A DOR population infertility risk assessment system based on multi-modal data

Abstract

This invention discloses a multimodal data-based infertility risk assessment system for the DOR (Disease of Infertility) population, including modules for data acquisition and preprocessing, feature screening and extraction, prediction information generation, and model training and validation. The system acquires clinical baseline data, Western medicine auxiliary examination data, traditional Chinese medicine four diagnostic methods and syndrome scales, and acupoint infrared thermography temperature data. Through univariate and multivariate Cox regression screening, age, antral follicle count, uterine coldness score, white and greasy tongue coating, thready pulse, and uterine acupoint temperature are selected and input into a trained XGBoost model to generate a predicted probability of natural pregnancy and a risk value or risk level for not achieving natural pregnancy, providing auxiliary information for infertility risk assessment in the DOR population.

Description

Technical Field

[0001] This invention relates to the fields of medical information technology and intelligent data processing technology, specifically to a system for preprocessing, feature screening, and machine learning analysis of multimodal data of individuals with diminished ovarian reserve (DOR) to generate pregnancy outcome prediction information within a preset follow-up period. Background Technology

[0002] Ovarian reserve reflects the quantity and quality of remaining follicles in a woman's ovaries, as well as the ovary's responsiveness to gonadotropin stimulation. Diminished ovarian reserve (DOR) refers to a reproductive dysfunction characterized by a reduced number of usable follicles and / or decreased oocyte quality, but without reaching ovarian failure. Individuals with DOR typically experience decreased fertility, prolonged time to conception, and increased uncertainty in pregnancy outcomes. Therefore, objectively analyzing the likelihood of pregnancy within a specific timeframe for individuals with DOR using available clinical data is a crucial issue that needs to be addressed in the field of reproductive medicine information processing.

[0003] Currently, the evaluation of ovarian reserve function typically utilizes clinical data and auxiliary Western medical examinations, including age, anti-Müllerian hormone (AMH), basal follicle-stimulating hormone (FSH), estradiol, and antral follicle count. AMH and FSH primarily reflect the number of follicles, while reproductive endocrine indicators such as FSH reflect ovarian function from different perspectives. While these indicators provide some basis for evaluating ovarian reserve function, ovarian reserve function is not entirely equivalent to actual pregnancy outcomes. Different individuals with similar ovarian reserve indicators may still experience differences in pregnancy outcomes within the same follow-up period. Relying solely on a single indicator, a fixed threshold, or a limited number of Western medical indicators is insufficient to fully characterize the correlation between factors such as age, reproductive endocrine status, follicle count, overall physiological state, and local temperature characteristics.

[0004] In clinical diagnosis and treatment, in addition to routine clinical data and Western medicine auxiliary examination data, we can also obtain tongue appearance, pulse appearance, TCM syndrome inquiry scales, uterine coldness scales, and infrared thermal imaging data of specific areas or acupoints of the body. Tongue appearance, pulse appearance, and inquiry scale data can reflect the overall condition of the subject from different dimensions, while infrared thermal imaging data can reflect the temperature distribution characteristics of the corresponding body surface area. The above data have different sources than information such as age and ovarian reserve indicators, and their data types, dimensions, value ranges, and representational meanings also differ, thus constituting multi-source heterogeneous clinical data.

[0005] However, in existing technologies, data from these different sources are typically stored and analyzed separately, lacking unified mechanisms for sample identification, data encoding, handling of missing values, dimensional transformation, and feature association. For example, age, hormone levels, antral follicle counts, and acupoint temperatures are continuous data, while tongue coating and pulse characteristics are usually categorical data, and syndrome scales may present as ordinal scoring data. If these data are directly used for statistical analysis or model training, differences in dimensions, missing values, and encoding methods between different data points may affect the analysis results, making it difficult to form stable and reproducible model inputs.

[0006] Furthermore, clinical data from individuals with diminished ovarian reserve typically contain a large number of candidate variables, which may be correlated or have overlapping information. Directly inputting all variables into the predictive model would not only increase model complexity but also introduce redundant information, making the model susceptible to variations in sample size, data partitioning, and random factors. Conversely, pre-selecting only a few variables based on human experience may miss features relevant to pregnancy outcomes, and the selections by different operators may be inconsistent. Therefore, it is necessary to obtain a set of features stably associated with pregnancy outcomes from multiple data modalities through a repeatable feature screening and testing process.

[0007] Current methods for analyzing pregnancy outcomes largely rely on traditional regression models. While these models can analyze the statistical associations between variables and outcomes, they typically require pre-defining the functional relationship between variables and outcomes. When there are non-linear relationships or interactions between age, follicle count, TCM syndrome characteristics, and acupoint temperature characteristics, traditional linear models may struggle to fully extract the relevant information. Although machine learning methods can handle more complex variable relationships, without proper data preprocessing, feature selection, model training, and performance validation, problems such as overfitting, unstable model output, or uninterpretable prediction results may still occur.

[0008] On the other hand, some existing predictive models are mainly aimed at identifying ovarian reserve function, assisted reproductive treatment outcomes, or pregnancy classification at a single time point. The data sources and output targets used by these models differ, making it difficult to directly use them to generate pregnancy outcome prediction information for individuals with diminished ovarian reserve within a preset follow-up period based on multi-source data from both traditional Chinese and Western medicine. Furthermore, there is a lack of a unified data analysis mechanism for processing and outputting the probability of pregnancy for the same individual at different follow-up time points.

[0009] In summary, existing technologies have at least the following shortcomings: a lack of unified data processing and feature representation methods for clinical data from different sources; the potential presence of redundant information and correlations among numerous candidate variables, making it difficult to obtain stable model input features; insufficient utilization of the nonlinear relationships and interactions between multimodal features by traditional analysis methods; and a lack of a computer-based system capable of performing machine learning analysis on screened multimodal features from both traditional Chinese and Western medicine, and generating pregnancy outcome prediction information within a preset follow-up period. Therefore, it is necessary to provide a DOR (Disease of Infertility) risk assessment system based on multimodal data to improve the consistency of processing multi-source heterogeneous clinical data, the reproducibility of the prediction process, and the objectivity of the prediction information. Summary of the Invention

[0010] The purpose of this invention is to provide a DOR population infertility risk assessment system based on multimodal data, in order to solve the problems in the prior art such as the difficulty in uniformly processing clinical data from different sources, the existence of information redundancy in candidate features, the inability of traditional analysis methods to fully utilize the nonlinear correlation between different modal features, and the lack of stability and repeatability in the pregnancy outcome prediction process.

[0011] To achieve the above objectives, in a first aspect of the present invention, a DOR (Disease of Infertility) risk assessment system based on multimodal data is provided, the system comprising: The multimodal data acquisition and preprocessing module is used to acquire multimodal data of people with diminished ovarian reserve, and to perform sample alignment, missing value processing, variable encoding and standardization on the multimodal data; the multimodal data includes at least clinical baseline data, Western medicine auxiliary examination data, traditional Chinese medicine four diagnostic methods and syndrome scale data, and acupoint infrared thermal imaging temperature data. The feature screening and extraction module is connected to the multimodal data acquisition and preprocessing module. It is used to perform univariate and multivariate analysis on the preprocessed multimodal data, screen predictors related to pregnancy outcomes from candidate features, and perform collinearity test and proportional hazards hypothesis test on the screened predictors to generate a target feature set for model input. The prediction information generation module, connected to the feature filtering and extraction module, is used to input the feature vector corresponding to the target feature set into the trained machine learning prediction model, extract the nonlinear correlation between different predictive factors through the machine learning prediction model, output the natural pregnancy prediction probability of the target object within a preset follow-up period, and generate an infertility risk assessment result based on the natural pregnancy prediction probability. The infertility risk assessment result includes at least one of the following: the risk value, risk level, and risk warning information of the target object not having a natural pregnancy within the preset follow-up period. The model training and validation module is connected to the multimodal data acquisition and preprocessing module, the feature selection and extraction module, and the prediction information generation module. It is used to train the machine learning prediction model based on sample data with pregnancy outcome labels, and to validate the performance of the trained machine learning prediction model by using discrimination evaluation, calibration evaluation, and decision curve analysis. The prediction result interpretation module, connected to the prediction information generation module, is used to determine the degree of contribution of each target feature to the prediction probability of pregnancy outcome, and to visualize the prediction results and the degree of contribution of the target features.

[0012] As a preferred embodiment of the present invention, the clinical baseline data includes at least one of age, body mass index, previous pregnancy history, delivery history, menstrual history, duration of infertility, and duration of trying to conceive.

[0013] The Western medicine auxiliary examination data includes at least one of anti-Müllerian hormone, follicle-stimulating hormone, luteinizing hormone, estradiol, thyroid-stimulating hormone, and antral follicle count.

[0014] The data from the four diagnostic methods of traditional Chinese medicine and the syndrome scales include at least one of the following: tongue appearance characteristics, pulse appearance characteristics, scores from the traditional Chinese medicine syndrome inquiry scale, and scores from the uterine cold scale.

[0015] The infrared thermal imaging temperature data of the acupoints includes the temperature of the middle jiao region, the temperature of the lower jiao region, and the body surface temperature value of the selected acupoints. The selected acupoints include at least one of the following: Dahe, Henggu, Qichong, Guilai, Shuidao, and Zigong.

[0016] As a preferred embodiment of the present invention, the multimodal data acquisition and preprocessing module is configured as follows: A unified sample identifier is set for different modal data from the same target object, and the different modal data are associated and aligned based on the sample identifier; Set pregnancy outcome labels for each sample within a preset follow-up period; The missing proportion of each candidate feature is calculated, and candidate features with a missing proportion exceeding a preset missing threshold are deleted. Missing values ​​are imputed for candidate features with a missing proportion not exceeding the preset missing threshold. Categorical variables are numerically encoded, and continuous variables are normalized or standardized to generate candidate feature vectors with a uniform data format.

[0017] Preferably, the preset missing threshold is 20%. For categorical variables, binary encoding or one-hot encoding is used for numerical processing; for continuous variables, min-max normalization or Z-score standardization is used.

[0018] In a preferred embodiment of the present invention, the feature filtering and extraction module is configured as follows: Univariate Cox proportional hazards regression analysis was used to determine the degree of association between each candidate feature and pregnancy outcome; Based on preset significance conditions, preliminary candidate prediction factors are obtained from the candidate features; The preliminary candidate predictors are input into a multivariate Cox proportional hazards regression model to eliminate confounding effects between different variables and obtain independent predictors. The independent predictors were subjected to variance inflation factor test, tolerance test, and proportional hazard hypothesis test. Independent predictors that satisfy the preset collinearity condition and proportional risk assumption are determined as the target feature set.

[0019] As a preferred embodiment of the present invention, the target feature set includes age, antral follicle count, uterine coldness score, white and greasy tongue coating, thin pulse, and infrared thermal imaging temperature value of the uterine acupoint.

[0020] Among them, the white greasy tongue coating and the thin pulse are classification features; the age, antral follicle count, uterine coldness score, and infrared thermal imaging temperature of the uterus are continuous features.

[0021] In a preferred embodiment of the present invention, the prediction information generation module is configured as follows: Receive a multidimensional feature vector consisting of the age, antral follicle count, uterine coldness score, white and greasy tongue coating marker, thin pulse marker, and infrared thermal imaging temperature value of the uterine acupoint. The multidimensional feature vector is input into the trained extreme gradient boosting model; The prediction residuals of the previous decision tree are fitted sequentially by multiple decision trees in the extreme gradient boosting model to obtain the output results of each decision tree. The outputs of each decision tree are weighted and combined to obtain the predicted probability of pregnancy outcome for the target subject within the preset follow-up period; The predicted pregnancy outcome probability is compared with a preset probability threshold to generate a corresponding pregnancy outcome prediction level or risk warning information.

[0022] As a preferred embodiment of the present invention, the preset follow-up period includes at least one of 3 months, 6 months, 9 months and 12 months.

[0023] The prediction information generation module is used to output the predicted probability of the target object having a natural pregnancy within the preset follow-up period, or to output at least two pregnancy outcome prediction levels obtained based on the predicted probability.

[0024] As a preferred embodiment of the present invention, the model training and verification module is configured as follows: The sample data labeled with pregnancy outcomes were divided into training datasets and validation datasets; Based on the training dataset, train at least two candidate prediction models from the following: Cox proportional hazards regression model, random forest model, support vector machine model, decision tree model, and extreme gradient boosting model. Calculate at least one of the following performance metrics for each candidate prediction model on the training and validation datasets: area under the receiver operating characteristic curve, accuracy, precision, recall, and F1 score. The consistency between the predicted probabilities of each candidate prediction model and the actual pregnancy outcomes was evaluated based on calibration curves. The net benefit of each candidate prediction model under different probability thresholds is determined based on decision curve analysis. Based on the performance indicators, the calibration curve, and the decision curve analysis results, a target prediction model for generating pregnancy outcome prediction information is determined from the candidate prediction models.

[0025] Preferably, the target prediction model is an extreme gradient boosting model.

[0026] As a preferred embodiment of the present invention, the model training and verification module is further configured as follows: In the training dataset, at least one of the following hyperparameters of the extreme gradient boosting model is optimized: learning rate, number of decision trees, maximum depth of decision trees, sample sampling ratio, and feature sampling ratio. The hyperparameters are determined using cross-validation or independent validation. The timing for stopping model iteration is determined based on performance metrics on the validation dataset to reduce the probability of model overfitting.

[0027] As a preferred embodiment of the present invention, the prediction result interpretation module is configured as follows: Calculate the feature importance or feature contribution value of each target feature in the target feature set; Based on the importance or contribution value of the features, determine the magnitude and direction of the influence of each target feature on the probability of pregnancy outcome prediction; The prediction results are presented in at least one of the following forms: feature importance ranking map, feature contribution map, individual prediction interpretation map, or prediction probability display interface.

[0028] Preferably, the prediction result interpretation module uses the SHAP value analysis method to calculate the contribution value of each target feature to the prediction probability of pregnancy outcome for a single target object.

[0029] In some implementations, the system can also generate nomograms based on a Cox proportional hazards regression model. The nomogram converts the values ​​of each target feature into corresponding feature scores, and generates the predicted probabilities of pregnancy outcomes for different follow-up periods based on the sum of these feature scores. The nomogram provides a visual reference for traditional statistical models, while the extreme gradient boosting model generates machine learning prediction results, thus avoiding directly equating the nonlinear extreme gradient boosting model with a nomogram constructed based on regression coefficients.

[0030] Through the above technical solution, the present invention has at least the following beneficial effects.

[0031] First, this invention transforms clinical baseline data, Western medicine auxiliary examination data, traditional Chinese medicine four diagnostic methods and syndrome scale data, and acupoint infrared thermal imaging temperature data into feature data that can be uniformly processed by computers by unifying sample identification, handling missing values, variable encoding, and standardization. This helps to reduce the impact of different data sources, data formats, and data dimensions on model input and improves the consistency and repeatability of multimodal data processing.

[0032] Secondly, this invention obtains a set of target features that are relevant to pregnancy outcomes and do not have significant redundancy from a large number of candidate features through univariate Cox proportional hazards regression analysis, multivariate Cox proportional hazards regression analysis, collinearity test, and proportional hazards hypothesis test. Therefore, it can reduce the dimensionality of the model input while retaining effective predictive information, and reduce the impact of irrelevant and redundant features on the model training results.

[0033] Furthermore, this invention inputs clinical and Western medicine auxiliary examination features such as age and antral follicle count, along with traditional Chinese medicine features such as uterine coldness score, white and greasy tongue coating, and thin pulse, as well as infrared thermal imaging temperature features of the uterine acupoint into an extreme gradient boosting model. This model can utilize complementary information between features from different sources and extract nonlinear and interactive relationships between different predictive factors, thereby generating a predicted probability of pregnancy outcome within a preset follow-up period.

[0034] Furthermore, by evaluating the discriminability, calibration, and decision curves of multiple candidate prediction models, this invention can verify the performance of prediction models from multiple aspects and determine the target prediction model based on the verification results, which helps to reduce the bias caused by selecting a model based on only a single performance index.

[0035] Finally, through feature importance analysis, SHAP value analysis, and visualization of prediction results, this invention enables the system not only to output the predicted probability of pregnancy outcome, but also to present the degree of contribution of each target feature to the predicted probability, thereby improving the interpretability of the prediction process and providing auxiliary reference for relevant personnel to understand and use the prediction information. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0037] Figure 1 The structural block diagram of the DOR population infertility risk assessment system based on multimodal data provided in this embodiment of the invention.

[0038] Figure 2 This is a comparison chart of the receiver operating characteristic curves of different candidate prediction models in the embodiments of the present invention.

[0039] Figure 3 This is a nomogram for pregnancy outcome prediction constructed based on a multivariate Cox proportional hazards regression model in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and beneficial effects of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the embodiments described below are only some embodiments of the present invention, and not all embodiments. Other implementation methods obtained by those skilled in the art based on the embodiments described in the present invention without creative effort are all within the protection scope of the present invention.

[0041] In this invention, unless otherwise specified, "connection" refers to a data connection that enables data transmission, which can be a data connection between different software modules or a communication connection between different hardware devices, servers or terminals.

[0042] Unless otherwise specified, the term "target subject" in this invention refers to a person who has been identified by a medical institution as having diminished ovarian reserve according to pre-set clinical criteria and is preparing for pregnancy. The system provided by this invention is not used to determine whether the target subject has diminished ovarian reserve, but rather to process the acquired multimodal data of the target subject using a computer to generate pregnancy outcome prediction information within a preset follow-up period.

[0043] The "pregnancy outcome label" referred to in this invention refers to marking whether a target subject has achieved natural pregnancy within a preset follow-up period according to pre-set follow-up rules. Samples that meet the preset criteria for natural pregnancy are marked as having a positive pregnancy outcome, while samples that do not meet the criteria are marked as having a negative pregnancy outcome. Preferably, intrauterine pregnancy confirmed by clinical examination can be used as the criterion for determining natural pregnancy.

[0044] Unless otherwise specified, the term "infertility risk assessment" in this invention refers to the use of a computer to process acquired multimodal data of a target subject and, based on the processing results, assess the likelihood that the target subject will not achieve a natural pregnancy within a preset follow-up period. The "infertility risk assessment result" in this invention includes at least one of the following: a risk value indicating that the target subject will not achieve a natural pregnancy within the preset follow-up period; a risk level determined based on the risk value; or corresponding risk warning information.

[0045] The "infertility risk assessment result" referred to in this invention is not used to determine whether the target subject has met the diagnostic criteria for clinical infertility, nor does it directly constitute a disease diagnosis conclusion or treatment plan. That is, the result is merely a calculation result of the absence of natural pregnancy within a specific time range, and not "infertility" in a clinical sense.

[0046] Example 1: Construction and validation of a DOR population infertility risk assessment system based on multimodal data This study utilizes data from 246 individuals with diminished ovarian reserve (DOR) collected by the Department of Reproductive Medicine at the First Affiliated Hospital of Tianjin University of Traditional Chinese Medicine between August 2024 and March 2025. It constructs and validates a multimodal data-based DOR population infertility risk assessment system. At the time of inclusion in the study, the timeframe for conception for each target individual did not exceed three months.

[0047] The follow-up period began with the date the target subjects entered the study, and followed up for a maximum of 12 months. For target subjects who met the pre-defined criteria for natural pregnancy during the follow-up period, the time from the start of the follow-up to the date the criteria were met was taken as the time of the pregnancy event. For target subjects who did not meet the criteria at the end of the follow-up, 12 months was used as the censoring period. For target subjects who were lost to follow-up during the follow-up period, the time when the last valid follow-up information was obtained was used as the censoring period.

[0048] Reference Figure 1 The DOR population infertility risk assessment system based on multimodal data provided in this embodiment includes: (1) Multimodal data acquisition and preprocessing module, used to acquire multimodal data of the target object, and perform data alignment, missing value processing, variable encoding and standardization processing on the multimodal data; (2) Feature screening and extraction module, connected to the multimodal data acquisition and preprocessing module, is used to screen target features related to pregnancy outcome from candidate features by using univariate Cox proportional hazards regression analysis, multivariate Cox proportional hazards regression analysis, collinearity test and proportional hazards hypothesis test; (3) Prediction information generation module, connected to the feature screening and extraction module, is used to input the feature vector composed of the target features into the trained XGBoost model, output the natural pregnancy prediction probability of the target object within the preset follow-up period, and generate an infertility risk assessment result based on the natural pregnancy prediction probability. The infertility risk assessment result includes at least one of the risk value, risk level or risk warning information of the target object not having a natural pregnancy within the preset follow-up period. (4) Model training and validation module, which is connected to the multimodal data acquisition and preprocessing module, the feature screening and extraction module and the prediction information generation module, is used to divide the training dataset and the validation dataset, train multiple candidate prediction models, and determine the target prediction model based on the model's discrimination, calibration and decision curve analysis results. (5) Prediction result interpretation module, which is connected to the prediction information generation module, is used to interpret and visualize the contribution of the target features and the model output results.

[0049] Example 1.1: Configuration of Multimodal Data Acquisition and Preprocessing Module To improve the consistency of data from different sources and to form input data that can be processed by the prediction model, this embodiment configures the multimodal data acquisition and preprocessing module in the following manner.

[0050] The multimodal data acquisition and preprocessing module is configured to acquire multimodal data of the target object, including general clinical data, Western medicine auxiliary examination data, traditional Chinese medicine four diagnostic methods and syndrome scale data, and acupoint infrared thermal imaging temperature data.

[0051] The general clinical data include at least one of the following: age, body mass index, previous pregnancy history, previous delivery history, menstrual history, duration of infertility, and duration of trying to conceive.

[0052] The Western medicine auxiliary examination data includes at least one of anti-Müllerian hormone, follicle-stimulating hormone, luteinizing hormone, estradiol, thyroid-stimulating hormone, and antral follicle count.

[0053] The data from the TCM four diagnostic methods and syndrome scales include tongue and pulse characteristics obtained through TCM four diagnostic methods information collection equipment, as well as scale scoring data obtained through the TCM syndrome inquiry scale and the uterine cold scale. Tongue characteristics include at least one of tongue coating color, tongue coating texture, and tongue body characteristics, and pulse characteristics include at least one of thready pulse, deep pulse, slow pulse, and weak pulse.

[0054] The infrared thermal imaging temperature data for acupoints includes the temperature of the middle jiao region, the temperature of the lower jiao region, and the surface temperature value of the selected acupoints. The selected acupoints include at least one of the following: Dahe, Henggu, Qichong, Guilai, Shuidao, and Zigong.

[0055] When collecting infrared thermal imaging temperature data, the target object is kept at rest under relatively stable environmental conditions. After the body surface temperature stabilizes, infrared thermal imaging data of the abdominal area of ​​the target object is collected using medical infrared thermal imaging equipment. According to the pre-set acupoint positioning rules, the target area of ​​the corresponding acupoint is determined in the infrared thermal image, and the average temperature of one or more pixels in the target area is used as the infrared thermal imaging temperature value of the corresponding acupoint.

[0056] To avoid mismatches between different modal data, the same sample number is assigned to different modal data from the same target object, and the sample number is used as an association identifier to align general clinical data, Western medicine auxiliary examination data, traditional Chinese medicine four diagnostic methods and syndrome scale data, acupoint infrared thermal imaging temperature data, and pregnancy outcome labels.

[0057] After data alignment, the missing value of candidate features is statistically analyzed. Candidate features with a missing value ratio greater than 20% are removed from subsequent feature selection; for candidate features with a missing value ratio of no more than 20%, multiple imputation is used to impute the missing values. During multiple imputation, multiple possible missing values ​​are generated based on other observed features, and the imputed values ​​for the corresponding missing features are obtained based on the analysis results of multiple imputed datasets.

[0058] For categorical variables, binary encoding or one-hot encoding is used to convert them into numerical variables. For example, when the target object has the vein feature, the vein feature is assigned a value of 1; when the vein feature is not present, the vein feature is assigned a value of 0. When the target object has the white greasy coating feature, the white greasy coating feature is assigned a value of 1; when the white greasy coating feature is not present, the white greasy coating feature is assigned a value of 0.

[0059] The white, greasy tongue coating characteristic can be obtained from the tongue coating color and texture classification results output by the TCM four diagnostic methods information collection device. When the tongue coating color is identified as white and the tongue coating texture is identified as greasy, the white, greasy tongue coating identifier is assigned a value of 1; otherwise, it is assigned a value of 0. The thin pulse characteristic can be obtained from the pulse classification results output by the pulse acquisition device, or from the pulse record formed according to a unified judgment standard.

[0060] For continuous variables such as age, antral follicle count, uterine coldness score, and acupoint infrared thermography temperature, the original measurements are retained and normalized or standardized to a numerical range suitable for model processing. In one implementation, a minimum-maximum normalization method is used for processing: in, Represents the original values ​​of continuous variables. and Let represent the minimum and maximum values ​​of the corresponding continuous variables in the training dataset, respectively. This represents the normalized variable value.

[0061] After dividing the training and validation datasets, the missing value imputation parameters, categorical variable encoding rules, and continuous variable normalization parameters are all determined based on the training dataset. The same parameters obtained from the training dataset are then applied to the validation dataset to reduce the possibility of information leakage from the validation data to the model training process.

[0062] Through the above preprocessing, data from different sources, with different dimensions, and with different variable types are transformed into structured candidate feature data that can be uniformly processed by statistical analysis models and machine learning models.

[0063] Example 1.2: Configuration of Feature Filtering and Extraction Module This embodiment yielded a total of 44 candidate features. To select target features related to pregnancy outcomes from these candidate features, the feature selection and extraction module was configured as follows.

[0064] First, univariate Cox proportional hazards regression analysis was performed on each candidate feature as an explanatory variable, with the time from the start of follow-up to the occurrence of a pregnancy event or data deletion as the outcome time, and whether or not a natural pregnancy occurred as the event state.

[0065] The univariate Cox proportional hazards regression model is expressed as: in, Indicating in candidate features Given the conditions, at time The instantaneous risk rate of a pregnancy event. Let represent the baseline risk function, and β represent the regression coefficient corresponding to the candidate feature.

[0066] Based on the results of univariate Cox proportional hazards regression analysis, candidate features that meet the preset significance criteria are identified as preliminary candidate predictors. In this embodiment, the preset significance criteria is (P<0.05).

[0067] In the specific implementation of this embodiment, through univariate Cox proportional hazards regression analysis, 15 preliminary candidate predictive factors were initially screened from 44 candidate features, including age, anti-Müllerian hormone, follicle-stimulating hormone, antral follicle count, kidney yang deficiency syndrome, uterine coldness score, white and greasy tongue coating, thready pulse, and infrared thermography temperature value of the uterine acupoint.

[0068] Subsequently, the preliminary candidate predictors were incorporated into a multivariate Cox proportional hazards regression model. By simultaneously calculating the regression coefficients of each variable, the confounding effects of other candidate predictors on the analysis results of individual variables were reduced.

[0069] Through multivariate Cox proportional hazards regression analysis, this embodiment obtained 6 independent predictors, namely age, antral follicle count, uterine coldness score, white and greasy tongue coating, thin pulse, and infrared thermography temperature value of the uterine acupoint.

[0070] The hazard ratios for each independent predictor were as follows: age was 0.899, antral follicle count was 1.450, uterine coldness score was 0.986, white and greasy tongue coating was 0.473, thin pulse was 0.511, and infrared thermography temperature of the uterine cavity was 2.959.

[0071] Since this embodiment treats natural pregnancy as an event in the Cox proportional hazards regression model, a hazard ratio greater than 1 indicates that an increase in the corresponding feature value is associated with an increased probability of pregnancy, while a hazard ratio less than 1 indicates that an increase in the corresponding feature value or the presence of the corresponding classification feature is associated with a decreased probability of pregnancy.

[0072] After obtaining the independent predictors, collinearity diagnosis was performed on each independent predictor. This embodiment uses variance inflation factor and tolerance to evaluate whether there is significant collinearity among the independent predictors. The detection results show that the variance inflation factor of each independent predictor is less than 10, and the tolerance is greater than 0.1, indicating that there is no significant collinearity among the independent predictors that affects model construction.

[0073] Furthermore, the proportional hazards hypothesis was tested on the multifactor Cox proportional hazards regression model. The chi-square value of the global proportional hazards hypothesis test in this embodiment was 3.516 (P=0.74), indicating that the model composed of the six independent predictors satisfies the proportional hazards hypothesis.

[0074] Therefore, age, antral follicle count, uterine coldness score, white and greasy tongue coating, thin pulse, and infrared thermal imaging temperature value of the uterine acupoint are determined as the target feature set, and the feature filtering and extraction module outputs a six-dimensional feature vector composed of the above six target features.

[0075] It should be noted that the feature selection process is used in the model training and construction phases. When using the already trained model to predict a single target object, there is no need to perform univariate Cox proportional hazards regression analysis or multivariate Cox proportional hazards regression analysis on that target object again. Instead, the six target features are directly obtained and used to form the model input feature vector.

[0076] Example 1.3: Configuration of the Prediction Information Generation Module The prediction information generation module is used to receive a six-dimensional feature vector consisting of age, antral follicle count, uterine coldness score, white and greasy tongue coating, thin pulse, and infrared thermal imaging temperature value of the uterine acupoint, and input the six-dimensional feature vector into the trained XGBoost model.

[0077] The XGBoost model is an ensemble learning model based on gradient boosting decision trees. The XGBoost model comprises multiple decision trees built sequentially. During model training, each newly added decision tree is used to fit the residual between the current model's prediction and the actual pregnancy outcome label. The final model output is obtained by weightedly combining the outputs of multiple decision trees.

[0078] In some implementations, the follow-up period corresponding to the predicted pregnancy outcome can be preset as needed. The preset follow-up period can be 3 months, 6 months, 9 months, or 12 months.

[0079] For any preset follow-up period, a corresponding binary label is assigned based on whether the target subject experienced a natural pregnancy within that period. Specifically, samples that experienced a natural pregnancy within the preset follow-up period are labeled as 1, and samples that did not experience a natural pregnancy within the preset follow-up period are labeled as 0. The XGBoost model is trained based on training samples with the corresponding binary labels to obtain the XGBoost model corresponding to the preset follow-up period.

[0080] During the model application phase, the six-dimensional feature vector of the target object is input into the XGBoost model corresponding to the preset follow-up period to output the predicted probability of the target object having a natural pregnancy within the preset follow-up period.

[0081] For multiple preset follow-up periods, binary labels can be set for each preset follow-up period, and XGBoost models corresponding to each preset follow-up period can be trained. Different XGBoost models can output the predicted probability of the target object having a natural pregnancy within 3 months, 6 months, 9 months or 12 months.

[0082] In this embodiment, whether the target subject experienced natural pregnancy within a 12-month follow-up period is used as the binary classification label of the XGBoost model. The sample label for natural pregnancy within 12 months is set to 1, and the sample label for no natural pregnancy within 12 months is set to 0.

[0083] The prediction function of the XGBoost model can be expressed as: in, Indicates the first The six-dimensional feature vector of each sample, Indicates the first A decision tree, Indicates the number of decision trees. This represents a mapping function that converts the model output into predicted probabilities between 0 and 1. Indicates the first The predicted probability of a sample becoming pregnant naturally within a 12-month follow-up period.

[0084] The objective function for model training includes a prediction loss term and a model complexity penalty term, which can be expressed as: in, Indicates the first The loss function between the actual pregnancy outcome label and the predicted probability for each sample Indicates the first A complexity penalty term is introduced for each decision tree. By introducing a complexity penalty term, the number of leaf nodes and the weights of the leaf nodes in the decision tree are limited, thereby reducing the possibility of model overfitting.

[0085] During model training, at least one of the following model parameters can be optimized: learning rate, number of decision trees, maximum depth of decision trees, sample sampling ratio, feature sampling ratio, and regularization coefficient. Model parameters can be determined based on cross-validation results within the training dataset or performance on the validation dataset.

[0086] The prediction information generation module converts the values ​​output by the XGBoost model into predicted probabilities of natural pregnancy within 12 months. It can also convert these predicted probabilities into different pregnancy outcome prediction levels based on pre-determined probability thresholds. For example, a classification threshold can be determined based on the receiver operating characteristic (ROC) curves of the training dataset, and the target population can be divided into a relatively high pregnancy probability group and a relatively low pregnancy probability group based on these thresholds.

[0087] The predicted probabilities and prediction levels mentioned above are only auxiliary prediction information generated based on existing multimodal data of the target object, and do not directly constitute a disease diagnosis conclusion or treatment plan.

[0088] Example 1.4: Configuration of the Model Training and Validation Module To evaluate the processing effect of different prediction models on the data of this embodiment, 246 target objects were randomly divided into training dataset and validation dataset in a ratio of 7:3. The training dataset included 172 samples, and the validation dataset included 74 samples.

[0089] Preferably, when dividing the data, stratified sampling is performed based on whether a natural pregnancy occurred within 12 months, so that the proportion of positive pregnancy outcome samples to negative pregnancy outcome samples in the training dataset and the validation dataset is basically the same.

[0090] Cox proportional hazards regression, random forest, support vector machine, decision tree, and XGBoost models were constructed on the training dataset. All candidate models used the same set of target features and were compared using the same training and validation datasets.

[0091] Among them, the Cox proportional hazards regression model is trained based on the occurrence time of pregnancy events and censoring information; the random forest model, support vector machine model, decision tree model and XGBoost model are trained using whether a natural pregnancy occurred within 12 months as a binary classification label.

[0092] For each candidate model, accuracy, precision, recall, F1 score, and area under the receiver operating characteristic (AUC) curve were calculated. Accuracy represents the proportion of correctly classified samples out of all samples; precision represents the proportion of samples predicted as having a positive pregnancy outcome that were actually positive; recall represents the proportion of samples with actual positive pregnancy outcomes that were correctly identified; the F1 score is the harmonic mean of precision and recall; and the AUC curve is used to evaluate the model's ability to distinguish between samples with positive and negative pregnancy outcomes.

[0093] In addition to discrimination evaluation, the consistency between the predicted pregnancy outcome probability output by the model and the actual pregnancy outcome rate is evaluated through calibration curve, and the net benefit obtained by using the model at different probability thresholds is evaluated through decision curve analysis.

[0094] The model comparison results in this embodiment show that, among the constructed candidate prediction models, the XGBoost model exhibits better overall performance on both the training and validation datasets.

[0095] In the training dataset, the XGBoost model achieved an accuracy of 0.901, a precision of 0.914, a recall of 0.951, an F1 score of 0.932, an area under the receiver operating characteristic curve of 0.973, and a 95% confidence interval of 0.952–0.993.

[0096] In the validation dataset, the XGBoost model achieved an accuracy of 0.919, a precision of 0.898, a recall of 1.000, an F1 score of 0.946, an area under the receiver operating characteristic curve of 0.982, and a 95% confidence interval of 0.958–1.000.

[0097] The calibration curve results show that the predicted probability curve of the XGBoost model is close to the ideal calibration line, indicating that the predicted probability of pregnancy outcome output by the model has good consistency with the actual pregnancy outcome rate in the sample of this embodiment.

[0098] The decision curve analysis results show that, within the probability threshold range evaluated in this embodiment, the XGBoost model can achieve a relatively high net benefit. Therefore, the trained XGBoost model is selected as the target prediction model in this embodiment.

[0099] It should be noted that the above performance data is used to illustrate the feasibility of the technical solution adopted in this embodiment and the processing effect on the dataset of this embodiment, and does not constitute a limitation on the model performance of this invention on other datasets.

[0100] After training, the model structure, decision tree parameters, leaf node weights, feature order, and data preprocessing parameters of the XGBoost model are stored for the prediction information generation module to call during the model application stage.

[0101] Example 1.5: Configuration of the Prediction Result Interpretation Module To improve the interpretability of the XGBoost model output, this embodiment configures the prediction result interpretation module as follows.

[0102] The prediction result interpretation module reads the trained XGBoost model and calculates the feature importance of age, antral follicle count, uterine coldness score, white and greasy tongue coating marker, thin pulse marker, and uterine cavity infrared thermography temperature value in the model. Feature importance can be determined based on the loss reduction brought about by each target feature when used as node splitting feature in the entire decision tree, or by using the SHAP value analysis method to calculate the contribution value of each target feature to the model prediction result. Through the contribution value, the direction and degree of influence of each target feature on increasing or decreasing the probability of pregnancy outcome prediction can be determined (this embodiment uses the latter). For a single target object, the model output can be expressed as: in, This represents the baseline output of the model. Indicates the first The contribution value of each target feature to the current prediction result. Indicates the number of target features.

[0103] The prediction result interpretation module can present the interpretation results using feature importance ranking charts, SHAP summary charts, single sample feature contribution charts, or probability display interfaces.

[0104] For pregnancy outcome prediction based on follow-up time, a nomogram can be constructed using the aforementioned multivariate Cox proportional hazards regression model. The nomogram maps age, antral follicle count, uterine coldness score, white and greasy tongue coating marker, thin pulse marker, and infrared thermography temperature value of the uterine cavity to corresponding scores. The scores for each target feature are summed to obtain a total score. Then, based on the relationship between the total score and the baseline survival function of the Cox proportional hazards regression model, the predicted probability of natural pregnancy within 3 months, 6 months, 9 months, and 12 months is obtained.

[0105] In the nomogram, an increase in antral follicle count and an increase in infrared thermography temperature of the uterine cavity correspond to an increased likelihood of pregnancy; while an increase in age, an increase in uterine coldness score, the presence of a greasy white tongue coating, and the presence of a thin pulse correspond to a decreased likelihood of pregnancy.

[0106] The nomogram is a time-outcome visualization tool built on the Cox proportional hazards regression model, and it is run separately from the XGBoost model used to output the predicted probability of natural pregnancy at 12 months. This avoids directly converting the nonlinear decision tree output from the XGBoost model into a nomogram that depends on the linear summation of regression coefficients.

[0107] Example 2: Application of the DOR population infertility risk assessment system based on multimodal data This embodiment illustrates the process of generating pregnancy outcome prediction information for a single target subject using the system constructed and trained in Embodiment 1.

[0108] First, obtain the age, antral follicle count, uterine coldness score, white and greasy tongue coating, thin pulse, and infrared thermal imaging temperature value of the uterine acupoint for the target subjects identified as having diminished ovarian reserve.

[0109] Among them, age, antral follicle count, uterine coldness score, and infrared thermography temperature value of uterine acupoint were normalized or standardized according to the continuous variable preprocessing parameters saved in Example 1; white greasy tongue coating and thin pulse were encoded according to the predetermined binary coding rules.

[0110] Subsequently, following the feature order used during model training, the processed age, antral follicle count, uterine coldness score, white and greasy tongue coating marker, thin pulse marker, and infrared thermal imaging temperature value of the uterine acupoint were combined into a six-dimensional feature vector and input into the XGBoost model trained in Example 1.

[0111] The XGBoost model sequentially calls each saved decision tree to process the six-dimensional feature vector, weights and combines the outputs of each decision tree, and converts the weighted combination result into a probability value between 0 and 1, which is used as the predicted probability of the target object having a natural pregnancy within 12 months.

[0112] The system can display the predicted probability on the interface, and can also compare the predicted probability with a predetermined classification threshold to generate a corresponding pregnancy outcome prediction level.

[0113] Meanwhile, the prediction result interpretation module can calculate the contribution value of each target feature to the prediction probability of the current target object, and sort the target features according to the degree of contribution to display the main features that cause the current prediction probability to increase or decrease.

[0114] When it is necessary to obtain the pregnancy probability corresponding to different follow-up time points, the six target features of the target object can be input into the nomogram or time outcome calculation module based on the multivariate Cox proportional hazards regression model to generate the predicted probability of natural pregnancy within 3 months, 6 months, 9 months and 12 months respectively.

[0115] Through the above process, the system completes the computer processing of the existing multimodal data of the target object and generates the corresponding pregnancy outcome prediction probability, prediction level and feature contribution information.

[0116] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art to the data preprocessing method, feature selection method, model parameters, prediction period, result presentation method, and system deployment method without departing from the technical concept of the present invention shall all fall within the scope of protection of the present invention.

Claims

1. A DOR population infertility risk assessment system based on multimodal data, characterized in that, include: The multimodal data acquisition and preprocessing module is used to acquire multimodal data of people with diminished ovarian reserve, and to perform sample alignment, missing value processing, variable encoding and standardization on the multimodal data. The multimodal data includes at least clinical baseline data, Western medicine auxiliary examination data, traditional Chinese medicine four diagnostic methods and syndrome scale data, and acupoint infrared thermal imaging temperature data. The feature screening and extraction module is connected to the multimodal data acquisition and preprocessing module. It is used to perform univariate Cox proportional hazards regression analysis and multivariate Cox proportional hazards regression analysis on the preprocessed multimodal data, screen predictors related to pregnancy outcomes from candidate features, and perform collinearity test and proportional hazards hypothesis test on the screened predictors to generate a target feature set. The prediction information generation module, connected to the feature filtering and extraction module, is used to input the feature vector corresponding to the target feature set into the trained machine learning prediction model, extract the nonlinear correlation between different predictive factors through the machine learning prediction model, output the natural pregnancy prediction probability of the target object within a preset follow-up period, and generate an infertility risk assessment result based on the natural pregnancy prediction probability. The infertility risk assessment result includes at least one of the following: the risk value, risk level, and risk warning information of the target object not having a natural pregnancy within the preset follow-up period. The model training and validation module, connected to the multimodal data acquisition and preprocessing module, the feature selection and extraction module, and the prediction information generation module, is used to train the machine learning prediction model based on sample data with pregnancy outcome labels, and to validate the performance of the trained machine learning prediction model through discrimination evaluation, calibration evaluation, and decision curve analysis; and The prediction result interpretation module, connected to the prediction information generation module, is used to determine the contribution of each target feature in the target feature set to the prediction probability of the pregnancy outcome, and to visualize the prediction probability of the pregnancy outcome and the contribution of each target feature.

2. The DOR population infertility risk assessment system based on multimodal data according to claim 1, characterized in that: The clinical baseline data includes at least one of the following: age, body mass index, previous pregnancy history, previous delivery history, menstrual history, duration of infertility, and duration of trying to conceive; The Western medicine auxiliary examination data includes at least one of anti-Müllerian hormone, follicle-stimulating hormone, luteinizing hormone, estradiol, thyroid-stimulating hormone, and antral follicle count; The data from the four diagnostic methods of traditional Chinese medicine and the syndrome scales include at least one of the following: tongue appearance characteristics, pulse appearance characteristics, scores from the traditional Chinese medicine syndrome inquiry scale and scores from the uterine cold scale. The infrared thermal imaging temperature data of the acupoints includes the temperature of the middle jiao region, the temperature of the lower jiao region, and the body surface temperature value of the selected acupoints. The selected acupoints include at least one of the following: Dahe, Henggu, Qichong, Guilai, Shuidao, and Zigong.

3. The DOR population infertility risk assessment system based on multimodal data according to claim 1, characterized in that, The multimodal data acquisition and preprocessing module is configured as follows: A unified sample identifier is set for different modal data from the same target object, and the different modal data are associated and aligned according to the sample identifier; Set pregnancy outcome labels for each sample within the preset follow-up period; The missing value ratio of each candidate feature is calculated, and candidate features with a missing value ratio greater than 20% are deleted. Then, multiple imputation method is used to imput missing values ​​for candidate features with a missing value ratio of no more than 20%. Binary encoding or one-hot encoding is used to quantify categorical variables, and min-max normalization or Z-score standardization is used to transform continuous variables to generate candidate feature vectors with a unified data format.

4. The DOR population infertility risk assessment system based on multimodal data according to claim 1, characterized in that, The feature filtering and extraction module is configured as follows: Univariate Cox proportional hazards regression analysis was used to determine the degree of association between each candidate feature and pregnancy outcome, and preliminary candidate predictive factors were obtained based on the pre-set significance conditions. The preliminary candidate predictors are input into a multivariate Cox proportional hazards regression model to obtain independent predictors. The independent predictors were subjected to variance inflation factor test, tolerance test, and proportional hazard hypothesis test. as well as Independent predictors that satisfy the preset collinearity condition and proportional risk assumption are identified as the target feature set.

5. The DOR population infertility risk assessment system based on multimodal data according to claim 1 or 4, characterized in that, The target feature set includes age, antral follicle count, uterine coldness score, white and greasy tongue coating, thin pulse, and infrared thermal imaging temperature value of the uterine acupoint. Among them, the white greasy tongue coating and the thin pulse are classification features, while the age, antral follicle count, uterine coldness score, and infrared thermal imaging temperature of the uterus are continuous features.

6. The DOR population infertility risk assessment system based on multimodal data according to claim 5, characterized in that, The machine learning prediction model is an extreme gradient boosting model, and the prediction information generation module is configured as follows: Receive a multidimensional feature vector consisting of the age, antral follicle count, uterine coldness score, white and greasy tongue coating marker, thin pulse marker, and infrared thermal imaging temperature value of the uterine acupoint. The multidimensional feature vector is input into the trained extreme gradient boosting model; The prediction residuals of the current model are sequentially fitted by multiple decision trees in the extreme gradient boosting model to obtain the output results of each decision tree; as well as The outputs of each decision tree are weighted and combined to obtain the predicted probability of pregnancy outcome for the target object within the preset follow-up period.

7. The DOR population infertility risk assessment system based on multimodal data according to claim 1 or 6, characterized in that: The preset follow-up period includes at least one of 3 months, 6 months, 9 months, and 12 months; The prediction information generation module is used to output the predicted probability of the target object having a natural pregnancy within the preset follow-up period, and / or compare the predicted probability with a preset probability threshold to generate at least two pregnancy outcome prediction levels.

8. The DOR population infertility risk assessment system based on multimodal data according to claim 1, characterized in that, The model training and validation module is configured as follows: The sample data labeled with pregnancy outcomes were divided into training datasets and validation datasets; Based on the training dataset, train at least two candidate prediction models from the following: Cox proportional hazards regression model, random forest model, support vector machine model, decision tree model, and extreme gradient boosting model. Calculate at least one of the following performance metrics for each candidate prediction model on the training dataset and the validation dataset: area under the receiver operating characteristic curve, accuracy, precision, recall, and F1 score. The consistency between the predicted probabilities of each candidate prediction model and the actual pregnancy outcomes was evaluated based on calibration curves, and the net benefit of each candidate prediction model at different probability thresholds was determined based on decision curve analysis. Based on the analysis results of the performance indicators, the calibration curve, and the decision curve, the target prediction model is determined from the candidate prediction models; In the training dataset, at least one of the hyperparameters of the target prediction model, namely the learning rate, number of decision trees, maximum depth of decision trees, sample sampling ratio, and feature sampling ratio, is optimized, and the point at which the model stops iterating is determined based on the cross-validation results or the performance metrics on the validation dataset. The target prediction model is an extreme gradient boosting model.

9. The DOR population infertility risk assessment system based on multimodal data according to claim 1, characterized in that, The prediction result interpretation module is configured as follows: The contribution of each target feature in the target feature set to the predicted probability of pregnancy outcome for a single target object is calculated using the SHAP value analysis method. The magnitude and direction of the influence of each target feature on the predicted probability of pregnancy outcome are determined based on the contribution value; and The predicted probability of the pregnancy outcome and the contribution of each target feature are presented in at least one of the following forms: feature importance ranking map, feature contribution map, individual prediction interpretation map, and prediction probability display interface.

10. The DOR population infertility risk assessment system based on multimodal data according to claim 5, characterized in that, The system also includes a column chart generation module, which is configured to: Based on the multifactor Cox proportional hazards regression model with the target feature set as input, the values ​​of each target feature are converted into corresponding feature scores; The total score is obtained by adding up the feature scores corresponding to each target feature; as well as Based on the total score and the baseline survival function of the multivariate Cox proportional hazards regression model, the predicted probability of the target subject having a natural pregnancy during at least one follow-up period of 3 months, 6 months, 9 months, and 12 months is generated.

Images