PDAC stage prediction method and device combining deep learning and machine learning

Abstract

The invention discloses a PDAC staging prediction method and device combining deep learning and machine learning. The method comprises the following steps: acquiring a serum metabolic spectrum data set; constructing different combination models randomly composed of a deep neural network, a convolutional neural network and a long short-term memory network; different hybrid models based on different combination models are constructed, the hybrid models adopt the combination models to realize feature extraction of the serum metabolism spectrum data, and classification is performed according to the extracted features by adopting a machine learning algorithm; training different hybrid models based on the data set, obtaining the trained different hybrid models, comparing the performance, and selecting the hybrid model with the best performance as a prediction model; and utilizing the prediction model to realize stage prediction of the PDAC. According to the method, different types of neural networks are constructed to perform staging prediction of different clinical stages of PDAC, and the optimal hybrid model is determined by comparing the classification performance of different clinical staging cases of PDAC by different hybrid models.

Description

Technical Field

[0001] The present invention relates to the technical field of early diagnosis of pancreatic cancer, and in particular to a PDAC staging prediction method combining deep learning and machine learning. Background Art

[0002] Pancreatic ductal adenocarcinoma (PDAC) is the most common pancreatic cancer, accounting for about 90% of pancreatic cancer, and is mostly discovered in the invasive and incurable stage. Due to the poor prognosis, strong invasiveness, treatment resistance and late diagnosis of pancreatic cancer, its incidence is almost equal to the mortality rate of pancreatic cancer patients. Therefore, achieving early diagnosis of pancreatic cancer is one of the important hot topics in the diagnosis and treatment of pancreatic cancer. At present, the screening methods for pancreatic cancer are mainly concentrated in medical imaging methods such as endoscopic ultrasound (EUS) and magnetic resonance imaging (MRI). In recent years, the identification of pancreatic cancer biomarkers and the analysis of changes in endogenous compounds in pancreatic cancer patients have also provided new ideas for the early diagnosis and treatment of pancreatic cancer.

[0003] Biomarkers of pancreatic cancer tumors and changes in endogenous compounds in pancreatic cancer patients can be obtained through metabolomics studies of body fluids such as blood and urine. Existing studies have analyzed the metabolic profiles of serum and urine of pancreatic cancer patients using MS and NMR techniques and discovered some potential biomarkers. However, current related studies mainly focus on the classification between pancreatic cancer patients and healthy people, and the study of specific stages of pancreatic cancer still faces challenges. Existing metabolomics methods are difficult to obtain significant inter-group differences in the metabolic profiles of PDAC at different stages, which limits the progress of promoting pancreatic cancer disease diagnosis through metabolomics. Metabolomics data has many characteristic variables and is characterized by high dimensionality and sparsity. Machine learning can process multivariate data, extract knowledge from data through training, and select variable features to promote decision-making. Although many studies have used machine learning algorithms to learn from metabolic profile data and achieve prediction of corresponding diseases, machine learning does not perform well for data that is difficult to distinguish. Deep learning, as an extension of machine learning technology, has strong learning ability and automated training performance, can find nonlinear relationships between variables, and achieve more accurate classification predictions. It has been widely used in various prediction problems. Summary of the invention

[0004] The purpose of the present invention is to solve the problems in the prior art.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide a PDAC staging prediction method combining deep learning and machine learning, comprising the following steps:

[0006] Collect serum samples and obtain serum metabolic profile data to form a data set;

[0007] Construct different combination models composed of deep neural networks, convolutional neural networks and long short-term memory networks for feature extraction of serum metabolic profile data;

[0008] Constructing different hybrid models based on different combination models, wherein the hybrid models include a feature extraction part and a classification part, wherein the feature extraction part uses the combination model to realize feature extraction of serum metabolic profile data, and the classification part uses a machine learning algorithm to classify according to the extracted features;

[0009] Train different hybrid models based on the data set, obtain the trained different hybrid models, compare the performance of different hybrid models, and select the hybrid model with the best performance as the prediction model;

[0010] Use the prediction model to predict the staging of PDAC.

[0011] Preferably, the data set is expanded using the ADASYN algorithm.

[0012] Preferably, the different combination models include:

[0013] A CNN-LSTM model consisting of a CNN layer and an LSTM layer was constructed by combining a convolutional neural network and a long short-term memory network. The CNN layer was used to extract local features of the serum metabolic profile, and the LSTM layer was used to process sequence data.

[0014] Combine deep neural networks and convolutional neural networks to build a DNN-CNN model consisting of DNN layers and CNN layers. Use the DNN layer to extract global features, and then use the CNN layer to extract local features.

[0015] A CNN-BiLSTM model consisting of a CNN layer and a bidirectional LSTM layer is constructed by combining a convolutional neural network and a bidirectional long short-term memory network. The CNN layer and the bidirectional LSTM layer are combined to process local features and context information.

[0016] A DNN-LSTM model consisting of a DNN layer and a LSTM layer is constructed by combining a deep neural network and a long short-term memory network. The DNN layer and the LSTM layer are combined to process high-dimensional features and time-dependent information.

[0017] Preferably, the machine learning classification algorithm includes support vector machine, random forest, gradient boosting decision tree and extreme gradient boosting.

[0018] Preferably, the training of different hybrid models based on the data set is specifically: optimizing the hyperparameters of the CNN-LSTM model, DNN-CNN model, CNN-BiLSTM model, DNN-LSTM model and machine learning algorithm through a grid search algorithm, including learning rate, number of training iterations, batch size, convolution kernel size, pooling kernel size, regularization coefficient and number of neurons in the LSTM layer, to obtain trained different hybrid models.

[0019] Preferably, comparing the performance of different hybrid models comprises:

[0020] Compare the ROC curve predictions of the hybrid model based on the CNN-LSTM model, the hybrid model based on the DNN-CNN model, the hybrid model based on the CNN-BiLSTM model, and the hybrid model based on the DNN-LSTM model to obtain the true positive rate and false positive rate of each hybrid model;

[0021] The confusion matrices of the hybrid model based on the CNN-LSTM model, the hybrid model based on the DNN-CNN model, the hybrid model based on the CNN-BiLSTM model, and the hybrid model based on the DNN-LSTM model were compared to obtain the classification accuracy of each hybrid model.

[0022] Preferably, the prediction model adopts a CNN-LSTM model, uses the CNN layer to extract local features of the serum metabolic spectrum, uses the LSTM layer to process the sequence data, and finally combines the fully connected layer and the ReLU activation function to perform feature expansion; the output of the model is then used as the input of the machine learning classification algorithm, and the machine learning classification algorithm is used to predict the PDAC stage.

[0023] Preferably, the CNN layer of the CNN-LSTM model includes multiple convolutional layers and pooling layers, wherein the convolutional layer includes a Tanh activation function; and the CNN-LSTM model includes multiple layers of LSTM layers.

[0024] The present invention also provides a PDAC staging prediction device combining deep learning and machine learning, comprising:

[0025] A data collection module collects serum samples and obtains serum metabolic profile data to form a data set;

[0026] A combined model building module, which builds different combined models composed of deep neural networks, convolutional neural networks and long short-term memory networks for feature extraction of serum metabolic profile data;

[0027] A hybrid model construction module is used to construct different hybrid models based on different combination models. The hybrid models include a feature extraction part and a classification part. The feature extraction part uses the combination model to extract features of serum metabolic profile data, and the classification part uses a machine learning algorithm to classify according to the extracted features.

[0028] The prediction model selection module trains different hybrid models based on the data set, obtains the trained different hybrid models, compares the performance of different hybrid models, and selects the hybrid model with the best performance as the prediction model;

[0029] The prediction model application module uses the prediction model to predict the staging of PDAC.

[0030] The present invention has the following beneficial effects:

[0031] (1) The examination data of the present invention uses serum metabolic profile data, which has a simple and low-invasive process and is easy for patients to accept. It also provides rich biochemical information, which helps to more accurately judge the disease status. In addition, the high-dimensional characteristics and complexity of serum metabolic profile data are very suitable for use in the CNN-LSTM-ML hybrid model, thereby giving full play to the advantages of deep learning and machine learning algorithms.

[0032] (2) The present invention innovatively combines convolutional neural networks (CNN) and long short-term memory networks (LSTM), which are used to extract local features of serum metabolic profile data and capture time series information, respectively, so as to comprehensively mine the complex features of the data; the model can process multi-dimensional serum metabolic profile data, and because the CNN-LSTM module can efficiently extract and represent complex features, these features can be seamlessly compatible with a variety of machine learning algorithm designs, allowing it to be combined with a variety of machine learning classification algorithms. Through this design, the generalization ability and prediction accuracy of the model are significantly improved, effectively enhancing the classification and prediction effects of complex metabolic profile data.

[0033] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A method step diagram of an embodiment of the present invention;

[0035] Figure 2 is a flow chart of a method according to an embodiment of the present invention;

[0036] Figure 3 It is a framework diagram of the CNN-LSTM-ML hybrid model of an embodiment of the present invention;

[0037] Figure 4 Schematic diagram of some parameter adjustment processes of PDAC dataset models for different networks;

[0038] Figure 5 ROC curves of different neural network models on the PDAC test set;

[0039] Figure 6 Schematic diagram of the test set confusion matrix of different PDAC prediction models;

[0040] Figure 7 The confusion matrix of the hybrid model of the embodiment of the present invention under different clinical stage test sets of PDAC;

[0041] Figure 8 Partial parameter adjustment process of the hybrid model in the malnutrition dataset according to an embodiment of the present invention;

[0042] Fig. 9 The confusion matrix of the hybrid model of the embodiment of the present invention under the test set of different clinical stages of malnutrition;

[0043] Fig.10 FIG. 4 is a diagram showing the structure of a device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] See also Figure 1 and Figure 2 The method step diagram and method flow chart of the embodiment of the present invention are shown, which include the following steps:

[0045] S101, collecting serum samples and obtaining serum metabolic profile data to form a data set;

[0046] S102, constructing different combination models composed of any deep neural network, convolutional neural network and long short-term memory network for feature extraction of serum metabolic profile data;

[0047] S103, constructing different hybrid models based on different combination models, wherein the hybrid models include a feature extraction part and a classification part, wherein the feature extraction part uses the combination model to extract features of the serum metabolic profile data, and the classification part uses a machine learning algorithm to classify according to the extracted features;

[0048] S104, training different hybrid models based on the data set, obtaining trained different hybrid models and comparing the performance of the different hybrid models, and selecting the hybrid model with the best performance as the prediction model;

[0049] S105, using the prediction model to predict the staging of PDAC.

[0050] Specifically, the machine learning algorithms of the embodiments of the present invention include support vector machines (SVM), random forests, gradient boosted decision trees (GBDT) and extreme gradient boosting (XGBoost), etc. These algorithms have their own advantages in processing data features, such as the excellent performance of SVM in high-dimensional space, which is particularly suitable for serum metabolic profile data with many feature variables and high-dimensional sparseness; since serum metabolic profile data in PDAC staging prediction has complex nonlinear relationships, random forests are selected to process such nonlinear relationships, which has a significant effect; GBDT can significantly improve the performance of the model in PDAC staging prediction through ensemble learning, especially when processing feature interactions and complex relationships; in addition, XGBoost performs outstandingly in processing large-scale serum metabolic profile data, especially in processing missing values ​​and achieving efficient parallel computing.

[0051] Specifically, the detailed process of the method in the embodiment of the present invention is described as follows.

[0052] (1) Data collection and preprocessing

[0053] The present invention uses two main data sets. The first data set is derived from serum samples of 106 patients with PDAC (pancreatic ductal adenocarcinoma) in stages I-IV, including 30 healthy cases and 76 PDAC disease groups (13 cases in stage I, 17 cases in stage II, 30 cases in stage III, and 16 cases in stage IV). These samples are processed by 1H-NMR (nuclear magnetic resonance spectroscopy) technology to eventually form a data set containing 3615 feature variables. The second data set is used to verify the performance of the model, and contains serum samples of 428 patients with varying degrees of malnutrition, with a total of 3445 feature variables. In order to improve the generalization ability of the model, indicators such as height, age, and weight are also added to the data set, and a data set of 3463 feature variables is constructed. At the same time, due to the data imbalance problem in the PDAC data set, the present invention uses the ADASYN (adaptive synthetic sampling) algorithm for data expansion. Through the adaptive oversampling technology, the data volume of PDAC cases in each stage is expanded to enhance the distinguishing ability of the model. The ADASYN algorithm generates new synthetic samples based on the distribution of minority class samples.

[0054] (2) Model construction

[0055] The present invention adopts a variety of deep learning models and traditional machine learning models for disease classification, including deep neural networks (DNN), convolutional neural networks (CNN), long short-term memory networks (LSTM) and their combined models. A hybrid prediction model is constructed by combining neural network feature extraction and machine learning classification algorithms. The model structure is implemented using Python3.6 and Keras libraries, and the deep learning open source library TensorFlow is used as a tool for calculation. In order to better extract the metabolic spectrum characteristics of PDAC, the present invention constructs a CNN-LSTM hybrid prediction model. The model uses the CNN layer to extract features, processes sequence data through the LSTM layer, and finally combines the Dense layer and ReLU activation function for feature expansion. The output of the model is then used as the input of the machine learning classification algorithm to predict the PDAC stage. Among them, the CNN part includes multiple convolutional layers and pooling layers, and the LSTM part includes multiple layers of LSTM units. Its specific structure is as follows Figure 3 shown.

[0056] (3) Hyperparameter Optimization

[0057] Hyperparameter optimization is achieved through a grid search algorithm. The optimized hyperparameters include learning rate, number of training iterations, batch size, convolution kernel size, pooling kernel size, regularization coefficient, and number of neurons in the LSTM layer. By adjusting these parameters, the model achieves the best performance during training. For example, the optimization range of learning rate is between 0.0001 and 0.01, the number of training iterations is between 50 and 200, the batch size is between 16 and 128, the convolution kernel size is between 3 and 7, the pooling kernel size is between 2 and 4, the regularization coefficient is between 0.0001 and 0.01, and the number of neurons in the LSTM layer is between 50 and 200. The final results of hyperparameter optimization are as follows: Figure 4 shown.

[0058] (4) Model training and evaluation

[0059] During model training, the cross entropy loss function and Adam optimizer are used to update parameters. To prevent overfitting, early stopping strategy and Dropout technology are used. Model evaluation uses indicators such as accuracy, precision, recall, and F1-score. The training data is divided into training set and validation set in a ratio of 8:2 for model training and performance evaluation.

[0060] (5) Model performance comparison and result analysis

[0061] The present invention compares the performance of the newly constructed CNN-LSTM-ML hybrid model with SVM (support vector machine), RF (random forest), LightGBM (lightweight gradient boosting machine), DNN, CNN, LSTM and CNN-LSTM models.

[0062] For details, see Figure 5 As shown, (a) to (d) are the ROC curves of different neural network models on the PDAC test set; (a) graphically depicts the ROC curve performance of the convolutional neural network (CNN) on the PDAC test set; (b) shows the ROC curve performance of the long short-term memory network (LSTM) on the PDAC test set, (c) provides the ROC curve performance of the hybrid model combining CNN and LSTM (CNN-LSTM) on the PDAC test set; (d) presents the ROC curve performance of the hybrid model combining convolutional neural network, long short-term memory network and machine learning method (CNN-LSTM-ML) on the PDAC test set. By comparison Figure 5 The ROC curves show that the CNN-LSTM-ML hybrid model exhibits a higher true positive rate and a lower false positive rate at different thresholds, indicating that the classification performance of this model in PDAC detection is better than that of a single CNN or LSTM model.

[0063] For details, see Figure 6 As shown, (a) to (b) are schematic diagrams of the test set confusion matrices of different PDAC prediction models; (a) shows the confusion matrix of the support vector machine (SVM) classification model on the PDAC test set, and the graph shows the comparison between the prediction results of the SVM model and the actual results; (b) provides the confusion matrix of the hybrid model (CNN-LSTM-ML) combining convolutional neural network, long short-term memory network and machine learning method on the PDAC test set, and the graph shows the comparison between the prediction results of the hybrid model and the actual results. Figure 6 From the confusion matrix in Figure 2, it can be concluded that the classification accuracy of the CNN-LSTM-ML hybrid model in PDAC detection is significantly higher than that of the SVM model, and the number of misclassifications is significantly reduced, indicating the superiority of the hybrid model in PDAC prediction.

[0064] For details, see Figure 7The figure shows the confusion matrix of the hybrid model under the test set of different clinical stages of PDAC. According to the figure (a), PDAC-Ⅰ cases can be accurately predicted, and only one PDAC-Ⅳ case is misclassified as PDAC-Ⅲ; while there is significant confusion between PDAC-Ⅱ and PDAC-Ⅲ cases, and one case is classified as PDAC-Ⅳ case each, indicating that the metabolic profiles of these two stages are slightly different and difficult to distinguish; (b) shows in detail the test set confusion matrix of the hybrid model when PDAC-Ⅱ and PDAC-Ⅲ are classified into two categories after excluding the serum metabolic profile data of PDAC-Ⅰ and PDAC-Ⅳ. The results show that although the hybrid model performs well in the four-classification task, it still has great difficulties in distinguishing PDAC-Ⅱ and PDAC-Ⅲ cases, and the degree of confusion in the prediction is relatively serious. This reveals the similarity between PDAC-Ⅱ and PDAC-Ⅲ at the metabolic level, indicating that other methods or additional indicators need to be considered to distinguish PDAC patients in these two clinical stages. This hybrid model performed well in predicting patients with PDAC-I and PDAC-IV and can be used as one of the reference indicators for diagnosing these two clinical stages.

[0065] According to the above analysis, the hybrid model optimized by the embodiment of the present invention has an accuracy of 90.00% in PDAC staging prediction, which is significantly better than the traditional machine learning classification model and the single neural network model. Specifically, the accuracy of SVM is 82.50%, RF is 85.00%, LightGBM is 87.50%, DNN is 88.00%, CNN is 86.50%, LSTM is 89.00%, and CNN-LSTM is 89.50%. The hybrid model performs well in PDAC staging prediction, especially in distinguishing the healthy group, PDAC-Ⅰ stage and PDAC-Ⅳ stage. The prediction accuracy of the healthy group reached 95.00%, 90.00% for PDAC-Ⅰ stage, and 92.50% for PDAC-Ⅳ stage. Although there are still some difficulties in distinguishing between PDAC-Ⅱ stage and PDAC-Ⅲ stage, the overall prediction performance is significantly improved. In order to verify the generalization performance of the hybrid model in the diagnosis of other diseases, the present invention applies it to malnutrition serum metabolic profile data. The results showed that the hybrid model achieved a classification accuracy of 84.66% in different clinical stages of malnutrition, further demonstrating the effectiveness and scalability of the model. Figure 8The figure shows part of the parameter adjustment process of the malnutrition dataset based on the hybrid model, where (a) is the learning rate parameter adjustment; (b) is the network training iteration number parameter adjustment; (c) is the batch size parameter adjustment; (d) is the overfitting coefficient parameter adjustment of the overfitting layer in the network structure; (e) is the number of neurons parameter adjustment. This dataset contains four labels: healthy group, mild malnutrition, moderate malnutrition and severe malnutrition. By training and adjusting the hybrid model, the hybrid model optimized for malnutrition data is obtained. See Fig. 9 The figure shows the confusion matrix of the mixed model under the test set of different clinical stages of malnutrition, which reveals that the healthy group and severe malnutrition cases can be clearly distinguished from other categories, while there are different degrees of confusion between mild and moderate malnutrition cases. This may be due to the fact that the difference between mild and moderate malnutrition cases is not obvious at the metabolic level, and it is not easy to find characteristic variables for distinction. In addition, there are also cases of mutual misclassification between the healthy group and mild malnutrition cases, indicating that there is a certain overlap in the serum metabolic spectrum data of healthy people and mild malnutrition patients.

[0066] In summary, the present invention achieves high-accuracy classification prediction of diseases such as PDAC and malnutrition by constructing and optimizing the CNN-LSTM-ML hybrid model. The results show that the hybrid model has high application potential and promotion value in early diagnosis of diseases. The performance of the model has been significantly improved through hyperparameter optimization and data expansion. Future work will further expand the scale of the data set and explore more deep learning and machine learning models to improve the accuracy and robustness of disease prediction.

[0067] See also Fig.10 FIG. 1 is a diagram showing a structure of a device according to an embodiment of the present invention, comprising:

[0068] Data collection module 1001, collects serum samples and obtains serum metabolic profile data to form a data set;

[0069] A combination model building module 1002 is used to build different combination models composed of any combination of deep neural network, convolutional neural network and long short-term memory network for feature extraction of serum metabolic profile data;

[0070] A hybrid model construction module 1003 is used to construct different hybrid models based on different combination models, wherein the hybrid models include a feature extraction part and a classification part. The feature extraction part uses the combination model to extract features of the serum metabolic profile data, and the classification part uses a machine learning algorithm to classify according to the extracted features.

[0071] A prediction model selection module 1004 is used to train different hybrid models based on the data set, obtain the trained different hybrid models, compare the performance of the different hybrid models, and select the hybrid model with the best performance as the prediction model;

[0072] The prediction model application module 1005 uses the prediction model to predict the staging of PDAC.

[0073] It can be seen that the present invention uses a deep learning method to construct different types of neural networks to learn and train the metabolic profile data of PDAC serum samples to perform staging predictions for different clinical stages of PDAC. The performance of the constructed hybrid model, neural network model, and machine learning model were compared to obtain the optimal classification method for PDAC cases at different clinical stages. This invention can provide a new method for clinical disease staging prediction based on serum metabolic profiles and provide a reference for early diagnosis of clinical diseases.

[0074] The above content describes the preferred embodiments of the present invention in detail, but the present invention is not limited to the specific details in the above embodiments. Without violating the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, and these modifications all belong to the protection scope of the present invention.

Claims

1. A PDAC staging prediction method combining deep learning and machine learning, characterized in that: The following steps are involved: Collect serum samples and obtain serum metabolic profile data to form a data set; Construct different combination models composed of deep neural networks, convolutional neural networks and long short-term memory networks for feature extraction of serum metabolic profile data; Constructing different hybrid models based on different combination models, wherein the hybrid models include a feature extraction part and a classification part, wherein the feature extraction part uses the combination model to realize feature extraction of serum metabolic profile data, and the classification part uses a machine learning algorithm to classify according to the extracted features; Train different hybrid models based on the data set, obtain the trained different hybrid models, compare the performance of different hybrid models, and select the hybrid model with the best performance as the prediction model; Use the prediction model to predict the staging of PDAC.

2. The PDAC staging prediction method combining deep learning and machine learning according to claim 1, characterized in that: The ADASYN algorithm is used to expand the dataset.

3. The PDAC staging prediction method combining deep learning and machine learning according to claim 1, characterized in that: The different combination models include: A CNN-LSTM model consisting of a CNN layer and an LSTM layer was constructed by combining a convolutional neural network and a long short-term memory network. The CNN layer was used to extract local features of the serum metabolic profile, and the LSTM layer was used to process sequence data. Combine deep neural networks and convolutional neural networks to build a DNN-CNN model consisting of DNN layers and CNN layers. Use the DNN layer to extract global features, and then use the CNN layer to extract local features. A CNN-BiLSTM model consisting of a CNN layer and a bidirectional LSTM layer is constructed by combining a convolutional neural network and a bidirectional long short-term memory network. The CNN layer and the bidirectional LSTM layer are combined to process local features and context information. A DNN-LSTM model consisting of a DNN layer and a LSTM layer is constructed by combining a deep neural network and a long short-term memory network. The DNN layer and the LSTM layer are combined to process high-dimensional features and time-dependent information.

4. The PDAC staging prediction method combining deep learning and machine learning according to claim 3, characterized in that: The machine learning classification algorithms include support vector machines, random forests, gradient boosting decision trees, and extreme gradient boosting.

5. The PDAC staging prediction method combining deep learning and machine learning according to claim 4, characterized in that: The training of different hybrid models based on the data set is specifically as follows: optimizing the hyperparameters of the CNN-LSTM model, the DNN-CNN model, the CNN-BiLSTM model, the DNN-LSTM model and the machine learning algorithm through a grid search algorithm, including the learning rate, the number of training iterations, the batch size, the convolution kernel size, the pooling kernel size, the regularization coefficient and the number of neurons in the LSTM layer, to obtain the trained different hybrid models.

6. The PDAC staging prediction method combining deep learning and machine learning according to claim 5, characterized in that: The performance of different hybrid models is compared, including: Compare the ROC curve predictions of the hybrid model based on the CNN-LSTM model, the hybrid model based on the DNN-CNN model, the hybrid model based on the CNN-BiLSTM model, and the hybrid model based on the DNN-LSTM model to obtain the true positive rate and false positive rate of each hybrid model; The confusion matrices of the hybrid model based on the CNN-LSTM model, the hybrid model based on the DNN-CNN model, the hybrid model based on the CNN-BiLSTM model, and the hybrid model based on the DNN-LSTM model were compared to obtain the classification accuracy of each hybrid model.

7. The PDAC staging prediction method combining deep learning and machine learning according to claim 6, characterized in that: The prediction model adopts a CNN-LSTM model, uses a CNN layer to extract local features of the serum metabolic spectrum, uses an LSTM layer to process sequence data, and finally combines a fully connected layer and a ReLU activation function to perform feature expansion; the output of the model is then used as an input to a machine learning classification algorithm, which performs PDAC staging prediction.

8. The PDAC staging prediction method combining deep learning and machine learning according to claim 7, characterized in that: The CNN layer of the CNN-LSTM model includes multiple convolutional layers and pooling layers, wherein the convolutional layer includes a Tanh activation function; the CNN-LSTM model includes multiple layers of LSTM layers.

9. A PDAC staging prediction device combining deep learning and machine learning, characterized in that: include: A data collection module collects serum samples and obtains serum metabolic profile data to form a data set; A combined model building module, which builds different combined models composed of deep neural networks, convolutional neural networks and long short-term memory networks for feature extraction of serum metabolic profile data; A hybrid model construction module is used to construct different hybrid models based on different combination models. The hybrid models include a feature extraction part and a classification part. The feature extraction part uses the combination model to extract features of serum metabolic profile data, and the classification part uses a machine learning algorithm to classify according to the extracted features. The prediction model selection module trains different hybrid models based on the data set, obtains the trained different hybrid models, compares the performance of different hybrid models, and selects the hybrid model with the best performance as the prediction model; The prediction model application module uses the prediction model to predict the staging of PDAC.

Images