Multi-modal feature importance guiding network based on deep learning
Through the multimodal feature importance guide network of deep learning, the problem of missing rate and interpretability in AML survival prediction is solved, and high-accurate survival prediction and biomarker discovery is achieved.
Patent Information
- Application Number
- CN202510484583.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-17
- Publication Date
- 2025-08-01
AI Technical Summary
The prior art has high deletion rate problems and insufficient interpretability of deep learning models in the survival prediction of acute myeloid leukemia (AML), resulting in degradation in model performance and difficulty in biomarker discovery.
A multimodal feature importance guide network based on deep learning, including feature reconstruction module and feature selection module, repair missing values through mask automatic encoder, and generate variable importance coefficients using self-attention mechanism to perform weighted and classifier prediction.
Improves the robustness and interpretability of the model under missing data, significantly improves the accuracy of survival prediction, and successfully identify key biomarkers.
Smart Images
Figure CN120413075A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning, and more specifically, it relates to a multi-modal feature importance-guided network based on deep learning. Background Art
[0002] Acute myeloid leukemia (AML) is one of the most common hematological malignancies and the most common type of acute leukemia in adults. It is a lethal hematological malignancy caused by the clonal expansion of myeloid progenitor cells in the bone marrow, peripheral blood, or extramedullary tissues, and its incidence and mortality rates are both on the rise globally.
[0003] In the clinical research of acute myeloid leukemia (AML), survival prediction and biomarker discovery based on multi-modal data are the core requirements of precision medicine. However, existing methods face significant technical bottlenecks. First, clinical data generally have a high missing rate problem, such as the missing of laboratory indicators or molecular biology data due to detection condition limitations or patient compliance differences. Traditional processing methods such as data deletion or simple imputation (mean, median filling) can partially alleviate the problem, but they will introduce noise or lose potential information, resulting in a significant decline in model performance.
[0004] In addition, due to the lack of interpretability, deep learning models are usually regarded as black boxes. In the medical field, quantitative and objective interpretability is crucial for disease modeling. Doctors not only need accurate predictions but also need to clearly understand which variables contribute significantly to the decision-making, because these variables (i.e., biomarkers) are crucial for personalized diagnosis and treatment. Due to this inherent defect of DL, most clinical multi-variable modeling studies tend to use traditional interpretable machine learning methods such as random forest (RF). These methods can facilitate the identification of biomarkers by learning coefficients reflecting the importance of each variable. However, these methods often rely on "linear" or strong assumptions to achieve interpretability, resulting in low representational ability and thus poor performance. Summary of the Invention
[0005] The object of the present invention is to address the deficiencies of the prior art and propose a multi-modal feature importance-guided network based on deep learning.
[0006] In the first aspect, a multi-modal feature importance-guided network based on deep learning is provided, including: a feature reconstruction module, a feature selection module, and a classifier;
[0007] The feature reconstruction module simulates and repairs the missing values in the input data through a masked autoencoder; the input data is the multi-modal clinical data of AML patients;
[0008] The feature selection module dynamically generates importance coefficients of the reconstruction variables based on the self-attention mechanism, and weights the reconstruction variables according to the importance coefficients to obtain weighted variables;
[0009] The classifier outputs a survival prediction result of AML patients according to the weighted variables.
[0010] Preferably, the multi-modal clinical data includes: demographic characteristics, laboratory test data, morphological examination data, immunophenotyping data, cytogenetic examination data, molecular biology examination data, and clinical intervention plans.
[0011] Preferably, the feature reconstruction module randomly masks some accessible variables of each sample by using a weighted mask strategy during the training process, and restores the masked values through the mean square error loss.
[0012] Preferably, the last layer of the feature selection module uses the Tanh activation function, and generates importance coefficients according to the Tanh activation function.
[0013] Preferably, exponential moving average processing is performed on the importance coefficients to obtain the importance ranking of biomarkers.
[0014] In a second aspect, a data processing method executed by the multi-modal feature importance-guided network described in the first aspect is provided, including:
[0015] S1. Obtain input data, and simulate and repair missing values in the input data through the feature reconstruction module; the input data is multi-modal clinical data of AML patients;
[0016] S2. Generate importance coefficients of the reconstruction variables through the feature selection module, and weight the reconstruction variables according to the importance coefficients to obtain weighted variables;
[0017] S3. Input the weighted features into the classifier, and output a survival prediction result of AML patients.
[0018] In a third aspect, a computer storage medium is provided, and a computer program is stored in the computer storage medium; when the computer program runs on a computer, the computer is enabled to execute the method described in the second aspect.
[0019] In a fourth aspect, an electronic device is provided, including:
[0020] A memory for storing a computer program;
[0021] A processor for executing the computer program to implement the method described in the second aspect.
[0022] The beneficial effects of the present invention are:
[0023] 1. The present invention proposes a data processing method performed by a multi-modal feature importance-guided network, which utilizes the powerful non-linear representation ability of deep learning while integrating the interpretability mechanism for biomarker discovery.
[0024] 2. The multi-modal feature importance-guided network proposed by the present invention includes a feature reconstruction module and a feature selection module. The feature reconstruction module is designed specifically to solve the common problem of missing values in clinical modeling. Even if there is a large amount of missing information in the data, this module can greatly improve the robustness of the model. The feature selection module is tailored for biomarker discovery within a unified deep learning framework. It also collaborates with the feature reconstruction module to further improve the model performance by guiding the feature reconstruction module to focus on key biomarkers for representation learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 is a flowchart of the multi-modal feature importance-guided network based on deep learning provided by the present invention;
[0026] Figure 2 is a schematic diagram of the feature reconstruction module and the feature selection module provided by the present invention;
[0027] Figure 3 is a flowchart of the operation of the feature reconstruction module provided by the present invention;
[0028] Figure 4 is a flowchart of the operation of the feature selection module provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0029] The present invention will be further described below in conjunction with embodiments. The description of the following embodiments is only for helping to understand the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several modifications can still be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
[0030] Embodiment 1:
[0031] As an embodiment, the multi-modal feature importance-guided network (MFIG-Net) based on deep learning provided by the present invention is used for binary classification of the survival period of patients with acute myeloid leukemia (AML) and the discovery of key biomarkers.
[0032] This application introduces two innovative modules in the MFIG-Net framework: one is the Feature Reconstruction Module (FRM) for solving the problem of missing information, and the other is the Feature Selection Module (FSM) for identifying key biomarkers closely related to the model results. This method effectively solves the problems of missing value sensitivity and insufficient model interpretability commonly existing in clinical data through the collaborative work of the feature reconstruction module and the feature selection module. The FRM uses a masked autoencoder to simulate and repair missing values, and the FSM dynamically identifies key biomarkers through the self-attention mechanism and optimizes the masking strategy of the FRM through a feedback mechanism.
[0033] Specifically, as Figure 1 shown, the multi-modal feature importance-guided network based on deep learning includes: a feature reconstruction module, a feature selection module, and a classifier.
[0034] The feature reconstruction module simulates and repairs the missing values in the input data through a masked autoencoder; the input data is the multi-modal clinical data of AML patients.
[0035] As Figure 2 and Figure 3 shown, the feature reconstruction module generates more realistic replacement values by modeling the non-linear relationships between variables instead of relying on simple interpolation techniques or strong linear assumptions. Specifically, it is designed as a masked autoencoder that adopts a random masking strategy during the training process to simulate missing values. Then, it attempts to recover these masked values. This masking-reconstruction learning process enables the model to adapt to potentially incomplete data in the real world while maintaining high accuracy.
[0036] The feature selection module dynamically generates importance coefficients for the reconstructed variables based on the self-attention mechanism and weights the reconstructed variables according to the importance coefficients to obtain weighted variables. As Figure 2 shown, the reconstructed variables are multiplied by the importance coefficients after identity mapping (i.e., keeping the original value passed) to obtain weighted variables.
[0037] In addition, as Figure 4 shown, the feature selection module is designed as a self-attention mechanism sub-network for finding biomarkers related to the model results. Similar to random forests, it learns coefficients to reflect the importance of each variable, thereby helping to identify biomarkers by sorting the variables according to the coefficients. More importantly, considering that the feature reconstruction module may abandon important features during the masking process and have an adverse impact on the model performance, this application further uses these coefficients as online feedback to guide the masking operation in the feature reconstruction module: making the mask biased towards unimportant variables, thereby alleviating this problem. This mechanism can guide the model to learn more informative variables, thereby further improving its performance.
[0038] The classifier outputs the survival prediction result of AML patients according to the weighted variables.
[0039] Example 2:
[0040] Based on Example 1, Example 2 of the present application provides a more specific multi-modal feature importance-guided network based on deep learning, including: a feature reconstruction module, a feature selection module, and a classifier.
[0041] A mini-batch of samples is input into the multi-modal feature importance-guided network. Specifically, first, the mini-batch of samples is input into the feature reconstruction module to recover potential missing values. Then, the feature selection module processes these samples to generate importance coefficients, weight the variables, and then enter the type classifier for final prediction.
[0042] For convenience, the neural network in the feature reconstruction module of the present application is called f frm (·|θ frm ), the neural network in the feature selection module is called f fsm (·|θ fsm ), and the classifier is called f cls (·|θ cls ), where θ * represents learnable parameters. The mini-batch of input variable missing values is represented as where B and N are the batch size and the number of variables per sample, respectively. The mini-batch with complete values output by the feature reconstruction module is represented as The importance coefficients and adjusted variables output by the feature selection module are where * represents element-wise multiplication. Then, the final prediction result can be expressed as follows:
[0043]
[0044] where, is the softmax activation function, used to generate the probability y ∈ R of binary classification (i.e., C = 2) of the survival period of AML patients B×C . At the same time, the importance coefficient ω can be used to determine biomarkers.
[0045] Specifically, the feature reconstruction module is a masked autoencoder, which randomly masks some accessible variables of each sample using a weighted masking strategy during the training phase. Then, it attempts to recover these masked values through the mean square error (MSE) loss. The masking operation for each sample can be formally expressed as:
[0046]
[0047] where, v mis the padding value used to replace the original value of the masked variable (default v m = -1). Indicates the i -th sample in a mini-batch of samples th and the j th -th variable. Ω i represents the set of variable indices obtained by the following method:
[0048]
[0049] where is the weighted random sampling algorithm. ρ = [ρ1, ρ2, …, ρ N are the weights for sampling, obtained from the FSM module. N m = Round(γN) is the number of masked variables selected, where γ ∈ [0.0, 1.0] is a hyperparameter used to control the masking ratio (default is 0.2). The MSE loss function for optimizing θ frm is defined as follows:[[ID=P27]] [[ID=P28]]
[0050] [[ID=P29]] [[ID=P30]] [[ID=P31]]
[0051] where represents the original variable value, is the sign function, indicating that only the masked variables are activated for reconstruction; is the predicted variable value obtained from .
[0052] It should be noted that for accessible variables, the final output of FRM retains its original value rather than the predicted value. This process can be expressed as:
[0053]
[0054] For the feature selection module, it adopts a self-attention mechanism sub-network, and the last layer uses Tanh activation to generate the importance coefficient ω for the reconstructed variable . Then, these coefficients are used to perform element-wise multiplication to produce the weighted variable and input it into the classifier. The relevant expressions are:
[0055]
[0056] The ω is obtained as follows:
[0057]
[0058] where σ tanh is the Tanh activation function. Then, is calculated as follows:
[0059]
[0060] Each coefficient ω i,j ∈[-1.0, 1.0] reflects the importance of the corresponding variable. The larger the value of |ω i,j |, the more important the variable x i,j . Therefore, the element-wise multiplication operation in the above formula can eliminate variables that are not important for classification. More importantly, each row in ω can be used to explain a specific individual prediction. For example, by sorting |ω i | in descending order, we can determine which variables contribute significantly to the prediction of the i th -th sample.
[0061] In addition, the present application applies exponential moving average (EMA) to ω to facilitate the identification of statistical biomarkers. The mathematical expression of the EMA operation is:
[0062]
[0063] where represents the EMA of the coefficients in the k th -th round of training. is initialized as a zero matrix. The hyperparameter α is default set to 0.99. In the test phase, statistical biomarkers can be identified by sorting |ω ema | in descending order.
[0064] To retain biomarkers in the masking operation of the FRM module, ω ema is further used to calculate the sampling weight ρ as follows:
[0065] ρ = 1.0 - σ softmax (|ω ema |)
[0066] The above formula indicates that the masking operation should be biased towards unimportant variables, which can guide the model to learn biomarker representations and thus further improve the model performance.
[0067] Finally, the following total loss is used to optimize the parameters of the entire model:
[0068]
[0069] where is the cross-entropy loss for the classification between the prediction y c and its true label .
[0070] It should be noted that the same or similar parts in this embodiment and Embodiment 1 can be referred to each other and will not be elaborated in this application.
[0071] Embodiment 3:
[0072] Based on Embodiment 2, Embodiment 3 of the present application provides a data processing method executed by a multi-modal feature importance guiding network, including:
[0073] S1. Obtain input data, and simulate and repair missing values in the input data through a feature reconstruction module; the input data is multi-modal clinical data of AML patients.
[0074] For training and evaluating the model, the present application obtained a dataset containing more than 1000 AML patients from a hospital, and normalized the dataset according to the guidance of doctors and divided the dataset into two categories: survival period ≥ 10 months and survival period < 10 months. The dataset contains a total of multiple types of variables:
[0075] Demographic characteristics: This information includes two basic patient characteristics, including age and gender
[0076] Laboratory tests: Include multiple laboratory test results, including white blood cell count (WBC), hemoglobin (Hb), platelet count (PLT);
[0077] Morphological examination: Bone marrow morphological examination (BMM), FAB classification;
[0078] Immunophenotyping: Flow cytometry (FCM);
[0079] Cytogenetic examination: Conditions of 24 pairs of chromosomes;
[0080] Molecular biology examination: Fusion genes, mutant genes, WT1 gene expression;
[0081] Clinical intervention plan: Drug treatment plan.
[0082] S2. Generate importance coefficients of the reconstructed variables through a feature selection module, and weight the reconstructed variables according to the importance coefficients to obtain weighted variables.
[0083] S3. Input the weighted features into a classifier, and output the survival prediction result of AML patients.
[0084] Experimental verification shows that on a multi-modal dataset containing more than 1000 AML patients, the AUC of the present invention under complete data reaches 95.40% ± 0.41%, significantly superior to random forest (AUC 82.3%) and traditional deep learning models (AUC 87.6%). At a 50% missing rate, the AUC of the present invention still remains at 93.94% ± 0.35%, while the performance of the comparative methods generally drops below 70%. In addition, through importance coefficient analysis, key prognostic factors are successfully identified, which are highly consistent with the conclusions of clinical medical research. These results verify the comprehensive advantages of the present invention in missing value processing, classification performance, and biomarker discovery.
[0085] It should be noted that the method provided in this embodiment is the method corresponding to the system provided in Embodiment 2. Therefore, for the parts that are the same or similar in this embodiment and Embodiment 2, reference can be made to each other and will not be elaborated in this application.
Claims
1. A multi-modal feature importance-guided network based on deep learning, characterized in that, Comprising: A feature reconstruction module, a feature selection module, and a classifier; The feature reconstruction module simulates and repairs missing values in the input data through a masked autoencoder; the input data is multi-modal clinical data of AML patients; The feature selection module dynamically generates importance coefficients of the reconstructed variables based on the self-attention mechanism, and weights the reconstructed variables according to the importance coefficients to obtain weighted variables; The classifier outputs a survival prediction result of AML patients based on the weighted variables.
2. The multi-modal feature importance-guided network based on deep learning according to claim 1, wherein The multi-modal clinical data includes: demographic characteristics, laboratory test data, morphological examination data, immunophenotyping data, cytogenetic examination data, molecular biology examination data, and clinical intervention plans.
3. The multi-modal feature importance guided network based on deep learning according to claim 2, wherein During the training process, the feature reconstruction module randomly masks partial accessible variables of each sample using a weighted masking strategy, and restores the masked values through the mean squared error loss.
4. The multi-modal feature importance-guided network based on deep learning according to claim 3, wherein The last layer of the feature selection module uses the Tanh activation function and generates importance coefficients according to the Tanh activation function.
5. The multi-modal feature importance-guided network based on deep learning according to claim 4, wherein An exponential moving average process is performed on the importance coefficients to obtain the importance ranking of biomarkers.
6. A data processing method performed by the multi-modal feature importance-guided network according to claim 1, characterized in that, Comprising: S1. Obtain the input data, and simulate and repair the missing values in the input data through the feature reconstruction module; the input data is multi-modal clinical data of AML patients; S2. Generate importance coefficients of the reconstructed variables through the feature selection module, and weight the reconstructed variables according to the importance coefficients to obtain weighted variables; S3. Input the weighted features into the classifier and output a survival prediction result of AML patients.
7. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program runs on a computer, the computer is caused to execute the method according to claim 6.
8. An electronic device, characterized in that, Comprising: A memory for storing a computer program; A processor for executing the computer program to implement the method according to claim 6.