowledge base constructed by integrating medical books, the experience of famous veteran TCM doctors and guidelines.

[0010] (3) Based on clinical data from a subset of datasets, an initial risk prediction model and a time-series risk prediction model are constructed. Through an adaptive weight fusion mechanism driven by model confidence, the prediction results of the two models are dynamically weighted and the full-cycle fusion risk prediction results are output to support daily real-time comprehensive risk assessment during ovulation induction. The output results provide quantitative decision-making basis for TCM syndrome differentiation intervention.

[0011] (4) Construct a dual-model bias prediction and early warning mechanism to avoid misjudgment caused by the limitation of data dimensions in a single model;

[0012] (5) After the prediction model is deployed in the hospital system, it automatically collects patient data and stores the collected patient data in blockchain encryption to ensure data privacy. At the same time, it outputs OHSS risk prediction results in real time to provide quantitative support for TCM syndrome differentiation and intervention programs.

[0013] This invention can achieve risk assessment and personalized TCM syndrome differentiation intervention throughout the entire ovulation induction cycle by integrating TCM and Western medicine data, using dynamic weighting of dual models and prior knowledge base support, while ensuring data privacy and real-time performance.

[0014] A device for predicting OHSS risk in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation is also provided, comprising:

[0015] The data acquisition and preprocessing module is configured to acquire clinical data as a dataset and perform preprocessing. The static data before ovulation induction and the dynamic data of the treatment timeline during ovulation induction are both subsets of this dataset.

[0016] The prior knowledge base construction module is configured to build a prior knowledge base, including a syndrome differentiation time series rule base built based on a subset of TCM syndrome differentiation time series data, and a TCM and Western medicine drug knowledge base built by integrating medical books, the experience of famous veteran TCM doctors and guidelines.

[0017] The model building module is configured to construct an initial risk prediction model and a time-series risk prediction model based on clinical data from a subset of datasets. Through an adaptive weight fusion mechanism driven by model confidence, the prediction results of the two models are dynamically weighted to output the full-cycle fused risk prediction results, thereby supporting daily real-time comprehensive risk assessment during ovulation induction. Its output results provide quantitative decision-making basis for TCM syndrome differentiation intervention.

[0018] The mechanism construction module is configured to build a dual-model bias prediction and early warning mechanism to avoid misjudgments caused by the limitation of data dimensions in a single model;

[0019] The prediction module, configured to predict the model's deployment within the hospital system, automatically collects patient data and stores it using blockchain encryption to ensure data privacy. Simultaneously, it outputs real-time OHSS risk prediction results, providing insights for TCM-based intervention programs.

[0020] Provide quantitative support. Attached Figure Description

[0021] Figure 1 A flowchart of the method for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to the present invention is shown.

[0022] Figure 2 A framework diagram of a system for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to the present invention is shown. Detailed Implementation

[0023] like Figure 1 As shown, a method for predicting OHSS risk in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation includes the following steps:

[0024] (1) Collect clinical data as a dataset and preprocess it. The static data before ovulation induction and the dynamic data of treatment time sequence during ovulation induction are both subsets of this dataset.

[0025] (2) Establish a prior knowledge base, including a diagnostic time series rule base based on the TCM diagnostic time series data of the subset dataset, and a TCM and Western medicine drug knowledge base constructed by integrating medical books, the experience of famous veteran TCM doctors and guidelines.

[0026] (3) Based on clinical data from a subset of datasets, an initial risk prediction model and a time-series risk prediction model are constructed. Through an adaptive weight fusion mechanism driven by model confidence, the prediction results of the two models are dynamically weighted and the full-cycle fusion risk prediction results are output to support daily real-time comprehensive risk assessment during ovulation induction. The output results provide quantitative decision-making basis for TCM syndrome differentiation intervention.

[0027] (4) Construct a dual-model bias prediction and early warning mechanism to avoid misjudgment caused by the limitation of data dimensions in a single model;

[0028] (5) After the prediction model is deployed in the hospital system, it automatically collects patient data and stores the collected patient data in blockchain encryption to ensure data privacy. At the same time, it outputs OHSS risk prediction results in real time to provide quantitative support for TCM syndrome differentiation and intervention programs.

[0029] This invention can achieve risk assessment and personalized TCM syndrome differentiation intervention throughout the entire ovulation induction cycle by integrating TCM and Western medicine data, using dynamic weighting of dual models and prior knowledge base support, while ensuring data privacy and real-time performance.

[0030] Preferably, in step (1), patient information is collected for preliminary diagnosis using the four diagnostic methods of traditional Chinese medicine (inspection, auscultation, inquiry, and palpation) and mobile terminal devices. The information collected by the four diagnostic methods includes the following: observation of spirit, facial complexion, lip color, tongue color, tongue body, tongue coating, body shape, posture, skin, and excretions; normal voice, pathological voice, body odor, secretion odor, excretion odor, and ward odor; basic medical history, symptom description, and questionnaire information; pulse waveform and palpation information.

[0031] Western medicine collects the following information: static data before ovulation induction, including: age, BMI, menstrual history, marital and reproductive history, genetic disease history, type of infertility, basal antral follicle information, basal sex hormone information, basal blood routine information, basic liver and kidney function information, and blood glucose information; dynamic data during ovulation induction treatment, including: ovulation induction protocol information, Gn medication information, hormone information, trigger drug information, and number of oocytes retrieved.

[0032] Preferably, in step (1), the visual diagnosis information is obtained by using computer vision technology and OpenCV library for image processing. The intensity of eye light reflection is obtained by calculating the proportion of bright spot pixels after grayscale image binarization. The high-definition image features of tongue color, tongue body and tongue coating are extracted by CLAHE histogram equalization + U-Net segmentation method. The RGB color values ​​of face color and lip color, body fat index, posture movement trajectory, skin texture pigment distribution and excretion morphology and color quantitative indicators are calculated by computer.

[0033] Auscultation information is obtained by collecting sound through a microphone and combining it with a multi-dimensional gas sensor array to collect volatile gases from the patient's body, secretions, excrement, and ward, and extracting voiceprint and odor molecule concentration characteristics;

[0034] The consultation information is collected by gathering basic medical history, symptom descriptions and questionnaire information, and the symptom descriptions are standardized and scored using a TCM symptom quantitative scoring scale.

[0035] Palpation information is obtained by collecting pulse waveform data at the cun, guan, and chi positions using standard instruments, as well as pressure-displacement feedback data from palpation, and converting them into a computable temporal feature vector.

[0036] The collected data underwent standardization processing, including filling in missing values ​​using bidirectional interpolation.

[0037] Z-score standardization is used for feature normalization, dynamic data is aligned with the ovulation induction start date as the time origin, and a dialectical time series matrix is ​​constructed to form a dataset for model training.

[0038] Preferably, in step (2), the prior knowledge base includes:

[0039] The syndrome differentiation time-series rule base is based on symptom description data in a subset of data. It standardizes symptom descriptions using a TCM symptom quantitative scoring table and generates symptom feature sequences according to treatment time. Natural language processing (NLP) techniques are used to parse and match syndrome criteria to extract syndrome differentiation features. The syndrome differentiation sequences are converted into numerical codes and generated evolution vectors according to time windows. The Apriori algorithm is then used to mine syndrome differentiation time-series association rules. The thresholds for the syndrome differentiation time-series rule base are support ≥30% and confidence ≥80%. Historical data is used for validation to ensure statistical significance. After selecting effective rules, they are structured and stored to generate a syndrome differentiation evolution rule table, forming prior rules for syndrome differentiation risk. These prior probabilities are then superimposed when the model outputs risk.

[0040] The Traditional Chinese Medicine (TCM) and Western Medicine knowledge base includes: collecting classic medical texts, the medication experience of renowned veteran TCM doctors, medication methods and contraindications; using entity recognition and relation extraction algorithms to store extracted entities as nodes and relations as edges; and using the graph database Neo4j to build the TCM knowledge base. The Western Medicine knowledge base includes: collecting medical guidelines issued by authoritative institutions, establishing a mapping relationship between indicators and ovulation induction protocols, common ovulation induction drugs' usage and contraindications, and prioritizing recommendations; and using the relational database MySQL to build the Western Medicine knowledge base, which is updated regularly.

[0041] Preferably, in step (3), the initial risk prediction model and the time-series risk prediction model include:

[0042] Initial risk prediction model: A subset of pre-ovulation induction static data is input, fusing Western medicine baseline indicators with Traditional Chinese Medicine (TCM) syndrome differentiation features to form a multi-dimensional input matrix. The model employs a method of embedding prior rules of syndrome differentiation risk into the objective function and introducing a knowledge base regularization term to enhance TCM syndrome differentiation.

[0043] The XGBoost algorithm is used to train the model, and the final output is the probability of OHSS risk before ovulation induction and intervention suggestions based on the dialectical results;

[0044] The time-series risk prediction model is constructed by inputting a subset of dynamic time-series data of ovulation induction treatment, and building a three-dimensional tensor containing the time-series characteristics of TCM syndrome differentiation and Western medicine monitoring indicators. The model adopts an LSTM + attention mechanism architecture, automatically focuses on key time-series features through attention weights, establishes a time-series risk prediction model, and introduces a dual-pathway correction mechanism. Combined with prior rules and knowledge base, it conducts daily real-time comprehensive risk assessment and dynamic intervention suggestions of TCM and Western medicine, including the adjustment of TCM prescriptions and the optimization of Western medicine dosages. The two are collaboratively decided through a multi-objective optimization algorithm.

[0045] Dual-pathway correction mechanism: Focusing on the internal optimization of the time series model, the model parameters are dynamically updated and the intervention strategy is optimized through dialectical feedback of the deviation between model prediction and actual dialectical analysis and intervention learning of historical intervention effects.

[0046] Model confidence-driven adaptive weight fusion mechanism:

[0047]

[0048] Wherein, W1 is the initial risk prediction model weight, W2 is the time-series risk prediction model weight, C1 is the confidence level of the initial risk prediction model, reflecting the reliability of the model in predicting baseline characteristics before ovulation induction, and C2 is the confidence level of the time-series risk prediction model, reflecting the reliability of the model in predicting changes in syndrome differentiation during treatment.

[0049] Preferably, in step (4), constructing the dual-model bias prediction and early warning mechanism includes:

[0050] Calculate the difference in OHSS risk probability between the two models: ΔR = |R1 - R2|, where R1 is the OHSS risk probability value output by the initial risk prediction model and R2 is the OHSS risk probability value output by the time series risk prediction model.

[0051] When ΔR > 0.2 and |Risklevel(R1) - Risklevel(R2)| ≥ 1, a Level 3 warning is triggered.

[0052] Level 1 Warning: A warning pop-up window will be pushed through the hospital information system, indicating that the model predictions show significant differences. Please review the TCM diagnosis to see if any concurrent symptoms have been missed.

[0053] Level 2 warning: The system will link with the knowledge base to generate a list of potential concurrent evidence and key points for differentiation.

[0054] Level 3 early warning: Automatically retrieves traditional Chinese and Western medicine intervention plans from similar historical medical records for reference.

[0055] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a device for predicting the risk of OHSS in assisted reproduction based on traditional Chinese medicine syndrome differentiation. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0056] The data acquisition and preprocessing module is configured to acquire clinical data as a dataset and perform preprocessing. The static data before ovulation induction and the dynamic data of the treatment timeline during ovulation induction are both subsets of this dataset.

[0057] The prior knowledge base construction module is configured to build a prior knowledge base, including a syndrome differentiation time series rule base built based on a subset of TCM syndrome differentiation time series data, and a TCM and Western medicine drug knowledge base built by integrating medical books, the experience of famous veteran TCM doctors and guidelines.

[0058] The model building module is configured to construct an initial risk prediction model and a time-series risk prediction model based on clinical data from a subset of datasets. Through an adaptive weight fusion mechanism driven by model confidence, the prediction results of the two models are dynamically weighted to output the full-cycle fused risk prediction results, thereby supporting daily real-time comprehensive risk assessment during ovulation induction. Its output results provide quantitative decision-making basis for TCM syndrome differentiation intervention.

[0059] The mechanism construction module is configured to build a dual-model bias prediction and early warning mechanism to avoid misjudgments caused by the limitation of data dimensions in a single model;

[0060] The prediction module, configured to predict the model after deployment in the hospital system, automatically collects patient data and stores the collected patient data in blockchain encryption to ensure data privacy. At the same time, it outputs OHSS risk prediction results in real time, providing quantitative support for TCM syndrome differentiation and intervention programs.

[0061] Preferably, in the data acquisition and preprocessing module, patient information is collected for preliminary diagnosis using the four diagnostic methods of traditional Chinese medicine (inspection, auscultation, inquiry, and palpation) and mobile terminal devices. The information collected by the four diagnostic methods includes: observation of spirit, facial complexion, lip color, tongue color, tongue body, tongue coating, body shape, posture, skin, and excretions; normal voice, pathological voice, body odor, secretion odor, excretion odor, and ward odor; basic medical history, symptom description, and questionnaire information; pulse waveform and palpation information.

[0062] The information collected by Western medicine includes the following: static data collected before ovulation induction, including: age, BMI, menstrual history, marital and reproductive history, genetic disease history, infertility type, basal antral follicle information, basal sex hormone information, basal blood routine information, basic liver and kidney function information, and blood glucose information; dynamic data collected during ovulation induction treatment, including: ovulation induction protocol information, Gn medication information, hormone information, trigger drug information, and number of oocytes retrieved.

[0063] The visual diagnosis information was obtained by acquiring facial and full-body images of the patient using standard instruments. The images were processed using the OpenCV library with computer vision technology. The intensity of the light reflection in the eyes was obtained by calculating the proportion of bright spots after binarization of the grayscale image. The high-resolution image features of tongue color, tongue body and tongue coating were extracted using CLAHE histogram equalization and U-Net segmentation. The RGB color values ​​of facial color and lip color, body fat index, posture and movement trajectory, skin texture pigment distribution and excretion morphology and color quantitative indicators were calculated by computer.

[0064] Auscultation information is obtained by collecting sound through a microphone and combining it with a multi-dimensional gas sensor array to collect volatile gases from the patient's body, secretions, excrement, and ward, and extracting voiceprint and odor molecule concentration characteristics;

[0065] The consultation information is collected by gathering basic medical history, symptom descriptions and questionnaire information, and the symptom descriptions are standardized and scored using a TCM symptom quantitative scoring scale.

[0066] Palpation information is collected using standard instruments to capture pulse waveform data at the cun, guan, and chi positions, as well as pressure displacement feedback data during palpation, and then converted into a computable temporal feature vector.

[0067] The collected data underwent standardization processing, including filling in missing values ​​using bidirectional interpolation.

[0068] Z-score standardization is used for feature normalization, dynamic data is aligned with the ovulation induction start date as the time origin, and a dialectical time series matrix is ​​constructed to form a dataset for model training.

[0069] Preferably, in the prior knowledge base construction module, the prior knowledge base includes:

[0070] The syndrome differentiation time-series rule base is based on symptom description data in a subset of data. It standardizes symptom descriptions using a TCM symptom quantitative scoring table and generates symptom feature sequences according to treatment time. Natural language processing (NLP) techniques are used to parse and match syndrome criteria to extract syndrome differentiation features. The syndrome differentiation sequences are converted into numerical codes and generated evolution vectors according to time windows. The Apriori algorithm is then used to mine syndrome differentiation time-series association rules. The thresholds for the syndrome differentiation time-series rule base are support ≥30% and confidence ≥80%. Historical data is used for validation to ensure statistical significance. After selecting effective rules, they are structured and stored to generate a syndrome differentiation evolution rule table, forming prior rules for syndrome differentiation risk. These prior probabilities are then superimposed when the model outputs risk.

[0071] The Traditional Chinese Medicine (TCM) and Western Medicine knowledge base includes: collecting classic medical texts, the medication experience of renowned veteran TCM doctors, medication methods and contraindications; using entity recognition and relation extraction algorithms to store extracted entities as nodes and relations as edges; and using the graph database Neo4j to build the TCM knowledge base. The Western Medicine knowledge base includes: collecting medical guidelines issued by authoritative institutions, establishing a mapping relationship between indicators and ovulation induction protocols, common ovulation induction drugs' usage and contraindications, and prioritizing recommendations; and using the relational database MySQL to build the Western Medicine knowledge base, which is updated regularly.

[0072] Preferably, in the model building module, the initial risk prediction model and the time-series risk prediction model include:

[0073] Initial risk prediction model: A subset of pre-ovulation induction static data is input, fused with Western medicine baseline indicators, and combined with TCM syndrome differentiation features to form a multi-dimensional input matrix. The model is trained using the XGBoost algorithm enhanced with TCM syndrome differentiation. This algorithm embeds prior rules of syndrome differentiation risk into the standard XGBoost objective function, introduces a knowledge base regularization term to integrate TCM prior knowledge, and finally outputs the pre-ovulation OHSS risk probability and intervention suggestions based on syndrome differentiation results.

[0074] The time-series risk prediction model is constructed by inputting a subset of dynamic time-series data of ovulation induction treatment, and building a three-dimensional tensor containing the time-series characteristics of TCM syndrome differentiation and Western medicine monitoring indicators. The model adopts an LSTM + attention mechanism architecture, automatically focuses on key time-series features through attention weights, establishes a time-series risk prediction model, and introduces a dual-pathway correction mechanism. Combined with prior rules and knowledge base, it conducts daily real-time comprehensive risk assessment and dynamic intervention suggestions of TCM and Western medicine, including the adjustment of TCM prescriptions and the optimization of Western medicine dosages. The two are collaboratively decided through a multi-objective optimization algorithm.

[0075] Dual-pathway correction mechanism: Focusing on the internal optimization of the time series model, the model parameters are dynamically updated and the intervention strategy is optimized through dialectical feedback of the deviation between model prediction and actual dialectical analysis and intervention learning of historical intervention effects.

[0076] Model confidence-driven adaptive weight fusion mechanism:

[0077]

[0078] Wherein, W1 is the initial risk prediction model weight, W2 is the time-series risk prediction model weight, C1 is the confidence level of the initial risk prediction model, reflecting the reliability of the model in predicting baseline characteristics before ovulation induction, and C2 is the confidence level of the time-series risk prediction model, reflecting the reliability of the model in predicting changes in syndrome differentiation during treatment.

[0079] Preferably, in the mechanism construction module, constructing the dual-model bias prediction and early warning mechanism includes:

[0080] Calculate the difference in OHSS risk probability between the two models: ΔR = |R1 - R2|, where R1 is the OHSS risk probability value output by the initial risk prediction model and R2 is the OHSS risk probability value output by the time series risk prediction model.

[0081] When ΔR > 0.2 and |Risklevel(R1) - Risklevel(R2)| ≥ 1, a Level 3 warning is triggered.

[0082] Level 1 Warning: A warning pop-up window will be pushed through the hospital information system, indicating that the model predictions show significant differences. Please review the TCM diagnosis to see if any concurrent symptoms have been missed.

[0083] Level 2 warning: The system will link with the knowledge base to generate a list of potential concurrent evidence and key points for differentiation.

[0084] Level 3 early warning: Automatically retrieves traditional Chinese and Western medicine intervention plans from similar historical medical records for reference.

[0085] The specific implementation of the present invention will be further explained below.

[0086] In the model confidence-driven adaptive weight fusion mechanism, C1 represents the accuracy of the initial risk prediction model on the validation set, reflecting the reliability of the model's prediction of baseline characteristics before ovulation induction. C2 represents the accuracy of the time-series risk prediction model on the validation set, reflecting the reliability of the model's prediction of changes in syndrome differentiation during treatment. The model confidence is derived based on the model's performance evaluation on the independent validation set, ensuring the reliability of the weights.

[0087] The initial risk prediction model and the time-series risk prediction model are constructed, and Western medicine baseline indicators are integrated with traditional Chinese medicine syndrome differentiation characteristics to form a multi-dimensional input matrix. The steps include:

[0088] Baseline indicators extracted from Western medicine: such as basic physiological indicators: age (A), BMI (B), basal antral follicle count (AFC), and basic sex hormone levels (such as FSH, LH, E2); medical history indicators: menstrual history code (M), marital and reproductive history code (P), genetic disease history markers (G), and infertility type code (I); laboratory indicators: basic blood routine parameters (WBC, RBC, PLT), liver and kidney function indicators (ALT, AST, Cr), and blood glucose (GLU); all Western medicine indicators were standardized into numerical vectors using Z-scores. in x is a standardized vector of Western medicine indicators with dimension n; i Let be the standardized value of the i-th indicator (i = 1, 2, 3, ..., n).

[0089] Traditional Chinese Medicine (TCM) diagnostic features extraction: Inspection features: eye light reflection intensity (grayscale + binarization), facial color RGB values ​​(R, G, B), tongue color, tongue texture, and tongue coating quantitative indicators (texture features T extracted through CLAHE+U-Net segmentation), body shape (s); Auscultation features: voiceprint feature vector (V), odor molecule concentration features (O); Inquiry features: symptom quantitative score (based on the standardized score of the "TCM Symptom Quantification Scoring Table"); Palpation features: pulse waveform parameters (pulse rate HR, pulse pressure difference PP, waveform feature parameter W), pressure-displacement features of palpation (P). These TCM diagnostic features are standardized to form numerical vectors. in, y is a standardized vector of TCM syndrome differentiation characteristics, with dimension m; j Let be the standardized value of the j-th dialectical feature (j = 1, 2, 3, ..., n).

[0090] Multidimensional input matrix: The Western medicine indicator vector and the traditional Chinese medicine syndrome differentiation vector are concatenated to form a multidimensional input matrix M. Where M is a multidimensional input matrix, k is the number of samples, and the dimension is k×(n+m); the i-th row corresponds to the feature vector of the i-th sample; the first n columns are Western medicine indicators, and the last m columns are traditional Chinese medicine syndrome differentiation features.

[0091] Sample alignment: Using the patient's medical record number as the key, the Western medicine indicators and TCM features of the same patient are matched according to the collection timestamp. If there is a time difference, the time point is filled in by interpolation. If the dimensions of the Western medicine indicator vector and the TCM syndrome differentiation feature vector are inconsistent, zero padding is used to ensure the feasibility of matrix splicing.

[0092] The initial risk prediction model and the time-series risk prediction model are constructed. Knowledge base regularization terms are introduced, and the XGBoost algorithm, enhanced by traditional Chinese medicine syndrome differentiation, is used for model training. The steps include:

[0093] Knowledge base regular expressions: Where k is the total number of samples, i is the sample index, and R0 i For the set of rules applicable to sample i (i.e., the features of sample i satisfy the conditions of rule r), expected r Let f(M) be the expected risk probability corresponding to rule r. i ;θ) is the predicted OHSS risk probability (range 0-1) for the i-th sample based on model parameters θ, where θ is the parameter vector of the XGBoost model; r is a single rule, and the expected value of rule r is defined based on the prior knowledge base, for example, "If the dialectical feature is C, then the OHSS risk probability should be ≥0.7"; ω r Let r be the confidence score of rule r, and let ω be the confidence score after filtering using the Apriori algorithm. r ≥80% indicates the reliability of the rule; λ1 and λ2 are regularization coefficients determined through cross-validation, used to balance the strength of rule constraints and the generalization ability of the model; This is the L2 regularization term, the L2 norm of the model parameters, which suppresses overfitting of the model parameters and ensures that the model complexity is controllable.

[0094] Integration with XGBoost objective function: Where Obj(θ) is the objective function used to optimize the model parameter θ; For the prediction loss term, k is the total number of samples, i is the sample index, and y is the index of the sample. i Let M be the true label of the i-th sample (0 indicates no OHSS risk, 1 indicates risk). i Let f(M) be the multidimensional input matrix of the i-th sample (containing Western medicine indicators and traditional Chinese medicine syndrome differentiation features). i ;θ) is the predicted OHSS risk probability value (range 0~1) for the i-th sample based on the model parameter θ, expected r Let ω be the expected risk probability corresponding to rule r, and I(·) be the logistic regression loss function, ensuring that the model fits the data and conforms to the prior knowledge of TCM syndrome differentiation. r Let r be the confidence score of rule r, and let ω be the confidence score after filtering using the Apriori algorithm. r ≥80% indicates the reliability of the rule; λ1 and λ2 are regularization coefficients determined through cross-validation, used to balance the strength of rule constraints and the generalization ability of the model; This is the L2 regularization term, the L2 norm of the model parameters, which suppresses overfitting of the model parameters and ensures that the model complexity is controllable.

[0095] Model training: The dataset was divided into training, validation, and test sets in a 6:2:2 ratio; the optimal regularization coefficients λ1 and λ2 were determined through 5-fold cross-validation; the XGBoost model was trained iteratively, and the objective function Obj(θ) was calculated in each round. The objective function was optimized through gradient descent so that the model prediction fits both the data and the prior rules; finally, the model's adherence to TCM syndrome differentiation rules and prediction accuracy were verified on the test set.

[0096] The steps for constructing an initial risk prediction model and a time-series risk prediction model, and training the models include:

[0097] Dataset partitioning: The clinical dataset was randomly divided into training, validation and test sets in a ratio of 6:2:2. The training set was used for model parameter learning, the validation set was used for hyperparameter tuning, and the test set was used for final performance evaluation.

[0098] Regularization coefficients are determined by selecting stable parameters through 5-fold cross-validation. The objective function is a weighted sum of "prediction loss + rule violation penalty" with a weight ratio of 7:3. The hyperparameter space (λ1∈[0.1,1.0], λ2∈[0.01,0.1]) is searched to select the optimal regularization coefficients λ1 and λ2 that minimize the objective function, thus balancing the strength of rule constraints with the model's generalization ability.

[0099] Iterative training process: The XGBoost algorithm based on TCM syndrome differentiation is used to iteratively train the model. The objective function Obj(θ) is calculated in real time during each round of decision tree generation. When splitting the XGBoost tree, the correlation between each feature and the prior rule is calculated, and features with high correlation are selected for splitting first.

[0100] Performance validation: Finally, the model performance is fully validated on the test set, including compliance with TCM syndrome differentiation rules: the proportion of the model's prediction results that conform to the prior rules; prediction accuracy: the reliability of the model in clinical applications is ensured by evaluating the accuracy, precision, and risk classification effect.

[0101] Construct a dual-model bias prediction and early warning mechanism, including:

[0102] Calculate the difference in OHSS risk probability between the two models: ΔR = |R1 - R2|, where R1 is the OHSS risk probability value output by the initial risk prediction model and R2 is the OHSS risk probability value output by the time series risk prediction model.

[0103] When ΔR > 0.2 and |Risklevel(R1) - Risklevel(R2)| ≥ 1, a Level 3 warning is triggered.

[0104] Risk level classification: Risk level classification is based on historical data statistics and clinical experience. In practical applications, parameters can be adjusted to suit different populations. For example, low risk: probability < 0.3 (Level 1); medium risk: 0.3 ≤ probability ≤ 0.6 (Level 2); high risk: probability > 0.6 (Level 3).

[0105] Level 1 warning: A warning pop-up message is pushed through the hospital information system, indicating "significant difference in model predictions. Please review whether any concurrent symptoms have been missed in the TCM diagnosis."

[0106] Level 2 warning: The system will link with the knowledge base to generate a list of potential concurrent evidence and key points for differentiation.

[0107] Level 3 early warning: Automatically retrieves traditional Chinese and Western medicine intervention plans from similar historical medical records for reference.

[0108] like Figure 2 As shown, a system for predicting OHSS risk in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation includes:

[0109] The data acquisition and preprocessing module is compatible with mobile terminal data interfaces. It collects basic patient data, treatment monitoring data, and other clinical data, and filters, processes, and unifies the collected clinical data to adapt to real-time data transmission and analysis.

[0110] The dataset generation module divides the original dataset into two subsets: a static data subset before ovulation induction and a dynamic data subset based on the treatment timeline during ovulation induction.

[0111] The prior knowledge base construction module includes a syndrome differentiation time series rule base built based on a subset of TCM syndrome differentiation time series data, and a TCM and Western medicine medication knowledge base built by integrating medical books, the experience of famous veteran TCM doctors and guidelines.

[0112] The model building module builds a risk prediction model based on the collected clinical data by integrating static data before ovulation induction and dynamic data of ovulation induction treatment.

[0113] The model training and validation module trains the model based on collected clinical data, using static patient data and dynamic treatment timeline data to achieve pre-ovulation induction prediction and daily real-time comprehensive prediction during ovulation induction. It validates the timeliness and accuracy of risk prediction, generating the final deployable risk prediction model component.

[0114] The intelligent generation module for TCM syndrome differentiation and intervention plans, based on a collaborative learning framework, integrates dynamic TCM syndrome differentiation, risk prediction output, historical treatment records, and real-time monitoring data for data fusion and decision collaboration, generating personalized TCM syndrome differentiation and intervention plans in real time; and utilizes treatment feedback to achieve dynamic data optimization.

[0115] The prediction module takes into account the patient's basic information and treatment monitoring information, and uses a prediction model to assess the patient's OHSS risk.

[0116] This invention constructs a multi-dimensional data input system by integrating quantitative information from the four diagnostic methods of Traditional Chinese Medicine (TCM) with Western medical testing indicators. This system captures individual differences and the evolution of time-series syndromes, improving the comprehensiveness and accuracy of OHSS risk prediction. Based on a priori knowledge base built from TCM syndrome differentiation time-series rules and TCM and Western medicine medication knowledge, classic medical theories, the experience of renowned TCM doctors, and authoritative guidelines are transformed into computable quantitative rules. This achieves synergistic processing of "data-prior knowledge-rules-decision," providing a scientific quantitative basis for TCM syndrome differentiation intervention. Simultaneously, by dynamically calculating the weights of the two models using information entropy, it effectively avoids prediction bias caused by the limitations of data dimensions in a single model, enhancing the stability of risk assessment. The original patient data is encrypted and stored using blockchain, and the model accesses standardized features through a de-identified interface, ensuring data privacy while supporting real-time prediction. This provides high-frequency and accurate feedback for clinical decision-making, improving treatment efficiency and ensuring patient safety.

[0117] The following details specific embodiments of the present invention.

[0118] Example 1

[0119] Reference Figure 1 As shown, a method for predicting OHSS risk in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation includes:

[0120] Clinical data were collected as a dataset and preprocessed, including static data before ovulation induction and dynamic data on the treatment sequence during ovulation induction.

[0121] Establish a prior knowledge base, including a diagnostic time series rule base based on a subset of TCM diagnostic time series data, and a TCM and Western medicine medication knowledge base constructed by integrating medical books, the experience of famous veteran TCM doctors and guidelines.

[0122] Based on clinical data from a subset of datasets, an initial risk prediction model and a time-series risk prediction model are constructed. Through an adaptive weight fusion mechanism driven by model confidence, the prediction results of the two models are dynamically weighted to output a full-cycle fused risk prediction result. This supports daily real-time comprehensive risk assessment during ovulation induction, and the output results provide a quantitative decision-making basis for TCM syndrome differentiation intervention.

[0123] A dual-model bias prediction and early warning mechanism is constructed to avoid misjudgments caused by the limitations of data dimensions in a single model.

[0124] Once deployed in the hospital system, the predictive model can automatically collect patient data and store it in encrypted form on the blockchain. Only authorized models can read standardized features through the desensitized interface, while the original data remains encrypted to ensure usability without visibility and to protect data privacy. At the same time, it outputs OHSS risk prediction results in real time, providing quantitative support for TCM syndrome differentiation and intervention programs.

[0125] As an optional implementation, in the first aspect of the invention, collecting clinical data as a dataset includes:

[0126] A preliminary diagnosis is made by collecting patient information through the four diagnostic methods of traditional Chinese medicine (inspection, auscultation and olfaction, inquiry, and palpation) and mobile terminal devices. The information collected through these four methods includes: observation of spirit, complexion, lip color, tongue color, tongue body, tongue coating, body shape, posture, skin condition, and excretions; normal voice, pathological voice, body odor, secretion odor, excretion odor, and ward odor; basic medical history, symptom description, and questionnaire information; pulse waveform and palpation information. The mobile terminal devices support real-time collection of these four diagnostic methods in remote medical scenarios, such as capturing tongue images via mobile phone camera or wearable devices, recording speech via microphone, and transmitting pulse data via portable pulse diagnostic instruments.

[0127] Western medicine collects the following information: static data before ovulation induction, such as age, BMI, menstrual history, marital and reproductive history, genetic disease history, infertility type, basal antral follicle information, basal sex hormone information, basal blood routine information, basic liver and kidney function information, and blood glucose information; dynamic data during ovulation induction treatment, such as ovulation induction protocol information, Gn medication information, hormone information, trigger drug information, and number of eggs retrieved.

[0128] The aforementioned information collected from both traditional Chinese and Western medicine constitutes a multimodal clinical dataset.

[0129] As an optional implementation, in the first aspect of the invention, the collection and preprocessing of clinical data includes:

[0130] By combining the four diagnostic methods of traditional Chinese medicine—inspection, auscultation and olfaction, inquiry, and palpation—patient information is dynamically collected and transformed into calculable diagnostic data.

[0131] The visual diagnosis information was processed using computer vision technology and the OpenCV library. The intensity of the light reflection in the eyes was obtained by calculating the proportion of bright spots after binarizing the grayscale image. The high-resolution image features of tongue color, tongue body and tongue coating were extracted using CLAHE histogram equalization and U-Net segmentation. The RGB color values ​​of facial color and lip color, body fat index, posture and movement trajectory, skin texture pigment distribution and excretion morphology and color quantification indicators were calculated by computer.

[0132] The system uses a microphone to collect sound and combines a multi-dimensional gas sensor array to collect volatile gases from the patient's body, secretions, excretions, and the ward, extracting voiceprint and odor molecule concentration characteristics.

[0133] The consultation information was collected by gathering basic medical history, symptom descriptions, and questionnaire information, and the symptom descriptions were standardized and scored using the "Traditional Chinese Medicine Symptom Quantitative Scoring Table".

[0134] Palpation information is obtained by collecting pulse waveform data from the three pulse positions (cun, guan, chi) using standard instruments, such as pulse rate, pulse pressure difference, waveform characteristic parameters, and pressure-displacement feedback data from palpation, such as pressure value changes at different positions, which are then converted into a calculable temporal feature vector.

[0135] The collected data is standardized to form a dataset for model training, including:

[0136] Sample alignment and interpolation: For the missing E2 value on day 3 of ovulation induction, a two-way interpolation method was used to fill the gap in the data within 48 hours before and after the induction. The calculation was as follows: assuming 1200 pg / mL on day 2 and 1800 pg / mL on day 4, the missing E2 value = 1200 + (1800 - 1200) / 2 = 1500 pg / mL.

[0137] Z-score standardization performs feature normalization: based on statistical principles, it maps the original data to a standard normal distribution space with a mean of 0 and a standard deviation of 1. Its calculation is as follows: Where X is the original feature value, μ is the mean of this feature in the historical dataset, and σ is the standard deviation. Taking the age feature as an example, assuming the historical dataset contains 1000 assisted reproductive cases, the estimated average age μ is 31.2 years, and the standard deviation σ is 3.5, if the age X is 30 years, then the standardized value is (30-31.2) / 3.5 = -0.34.

[0138] Using the ovulation induction start date as the time origin, dynamic data were aligned and a dialectical time series matrix was constructed: taking the ovulation induction start date as the time origin, a dialectical time series matrix of time points was constructed in chronological order, with each time point containing characteristics of both traditional Chinese and Western medicine dimensions.

[0139] As an optional implementation, in the first aspect of the present invention, the establishment of the prior knowledge base includes:

[0140] Syndrome Differentiation Time Sequence Rule Base: Natural language processing is used to analyze the symptom quantitative scoring sequences of historical cases to identify syndrome labels such as "Yin Deficiency Syndrome" and "Phlegm-Dampness Syndrome," and syndrome evolution vectors are generated according to time windows ("Yin Deficiency turning into Yin Deficiency with Excessive Fire"). The Apriori algorithm is used to mine association rules, with a support of ≥30% and a confidence of ≥80%. The rule "Yellow and greasy tongue coating score ≥3 points + pulse rate ≥85 beats / min from day 5 to day 7 of ovulation induction, indicating high risk" is selected and stored in the "Syndrome Differentiation Evolution Rule Table."

[0141] Traditional Chinese Medicine and Western Medicine Drug Knowledge Base: The Traditional Chinese Medicine knowledge base is built on the Neo4j graph database. Nodes include Chinese herbs (e.g., "Astragalus membranaceus"), efficacy (e.g., "Qi tonification"), and syndromes (e.g., "Spleen deficiency syndrome"). Edge relationships are "Astragalus membranaceus - Indications / Efficacy - Spleen deficiency syndrome," and it stores empirical formula compatibility rules. The Western Medicine knowledge base is built on a MySQL database and stores dose-risk mapping relationships for ovulation-inducing drugs (e.g., risk level +1 when daily dose > 300 IU), trigger drug contraindications (e.g., contraindicated in patients with HCG allergy), and other guideline recommendations.

[0142] As an optional implementation, in the first aspect of the present invention, the construction of the initial risk prediction model and the time-series risk prediction model includes:

[0143] Initial risk prediction model: A subset of static data before ovulation induction is input, and Western medicine baseline indicators are fused with traditional Chinese medicine (TCM) syndrome differentiation features to form a multi-dimensional input matrix. The model is trained using the XGBoost algorithm, which embeds "syndrome differentiation-risk" prior rules into the objective function and introduces a knowledge base regularization term to enhance TCM syndrome differentiation. The final output is the OHSS risk probability before ovulation induction and intervention recommendations based on the syndrome differentiation results.

[0144] Taking a 30-year-old female patient as an example, undergoing assisted reproductive technology, with personalized variable characteristics: Age (A): 30 years old, BMI: 26 kg / m² 2 AFC: 12, FSH: 7.2 mIU / mL, ALT: 28 U / L, WBC: 5.204 × 10⁻⁶ 9 / L, serum creatinine: 68μmol / L, risk prediction was performed.

[0145] Western medicine baseline indicators: such as age (A): 30 years old, BMI: 26 kg / m² 2 AFC: 12, FSH: 7.2 mIU / mL, ALT: 28 U / L, WBC: 5.204 × 10⁻⁶ 9 / L, serum creatinine: 68μmol / L; Western medicine baseline indicators extracted: such as age (A)x1=-0.34, BMI (B)x2=2.28, AFCx3=0.8, FSHx4=0.3, ALTx5=-0.1, menstrual historyx6 “oligomenorrhea” one-hot code is [0,0,1], infertility typex7 “tubal infertility” corresponding to one-hot code is [0,0,1], genetic historyx8 polycystic ovary syndrome family history G=1, marital and reproductive historyx9 “first marriage without children” corresponding to one-hot code is [1,0], WBCx 10 =-1.08, serum creatinine x 11 = -0.875;

[0146] Traditional Chinese Medicine diagnostic characteristics: Inspection: Measured value of light reflection intensity in the eyes is 65%; Facial complexion RGB values: Measured values ​​R=220, G=180, B=150; Tongue color, tongue body, and tongue coating thickness characteristics T=0.72; Body mass index (standard normal range is 18.5-24) BMI: 26 kg / m² 2Auscultation: Voiceprint characteristics [0.25, 0.18, 0.32]; Odor molecule concentration characteristics 2.3ppm; Inquiry: Based on the "Traditional Chinese Medicine Symptom Quantification Scoring Table", the patient's chief complaint was "dry mouth and tongue 4 points" and "five-center heat 3 points"; Palpation: Pulse rate (HR): 78 beats / min, pulse pressure difference (PP): 42mmHg, waveform characteristic parameters (W): the main peak amplitude ratio extracted by time domain analysis was 0.65; the slope of the pressure-displacement curve of the palpation was 0.8N / mm; After the TCM syndrome differentiation characteristics were extracted, they were converted into numerical vectors by Z-score standardization: Inspection: The original value of the light reflection intensity of the eyes was 65%, the historical mean was 60%, and the standard deviation was 10%, so the standardized value y1 = 0.65, indicating that the eyes were bright, the proportion of reflected light spots was high, the spirit was sufficient, and the pathogenic factors were still shallow; The RGB values ​​of the face were R = 0.86, G = 0.71, B = 0.59, indicating that the R value was high and the face was red, indicating heat syndrome or yin deficiency. The patient presents with excessive internal heat; a slightly yellow B-value, indicating dampness; tongue color, tongue body, and tongue coating thickness (y4 = 0.72) suggest a thick tongue coating (normal T < 0.5), indicating internal phlegm and dampness; combined with a red complexion, this suggests phlegm and dampness transforming into heat; body shape index (y5 = 0.91) indicates a plump physique, belonging to the "phlegm-dampness constitution" in Traditional Chinese Medicine, prone to generating dampness and turbidity, which matches the thick, greasy tongue coating; auscultation: voiceprint characteristics [1.0, -0.5, 0.67], indicating the second dimension is -0.5 (below the mean), indicating poor voice quality. The slightly low odor may be related to phlegm-dampness obstructing the flow of Qi; the odor molecule concentration characteristic of 0.46 indicates a slightly abnormal breath, between normal and noticeably sour and putrid, mainly indicating damp-heat in the spleen and stomach or food stagnation; inquiries revealed dry mouth and tongue, and five-center heat (y7 = 0.7), suggesting Yin deficiency with excessive fire, forming a syndrome chain with symptoms such as a reddish complexion and a hot tongue coating; palpation revealed a pulse rate of y8 = 1.0 and a pulse pressure difference of y9 = 1.0, indicating a fast pulse rate and increased pulse pressure difference, mainly indicating a heat syndrome; waveform characteristic parameters y 10 =0.65, indicating: W>0.5, consistent with the characteristics of a slippery pulse, mainly indicating phlegm-dampness or phlegm-heat; pressure-displacement characteristics during palpation y 11 =1.5, indicating reduced tissue elasticity and high resistance when pressed, mainly based on empirical evidence;

[0147]

[0148] Multidimensional input matrix: If the dimensions of Western medicine indicator vectors and traditional Chinese medicine syndrome differentiation feature vectors are inconsistent, zero-padding is used to make them the same dimension to ensure the feasibility of matrix splicing. The XGBoost model learns through training that the combination of "T≥0.6 + BMI>25 + fast pulse rate + reddish complexion" is strongly associated with high risk. Assuming the model contains one decision tree, the splitting rule is: if the standardized BMI value > 2.0 (corresponding to BMI > 25), proceed to the right node; the right node then splits according to the tongue appearance T value: if T≥0.6, the leaf node score = 0.6; after sigmoid transformation: 1 / (1+e -0.6 The initial risk prediction probability is 0.65, approximately 0.65.

[0149] The time-series risk prediction model is constructed by inputting a subset of dynamic time-series data from ovulation induction treatment, and building a three-dimensional tensor that includes TCM syndrome differentiation time-series characteristics and Western medicine monitoring indicators. Using an LSTM+attention mechanism architecture, key time-series features are automatically focused through attention weights to establish a time-series risk prediction model. A dual-pathway correction mechanism is introduced, and combined with prior rules and a knowledge base, daily real-time comprehensive risk assessment and dynamic TCM and Western medicine intervention suggestions are made, including adjustments to TCM prescriptions and optimization of Western medicine dosages. The two are collaboratively decided through a multi-objective optimization algorithm.

[0150] Inputting time-series characteristic data from day 5 of ovulation induction (tongue coating yellow and greasy score, pulse rate, E2, etc.), the data was analyzed using an LSTM combined with an attention mechanism model. The model, through adaptive weight allocation, focused 45% of its attention weight on two key features: "trend of change in yellow and greasy tongue coating" and "sudden increase in E2 level," outputting a 45% OHSS risk probability for that day. Simultaneously, personalized intervention suggestions were generated: adjusting the dosage of Gn medication and combining it with traditional Chinese medicine formulas for clearing heat and promoting diuresis as adjunctive treatment.

[0151] Dual-pathway correction mechanism: Focusing on the internal optimization of the time series model, the model parameters are dynamically updated and the intervention strategy is optimized through dialectical feedback of the deviation between model prediction and actual situation and intervention learning of historical intervention effects to feed back to the model parameters.

[0152] Model confidence-driven adaptive weight fusion mechanism: Wherein, W1 is the initial risk prediction model weight, W2 is the time-series risk prediction model weight, C1 is the confidence level of the initial risk prediction model (e.g., the model's accuracy on the validation set), reflecting the model's reliability in predicting baseline characteristics before ovulation induction, and C2 is the confidence level of the time-series risk prediction model (e.g., the model's accuracy on the validation set), reflecting the model's reliability in predicting changes in syndrome differentiation during treatment.

[0153] Let the accuracy of the initial risk prediction model on the validation set be C1 = 0.85, and the accuracy of the time series risk prediction model on the validation set be C2 = 0.90. Then...

[0154] Taking a 30-year-old female patient undergoing assisted reproductive technology (ART) as an example, with personalized variable features: the initial risk prediction model achieved an accuracy of C1 = 0.85 on the validation set, and the time-series risk prediction model achieved an accuracy of C2 = 0.90 on the validation set, performing a full-cycle fusion risk prediction. Using... Formula calculation, The initial risk probability R1 = 0.65, the time-series risk probability R2 = 0.45, and the full-cycle fusion risk = 0.486 × 0.65 + 0.514 × 0.45 ≈ 0.3159 + 0.2313 = 0.5472.

[0155] As an optional implementation, in the first aspect of the present invention, the construction of the dual-model deviation prediction and early warning mechanism includes:

[0156] Calculate the difference in OHSS risk probability between the two models: ΔR = |R1 - R2|, where R1 is the OHSS risk probability value output by the initial risk prediction model and R2 is the OHSS risk probability value output by the time series risk prediction model.

[0157] When ΔR > 0.2 and |Risklevel(R1) - Risklevel(R2)| ≥ 1, a Level 3 warning is triggered.

[0158] Risk level classification: Risk level classification is based on historical data statistics and clinical experience. In practical applications, parameters can be adjusted to suit different populations. For example, low risk: probability < 0.3 (Level 1); medium risk: 0.3 ≤ probability ≤ 0.6 (Level 2); high risk: probability > 0.6 (Level 3).

[0159] Level 1 warning: A warning pop-up window is pushed through the hospital information system, prompting "The model prediction shows a significant difference. Please review whether any concurrent symptoms have been missed in the TCM diagnosis"; Level 2 warning: The knowledge base is linked to generate a list of potential concurrent symptoms and key points for diagnosis; Level 3 warning: The TCM and Western medicine intervention plans of similar historical medical records are automatically retrieved for reference.

[0160] If the difference in risk probability between the two models is greater than 0.2 on day 7, the system will trigger a secondary early warning mechanism: automatically retrieve the diagnostic criteria for "phlegm-dampness syndrome with yin deficiency" from the knowledge base, generate a list of potential concurrent syndrome diagnostic suggestions, and push it to the clinical decision support system. After the physician confirms the diagnosis by combining tongue diagnosis images, if the patient meets the syndrome characteristics of "phlegm-dampness with heat", the system recommends using "Cangfu Daotan Wan combined with Zhibai Dihuang Wan" as the base formula, and then adjusts the traditional Chinese medicine prescription accordingly. For example, if the initial risk is 65% (high risk), add 15g of Poria cocos and 6g of Rehmannia glutinosa to the base formula; if the risk drops to 45% (medium risk), reduce Phellodendron chinense to 6g.

[0161] Example 2

[0162] This embodiment provides an assisted reproductive OHSS risk prediction system based on traditional Chinese medicine syndrome differentiation, which corresponds to the artificial intelligence method in Embodiment 1, and combines... Figure 2 As shown, through multimodal data acquisition, temporal feature modeling, and the integration of traditional Chinese and Western medicine knowledge, efficient and accurate prediction of module operation is achieved, including:

[0163] The data acquisition and preprocessing module is compatible with mobile terminal data interfaces. It collects basic patient data, treatment monitoring data, and other clinical data, and filters, processes, and unifies the collected clinical data to adapt to real-time data transmission and analysis.

[0164] The dataset generation module divides the original dataset into two subsets: a static data subset before ovulation induction and a dynamic data subset based on the treatment timeline during ovulation induction.

[0165] The prior knowledge base construction module includes a syndrome differentiation time series rule base built based on a subset of TCM syndrome differentiation time series data, and a TCM and Western medicine medication knowledge base built by integrating medical books, the experience of famous veteran TCM doctors and guidelines.

[0166] The model building module builds a risk prediction model based on the collected clinical data by integrating static data before ovulation induction and dynamic data of ovulation induction treatment.

[0167] The model training and validation module trains the model based on collected clinical data, using static patient data and dynamic treatment timeline data to achieve pre-ovulation induction prediction and daily real-time comprehensive prediction during ovulation induction. It validates the timeliness and accuracy of risk prediction, generating the final deployable risk prediction model component.

[0168] The intelligent generation module for TCM syndrome differentiation and intervention plans, based on a collaborative learning framework, integrates dynamic TCM syndrome differentiation, risk prediction output, historical treatment records, and real-time monitoring data for data fusion and decision collaboration, generating personalized TCM syndrome differentiation and intervention plans in real time; and utilizes treatment feedback to achieve dynamic data optimization.

[0169] The prediction module takes into account the patient's basic information and treatment monitoring information, and uses a prediction model to assess the patient's OHSS risk.

[0170] As an optional implementation, unified deployment and security include:

[0171] It supports multiple terminals, including mobile devices, smart medical devices, hospital systems, etc., and enables real-time data transmission through RESTful API interfaces.

[0172] Deployed on an NIDIA DGX A100 server cluster, using the TensorFlow framework and equipped with high-performance GPUs and accelerators, it improves the computational efficiency of the model, shortens the daily risk prediction response time to ≤30 seconds, and enables data preprocessing, model training and real-time inference.

[0173] Patient clinical data is encrypted and stored using Hyperledger Fabric blockchain technology. Blockchain nodes are deployed on internal hospital servers and are accessible only to doctors and researchers via digital certificates.

[0174] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting the risk of OHSS in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation, characterized by: It includes the following steps: (1) Collect clinical data as a dataset and preprocess it. The static data before ovulation induction and the dynamic data of treatment time sequence during ovulation induction are both subsets of this dataset. (2) Establish a prior knowledge base, including a diagnostic time series rule base based on the TCM diagnostic time series data of the subset dataset, and a TCM and Western medicine drug knowledge base constructed by integrating medical books, the experience of famous veteran TCM doctors and guidelines. (3) Based on clinical data from a subset of datasets, an initial risk prediction model and a time-series risk prediction model are constructed. Through an adaptive weight fusion mechanism driven by model confidence, the prediction results of the two models are dynamically weighted and the full-cycle fusion risk prediction results are output to support daily real-time comprehensive risk assessment during ovulation induction. The output results provide quantitative decision-making basis for TCM syndrome differentiation intervention. (4) Construct a dual-model bias prediction and early warning mechanism to avoid misjudgment caused by the limitation of data dimensions in a single model; (5) After the prediction model is deployed in the hospital system, it automatically collects patient data and stores the collected patient data in blockchain encryption to ensure data privacy. At the same time, it outputs OHSS risk prediction results in real time to provide quantitative support for TCM syndrome differentiation and intervention programs.

2. The method for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 1, characterized in that: In step (1), patient information is collected for preliminary diagnosis using the four diagnostic methods of traditional Chinese medicine (inspection, auscultation, inquiry, and palpation) and mobile terminal devices. The information collected by the four diagnostic methods includes the following: observation of spirit, facial complexion, lip color, tongue color, tongue body, tongue coating, body shape, posture, skin, and excretions; normal voice, pathological voice, body odor, secretion odor, excretion odor, and ward odor. Basic medical history information, symptom description information, and questionnaire information; Pulse waveform information and palpation information; Western medicine collects the following information: static data before ovulation induction, including: age, BMI, menstrual history, marital and reproductive history, genetic disease history, type of infertility, basal antral follicle information, basal sex hormone information, basal blood routine information, basic liver and kidney function information, and blood glucose information; dynamic data during ovulation induction treatment, including: ovulation induction protocol information, Gn medication information, hormone information, trigger drug information, and number of oocytes retrieved.

3. The method for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 2, characterized in that: In step (1), the visual diagnosis information is processed using the OpenCV library with computer vision technology. The intensity of the eye light reflection is obtained by calculating the proportion of bright spot pixels after the grayscale image is binarized. The high-definition image features of tongue color, tongue body and tongue coating are extracted by CLAHE histogram equalization + U-Net segmentation. The RGB color values ​​of face color and lip color, body fat index, posture movement trajectory, skin texture pigment distribution and excretion morphology and color quantitative indicators are calculated by computer. Auscultation information is obtained by collecting sound through a microphone and combining it with a multi-dimensional gas sensor array to collect volatile gases from the patient's body, secretions, excrement, and ward, and extracting voiceprint and odor molecule concentration characteristics; The consultation information is collected by gathering basic medical history, symptom descriptions and questionnaire information, and the symptom descriptions are standardized and scored using a TCM symptom quantitative scoring scale. Palpation information is obtained by collecting pulse waveform data at the cun, guan, and chi positions using standard instruments, as well as pressure-displacement feedback data from palpation, and converting them into a computable temporal feature vector. The collected data were standardized, including using bidirectional interpolation to fill missing values, Z-score standardization to normalize features, aligning dynamic data with the ovulation induction start date as the time origin, and constructing a dialectical time series matrix to form a dataset for model training.

4. The method for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 3, characterized in that: In step (2), the prior knowledge base includes: The syndrome differentiation time-series rule base, based on symptom description data in a subset, standardizes symptom descriptions using a TCM symptom quantitative scoring table and generates symptom feature sequences according to treatment time order. Natural language processing (NLP) techniques are used to parse and match syndrome criteria to extract syndrome differentiation features. The syndrome differentiation sequences are converted into numerical codes and generated evolution vectors according to time windows. The Apriori algorithm is then used to mine syndrome differentiation time-series association rules. The threshold support of the syndrome differentiation time-series rule base is ≥30%, and the confidence level is ≥80%. Validation based on historical data ensures that the rules are statistically significant. After selecting effective rules, they are structured and stored to generate a syndrome differentiation evolution rule table, forming prior rules for syndrome differentiation risk. These prior probabilities are then superimposed when the model outputs risk. The Traditional Chinese Medicine (TCM) and Western Medicine knowledge base includes: collecting classic medical texts, the medication experience of renowned veteran TCM doctors, medication methods and contraindications; using entity recognition and relation extraction algorithms to store extracted entities as nodes and relations as edges; and using the graph database Neo4j to build the TCM knowledge base. The Western Medicine knowledge base includes: collecting medical guidelines issued by authoritative institutions, establishing a mapping relationship between indicators and ovulation induction protocols, common ovulation induction drugs' usage and contraindications, and prioritizing recommendations; and using the relational database MySQL to build the Western Medicine knowledge base, which is updated regularly.

5. The method for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 4, characterized in that: In step (3), the initial risk prediction model and the time-series risk prediction model include: Initial risk prediction model: Input a subset of static data before ovulation induction, integrate Western medicine baseline indicators and TCM syndrome differentiation characteristics to form a multi-dimensional input matrix, use the XGBoost algorithm with TCM syndrome differentiation enhancement that embeds syndrome differentiation risk prior rules into the objective function and introduces knowledge base regularization terms to train the model, and finally output the OHSS risk probability before ovulation induction and intervention suggestions based on syndrome differentiation results; The time-series risk prediction model is constructed by inputting a subset of dynamic time-series data of ovulation induction treatment, and building a three-dimensional tensor containing the time-series characteristics of TCM syndrome differentiation and Western medicine monitoring indicators. The model adopts an LSTM + attention mechanism architecture, automatically focuses on key time-series features through attention weights, establishes a time-series risk prediction model, and introduces a dual-pathway correction mechanism. Combined with prior rules and knowledge base, it conducts daily real-time comprehensive risk assessment and dynamic intervention suggestions of TCM and Western medicine, including the adjustment of TCM prescriptions and the optimization of Western medicine dosages. The two are collaboratively decided through a multi-objective optimization algorithm. Dual-pathway correction mechanism: Focusing on the internal optimization of the time series model, the model parameters are dynamically updated and the intervention strategy is optimized through dialectical feedback of the deviation between model prediction and actual dialectical analysis and intervention learning of historical intervention effects. Model confidence-driven adaptive weight fusion mechanism: Wherein, W1 is the initial risk prediction model weight, W2 is the time-series risk prediction model weight, C1 is the confidence level of the initial risk prediction model, reflecting the reliability of the model in predicting baseline characteristics before ovulation induction, and C2 is the confidence level of the time-series risk prediction model, reflecting the reliability of the model in predicting changes in syndrome differentiation during treatment.

6. The method for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 5, characterized in that: In step (4), the construction of the dual-model deviation prediction and early warning mechanism includes: Calculate the difference in OHSS risk probability between the two models: ΔR = |R1 - R2|, where R1 is the OHSS risk probability value output by the initial risk prediction model and R2 is the OHSS risk probability value output by the time series risk prediction model. When ΔR>0.2 and |Risklevel(R1)-Risklevel(R2)|≥1, a three-level warning is triggered: Level 1 warning: A warning pop-up window is pushed through the hospital information system, indicating that the model prediction difference is significant. Please check whether the TCM syndrome differentiation has missed any concurrent syndromes. Level 2 warning: The system will link with the knowledge base to generate a list of potential concurrent evidence and key points for differentiation. Level 3 early warning: Automatically retrieves traditional Chinese and Western medicine intervention plans from similar historical medical records for reference.

7. A device for predicting the risk of OHSS in assisted reproductive technology based on traditional Chinese medicine syndrome differentiation, characterized in that: It includes: The data acquisition and preprocessing module is configured to acquire clinical data as a dataset and perform preprocessing. The static data before ovulation induction and the dynamic data of the treatment timeline during ovulation induction are both subsets of this dataset. The prior knowledge base construction module is configured to build a prior knowledge base, including a syndrome differentiation time series rule base built based on a subset of TCM syndrome differentiation time series data, and a TCM and Western medicine drug knowledge base built by integrating medical books, the experience of famous veteran TCM doctors and guidelines. The model building module is configured to construct an initial risk prediction model and a time-series risk prediction model based on clinical data from a subset of datasets. Through an adaptive weight fusion mechanism driven by model confidence, the prediction results of the two models are dynamically weighted to output the full-cycle fused risk prediction results, thereby supporting daily real-time comprehensive risk assessment during ovulation induction. Its output results provide quantitative decision-making basis for TCM syndrome differentiation intervention. The mechanism construction module is configured to build a dual-model bias prediction and early warning mechanism to avoid misjudgments caused by the limitation of data dimensions in a single model; The prediction module, configured to predict the model after deployment in the hospital system, automatically collects patient data and stores the collected patient data in blockchain encryption to ensure data privacy. At the same time, it outputs OHSS risk prediction results in real time, providing quantitative support for TCM syndrome differentiation and intervention programs.

8. The device for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 7, characterized in that: In the data acquisition and preprocessing module, patient information is collected for preliminary diagnosis using the four diagnostic methods of traditional Chinese medicine (inspection, auscultation, inquiry, and palpation) and mobile terminal devices. The information collected by the four diagnostic methods includes: observation of spirit, facial complexion, lip color, tongue color, tongue body, tongue coating, body shape, posture, skin, and excretions; normal voice, pathological voice, body odor, secretion odor, excretion odor, and ward odor. Basic medical history information, symptom description information, and questionnaire information; Pulse waveform information and palpation information; The information collected by Western medicine includes the following: static data collected before ovulation induction, including: age, BMI, menstrual history, marital and reproductive history, genetic disease history, infertility type, basal antral follicle information, basal sex hormone information, basal blood routine information, basic liver and kidney function information, and blood glucose information; dynamic data collected during ovulation induction treatment, including: ovulation induction protocol information, Gn medication information, hormone information, trigger drug information, and number of oocytes retrieved. The visual diagnosis information was obtained by acquiring facial and full-body images of the patient using standard instruments. The images were processed using the OpenCV library with computer vision technology. The intensity of the light reflection in the eyes was obtained by calculating the proportion of bright spots after binarization of the grayscale image. The high-resolution image features of tongue color, tongue body and tongue coating were extracted using CLAHE histogram equalization and U-Net segmentation. The RGB color values ​​of facial color and lip color, body fat index, posture and movement trajectory, skin texture pigment distribution and excretion morphology and color quantitative indicators were calculated by computer. Auscultation information is obtained by collecting sound through a microphone and combining it with a multi-dimensional gas sensor array to collect volatile gases from the patient's body, secretions, excrement, and ward, and extracting voiceprint and odor molecule concentration characteristics; The consultation information is collected by gathering basic medical history, symptom descriptions and questionnaire information, and the symptom descriptions are standardized and scored using a TCM symptom quantitative scoring scale. Palpation information is collected using standard instruments to capture pulse waveform data at the cun, guan, and chi positions, as well as pressure displacement feedback data during palpation, and then converted into a computable temporal feature vector. The collected data were standardized, including using bidirectional interpolation to fill missing values, Z-score standardization to normalize features, aligning dynamic data with the ovulation induction start date as the time origin, and constructing a dialectical time series matrix to form a dataset for model training.

9. The device for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 8, characterized in that: The prior knowledge base construction module includes: The syndrome differentiation time-series rule base is based on symptom description data in a subset of data. It standardizes symptom descriptions using a TCM symptom quantitative scoring table and generates symptom feature sequences according to treatment time. Natural language processing (NLP) techniques are used to parse and match syndrome criteria to extract syndrome differentiation features. The syndrome differentiation sequences are converted into numerical codes and generated evolution vectors according to time windows. The Apriori algorithm is then used to mine syndrome differentiation time-series association rules. The thresholds for the syndrome differentiation time-series rule base are support ≥30% and confidence ≥80%. Historical data is used for validation to ensure statistical significance. After selecting effective rules, they are structured and stored to generate a syndrome differentiation evolution rule table, forming prior rules for syndrome differentiation risk. These prior probabilities are then superimposed when the model outputs risk. The Traditional Chinese Medicine (TCM) and Western Medicine knowledge base includes: collecting classic medical texts, the medication experience of renowned veteran TCM doctors, medication methods and contraindications; using entity recognition and relation extraction algorithms to store extracted entities as nodes and relations as edges; and using the graph database Neo4j to build the TCM knowledge base. The Western Medicine knowledge base includes: collecting medical guidelines issued by authoritative institutions, establishing a mapping relationship between indicators and ovulation induction protocols, common ovulation induction drugs' usage and contraindications, and prioritizing recommendations; and using the relational database MySQL to build the Western Medicine knowledge base, which is updated regularly.

10. The device for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 9, characterized in that: The model building module includes an initial risk prediction model and a time-series risk prediction model, which are: Initial risk prediction model: Input a subset of static data before ovulation induction, integrate Western medicine baseline indicators and TCM syndrome differentiation characteristics to form a multi-dimensional input matrix, use the XGBoost algorithm with TCM syndrome differentiation enhancement that embeds syndrome differentiation risk prior rules into the objective function and introduces knowledge base regularization terms to train the model, and finally output the OHSS risk probability before ovulation induction and intervention suggestions based on syndrome differentiation results; The time-series risk prediction model is constructed by inputting a subset of dynamic time-series data of ovulation induction treatment, and building a three-dimensional tensor containing the time-series characteristics of TCM syndrome differentiation and Western medicine monitoring indicators. The model adopts an LSTM + attention mechanism architecture, automatically focuses on key time-series features through attention weights, establishes a time-series risk prediction model, and introduces a dual-pathway correction mechanism. Combined with prior rules and knowledge base, it conducts daily real-time comprehensive risk assessment and dynamic intervention suggestions of TCM and Western medicine, including the adjustment of TCM prescriptions and the optimization of Western medicine dosages. The two are collaboratively decided through a multi-objective optimization algorithm. Dual-pathway correction mechanism: Focusing on the internal optimization of the time series model, the model parameters are dynamically updated and the intervention strategy is optimized through dialectical feedback of the deviation between model prediction and actual dialectical analysis and intervention learning of historical intervention effects. Model confidence-driven adaptive weight fusion mechanism: Wherein, W1 is the initial risk prediction model weight, W2 is the time-series risk prediction model weight, C1 is the confidence level of the initial risk prediction model, reflecting the reliability of the model in predicting baseline characteristics before ovulation induction, and C2 is the confidence level of the time-series risk prediction model, reflecting the reliability of the model in predicting changes in syndrome differentiation during treatment.

11. The device for predicting OHSS risk in assisted reproduction based on traditional Chinese medicine syndrome differentiation according to claim 10, characterized in that: The mechanism construction module includes the following steps for constructing a dual-model bias prediction and early warning mechanism: Calculate the difference in OHSS risk probability between the two models: ΔR = |R1 - R2|, where R1 is the OHSS risk probability value output by the initial risk prediction model and R2 is the OHSS risk probability value output by the time series risk prediction model. When ΔR>0.2 and |Risklevel(R1)-Risklevel(R2)|≥1, a three-level warning is triggered: Level 1 warning: A warning pop-up window is pushed through the hospital information system, indicating that the model prediction difference is significant. Please check whether the TCM syndrome differentiation has missed any concurrent syndromes. Level 2 warning: The system will link with the knowledge base to generate a list of potential concurrent evidence and key points for differentiation. Level 3 early warning: Automatically retrieves traditional Chinese and Western medicine intervention plans from similar historical medical records for reference.

Images