Lumbar bone mineral density index prediction method and system based on CNN-LSTM hybrid model

Abstract

The invention discloses a lumbar vertebra bone mineral density index prediction method and system based on a CNN-LSTM hybrid model, belongs to the technical field of medical data processing, and is used for accurately predicting lumbar vertebra bone mineral density to assist clinical decision making. The system is divided into four layers of architectures, wherein a data input layer integrates 42 clinical features and a lumbar vertebra bone mineral density true value detected by DXA; the data preprocessing layer purifies data through missing value and abnormal value processing and MIC algorithm feature screening; the model prediction layer extracts local features through CNN, models a sequential relationship through LSTM, and optimizes parameters in combination with an Adam optimizer and an early stop strategy; and the result output layer identifies the delivery result through prediction calculation, visualization and high-risk population. Through cooperation of multi-source data processing and the mixed model, the prediction precision is improved, and support is provided for osteoporosis screening and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical data processing, and particularly relates to a lumbar bone mineral density index prediction method and system based on a CNN-LSTM hybrid model. BACKGROUND

[0002] Lumbar bone mineral density (BMD) is a key indicator for assessing human bone health and plays an irreplaceable role in early diagnosis, disease monitoring, and treatment effect evaluation of bone diseases such as osteoporosis. Osteoporosis is a systemic bone disease characterized by reduced bone mass and damaged bone microstructure, which easily leads to increased bone fragility and elevated risk of fractures.

[0003] Currently, the commonly used method for detecting lumbar bone mineral density in clinics is dual-energy X-ray absorptiometry (DXA). The bone density detection technology represented by DXA calculates bone density by emitting two different energy X-rays to the human body according to the attenuation difference of the X-rays penetrating the bone and soft tissue. This method is considered as the "gold standard" for bone density measurement, but it relies on professional equipment, and the cost of equipment purchase and maintenance is high, with a single detection device usually exceeding one million yuan. Moreover, the operation process is complex, and professional personnel are required for operation. In addition, radiation is generated during the detection process, and each detection generates about 0.01-0.02 mSv of radiation, which is not suitable for special groups such as pregnant women and children, and cannot meet the needs of rapid screening and high-frequency monitoring of large-scale populations. At the same time, the detection sensitivity of DXA for early bone mass loss is low, and abnormalities are often not found until the patient's bone mass has significantly decreased, missing the opportunity for early intervention.

[0004] In recent years, deep learning technology has rapidly developed in the medical field, providing a new approach to solving complex medical problems. Convolutional neural networks (CNN) have strong local feature extraction capabilities and have performed outstandingly in image recognition and medical data feature mining tasks. Long short-term memory networks (LSTM) can effectively process time series data and capture long-term dependencies in the data, and have an advantage in trend prediction tasks. However, in the existing technology, using CNN or LSTM model alone to predict lumbar spine bone density cannot simultaneously consider the mining of local features and time series relationships, resulting in difficulty in meeting the clinical demand for prediction accuracy. For example, some studies use CNN to extract local features (such as the correlation between laboratory indicators) in clinical indicators to predict bone density, but ignore the time series characteristics of some indicators (such as hormone levels that change over time), resulting in insufficient ability of the model to predict long-term bone mass trends. Some solutions use LSTM to process longitudinal clinical data (such as multiple follow-up detection indicators), but cannot effectively extract local key features within the indicators (such as the synergistic effect of height and weight), and the model converges slowly and is prone to overfitting. In addition, lumbar spine bone density related data (such as age, height, weight, laboratory indicators, and lifestyle habits) often have missing values, outliers, and high feature dimensions. Existing data preprocessing methods often use a single method to handle data problems (such as simple mean filling and single outlier deletion), which cannot specifically address the quality problems of different types of data (such as continuous age and categorical smoking status), and do not combine the characteristics of the bone density prediction task for feature selection, resulting in redundant features interfering with model training.

[0005] In view of the above, there is an urgent need to provide a lumbar spine bone density index prediction scheme that is radiation-free, low-cost, and easy to popularize.

[0006] The information disclosed in this Background section is only intended to increase an understanding of the general context in which the present application can be practiced. It should not be taken as an acknowledgement or any form of suggestion that this information forms prior art that is already widely known and available to the general public before the application was made. SUMMARY

[0007] The purpose of the present application is to provide a lumbar spine bone density index prediction method and system based on a CNN-LSTM hybrid model, which can improve the accuracy and fitting effect of lumbar spine bone density prediction; solve the problems of missing values, outliers, and feature redundancy, improve data quality, and provide reliable input for model training; realize early identification of high-risk groups of osteoporosis, provide technical support for clinicians to develop personalized intervention strategies, and reduce the risk of fractures.

[0008] To achieve the above-mentioned purpose, the present application provides a lumbar spine bone density index prediction method based on a CNN-LSTM hybrid model, comprising the following steps:

[0009] Collecting lumbar spine bone density clinical feature data;

[0010] The lumbar vertebrae density clinical feature data is classified and preprocessed to obtain a preprocessed effective feature set, the effective feature set including missing values, abnormal values and feature values;

[0011] The preprocessed effective feature set is divided into a training set and a test set;

[0012] A CNN-LSTM hybrid model is constructed, wherein the CNN subnetwork includes a convolutional layer, a pooling layer and a flattening layer, the LSTM subnetwork includes an LSTM layer and a fully connected layer, and the training set is used to train the CNN-LSTM hybrid model;

[0013] The trained CNN-LSTM hybrid model is evaluated using the test set, and the evaluation indicators include mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 ;

[0014] The prediction result of the lumbar vertebrae density value is output, including the lumbar vertebrae density prediction value, the visualization chart of the prediction result and the marking of the high-risk population corresponding to the lumbar vertebrae density value.

[0015] Optionally, the bone density clinical feature data is classified and preprocessed, specifically including:

[0016] The missing value processing adopts a combination of moving median, K-nearest neighbor algorithm and random filling;

[0017] The abnormal value processing adopts a combination of linear interpolation, mean smoothing and isolation forest algorithm;

[0018] The feature screening adopts MIC threshold screening, and the MIC threshold is 0.12.

[0019] Optionally, in the missing value processing, the moving window size is set to 5, and the K value of the K-nearest neighbor algorithm is 3;

[0020] In the abnormal value processing, the abnormal values are identified by a box plot, a straight line is fitted based on the 5 normal data points before and after the abnormal values, and the window size of the mean smoothing is 3.

[0021] Optionally, the bone density clinical feature data is classified and preprocessed, specifically including:

[0022] The missing value processing adopts multiple imputation by chained equations (MICE), and multiple missing value filling schemes are generated by constructing a regression model;

[0023] The feature screening adopts a combination of recursive feature elimination (RFE) and MIC, redundant features are deleted by RFE, and the feature relevance is verified by MIC.

[0024] Optionally, the CNN-LSTM hybrid model adopts a serial architecture, first extracting local features through CNN, and then processing temporal relationships through LSTM.

[0025] Optionally, in the CNN sub-network, the number of convolutional layers is 2, the number of convolutional kernels in each convolutional layer is 32 and 64 respectively, and the activation function is ReLU; the pooling layer is a max pooling layer with a pooling kernel size of 2×2; the flattening layer converts the feature map output by the pooling layer into a one-dimensional vector;

[0026] In the LSTM subnetwork, the LSTM layer has 128 units, the activation function is tanh, and the fully connected layer has 1 output unit.

[0027] Optionally, in the step of training the CNN-LSTM hybrid model using the training set, the Adam algorithm is used, the initial learning rate is set to 0.001, the maximum number of training epochs is set to 300, and an early stopping strategy is adopted, where training is stopped when the validation set error does not decrease for 10 consecutive epochs.

[0028] Optionally, the visualization charts include a line graph comparing the predicted and actual values ​​of lumbar spine bone mineral density, a residual distribution histogram, and a prediction interval graph for different confidence levels;

[0029] The high-risk population marker is based on a threshold division of lumbar spine bone mineral density (BSMD) values. When the BSMD value is lower than the preset percentile of the reference BSMD value for people of the same sex and age, the person is marked as a high-risk population.

[0030] Optionally, the CNN-LSTM hybrid model adopts a parallel architecture, where the CNN and LSTM process the same effective feature set respectively, are then fused through an attention mechanism, and then input into the fully connected layer to predict bone density.

[0031] A second aspect of the present invention also provides a lumbar spine bone mineral density index prediction system based on a CNN-LSTM hybrid model, which is based on the lumbar spine bone mineral density index prediction method based on a CNN-LSTM hybrid model described in the first aspect of the present invention. The system includes:

[0032] The data input layer is used to collect clinical characteristic data of lumbar spine bone mineral density.

[0033] The data preprocessing layer is used to perform categorical preprocessing on the clinical feature data of lumbar spine bone mineral density to obtain a preprocessed effective feature set, which includes missing values, outliers, and feature values.

[0034] The model prediction layer is used to divide the preprocessed effective feature set into a training set and a test set to construct a CNN-LSTM hybrid model. The CNN sub-network includes convolutional layers, pooling layers, and flattening layers, and the LSTM sub-network includes LSTM layers and fully connected layers. The CNN-LSTM hybrid model is trained using the training set. The trained CNN-LSTM hybrid model is evaluated using the test set, and the evaluation metrics include mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²).

[0035] The results output layer is used to output the predicted results of lumbar spine bone mineral density values. The predicted results include the predicted lumbar spine bone mineral density value, a visualization chart of the predicted results, and a marker of the high-risk groups corresponding to the lumbar spine bone mineral density values.

[0036] Compared with the prior art, the lumbar spine bone mineral density index prediction method and system based on the CNN-LSTM hybrid model according to the present invention has the following advantages or beneficial effects;

[0037] The clinical characteristic data of lumbar spine bone mineral density were preprocessed by classification, and differentiated missing value imputation methods were designed for different types of characteristics, including moving median, K-nearest neighbor, and random imputation. Outliers were handled by combining linear interpolation, mean smoothing, and isolated forest methods, thus improving data quality. Furthermore, lumbar spine bone mineral density can be predicted based solely on clinical indicators without relying on DXA equipment, posing no radiation risk. The equipment investment only requires a regular computer, reducing costs by more than 90%, enabling large-scale application in community hospitals, health check centers, and other primary care settings.

[0038] Features strongly correlated with lumbar spine bone density are selected based on the maximum information coefficient (threshold 0.12), eliminating redundant information, reducing model computation, and improving generalization ability.

[0039] A CNN-LSTM hybrid model architecture was constructed. By connecting CNN and LSTM in series or in parallel, CNN extracts local features (such as the synergistic relationship between indicators), and LSTM captures temporal dependencies (such as long-term bone mass change trends), thus solving the functional limitations of a single model.

[0040] The Adam optimizer (initial learning rate 0.001) and early stopping strategy (patience=10) are used to update the CNN and LSTM parameters simultaneously, ensuring that the model converges quickly and avoiding overfitting.

[0041] This invention also realizes a clinically adapted result output system that integrates predicted value calculation, visualization, and high-risk population identification, providing clinicians with intuitive and interpretable diagnostic support and adapting to the needs of primary healthcare scenarios. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating a method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to an embodiment of the present invention.

[0043] Figure 2 This is a structural block diagram of a lumbar spine bone mineral density index prediction system based on a CNN-LSTM hybrid model according to an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram illustrating the feature selection process according to an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the CNN-LSTM hybrid model structure according to an embodiment of the present invention;

[0046] Figure 4a for Figure 4 A schematic diagram of the structure of the LSTM model;

[0047] Figure 5 This is a schematic diagram comparing the prediction results of three detection modes based on a CNN model, an LSTM model, and a CNN-LSTM hybrid model according to one embodiment of the present invention.

[0048] Figure 6 This is a schematic diagram illustrating the evaluation of the prediction residuals using the CNN-LSTM method according to one embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0051] It should be understood that in the various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0052] It should be understood that in this invention, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0053] It should be understood that in this invention, "multiple" refers to two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, "and / or B" can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "Contains A, B, and C", "Contains A, B, and C" means that all three A, B, and C are contained; "Contains A, B, or C" means that one of A, B, and C is contained; "Contains A, B, and / or C" means that any one, two, or three of A, B, and C are contained.

[0054] It should be understood that in this invention, "B corresponding to A", "B corresponding to A", "A and B correspond", or "B and A correspond" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. Matching A and B is defined as a similarity between A and B that is greater than or equal to a preset threshold.

[0055] Depending on the context, "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection."

[0056] Traditional methods for detecting lumbar spine bone mineral density (DXA) suffer from radiation exposure, high equipment costs, and complex operation, making them difficult to implement in large-scale population screening and primary healthcare settings. Furthermore, they have low sensitivity to early bone loss, failing to identify high-risk groups for osteoporosis in a timely manner. Existing single deep learning models (such as CNN or LSTM alone) have limitations in lumbar spine bone mineral density prediction: CNN excels at extracting local features but cannot capture temporal dependencies in the data; while LSTM can process temporal data, its ability to extract key local features is weak, resulting in poor prediction accuracy and fitting performance. Lumbar spine bone mineral density data suffers from quality issues such as missing values ​​and outliers, and its high feature dimensionality means that directly using it for model training reduces prediction performance. Existing data preprocessing methods are insufficient to comprehensively address these issues to meet the demands of high-precision prediction. This invention constructs a CNN-LSTM hybrid model, integrating the local feature extraction capabilities of CNN and the temporal modeling capabilities of LSTM to improve the accuracy and fitting performance of lumbar spine bone mineral density prediction. A targeted data preprocessing workflow was designed to address issues such as missing values, outliers, and feature redundancy, thereby improving data quality and providing reliable input for model training. Furthermore, it enables early identification of high-risk groups for osteoporosis, providing technical support for clinicians to develop personalized intervention strategies and reduce fracture risk.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Terminology Explanation:

[0059] MIC (Maximum Information Coefficient): A metric that measures the degree of association between two variables. It ranges from 0 to 1, with a larger value indicating a stronger association. It can capture both linear and nonlinear relationships.

[0060] MSE (mean squared error): the average of the squares of the differences between the predicted and actual values, reflecting the overall level of prediction error; the smaller the value, the higher the accuracy.

[0061] RMSE (Root Mean Square Error): The square root of MSE, reflecting the average error between the predicted and actual values.

[0062] MAE (Mean Absolute Error): The average of the absolute errors between the predicted and actual values.

[0063] MAPE (Mean Absolute Percentage Error): The percentage of the absolute error between the predicted and actual values.

[0064] R² (coefficient of determination): measures how well a model fits the data. It ranges from -∞ to 1. The closer the value is to 1, the better the fit. A negative value indicates that the model's predictive ability is weaker than that of simple mean prediction.

[0065] like Figure 1 As shown, the lumbar spine bone mineral density index prediction method based on a CNN-LSTM hybrid model according to a preferred embodiment of the present invention includes the following steps:

[0066] It should be noted that the clinical characteristic data of lumbar spine bone mineral density collected includes the collected clinical data and the actual value of lumbar spine bone mineral density measured by DXA. The collected clinical data includes 42 input features related to lumbar spine bone mineral density, such as age, height, weight, age at menarche, age at menopause, smoking status, and laboratory indicators.

[0067] The clinical feature data of lumbar spine bone mineral density are preprocessed by classification to obtain a preprocessed effective feature set, which includes missing values, outliers and feature values.

[0068] The preprocessed effective feature set is divided into a training set and a test set;

[0069] Construct a CNN-LSTM hybrid model, wherein the CNN sub-network includes convolutional layers, pooling layers, and flattening layers, and the LSTM sub-network includes LSTM layers and fully connected layers, and train the CNN-LSTM hybrid model using the training set;

[0070] The trained CNN-LSTM hybrid model was evaluated using the test set, and the evaluation metrics included mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ;

[0071] Output the predicted lumbar spine bone mineral density value, which includes the predicted lumbar spine bone mineral density value, a visualization chart of the predicted result, and a marker of the high-risk group corresponding to the lumbar spine bone mineral density value.

[0072] In this embodiment of the invention, the clinical bone mineral density data undergoes categorized preprocessing, specifically including:

[0073] Missing value handling employs a combination of moving median, K-nearest neighbor (KNN) algorithm, and random imputation. It's important to note that in the missing value handling steps, for continuous variables (such as age, height, and weight), the moving median method is first used to imput missing values ​​(with a moving window size of 5, using the median of adjacent data to reduce the impact of extreme values). The remaining unimpeded missing values ​​are then filled using the KNN algorithm (K=3) to ensure data continuity. After imputation, age data is integerized to ensure clinical relevance. For categorical variables (such as smoking status, 1=smoker, 2=non-smoker), missing values ​​are randomly imputed based on population prevalence (80% of the dataset consists of non-smokers) to avoid data bias caused by a single imputation method. For other variables (such as age at menarche and number of pregnancies), the handling logic for continuous or categorical variables is followed, and the moving median or KNN algorithm is selected based on the variable type to ensure a 100% imputation rate for all missing values.

[0074] Outlier handling employs a combination of linear interpolation, mean smoothing, and the Isolation Forest algorithm. Specifically, the outlier handling step uses a combination of linear interpolation and mean smoothing: first, outliers are identified using box plots (interquartile range IQR = 1.5). Linear interpolation (fitting a straight line based on the five normal data points before and after the outlier) is then used for initial correction of the outlier's location. Next, mean smoothing (with a window size of 3) is applied to the corrected data to further reduce the impact of outliers on the data distribution. In the outlier removal step, the Isolation Forest algorithm is used, with an outlier threshold of 5%. Extreme outliers that cannot be corrected by interpolation (such as unreasonable data with heights <100cm or >220cm) are removed to ensure the validity of the dataset.

[0075] It should be further noted that the interquartile range (IQR) is a core indicator for measuring the dispersion of data. Its calculation and application require consideration of the data's quartiles (three key nodes, denoted as Q1, Q2, and Q3, after sorting the data from smallest to largest, based on the 25th, 50th, and 75th quartiles). This invention addresses outlier handling scenarios for lumbar spine bone mineral density-related data (such as age, height, weight, and laboratory indicators), selecting an IQR of 1.5 as the initial outlier identification standard, primarily based on considerations of data characteristics, task requirements, and technical adaptability. The lumbar spine bone mineral density input data includes continuous physiological indicators (such as age 45-55 years, height 145-175cm, and weight 40-80kg) and laboratory test indicators (such as osteocalcin and parathyroid hormone). This type of data conforms to an "approximately normal distribution." The physiological indicators (such as height and weight) and laboratory indicators of most healthy individuals tend to cluster around the mean. The interim quality range (IQR) of the middle 50% of the data effectively represents the fluctuation range of normal data. Outliers in this type of data are mostly "extreme deviation values": such as height < 100cm or > 220cm, weight < 30kg or > 150kg, values ​​that significantly exceed medical common sense. These are often due to data collection errors (such as unit recording bias or entry mistakes) or extreme individual cases, requiring screening through a clearly defined threshold. The standard of "IQR = 1.5" can accurately identify such extreme outliers without excessively rejecting marginal data close to the normal range (such as individual data like a height of 140cm or a weight of 85kg, which, although slightly deviating from the mean, may still exist), thus avoiding data loss due to outlier processing. This invention does not rely solely on "IQR=1.5" to screen outliers, but rather uses it as an "initial identification step"—first, potential outliers are marked using this standard, then "linear interpolation + mean smoothing" is used to correct correctable marginal outliers, and finally, "Isolation Forest algorithm (outlier proportion threshold of 5%)" is used to delete uncorrectable extreme outliers. The intermediate threshold characteristic of "IQR=1.5" provides a reasonable set of outlier candidates for the subsequent two processing steps, ensuring a balance between data quality and sample size.

[0076] Feature selection employed a MIC threshold of 0.12. It's important to note that the Maximum Information Coefficient (MIC) was used to quantify the correlation between features and lumbar spine bone mineral density (the target variable). MIC effectively captures both linear and nonlinear relationships. By setting the MIC threshold to 0.12, features with MIC values ​​≥ 0.12 (such as osteocalcin and age at first pregnancy) were retained, while redundant features (such as some lifestyle indicators with weak correlations to bone mineral density) were removed. This resulted in a final selection of 18-22 core features, reducing model computation and improving generalization ability. The core logic of feature selection is to retain features strongly correlated with lumbar spine bone mineral density (the target variable) and eliminate redundant and weakly correlated information. The range of 18-22 features is directly determined by the selection result of "MIC threshold = 0.12". The Maximum Information Coefficient (MIC) is a quantitative indicator (ranging from 0 to 1) that measures linear and nonlinear relationships between variables; a higher value indicates a stronger correlation. This invention assesses the correlation between 42 initial characteristics (including age, height, weight, osteocalcin, smoking status, etc.) and bone mineral density using MIC calculations, specifically as follows: Figure 3 As shown, the horizontal axis represents predictors, and the vertical axis represents the importance estimates of the predictors. When the MIC threshold is set to 0.12 (features below this threshold are considered "weakly associated"), the number of strongly associated features selected ultimately falls within the range of 18-22. The reasonable ratio (approximately 1:2:2) of the three categories of indicators—physiological baseline indicators (e.g., age, height, weight, 3-4 items), laboratory test indicators (e.g., osteocalcin, parathyroid hormone, vitamin D, 8-10 items), and clinical behavioral indicators (e.g., age at menarche, age at menopause, smoking status, 7-8 items)—determines that the number of features should be maintained between 18-22. Too few features will result in the loss of certain key indicators, while too many will introduce redundant information of the same type.

[0077] In this embodiment of the invention, the clinical bone mineral density data undergoes categorized preprocessing, specifically including:

[0078] Missing value handling employs the Multiple Imputation Method (MICE), which generates multiple missing value imputation schemes by constructing a regression model, reducing the uncertainty of a single imputation, and is suitable for datasets with a missing value rate exceeding 20%.

[0079] Feature selection employs a combination of recursive feature elimination (RFE) and feature identification (MIC). First, redundant features are removed using RFE, and then feature correlation is verified using MIC, further improving the accuracy of feature selection. This approach is suitable for scenarios with extremely high feature dimensions (e.g., more than 60 items).

[0080] In this embodiment of the invention, the CNN-LSTM hybrid model adopts a sequential architecture, first extracting local features through CNN, and then processing temporal relationships through LSTM. Specifically, as follows... Figure 4 As shown, the CNN-LSTM hybrid model follows a process of "input layer → multi-layer CNN feature extraction → LSTM sequence modeling → fully connected layer output prediction." It leverages the local feature extraction capabilities of CNNs and the long-sequence dependency modeling capabilities of LSTMs to solve the problem of learning complex patterns from multi-dimensional features and predicting continuous values. The CNN extracts features from local to high-level abstraction through a cycle of "convolution → activation → pooling." These features are then converted into 1-dimensional vectors by a flattening layer and input to the LSTM. The 128 hidden units of the LSTM layer represent the internal memory capacity of the LSTM; tanh activation solves the gradient vanishing problem of RNNs, allowing the model to learn long-term dependencies. The fully connected layer maps the "sequence features" output by the LSTM to the final predicted value.

[0081] In this embodiment of the invention, the CNN sub-network has two convolutional layers, each with 32 and 64 kernels respectively, and uses ReLU activation. The pooling layers are max-pooling layers with a 2×2 kernel size. The flattening layer converts the feature map output by the pooling layers into a one-dimensional vector. It should be noted that the CNN sub-network contains two convolutional layers, two max-pooling layers, and one flattening layer. The first convolutional layer has 32 kernels, a 3×3 kernel size, and a stride of 1, using the ReLU activation function. The second convolutional layer has 64 kernels, a 3×3 kernel size, and a stride of 1, also using the ReLU activation function. Sliding convolution captures local correlations between features (such as the synergistic effect between height and weight, and the interaction between laboratory indicators). Each convolutional layer is followed by a max-pooling layer (pool size = 2×2, stride = 2). Max pooling reduces the feature dimensionality, decreases overfitting, and preserves key local features. The flattening layer converts the output of the last pooling layer into a one-dimensional feature vector, providing input for the LSTM sub-network. It's worth noting that the CNN sub-network contains two convolutional layers, requiring multi-layer gradient backpropagation to optimize parameters. The ReLU function expression is "f(x)=max(0,x)", where the gradient is always 1 when the input x>0, ensuring stable gradient propagation across multiple convolutional layers. This allows the CNN to effectively learn complex local relationships between features (such as the synergistic effect of osteocalcin and parathyroid hormone), avoiding insufficient feature extraction due to gradient vanishing. On the other hand, among the 42 initial features of lumbar spine bone mineral density, only some features (such as age, weight, and osteocalcin) actually contribute to local relationship extraction; redundant features need to be "suppressed." ReLU automatically "shuts down" feature channels that are meaningless for prediction (such as the convolutional output corresponding to lifestyle indicators with very weak correlation to bone mineral density) by directly mapping negative input values ​​to 0, retaining only the key feature signals corresponding to positive inputs, forming a sparse feature representation. This sparsity can both strengthen the weight of key local features and reduce the interference of irrelevant information on subsequent LSTM subnetworks, thereby improving the feature utilization efficiency of the overall model.

[0082] The LSTM subnetwork consists of one LSTM layer and one fully connected layer. Specifically, as follows... Figure 4aAs shown, it should be noted that in the LSTM layer, the number of hidden units is set to 128, and the tanh activation function is used. Information flow is regulated through the input gate, forget gate, and output gate: the forget gate removes outdated feature information (such as short-term fluctuation data unrelated to bone density), the input gate stores new key features, and the output gate controls the transmission of information to the hidden state, effectively capturing the temporal dependencies of features (such as the trend of bone density with age). t and c t−1 W represents the memory cell state vectors at times t and t-1, respectively. f W i W c These are the corresponding calculated weight matrices, b f b i b c For the corresponding bias term, h t−1 Indicates the unit output at the previous time step, x t This represents the current unit input; the fully connected layer has a set output unit count of 1, uses a linear activation function, and outputs the final predicted lumbar spine bone density value.

[0083] It should be further explained that LSTM regulates information flow through the synergistic effect of the input gate, forget gate, and output gate, requiring the activation function to output positive and negative values ​​to distinguish between "information enhancement" and "information inhibition." The output range of tanh is [-1, 1], which can map the numerical values ​​of LSTM cell states to this range—positive values ​​represent key temporal information that needs to be retained (such as the trend of rapid decline in bone mineral density after menopause), while negative values ​​represent redundant information that needs to be weakened (such as short-term fluctuations in laboratory indicator noise), which can accurately support LSTM in filtering and transmitting temporal information. On the other hand, although bone mineral density values ​​have a clinically common range (such as approximately 0.8-1.2 g / cm² in healthy individuals and <0.6 g / cm² in individuals with osteoporosis), individual differences may cause values ​​to exceed the conventional range (such as bone mineral density >1.3 g / cm² in individuals with special physical conditions). Although the output range of tanh [-1,1] has boundaries, it can be mapped to any range of bone density values ​​through linear transformations of subsequent fully connected layers (such as multiplying by a scaling factor and adding an offset). Unlike sigmoid, tanh does not cause the predicted value to be "truncated" due to the upper limit (value of 1) or lower limit (value of 0) of the output (such as being unable to predict bone density higher than 1.2 g / cm²), thus ensuring the integrity of the regression results.

[0084] In this embodiment of the invention, in the step of training the CNN-LSTM hybrid model using the training set, the Adam algorithm is employed, with an initial learning rate of 0.001. The convergence speed is improved through adaptive learning rate adjustment. The maximum number of training epochs is set to 300, and an early stopping strategy (patience=10) is used. Training stops when the validation set error does not decrease for 10 consecutive epochs to avoid overfitting. Excessive bias towards a particular feature class can lead to a decrease in generalization ability.

[0085] In this embodiment of the invention, in the step of dividing the preprocessed effective feature set into a training set and a test set, the training set accounts for 80% and the test set accounts for 20%, and stratified sampling is used to ensure that the feature distribution of the training set and the test set is consistent.

[0086] In this embodiment of the invention, mean squared error (MSE) is used as the loss function, and the parameters of CNN and LSTM are updated simultaneously through backpropagation to minimize the difference between the predicted value and the true value.

[0087] In this embodiment of the invention, the visualization charts include a line graph comparing the predicted and actual lumbar spine bone mineral density values, a residual distribution histogram, and prediction interval graphs at different confidence levels. It should be noted that the CNN-LSTM hybrid model outputs a predicted lumbar spine bone mineral density value (unit: g / cm²), and simultaneously calculates and outputs evaluation indicators such as MSE, RMSE, MAE, and R², facilitating clinicians' assessment of prediction reliability. See Table 1 below for details. ,

[0088] By comparing the performance differences of the three models on the training and test sets, CNN and LSTM exhibit severe overfitting. While their performance on the training set is acceptable, their error on the test set skyrockets, resulting in negative R² and a complete lack of generalization ability. CNN-LSTM, on the other hand, is the best performing model among the three. It demonstrates strong fitting ability on the training set, low error on the test set, and high R², indicating that it effectively combines the spatial feature extraction capabilities of CNN with the time series modeling advantages of LSTM, overcoming the shortcomings of a single model and achieving a balance between strong fitting and strong generalization.

[0089] Furthermore, a comparison curve between predicted and actual values, a histogram of residual distribution, and prediction interval plots for different confidence levels (5%-95%) are generated, as detailed below. Figure 5 and Figure 6 As shown, the model's performance is presented intuitively to assist doctors in interpreting the results. Figure 5In the diagram, the horizontal axis (Sample points) represents the index of the data sample points (100 sample points in total), and the vertical axis (T-value) represents the value of the target indicator to be detected. The blue line with an asterisk represents the true value, the orange solid line represents the prediction result of the CNN model, the yellow solid line represents the prediction result of the LSTM model, and the purple solid line represents the prediction result of the CNN-LSTM model (Predicted-CNN-LSTM). By observing the degree of fit between the different colored lines and the true values, the model performance can be compared intuitively. The prediction trend of the CNN model is generally consistent with the true value, but there are significant deviations from the true value in local details. The prediction volatility of the LSTM model is significantly stronger than other models, with a significant "peak" in the 40-50 sample point range, much higher than the true value, and then quickly falling back, showing a large deviation from the true value. The prediction result of the CNN-LSTM model has the highest degree of fit with the true value, not only with an overall fluctuation trend almost consistent with the true value, but also with a high degree of overlap with the true value in most sample points (such as the sample point ranges of 10-30 and 60-90). It is evident that the CNN-LSTM combined model performs optimally in this time-series detection task, with the prediction results being closest to the true values.

[0090] Figure 6 The image shows a visualization of the evaluation results of the CNN-LSTM method prediction residuals. It analyzes the characteristics of the model's prediction error from two perspectives: the temporal / sampling distribution of the residuals and the statistical distribution characteristics of the residuals, helping to judge the model's fit. The left figure is a sequence distribution plot of residuals as a function of data sample points, where MSE = 0.07152 (mean squared error, measuring the average level of the squared residuals; the smaller the value, the smaller the error), and RMSE = 0.26743 (root mean squared error, the square root of MSE, consistent with the original data units, more intuitively reflecting the error magnitude). The right figure shows the frequency distribution histogram and kernel density estimation curve of the residuals. The histogram shows the distribution density of the residuals in different numerical intervals, and the blue curve represents the kernel density estimation (KDE), smoothly depicting the distribution shape of the residuals. The average residual is 0.0079331, and the standard deviation of residuals is 0.26866. The two graphs verify the predictive residual characteristics of the CNN-LSTM model from two dimensions: "temporal randomness of the residuals" and "reasonableness of the statistical distribution of the residuals." The residual sequence in the left graph shows no obvious trend, indicating that the CNN-LSTM model has no systematic bias in the error of different sampling points; the residual mean in the right graph is close to 0 and the distribution is reasonable, indicating that the CNN-LSTM model has no significant systematic bias and the randomness of the error is as expected.

[0091] The high-risk population marker is divided based on the threshold of lumbar spine bone mineral density value. When the lumbar spine bone mineral density value is lower than the preset percentile of the bone mineral density reference value of the population of the same gender and the same age group, it is marked as a high-risk population. It should be noted that according to the predicted bone mineral density value, combined with the WHO osteoporosis diagnosis standard (T value ≤ -2.5 is osteoporosis, -2.5 < T value < -1 is osteopenia), the high-risk population is automatically marked and intervention suggestions are output, such as supplementing calcium and vitamin D, increasing exercise, etc.

[0092] In the embodiment of the present invention, the CNN-LSTM hybrid model adopts a parallel architecture. CNN and LSTM respectively process the same effective feature set, and then are fused through an attention mechanism, and then input into the fully connected layer to predict bone mineral density. This solution can further strengthen the synergistic effect of local features and temporal features and is applicable to scenarios with higher feature dimensions (such as more than 50 items). It should be noted that through the attention mechanism, dynamic weight allocation is performed on the local features extracted by CNN and the temporal features captured by LSTM to strengthen the contribution of key features to bone mineral density prediction. Scaled Dot-Product Attention is selected as the self-attention weight calculation method to adapt to the feature dimension (d = 64) and the clinical data volume (529 samples). And the parameters of the attention layer are initialized to a normal distribution (mean 0, standard deviation 0.01), which is consistent with the initialization method of the LSTM sub-network parameters to ensure the stability of the gradients of each layer in the initial stage of training.

[0093] Based on the 529 perimenopausal women data of the present invention, the performance of "parallel CNN-LSTM + self-attention" and "serial CNN-LSTM" is compared to verify the effectiveness of the implementation of the attention mechanism. The results are as follows: ,

[0094] The results show that after adding the self-attention mechanism, the model MSE is further reduced (<0.07), R² is increased to more than 93%, and the accuracy of high-risk population identification is increased by 2.7 percentage points. At the same time, the prediction time for a single sample is still controlled within 0.1 second (meeting the clinical requirements), which proves that this implementation method of the attention mechanism can effectively strengthen the synergistic effect of features and adapt to the lumbar spine bone mineral density prediction task of the present invention.

[0095] In the embodiment of the present invention, LSTM can be replaced by a gated recurrent unit (GRU). GRU simplifies the gating structure of LSTM (only including an update gate and a reset gate), and the calculation amount is reduced by 30%. In scenarios with a small amount of data (such as the number of samples < 300), the training speed can be improved on the premise of ensuring the prediction accuracy (MSE < 0.08).

[0096] In this embodiment of the invention, the CNN-LSTM hybrid model can be deployed to edge computing devices (such as embedded terminals and smartphones). Model quantization (such as INT8 quantization) reduces the model size to about one-quarter of its original size, enabling real-time prediction on edge devices. This is suitable for scenarios such as outdoor screening and medical care in remote areas. It should be noted that, based on the test set of this invention (106 perimenopausal women samples, original FP32 model performance: MSE=0.07152, R²=92.4488%, MAE=0.21353), after implementing INT8 quantization using TensorRT, the accuracy changes are shown in the table below. As can be seen from the table, the accuracy loss is controllable, and the key indicators still meet clinical needs. ,

[0097] By comparing the performance of "uncalibrated INT8 quantization" and "INT8 quantization calibrated according to this invention", the effect of the calibration strategy on controlling accuracy loss can be verified:

[0098] Uncalibrated INT8 quantization (default range [-127, 127]): MSE = 0.0892, R² = 89.1%, MAE = 0.247, with accuracy loss significantly exceeding the clinically acceptable range;

[0099] The INT8 quantization after calibration in this invention has the following results: MSE=0.07586, R²=91.67%, MAE=0.22417, and the accuracy loss rate is controlled within 6%, proving that the Min-Max calibration and hierarchical quantization strategy can effectively suppress the loss.

[0100] Furthermore, the core objective of choosing INT8 quantization is to achieve real-time prediction on edge devices (such as embedded terminals and smartphones). INT8 quantization compresses the model size from 12.8MB in FP32 to 3.2MB (compression ratio 4:1), and reduces the single-sample prediction time from 0.032 seconds (FP32, CPU i7) to 0.018 seconds (INT8, embedded ARM CPU). This meets the "low hardware dependence + fast response" requirements of scenarios such as outdoor screening and medical care in remote areas, and the loss of accuracy (such as a decrease in R² of 0.7865 percentage points) does not affect the core clinical functions (such as the accuracy of distinguishing between osteoporosis and osteopenia is still > 88%).

[0101] Specifically, in the actual application test scenario, the training data came from 529 perimenopausal women in the menopause clinic of a tertiary hospital, including 42 characteristics such as age (45-55 years), height (145-175cm), weight (40-80kg), laboratory indicators (such as osteocalcin and parathyroid hormone), and lifestyle habits (smoking, drinking).

[0102] This invention presents a lumbar spine bone mineral density (BSD) prediction method based on a CNN-LSTM hybrid model. It designs differentiated missing value imputation methods (moving median, K-nearest neighbors, random imputation) for different feature types (continuous and categorical), and combines "linear interpolation + mean smoothing + isolated forest" to handle outliers, improving data quality. Features strongly correlated with lumbar spine BSD are selected based on the maximum information coefficient (threshold 0.12), eliminating redundant information, reducing model computation, and improving generalization ability. The CNN and LSTM are concatenated; the CNN extracts local features (such as synergistic relationships between indicators), and the LSTM captures temporal dependencies (such as long-term bone mass change trends), overcoming the functional limitations of a single model. The Adam optimizer (initial learning rate 0.001) and early stopping strategy (patience=10) are used to simultaneously update CNN and LSTM parameters, ensuring rapid model convergence and avoiding overfitting. Furthermore, the method integrates predicted value calculation, visualization, and high-risk population identification, providing clinicians with intuitive and interpretable diagnostic support, adapting to the needs of primary healthcare scenarios.

[0103] This invention also provides an embodiment of a lumbar spine bone mineral density index prediction system based on a CNN-LSTM hybrid model, specifically as follows: Figure 2 As shown, based on the above-mentioned CNN-LSTM hybrid model-based method for predicting lumbar spine bone mineral density index, the system includes:

[0104] The data input layer is used to collect clinical characteristic data of lumbar spine bone mineral density. Specifically, the data input layer includes a clinical data acquisition module and a bone mineral density data acquisition module. The clinical data acquisition module collects 42 clinical characteristics related to bone mineral density (such as age, height, weight, hormone levels, etc.), covering multi-dimensional data such as physiological indicators and demographic information, reflecting an individual's basic health status. The bone mineral density data acquisition module obtains the true value of lumbar spine bone mineral density (the gold standard for clinical bone mineral density testing) through DXA (dual-energy X-ray absorptiometry) as a supervised learning label for model training.

[0105] The data preprocessing layer is used to perform categorical preprocessing on the lumbar spine bone mineral density clinical feature data to obtain a preprocessed effective feature set, which includes missing values, outliers, and eigenvalues. Specifically, the data preprocessing layer includes a missing value handling unit, an outlier handling unit, and a feature selection unit. The missing value handling unit addresses potential missing items in the clinical data (such as a patient's height / weight not being recorded) by using methods such as interpolation, statistical imputation (mean / median), or deletion to ensure data integrity. The outlier handling unit identifies and corrects abnormal data (such as extreme values ​​of age or weight that significantly deviate from the reasonable range) to avoid outliers interfering with model training (e.g., by detecting and correcting outliers using box plot methods). The feature selection unit uses the maximum information coefficient (MIC) algorithm to select the feature subset with the "highest correlation" with lumbar spine bone mineral density from 42 clinical features.

[0106] The model prediction layer is used to divide the preprocessed effective feature set into a training set and a test set, and construct a CNN-LSTM hybrid model. The CNN sub-network includes convolutional layers, pooling layers, and flattening layers, and the LSTM sub-network includes LSTM layers and fully connected layers. The CNN-LSTM hybrid model is trained using the training set; the trained CNN-LSTM hybrid model is evaluated using the test set. Evaluation metrics include mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Specifically, the model prediction layer includes a CNN sub-network, an LSTM sub-network, and a parameter optimization unit. The CNN sub-network utilizes the "sliding computation" property of convolutional kernels to capture local correlation patterns between clinical features, efficiently extracting key local features from high-dimensional data. The LSTM sub-network, through the gating mechanism of Long Short-Term Memory (LSTM) networks, learns the temporal dependence of clinical data (if multiple physical examination records exist, it can capture the long-term trend of bone density changes), solving the gradient vanishing problem of traditional Recurrent Neural Networks (RNNs) and modeling "long-term sequence memory." The parameter optimization unit includes the Adam optimizer, which uses an adaptive learning rate gradient descent algorithm to dynamically adjust the parameter update step size, accelerating model convergence. It also employs an early stopping strategy, terminating training when validation set performance stops improving to avoid overfitting, while preserving the model weights with the "optimal generalization ability," balancing training efficiency and model robustness.

[0107] The results output layer outputs the predicted lumbar spine bone mineral density (BSD) values, including the predicted BSD value, a visualization chart of the prediction results, and markers for high-risk groups corresponding to the BSD values. Specifically, the results output layer includes a prediction result calculation unit, a visualization unit, and a high-risk group identification unit. The prediction result calculation unit integrates the output of the CNN-LSTM to generate the predicted BSD value and calculates the error indices (such as MAE and RMSE) between the predicted and actual values ​​to quantify the model's accuracy. The visualization unit displays comparison curves and residual plots. The comparison curves visually present the trend of change between "predicted bone mineral density" and "actual bone mineral density," verifying the consistency and reliability of the model's predictions. The residual plots show the error distribution between the predicted and actual values, assisting in analyzing the model's prediction accuracy across different bone mineral density ranges. The high-risk group identification unit, based on the predicted bone mineral density values ​​(such as the osteoporosis threshold of T-score < -2.5), screens out high-risk individuals with low bone mineral density, providing decision support for clinical screening and intervention (such as calcium supplementation and exercise guidance).

[0108] It should be noted that this invention can be applied to large-scale population osteoporosis screening (such as community elderly physical examinations), dynamic monitoring of bone mineral density in perimenopausal women (predicted every 6-12 months to track changes in bone mass), and preliminary diagnosis of osteoporosis in primary hospitals (to assist doctors in determining whether further DXA testing is needed).

[0109] In summary, the embodiments of this invention do not rely on DXA equipment; they can predict lumbar spine bone mineral density solely through clinical indicators, posing no radiation risk. The equipment required is only a standard computer (hardware configuration: CPU i7 or higher, GPU GTX 1080 or higher), reducing costs by over 90%. This allows for large-scale application in grassroots settings such as community hospitals and health check centers. The CNN-LSTM hybrid model on the test set exhibits a low MSE of 0.07152, an RMSE of 0.26743, and an R² of 92.4488%, significantly outperforming both standalone CNN (MSE=1.0457, R²=-0.55676%) and standalone LSTM (MSE=1.3717, R²=-25.1798%), accurately capturing the patterns of bone mineral density changes.

[0110] Furthermore, the CNN-LSTM hybrid model significantly improves the sensitivity to early bone loss (MAE=0.21353), identifying individuals with osteopenia 2-3 years earlier, providing a window for early intervention (such as dietary adjustments and exercise guidance) and reducing fracture risk. Targeted preprocessing can handle different types of missing and outlier values, and MIC feature selection is adaptable to clinical data from different populations (such as premenopausal and postmenopausal women), demonstrating high model robustness. The output includes visual charts and high-risk group markers, which physicians can interpret without requiring specialized deep learning knowledge, and the prediction time is short (<0.1 seconds per sample), meeting the needs of rapid clinical diagnosis.

[0111] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model, characterized in that, Includes the following steps: Collect clinical characteristic data of lumbar spine bone mineral density; The clinical feature data of lumbar spine bone mineral density are preprocessed by classification to obtain a preprocessed effective feature set, which includes missing values, outliers and feature values. The preprocessed effective feature set is divided into a training set and a test set; Construct a CNN-LSTM hybrid model, wherein the CNN sub-network includes convolutional layers, pooling layers, and flattening layers, and the LSTM sub-network includes LSTM layers and fully connected layers, and train the CNN-LSTM hybrid model using the training set; The trained CNN-LSTM hybrid model was evaluated using the test set, and the evaluation metrics included mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ; Output the predicted lumbar spine bone mineral density value, which includes the predicted lumbar spine bone mineral density value, a visualization chart of the predicted result, and a marker of the high-risk group corresponding to the lumbar spine bone mineral density value.

2. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 1, characterized in that, The clinical bone mineral density data were subjected to categorical preprocessing, which included missing value handling, outlier handling, and feature selection, wherein: Missing values ​​are handled using a combination of moving median, K-nearest neighbor algorithm, and random imputation. Outlier handling employs a combination of linear interpolation, mean smoothing, and the isolated forest algorithm; Feature selection uses the MIC threshold, with a threshold value of 0.

12.

3. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 2, characterized in that, When handling missing values, the moving window size is set to 5, and the K value of the K nearest neighbor algorithm is set to 3; When handling outliers, outliers are identified by box plots, and straight lines are fitted based on the five normal data points before and after the outlier. The window size for mean smoothing is 3.

4. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 1, characterized in that, The clinical bone mineral density data are preprocessed by classification, specifically including: Missing value handling employs the Multiple Imputation Method (MICE), which generates multiple missing value imputation schemes by constructing a regression model. Feature selection employs a combination of recursive feature elimination (RFE) and feature filtering (MIC). First, redundant features are removed using RFE, and then feature correlation is verified using MIC.

5. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 1, characterized in that, The CNN-LSTM hybrid model adopts a serial architecture, first extracting local features through CNN, and then processing temporal relationships through LSTM.

6. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In the CNN subnetwork, there are 2 convolutional layers, with 32 and 64 convolutional kernels in each layer, and the activation function is ReLU; the pooling layer is a max pooling layer with a 2×2 kernel size; the flattening layer converts the feature map output by the pooling layer into a one-dimensional vector. In the LSTM subnetwork, the LSTM layer has 128 units, the activation function is tanh, and the fully connected layer has 1 output unit.

7. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 6, characterized in that, In the step of training the CNN-LSTM hybrid model using the training set, the Adam algorithm is used, the initial learning rate is set to 0.001, the maximum number of training epochs is set to 300, and an early stopping strategy is adopted, which stops training when the validation set error does not decrease for 10 consecutive epochs.

8. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 1, characterized in that, The visualization charts include a line graph comparing the predicted and actual values ​​of lumbar spine bone mineral density, a residual distribution histogram, and a prediction interval chart for different confidence levels. The high-risk population marker is based on a threshold division of lumbar spine bone mineral density (BSMD) values. When the BSMD value is lower than the preset percentile of the reference BSMD value for people of the same sex and age, the person is marked as a high-risk population.

9. The method for predicting lumbar spine bone mineral density index based on a CNN-LSTM hybrid model according to claim 1, characterized in that, The CNN-LSTM hybrid model adopts a parallel architecture, where CNN and LSTM process the same effective feature set respectively, and then fuse them through an attention mechanism before being input into the fully connected layer to predict bone density.

10. A lumbar spine bone mineral density index prediction system based on a CNN-LSTM hybrid model, comprising the lumbar spine bone mineral density index prediction method based on a CNN-LSTM hybrid model as described in any one of claims 1 to 9, characterized in that, The system includes: The data input layer is used to collect clinical characteristic data of lumbar spine bone mineral density. The data preprocessing layer is used to perform categorical preprocessing on the clinical feature data of lumbar spine bone mineral density to obtain a preprocessed effective feature set, which includes missing values, outliers, and feature values. The model prediction layer is used to divide the preprocessed effective feature set into a training set and a test set to construct a CNN-LSTM hybrid model. The CNN sub-network includes convolutional layers, pooling layers, and flattening layers, and the LSTM sub-network includes LSTM layers and fully connected layers. The CNN-LSTM hybrid model is trained using the training set. The trained CNN-LSTM hybrid model is evaluated using the test set, and the evaluation metrics include mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The results output layer is used to output the predicted results of lumbar spine bone mineral density values. The predicted results include the predicted lumbar spine bone mineral density value, a visualization chart of the predicted results, and a marker of the high-risk groups corresponding to the lumbar spine bone mineral density values.

Images