Elderly hip bone mechanics characteristic analysis method and system based on progressive fusion

Abstract

The invention provides an old-age hip bone mechanics characteristic analysis method and system based on progressive fusion, and the method comprises the steps: building and training machine learning models based on use data, enabling each part of model to only correspond to the individual data characteristics of the part, completing the combination of all sub-models through a residual link mode, and obtaining a training model; and a final progressive fusion model is formed. And determining the specific sequence of the sub-models of the progressive fusion model based on the medical theory and traversal so as to achieve the optimal bone mechanics characteristic analysis and risk prediction effect. According to the scheme, the progressive fusion model chain can be cut off and output at any position, the model is independently selected according to the data owned by the user, the prediction flexibility and accuracy are realized, and meanwhile, the problems of data missing and incapability of aligning each feature quantity in the old hip bone mechanics feature analysis can be effectively solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, data processing, and bone mechanics, and in particular to a method and system for analyzing the bone mechanics characteristics of the hip in the elderly based on progressive fusion. Background Technology

[0002] In the field of precision medicine, with the rapid development of artificial intelligence technology, research based on small-sample, refined medical data has gradually become a focus. However, such research still faces many challenges in practical applications, such as data gaps, insufficient standardization, poor integration, and individual differences, among which data gaps are particularly critical. Data gaps are not only a problem in statistics but are especially prominent in the field of medical data processing, particularly the phenomenon of misalignment or data loss due to individual data differences. These problems seriously affect the effectiveness and application of artificial intelligence technology in precision medical data processing.

[0003] Against this backdrop, the analysis and prediction of hip bone biomechanics in the elderly has become a critical and urgent issue. Hip fractures are a common and serious skeletal problem in the elderly, attracting significant attention due to their high rates of disability and mortality, as well as the burden of medical costs. With the increasing aging of the population, hip fractures in the elderly have become a serious public health problem. How to better and more effectively analyze the hip bone biomechanics in the elderly and further apply it to the prediction of hip fracture risk has become a major challenge in this field.

[0004] Currently, several tools are widely used in assessing hip osteomechanical analysis in older adults: FRAX (Fracture Risk Assessment Tool), developed by the WHO Collaborating Centre for Metabolic Bone Diseases, is an osteoporotic fracture prediction tool that integrates multiple risk factors, such as age, sex, weight, height, history of fractures, and bone mineral density information, to predict the risk of osteoporotic fractures of the hip and other sites within the next 10 years; QFracture, a risk prediction tool developed based on the UK General Practitioner Research Database, does not rely on bone mineral density measurements. Although QFracture has received high ratings in several meta-analyses, its applicability to the Chinese population requires further validation. The Garvan tool focuses on factors such as sex, age, number of fractures since age 50, number of falls in the past 12 months, and weight to predict the absolute risk of hip fractures and osteoporotic fractures of any site within the next 5 or 10 years. The Garvan tool considers the impact of the number of falls on hip fractures.

[0005] However, in actual medical diagnostic scenarios, most individual data will have modular missing data. In this case, if more data features are used, the amount of available data will be reduced. As a result, traditional tools will be ineffective or even fail. How to complete the analysis of hip mechanics in the elderly and subsequent fracture risk prediction with limited data and when the amount of data and the number of data features cannot be balanced has become an urgent problem to be solved. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention introduces the concept of data fusion into the analysis and prediction of hip mechanics in the elderly. The core idea of ​​this scheme is to treat the samples as samples with different levels of supplementary data. In this scheme, samples with complete data are considered to have additional supplementary data, while samples with missing data are grouped according to the degree of missing data.

[0007] The training process of the progressive fusion model includes multiple stages. In the first stage, the model is trained using the most basic dataset, providing a preliminary mechanical analysis for samples with only basic information. As training progresses, more supplementary data is gradually incorporated into the model training to improve the model's predictive accuracy and robustness. Thus, for samples with more complete data, the model obtained through multiple progressive fusion training iterations can be used for prediction, resulting in more accurate mechanical analysis results; while for samples with a significant amount of missing data, a model with fewer fusion iterations can be used for prediction, providing basic mechanical analysis information.

[0008] Specifically, the following technical solutions are provided:

[0009] On the one hand, this invention provides a method for analyzing the osteomechanical characteristics of the hip in the elderly based on progressive fusion, the method comprising:

[0010] S1. Collect sample data on hip bone mechanics in the elderly, preprocess the sample data, and classify it into multiple types of feature data according to data type;

[0011] S2. Construct a progressive fusion model containing multiple sub-models, determine the order of the sub-models, and establish chain connections between the sub-models in sequence through residuals.

[0012] S3. During the model training phase, based on the number of samples containing feature data of each type, the corresponding sub-model is trained using feature data of a single type and the labels of the corresponding samples. The residual between the prediction result of the previous sub-model and the true label of the sample is used as the new label of the sample of the next sub-model for training the next sub-model.

[0013] S4. In the model prediction stage, the data to be predicted is divided into multiple types of feature data in the manner described in S1. In the reverse order of the model training stage, the feature data of the last type is input into the sub-model corresponding to that type. The prediction result is used as the residual of the previous sub-model to correct the prediction result of the previous sub-model. This process continues until the prediction results of all types of feature data are corrected, and the final prediction result is output.

[0014] Preferably, in step S2, the order of the sub-models is determined based on the number of samples with corresponding feature data types. The feature data with the largest number of samples of the corresponding feature data type is used as the input data of the first-level sub-model, and its corresponding sub-model is used as the first-level sub-model. The order of subsequent feature data is calculated as follows:

[0015]

[0016] Where I represents the total number of feature data types, To simultaneously possess ω i and ω j Samples of feature data For the sample with ω i The sample size with the largest sample size among the remaining feature data; The larger the value, the higher the order of the corresponding sub-model.

[0017] Preferably, in step S3, during the model training phase, when the difference between the number of positive samples and the number of negative samples is large, the sample size is oversampled in the following manner:

[0018] First, the larger of the number of positive samples and the number of negative samples is taken as the unified number N of the number of negative samples and the number of positive samples after sampling, that is:

[0019] N = max(N positive N negative )

[0020] Where, N positive N represents the number of positive samples. negative Indicates the number of negative samples;

[0021] Secondly, for sample data where feature data is discontinuous due to missing data, first copy the sample that needs to be supplemented with feature data, and then calculate the mean μ and variance τ of the feature data of all samples.

[0022] Next, discontinuous sample data are supplemented into continuous variables in the following way:

[0023] X new =X old +N(μ,τ)

[0024] Among them, X new For the new sample, X old The original sample is N(μ,τ), which is a random number that conforms to a normal distribution with mean μ and variance τ. Oversampling is then performed.

[0025] Preferably, in step S3, the new label is calculated as follows:

[0026] First, calculate the residuals of the first sub-model:

[0027] E 1m =C(P) m y m m = 1, 2, ..., K

[0028] Where K represents the total number of samples input to the current sub-model, C represents the residual operation, and E 1m Let C represent the residual of the first sub-model with respect to the m-th sample, where C is calculated as follows:

[0029]

[0030] Among them, P m Y represents the prediction result for sample m. m This represents the true label of sample m;

[0031] Secondly, E 1m The new labels are assigned to each corresponding sample and then used as the second set of feature data and the new labels for each corresponding sample to train the next level sub-model.

[0032] Preferably, after each sub-model completes training, a model chain of progressively fused models is formed:

[0033] M = M1←M2←…←M p

[0034] Where ← represents the chained correction of the model, and p represents the total number of sub-models.

[0035] Preferably, in step S3, during the model training phase, the order of the sub-models in the progressive fusion model is consistent with the order of the feature data, with one type of feature data corresponding to one sub-model.

[0036] Preferably, in step S4, the method for updating the prediction results of the previous-level sub-model is as follows:

[0037] First, input the feature data corresponding to the previous sub-model into this sub-model to obtain the prediction result P1;

[0038] Secondly, the feature data corresponding to the current sub-model is input into the current sub-model, and the resulting prediction result E1 is used as the residual of the previous sub-model.

[0039] Next, the prediction result P1 of the previous sub-model is corrected using the residual E1:

[0040]

[0041] Among them, P new1 -P1 represents the corrected prediction result, indicating a label that is the opposite of the P1 prediction result.

[0042] Preferably, when the feature data type of the data to be predicted is missing, causing the sub-models of the corresponding prediction steps to be not directly continuous, the prediction result of the sub-model corresponding to the missing feature data type is set to 0 by default, so that its prediction result indicates that there is no need to correct the prediction result of the preceding sub-model.

[0043] Preferably, in S2, the order of the sub-models is determined based on the number of samples with corresponding feature data types, the importance of the feature data, and the correlation between the feature data: First, the correlation of the feature data is calculated: the Pearson coefficient between each feature data and the Pearson coefficient between the training labels are calculated, and the models are sorted according to the Pearson coefficient between the feature data and the Pearson coefficient between the training labels, and the correlation score of the sub-model corresponding to the feature data is determined based on the ranking.

[0044] Next, the importance of each feature data is calculated based on the XGBoost model and sorted to calculate the importance score of the sub-module corresponding to the feature data.

[0045] The order of sub-models is determined by the sum of their relevance and importance scores. When the sum of the relevance and importance scores of multiple sub-models is equal, the larger the amount of feature data corresponding to the sub-model, the higher the priority of the sub-model in the ranking.

[0046] Preferably, the types of the feature data include: basic information data, bone density data, bone marker test data, biochemical data, hemoglobin data, and glycated hemoglobin data.

[0047] Secondly, the present invention also provides a progressive fusion-based system for analyzing the biomechanical characteristics of the hip in the elderly, the system comprising:

[0048] The data preprocessing module collects sample data on hip bone mechanics in the elderly, preprocesses the sample data, and divides it into multiple types of feature data according to data type; and preprocesses the data to be predicted and divides it into multiple types of feature data according to data type.

[0049] The model building module constructs a progressive fusion model containing multiple sub-models, determines the order of the sub-models, and establishes chain connections between the sub-models in sequence through residuals.

[0050] The model training module is used to train the corresponding sub-model during the model training phase, based on the number of samples containing feature data of each type, using feature data of a single type and the labels of the corresponding samples, and using the residual between the prediction result of the previous sub-model and the true label of the sample as the new label of the sample of the next sub-model for the training of the next sub-model.

[0051] The online prediction module, based on the feature data of the data to be predicted, inputs the feature data of the last type into the sub-model corresponding to that type in the reverse order of the model training phase. The prediction result is used as the residual of the previous sub-model to correct the prediction result of the previous sub-model. This process continues until the prediction results of all types of feature data have been corrected, and then the final prediction result is output.

[0052] Thirdly, the present invention also provides a progressive fusion-based hip biomechanical characteristic analysis device for the elderly, the device comprising: a processor and a memory, wherein the processor calls computer instructions in the memory to execute the progressive fusion-based hip biomechanical characteristic analysis method for the elderly as described above.

[0053] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects:

[0054] This solution employs a sub-model selection method for multi-source, multi-modal data. It can accurately classify datasets, including limited individual data samples and missing data, to determine the most suitable sub-model for each data type. This effectively improves the accuracy and efficiency of bone mechanics data analysis and enables deeper exploration of the data's underlying structure.

[0055] Meanwhile, this invention constructs a progressive fusion method using residual learning of the model, which uses the residual between the prediction results of the preceding model and the true label as the label of the subsequent model, thereby achieving effective chain correction and optimization between models. The model in the previous stage can guarantee the most basic analysis results, and the subsequent model can further correct the analysis effect of the previous model, thus ensuring the accuracy of data analysis and prediction.

[0056] This solution employs a progressive fusion approach to analyze bone mechanics data. Firstly, it avoids issues such as missing test results and misaligned data structures that could affect the accuracy of the analysis. The progressive fusion model architecture is better able to uncover and capture the interactions between data from different modules in the bone mechanics analysis and prediction of hip fractures. Compared to traditional machine learning models that simply list data side-by-side, the progressive fusion model architecture offers greater accuracy and depth in its analysis. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of the system architecture design according to an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the bone biomechanical feature analysis method according to an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram illustrating the results of the model under non-progressive fusion conditions when the data features are most complete according to an embodiment of the present invention.

[0061] Figure 4 This represents the relative importance of each feature of the model in this embodiment of the invention;

[0062] Figure 5 This is a schematic diagram comparing the prediction performance of this solution in this embodiment of the invention with that of six other common models;

[0063] Figure 6 This is a schematic diagram of the progressive fusion model chain in an embodiment of the present invention. Detailed Implementation

[0064] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0065] Those skilled in the art should understand that the following specific embodiments or implementation methods are a series of optimized configurations listed to further explain the specific content of the invention. These configuration methods can be combined or used in conjunction with each other, unless the invention explicitly states that some or a specific embodiment or implementation method cannot be associated with or used in conjunction with other embodiments or implementation methods. Furthermore, the following specific embodiments or implementation methods are merely optimized configurations and are not intended to limit the scope of protection of the invention.

[0066] Figure 1 This is a schematic diagram of the system architecture for a progressive fusion-based hip biomechanical characteristic analysis scheme for the elderly, provided as an embodiment of this specification.

[0067] like Figure 1 As shown in this embodiment, during the model training phase, real-world hip data from elderly individuals is used to form a training dataset. Individual data includes basic information data, bone mineral density data, bone marker test data, biochemical data, hemoglobin data, and glycated hemoglobin data. Basic information data may include, for example, gender, age, and fracture history data; bone mineral density data may include, for example, femoral neck density, femoral neck density t-value, total hip density, total hip density t-value, L-total bone mineral density t-value, mechanical index, curvature ratio, and section modulus; bone marker test data may include, for example, whole-segment parathyroid hormone data, 25-hydroxyvitamin D data, total type I procollagen N-terminal elongated peptide data, type I collagen C-terminal peptide β-specific sequence data, and N-terminal osteocalcin data; biochemical data includes, for example, albumin data, total cholesterol data, low-density lipoprotein cholesterol data, high-density lipoprotein cholesterol data, and triglyceride data.

[0068] For the data mentioned above, before model training, original labels can be set to annotate the collected individual data. Individuals who have experienced a hip fracture within the past five years are labeled as predicted positive, while the original labels for the remaining individual data are labeled as predicted negative.

[0069] Next, the features in the collected dataset undergo a unified standardization process. This process primarily targets basic information data, such as:

[0070] (1) Set corresponding values ​​for gender: for example, set male to 1 and female to 2.

[0071] (2) Set corresponding values ​​according to age group: for example, set 1 for those under 60 years old, set 2 for those between 60 and 70 years old, etc.

[0072] (3) History of fracture: 1 and 0 are used to indicate whether the individual has a history of fracture.

[0073] Other data items in the basic information data can be configured in a similar way as described above.

[0074] In the medical context of hip biomechanics data analysis and risk prediction in the elderly, data gaps generally lead to missing data modules, such as missing bone mineral density data or bone marker test data, which further means that more data is needed to train the model, resulting in fewer usable features. If more features are used, the corresponding number of individuals will decrease, making it difficult to achieve a balance between the number of features and the amount of individual data.

[0075] Therefore, this solution adopts a progressive fusion-based approach to analyze the biomechanical characteristics of the hip in the elderly. Based on the input multi-source and multi-type data, the solution selects the corresponding machine learning sub-models for analysis and prediction, and uses residual linking to form a complete model chain from multiple sub-models.

[0076] Figure 2 This is a flowchart illustrating a progressive fusion-based method and system for analyzing the biomechanical characteristics of the hip in the elderly, as described in this specification. Figure 2 and Figure 6 The progressive fusion model used in this embodiment can fuse missing feature data. In the experiment, the progressive fusion model incorporates all non-nullable data into the model training. In this embodiment, we divide the processed data into six parts according to features. Here, features refer to the six aspects of data contained in the data, namely, basic information data, bone density data, bone marker test data, biochemical data, hemoglobin data, and glycated hemoglobin data. In practical applications, basic information data is generally the most complete data. Of course, when faced with different feature data, we generally select the most complete feature data as the first part of the progressive fusion model chain input data. In this embodiment, basic information data is used as the first part of the progressive fusion model chain, and the remaining parts are first analyzed according to the specific data structure and missing data.

[0077] In this embodiment, the sample contains six types of feature data: ω1, ω2, ω3, ω4, ω5, and ω6. The feature data with the most complete data in the sample, i.e., the basic information data, is used as the first part of the model chain to obtain the first model. The subsequent feature data selection can preferably be determined based on the following formula:

[0078]

[0079] Where I represents the total number of feature data types, To simultaneously possess ω i and ω j Samples of feature data For the sample with ω i The sample size with the largest number of samples among the remaining feature data (i.e., the sample size with the largest number of samples among samples with two feature data), sum is the sample size of all samples with ω. i and ω j The number of samples for a given feature. (This will be discussed at this point.) The corresponding feature ω in the case j This serves as the sample input feature data for the next sub-model in the progressive fusion process. For example, after the first feature data is determined to be ω1, the calculation... The process proceeds sequentially, determining the number of samples possessing both ω1 and ω2, the number of samples possessing both ω1 and ω3, the number of samples possessing both ω1 and ω4, and so on. Then, it identifies the combination with the largest number of samples, for example, the combination of ω1 and ω2, where the number of samples containing both ω1 and ω2 is the largest. Therefore, ω2 is used as the input feature data for the second part of the model chain. This process continues in the same manner. When determining the input feature data for the third part of the model chain, the same method is used: i.e., calculating... Once the second feature data is determined to be ω2, the number of samples that simultaneously possess ω2 and ω3, the number of samples that simultaneously possess ω2 and ω4, the number of samples that simultaneously possess ω2 and ω5, and so on, is determined sequentially. Then, the combination with the largest number of samples is identified. For example, the combination of ω2 and ω3, i.e., the combination containing both ω2 and ω3, is considered the largest. Therefore, ω3 is used as the input feature data for the third part of the model chain. The remaining determination methods are the same and will not be elaborated further.

[0080] When the number of samples with the same amount of feature data is close to or equal, the model can be trained first, and the accuracy of the overall progressive fusion model at this point can be used to determine which feature to use as input for the next step. In another implementation, the above method can be used as one factor in determining the training order. We can add other factors to consider when determining the order of feature data, such as combining correlation and importance analysis, to obtain the optimal training strategy for the progressive fusion model, or the optimal order of sub-models.

[0081] In a more preferred embodiment, we can introduce correlation analysis of the feature data. Figure 3 This embodiment uses the most comprehensive set of data features during the experiment, specifically 800 data points to train the XGBoost model in a non-progressive fusion scenario. Correlation analysis was performed on some feature indicators, and the results are as follows: Figure 3As shown, among the characteristic indicators of hip fractures, there is a strong positive correlation between total type I procollagen N-terminal elongated peptide, the specific sequence of type I collagen C-terminal peptide β, and N-osteocalcin within the bone marker test data; within the bone mineral density data, there is a strong positive correlation between femoral neck density, total hip density, and L total bone mineral density t-value; among the various characteristic data, there are significant negative correlations between gender and smoking / drinking history, femoral neck density, total hip density, L total bone mineral density t-value, and cross-sectional modulus. In the correlation analysis between characteristic indicators and hip fractures in the elderly, there is a strong negative correlation between femoral neck density, femoral neck density t-value, total hip density, total hip density t-value, L total bone mineral density t-value, and cross-sectional modulus, indicating that bone mineral density data is one of the important indicators for predicting hip fractures in the elderly.

[0082] For correlation analysis between feature data, this embodiment mainly adopts the following method: Based on the training data, calculate the Pearson coefficient between each feature and the Pearson coefficient with the training label to perform preliminary data analysis. Further, rank the correlation between each feature and the label, awarding one point to the corresponding data module for each correlation. The final order of data modules in model training is determined by the total score. In the formula, PS represents the data module score, which is calculated by summing the number of features of the module with the highest importance in the top 20.

[0083] PS = num(top(x))

[0084] After completing the correlation analysis, we need to further determine the relative importance of each feature data, and then combine the importance with the number of sample data mentioned above to optimize the input order of feature data in the model chain.

[0085] Figure 4 Taking the XGBoost model as an example, we obtain the relative importance of various feature data in the prediction of this model. Figure 4 The document displays the top 20 features by importance. Overall, 12 of these top 20 features are derived from basic information and bone mineral density (BMD) data. Using only basic information data yields high prediction accuracy, and BMD data further enhances prediction accuracy. Additionally, alcohol consumption history, hip BMD, gender, age, and fracture history are also significant influencing factors. During model training, the top 20 features in each module are grouped according to their importance, with each feature contributing an additional point to that module. The final order of data modules used in model training is determined by these scores. The IS (Important Scale) is calculated as follows:

[0086] IS = num(top(x))

[0087] In the formula, IS represents the data module score, and the formula is to sum the number of features of the module in the top 20 importance.

[0088] Figure 3 and Figure 4 This embodiment demonstrates that in completing Figure 2 Step S101 in the process. The progressive fusion model chain can effectively explore the interactions between different features in completing the prediction; therefore, through... Figure 3 Correlation analysis between feature groups can preliminarily determine which features should be placed in adjacent training sequences, while correlation analysis between feature groups and hip fractures can help determine which features should be placed in earlier positions during training. For example, bone mineral density data has a higher correlation with hip fractures in the elderly; therefore, when arranging the model order, placing bone mineral density data earlier will provide higher prediction accuracy. Figure 4 This more intuitively demonstrates which feature data should be used as the initial input to the progressively fused model, and which features are of relatively lower importance and need to be placed in later models for fine-tuning. The decision of which data module should be placed before training is based on a combination of relevance and importance scores. The total score S is calculated as follows:

[0089] S = IS + PS

[0090] Where S is the total score of the data module. The larger the S is, the earlier the module is in the order. If the S values ​​are close, the order can be fine-tuned based on the amount of data. For example, modules with larger amounts of data can be placed earlier.

[0091] As shown above, this embodiment, taking into account the number of sample features, feature correlation, and feature importance, determined the basic information data, bone density data, bone marker test data, biochemical data, hemoglobin data, and glycated hemoglobin data as the final optimized input order of feature data.

[0092] In another specific implementation, during the training process of the progressive fusion model, due to the screening of missing values, a situation may arise where the number of positive samples and negative samples differs too much in a specific training step. When this situation occurs, we oversample the training data at each fusion step (i.e., during the training of each sub-model part of the model chain), as follows:

[0093] When this step is first checked, the number of positive samples N is... positive and the number of negative samples N negative When the difference between the two is too large, oversampling is required. For example, in this instance with a total sample size of 1287, when |N positive -Nnegative When |≥100, we can consider the difference between the positive and negative sample sizes to be too large.

[0094] N = max(N positive N negative )

[0095] N is the larger of the number of positive samples and the number of negative samples.

[0096] Since the input data in this embodiment has discontinuous features, such as the basic information data, when oversampling this part of the data, the existing samples of labels that need to be supplemented with feature data are first copied, and then the corresponding discontinuous feature data are randomly modified to construct new positive or negative samples.

[0097] When oversampling continuous features, we first calculate the mean μ and variance τ of all training samples based on the existing data. During oversampling, the corresponding positive or negative samples that need to be added are first copied, and the continuous variables are modified using the following formula:

[0098] X new =X old +N(μ,τ)

[0099] Among them, X new For the new sample, X old The original sample is N(μ,τ), which is a random number that conforms to a normal distribution with mean μ and variance τ. This random number is used as perturbation noise to complete the oversampling of the sample.

[0100] In this embodiment, the established progressive fusion model structure is as follows: Figure 6 As shown, the progressive fusion model in this embodiment includes six sub-models, corresponding to six different sets of feature data used in this embodiment: basic information data, bone density data, bone marker test data, biochemical data, hemoglobin data, and glycated hemoglobin data. It should be noted that the number of sub-models (i.e., the number of fusion stages) in the progressive fusion model can be adjusted based on the number of feature data categories, or according to the actual accuracy requirements of identification. For example, the number of sub-models can be equal to the number of feature data (i.e., the method used in this embodiment), or the number of sub-models can be less than the number of feature data, that is, some feature data can be combined together as input data for a certain sub-model, or the number of sub-models can be greater than the number of feature data, that is, one or more sets of feature data can be input into different sub-models to obtain the final result. All of these adjustments should be considered to fall within the protection scope of this invention.

[0101] according to Figure 2The process is analyzed as follows: First, based on the missing data of these six sets of features, we determined the training path as shown above and completed step S101; then, based on the structure of these six sets of features, the sub-models in this embodiment all adopted the XGBoost model and completed step S102; then, for the input data of each step, we checked the balance of the corresponding positive and negative labels, and for those with too large a difference, we used the above method to oversample and completed step S103; next, we will further complete step S104.

[0102] In step S104, simultaneously combined Figure 6 This progressive fusion construction method based on residual learning utilizes the residual between the prediction result of the previous sub-model and the true label as the label of the subsequent sub-model, thereby achieving effective chain-like correction and optimization between sub-models. The core advantage of this approach lies in its adaptability to datasets with varying degrees of missing data, incorporating any level of missing data into the entire model chain for comprehensive prediction. In practical applications, this approach first uses a base sub-model to make preliminary predictions on the data, then calculates the residual between the predicted result and the actual observed value. This residual actually reflects the prediction error of the sub-model and can therefore be considered a valuable source of information. Using the residual as input to subsequent sub-models helps them learn data features that previous sub-models failed to capture, thereby correcting and refining the prediction results.

[0103] Combination Figure 6 As shown, the progressive fusion model comprises multiple interconnected sub-models. The residual between the prediction result of the previous sub-model and the true label serves as the new label for the sample data input to the next sub-model. The next sub-model is trained based on the labeled sample data to obtain a prediction result. This prediction result is then recalculated to obtain the residual between itself and the true label, and this process continues, updating the labels of the sample data for the next-level sub-model, and so on, until the entire model is trained, forming the progressive fusion model. This ensures that the training of subsequent sub-models does not affect the parameters already trained in the previous sub-model. For the trained progressive fusion model, when predicting real-world data, the individual data to be predicted is first classified according to different feature data. The corresponding feature data is then used as the input to the corresponding sub-model. The final prediction result is obtained through a backward correction process, which will be explained in detail later.

[0104] The following explanation uses four types of feature data as an example. Assume that different samples contain four different test data categories, namely the four types of feature data mentioned earlier: ω1, ω2, ω3, and ω4, corresponding to four different samples X1, X2, X3, and X4. To simulate the data gap problem in small-sample medical data modeling, we assume that different samples have different types of data. X1 has four types of feature data, and the others decrease in number. For example, sample X1 has the following features: ω1, ω2, ω3, and ω4; sample X2 has the following features: ω1, ω2, and ω3; sample X3 has the following features: ω1 and ω2; and sample X4 has the following feature: ω1. In the first round of model training, since the number of samples containing feature data ω1 is the largest, feature ω1 can be used as input, and the labels Y1, Y2, Y3, and Y4 of the four samples can be used as output Y to train the first basic sub-model M1.

[0105] Y = M1(ω1)

[0106] Since X4 only contains feature ω1, the data contained in X4 will not be used in subsequent training. In the second round of model training, a progressive fusion residual learning mechanism is required. The model uses the residual learning mechanism to allow subsequent sub-models to correct the preceding sub-models. Specifically:

[0107] In the second round of training, the three samples participating in the training (i.e., X1, X2, X3) all have two features, ω1 and ω2. Therefore, the first set of features ω1 (of course, another set of features ω2 can also be used as the first set of features, this is just an example) is first input into the model M1 to obtain the prediction result P. m Using P m And the real label Y m The residuals are calculated as follows:

[0108] E 1m =C(P) m Y m m = 1, 2, ..., K

[0109] Where K represents the total number of samples input to the current sub-model, C represents the residual operation, and E 1m This represents the residual of submodel 1 with respect to the m-th sample.

[0110]

[0111] E 1m As the new labels for the corresponding samples during the training of the second sub-model M2, the labels of the sample data input to the sub-model M2 are updated. At this time, it is not necessary to train the first sub-model M1 again. Instead, the additional feature ω2 is used to complete the training of the second sub-model M2.

[0112] E1=M2(ω2)

[0113] E1 represents the total residual obtained from sub-model M1, i.e., E 1m The overall residual from the first step after merging is used to train the second model M2. The overall progressive fusion model at this point can be expressed as:

[0114] M = M1←M2

[0115] This process is repeated to obtain a complete model chain. When there are 6 sub-models in the model chain, step S104 is completed, and the complete model chain is obtained.

[0116] M=M1←M2←M3←M4←M5←M6

[0117] Where ← represents the chained correction of the model.

[0118] During training, the model chain is trained sequentially based on data with different missing values. During prediction, the model uses chained correction, i.e., prediction and correction are performed in reverse order, with the total number of prediction steps determined by the number of features it possesses. Assuming a sample has three sets of data features ω1, ω2, and ω3, then:

[0119] First, its feature ω1 is used as the input to model M1 to obtain the first step prediction result P1:

[0120] P1=M1(ω1)

[0121] If subsequent features are needed to optimize the prediction, then feature ω2 is used as the input to sub-model M2. During ordered training, the labels of sub-model M2 are the residuals. Therefore, during prediction, the output of sub-model M2 is the residual between sub-model M1 and the true labels.

[0122] E1=M2(ω2)

[0123] This residual can be used to correct the output P1 of the first step model:

[0124]

[0125] Where P new1 To correct the prediction result after one iteration, -P1 represents the label opposite to the prediction result of P1. In this case, P can be... new1 The prediction results of the progressive fusion two-step model are output as the result of the prediction. Similarly, when a new feature is added, the subsequent sub-model is used to obtain a new prediction residual, which is then used to correct the prediction results of the previous sub-model step by step, progressively correcting each step to obtain the corresponding progressive fusion model prediction result. This process continues until all data predictions are completed, thus completing step S106.

[0126] If, during prediction, there are gaps between individual patient data and training model data, assuming the model only has data features corresponding to sub-models 1, 2, and 5, while data features corresponding to sub-models 3 and 4 are missing, this solution addresses this issue as follows:

[0127] The progressive fusion model's prediction is a step-by-step prediction. If the result predicted by the next model is 0, it means that the subsequent model has determined that the result of the preceding model does not need to be modified. If the prediction result is 1, it means that the subsequent model determines that the prediction result of the preceding model needs to be modified. Applying this to the prediction of missing features, we can adopt the following strategy: After models 1 and 2 complete their predictions, we do not have the data needed by models 3 and 4, meaning that the models have not yet corrected the situation of the preceding models. Therefore, during prediction, we can manually set the prediction results of models 3 and 4 to 0. Model 5 then faces the prediction result still from model 2, meaning that model 5 directly corrects model 2, resolving the data missing situation during prediction. That is: let the features of the sample be ω1, ω2, and ω5, and the models be M1, M2, M3, M4, and M5 respectively.

[0128] Then the prediction result of the second sub-model is:

[0129] P2=M2(ω2)

[0130] By manually setting the predictions P3 and P4 of sub-models 3 and 4 to 0, the prediction model chain changes, thus resolving the issue of missing data features.

[0131] M=M1←M2←M3(P3=0)←M4(P4=0)←M5

[0132] This approach can achieve different levels of prediction accuracy based on the amount of feature data available, thus avoiding model training problems caused by the lack of data structuring.

[0133] In addition, for datasets with many data features, there may be multiple prediction orders. In this case, it is necessary to conduct preliminary experiments to determine the model performance under different prediction orders before determining the final progressive fusion order.

[0134] Furthermore, in this embodiment, different sub-models suitable for different data types can be used for matching. Various machine learning methods can be employed to model different data based on the data type, thereby constructing various sub-models in the model chain. In this embodiment, each sub-model can adopt the following sub-models according to its corresponding type. The selection of sub-models can be based on the continuity or discontinuity of the data itself. Preliminary experiments can be conducted to determine which model will have better performance for the data at this stage, and all can be directly connected to the progressive fusion model:

[0135] 1. Logistic Regression Model: In this embodiment, a binary classification method is used for logistic regression. The logistic regression model used in this embodiment is set as follows:

[0136]

[0137] Where: P(y=1|x) represents the probability of event y=1, β0, β1, ..., β n These represent the corresponding parameters in the logistic regression model.

[0138] 2. Support Vector Machine (SVM): In this embodiment, the decision function of the linear support vector machine is:

[0139] f(x)=sign(w·x+b) (2)

[0140] Where w is the weight vector and b is the bias parameter.

[0141] 3. Decision Tree: In this embodiment, each internal node of the decision tree represents a test for a feature, and each leaf node represents a category. This embodiment combines decision tree classification with random forest, and its implementation is as follows: Information Gain (used to select the best feature):

[0142] IG(T,X)=H(T)-H(T|X) (3)

[0143] Gini index (used to select the best feature):

[0144]

[0145] p i is the probability of category i in the dataset, c represents the total number of categories, T represents the total sample set, and X represents the optimal feature set.

[0146] 4. Gradient Boosting Decision Tree: Configure multiple decision trees to build a random forest. The training set is randomly generated using the Bagging (Bootstrap Aggregating) method. The prediction formula is as follows:

[0147]

[0148] Where: h k (x) represents the prediction result of the k-th tree. This represents the overall prediction result, K represents the total number of decision trees, and mode represents the ensemble decision algorithm.

[0149] The decision tree learning process is combined with improvements to the overall model's predictive capabilities.

[0150] Updates at each step:

[0151] F m (x)=F m-1 (x)+η·l m (x) (6)

[0152] Wherein: F m (x) represents the model at the m-th iteration, η is the learning rate, and l m (x) is the newly added weak learner.

[0153] In the above embodiments, x represents the input data in the corresponding sub-model.

[0154] Furthermore, in this embodiment, other types of sub-models can be added, such as KNN classifiers, Naive Bayes classifiers, and XGBoost classifiers. Selecting the optimal model for training based on different application data types is fundamental to achieving the analytical effect of progressive fusion models.

[0155] Figure 5 This embodiment uses 1287 elderly hip fracture samples extracted and screened for training and testing the progressive fusion model. The dataset is divided into training and testing sets in a 7:3 ratio, with 900 and 287 data points respectively. XGBoost, Random Forest, Logistic Regression, Naive Bayes, KNN, Support Vector Machine, and the progressive fusion model were trained. The traditional machine learning model was trained using basic information features without missing data. The performance of the progressive fusion model and the other six common machine learning models (based on the largest available data) in predicting the risk of hip fractures in the elderly was compared. Figure 5 The curves show the actual prediction results of different models.

[0156] The progressive fusion model uses a specific order, which in this embodiment is set as basic information data, bone density data, bone marker test data, biochemical data, hemoglobin data, and glycated hemoglobin data.

[0157] Analyzing risk prediction from the perspectives of prediction accuracy, precision, recall, and F1 score, as shown in Table 1, the progressive fusion model significantly outperforms the other six machine learning models in accuracy, precision, and F1 score, while ranking second in recall among all models. Compared to traditional machine learning models that only use basic information features to obtain the maximum amount of data possible, the experimental results demonstrate the superiority of the progressive fusion model in predicting the risk of hip fractures in the elderly.

[0158] Table 1 Comparison of Risk Prediction Effectiveness

[0159]

[0160] Further plotting the ROC graphs for each model revealed that all six models achieved similar results at the end of the curve, with the progressive fusion model showing the best performance in the remaining portions. From an AUC perspective, the progressive fusion model achieved a predicted AUC of 0.9287, significantly outperforming the other machine learning models. This further confirms the effectiveness of the progressive fusion model in addressing the problem of missing data.

[0161] In yet another embodiment, this solution can also be implemented systematically, the system comprising:

[0162] The data preprocessing module collects sample data on hip bone mechanics in the elderly, preprocesses the sample data, and divides it into multiple types of feature data according to data type; and preprocesses the data to be predicted and divides it into multiple types of feature data according to data type.

[0163] The model building module constructs a progressive fusion model containing multiple sub-models, determines the order of the sub-models, and establishes chain connections between the sub-models in sequence through residuals.

[0164] The model training module is used to train the corresponding sub-model during the model training phase, based on the number of samples containing feature data of each type, using feature data of a single type and the labels of the corresponding samples, and using the residual between the prediction result of the previous sub-model and the true label of the sample as the new label of the sample of the next sub-model for the training of the next sub-model.

[0165] The online prediction module, based on the feature data of the data to be predicted, inputs the feature data of the last type into the sub-model corresponding to that type in the reverse order of the model training phase. The prediction result is used as the residual of the previous sub-model to correct the prediction result of the previous sub-model. This process continues until the prediction results of all types of feature data have been corrected, and then the final prediction result is output.

[0166] Of course, the above system module settings can also be adjusted appropriately based on actual operating requirements. For example, a single function of different modules can be adjusted or merged, or merged with other modules.

[0167] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0168] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside on local and remote computer storage media, including storage devices.

[0169] The various embodiments in this specification are described in a progressive manner. For the same or similar parts between the various embodiments, please refer to each other. Each embodiment focuses on describing the differences from other embodiments.

[0170] The above description is merely an embodiment of this specification and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for analyzing the biomechanical characteristics of the hip in the elderly based on progressive fusion, characterized in that, The method includes: S1. Collect sample data on hip bone mechanics in the elderly, preprocess the sample data, and classify it into multiple types of feature data according to data type; S2. Construct a progressive fusion model containing multiple sub-models, determine the order of the sub-models, and establish chain connections between the sub-models in sequence through residuals. S3. During the model training phase, based on the number of samples containing feature data of each type, the corresponding sub-model is trained using feature data of a single type and the labels of the corresponding samples. The residual between the prediction result of the previous sub-model and the true label of the sample is used as the new label of the sample of the next sub-model for training the next sub-model. S4. In the model prediction stage, the data to be predicted is divided into multiple types of feature data in the manner described in S1. In the reverse order of the model training stage, the feature data of the last type is input into the sub-model corresponding to that type. The prediction result is used as the residual of the previous sub-model to correct the prediction result of the previous sub-model. This process continues until the prediction results of all types of feature data are corrected, and the final prediction result is output.

2. The method according to claim 1, characterized in that, In step S2, the order of the sub-models is determined based on the number of samples with corresponding feature data types. The feature data with the largest number of samples of the corresponding feature data type is used as the input data of the first-level sub-model, and its corresponding sub-model is used as the first-level sub-model. The order of subsequent feature data is calculated as follows: Where I represents the total number of feature data types, To simultaneously possess ω i and ω j Samples of feature data For the sample with ω i The sample size with the largest sample size among the remaining feature data; The larger the value, the higher the order of the corresponding sub-model.

3. The method according to claim 1, characterized in that, In S3, during the model training phase, when the difference between the number of positive samples and the number of negative samples is large, the sample size is oversampled in the following way: First, the larger of the number of positive samples and the number of negative samples is taken as the unified number N of the number of negative samples and the number of positive samples after sampling, that is: N = max(N) positive N negative ) Where, N positive N represents the number of positive samples. negative Indicates the number of negative samples; Secondly, for sample data where feature data is discontinuous due to missing data, first copy the sample that needs to be supplemented with feature data, and then calculate the mean μ and variance τ of the feature data of all samples. Next, discontinuous sample data are supplemented into continuous variables in the following way: X new =ω old +N(μ,τ) Among them, X new For the new sample, X old The original sample is N(μ,τ), which is a random number that conforms to a normal distribution with mean μ and variance τ. Oversampling is then performed.

4. The method according to claim 1, characterized in that, In step S3, the new label is calculated as follows: First, calculate the residuals of the first sub-model: E 1m =C(P m ,Y m ),m=1、2、…、K Where K represents the total number of samples input to the current sub-model, C represents the residual operation, and E 1m Let C represent the residual of the first sub-model with respect to the m-th sample, where C is calculated as follows: Among them, P m Y represents the prediction result for sample m. m This represents the true label of sample m; Secondly, E 1m The new label is assigned to each corresponding sample and then used as the second set of feature data and the new label to train the next sub-model.

5. The method according to claim 4, characterized in that, Once each sub-model has completed training, a model chain is formed that progressively merges the models. M=M1←M2←…←M p Where ← represents the chained correction of the model, and p represents the total number of sub-models.

6. The method according to claim 1, characterized in that, In S3, during the model training phase, the order of the sub-models in the progressive fusion model is consistent with the order of the feature data, with one type of feature data corresponding to one sub-model.

7. The method according to claim 1, characterized in that, In step S4, the prediction results of the previous sub-model are updated in the following way: First, input the feature data corresponding to the previous sub-model into this sub-model to obtain the prediction result P1; Secondly, the feature data corresponding to this level sub-model is input into this level sub-model, and the resulting prediction result E1 is used as the residual of the previous level sub-model. Next, the prediction result P1 of the previous sub-model is corrected using the residual E1: Among them, P new1 -P1 represents the corrected prediction result, indicating a label that is the opposite of the P1 prediction result.

8. The method according to claim 7, characterized in that, When the feature data type of the data to be predicted is missing, causing the sub-models of the corresponding prediction steps to be not directly continuous, the prediction result of the sub-model corresponding to the missing feature data type is set to 0 by default, so that its prediction result indicates that there is no need to correct the prediction result of the preceding sub-model.

9. The method according to claim 1, characterized in that, In S2, the order of the sub-models is determined based on the number of samples with corresponding feature data types, the importance of the feature data, and the correlation between the feature data: First, the correlation of the feature data is calculated: the Pearson coefficient between each feature data and the Pearson coefficient between the training labels are calculated, and the models are sorted according to the Pearson coefficient between the feature data and the Pearson coefficient between the training labels respectively. The correlation score of the sub-model corresponding to the feature data is determined based on the ranking. Next, the importance of each feature data is calculated based on the XGBoost model and sorted to calculate the importance score of the sub-module corresponding to the feature data. The order of sub-models is determined by the sum of their relevance and importance scores. When the sum of the relevance and importance scores of multiple sub-models is equal, the larger the amount of feature data corresponding to the sub-model, the higher the priority of the sub-model in the ranking.

10. A progressive fusion-based system for analyzing the biomechanical characteristics of the hip in the elderly, characterized in that, The system includes: The data preprocessing module collects sample data on hip bone mechanics in the elderly, preprocesses the sample data, and divides it into multiple types of feature data according to data type; and preprocesses the data to be predicted and divides it into multiple types of feature data according to data type. The model building module constructs a progressive fusion model containing multiple sub-models, determines the order of the sub-models, and establishes chain connections between the sub-models in sequence through residuals. The model training module is used to train the corresponding sub-model during the model training phase, based on the number of samples containing feature data of each type, using feature data of a single type and the labels of the corresponding samples, and using the residual between the prediction result of the previous sub-model and the true label of the sample as the new label of the sample of the next sub-model for the training of the next sub-model. The online prediction module, based on the feature data of the data to be predicted, inputs the feature data of the last type into the sub-model corresponding to that type in the reverse order of the model training phase. The prediction result is used as the residual of the previous sub-model to correct the prediction result of the previous sub-model. This process continues until the prediction results of all types of feature data have been corrected, and then the final prediction result is output.

Images