Deep learning-based unspecified diabetes analysis method, device, equipment and medium

Abstract

This invention relates to the field of bioinformatics technology and discloses a method, apparatus, device, and medium for analyzing unspecified diabetes mellitus based on deep learning. The method involves acquiring an in vitro sample to be analyzed, which includes multiple protein expression features, including at least one important protein feature corresponding to unspecified diabetes mellitus. The in vitro sample is input into a trained deep learning model to obtain the real-valued logarithmic probability corresponding to the matching result between the in vitro sample and the binary classification result. A sigmoid activation function is used to map the matching probability value between the in vitro sample and the binary classification result. Based on the matching probability value, the prediction result for unspecified diabetes mellitus corresponding to the in vitro sample is determined. Based on deep learning technology, important protein features related to unspecified diabetes mellitus can be automatically identified from multiple protein expression features of the in vitro sample, resulting in a biologically interpretable deep learning model that significantly improves the accuracy and stability of model predictions.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method, apparatus, device, and medium for analyzing unspecified diabetes based on deep learning. Background Technology

[0002] Proteins are effector molecules of gene expression, and their circulating levels are regulated by genetic variation and environmental factors, accurately reflecting the activity of biological pathways in the body. Proteomics not only provides rich information on protein post-translational modifications, protein-protein interactions, and signal transduction, but also provides a basis for disease mechanism research and the discovery of novel drug targets. With the development of proteomics technology, more and more circulating proteins have been found to be closely related to the occurrence and development of diabetes, providing new molecular markers for early disease warning. However, how to effectively extract key features from massive protein expression data to build reliable predictive models and improve the accuracy of model predictions remains a challenge in current research. Summary of the Invention

[0003] The purpose of this invention is to provide at least one method for analyzing unspecified diabetes based on deep learning, which can at least solve the technical problem of how to effectively extract key features from massive protein expression data and establish a reliable prediction model, and improve the accuracy of model prediction, and at least achieve the goal of establishing a reliable prediction model and obtaining reliable prediction results.

[0004] To address the aforementioned technical problems, at least one embodiment of this application provides a deep learning-based method for analyzing unspecified diabetes, comprising: acquiring an in vitro sample to be analyzed, the in vitro sample including multiple protein expression features, the multiple protein expression features including at least one important protein feature corresponding to unspecified diabetes; inputting the in vitro sample into a trained deep learning model to obtain the real-valued logarithmic probability corresponding to the matching result between the in vitro sample and the binary classification result, and obtaining the matching probability value between the in vitro sample and the binary classification result through a Sigmoid activation function; and determining the predicted result of unspecified diabetes corresponding to the in vitro sample based on the matching probability value.

[0005] Based on deep learning technology, this method can automatically identify important protein features associated with unspecified diabetes from multiple protein expression characteristics in ex vivo samples, resulting in a biologically interpretable deep learning model. The trained deep learning model is then used to calculate the matching probability of an ex vivo sample belonging to unspecified diabetes. Compared to traditional methods, this approach more comprehensively integrates multidimensional protein information, effectively captures complex nonlinear relationships, and significantly improves the accuracy and stability of model predictions.

[0006] In some examples, the key protein features are determined through interpretive analysis after the deep learning model has been trained.

[0007] In some examples, interpretive analysis of the trained deep model is used to determine important protein features, including: constructing an ensemble model for interpretive analysis and training the ensemble model using the training set used to train the deep learning model to obtain a trained ensemble model; based on the trained ensemble model, calculating a first important feature evaluation index and a second important feature evaluation index for each protein feature; the first important feature evaluation index is used to indicate the structural contribution of the protein feature in class segmentation, and the second important feature evaluation index is used to characterize the criticality of the protein feature in the prediction process; and determining the important protein feature based on the consistency between the first important feature evaluation index and the second important feature evaluation index corresponding to the protein feature.

[0008] In some examples, when the ensemble model is a random forest model, calculating the first important feature evaluation index includes: obtaining the total number of trees in the ensemble model; for each protein feature, obtaining all nodes t that split on the protein feature in each tree structure, the sample weight of the node t, and the gain of the node t; determining the importance of the protein feature in the tree based on the sample weight and gain of the node t; and determining the first important feature evaluation index of the protein feature based on the average importance of the protein feature in multiple trees.

[0009] In some examples, calculating the second important feature evaluation index includes: obtaining the baseline performance index corresponding to the validation set of the ensemble model; performing a scrambling operation R times on each protein feature to obtain the scrambling performance index after each scrambling of the protein feature; and determining the second important feature evaluation index corresponding to the protein feature based on the baseline performance index and the scrambling performance index.

[0010] In some examples, determining the important protein feature based on the consistency of the first important feature evaluation index and the second important feature evaluation index corresponding to the protein feature includes: calculating the importance consistency score corresponding to the protein feature according to the first important feature evaluation index and the second important feature evaluation index; when the importance consistency score is greater than or equal to a preset consistency threshold, determining the importance ranking of the protein feature according to the first important feature evaluation index and the second important feature evaluation index; and determining at least one protein feature whose importance ranking satisfies a preset number of such protein features as the important protein feature.

[0011] At least one embodiment of this application also provides a deep learning-based device for analyzing unspecified diabetes, comprising: an acquisition unit for acquiring an in vitro sample to be analyzed, the in vitro sample including multiple protein expression features, the multiple protein expression features including at least one important protein feature corresponding to unspecified diabetes; an input unit for inputting the in vitro sample into a trained deep learning model to obtain a real-valued logarithmic probability corresponding to the matching result between the in vitro sample and the binary classification result, and obtaining a matching probability value between the in vitro sample and the binary classification result through a Sigmoid activation function; and a determination unit for determining the unspecified diabetes prediction result corresponding to the in vitro sample based on the matching probability value.

[0012] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described deep learning-based method for analyzing unspecified diabetes.

[0013] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described deep learning-based method for analyzing unspecified diabetes. Attached Figure Description

[0014] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0015] Figure 1 This is a flowchart illustrating a deep learning-based method for analyzing unspecified diabetes, provided in one embodiment of this application. Figure 2 This is a schematic flowchart illustrating the process of determining important protein features according to an embodiment of this application; Figure 3 This is a schematic diagram of ten key protein features associated with unspecified diabetes, provided in one embodiment of this application; Figure 4 This is a schematic diagram illustrating the relationship between epoch and loss value, provided in another embodiment of this application; Figure 5 This is a schematic diagram of the structure of a non-specific diabetes analysis device based on deep learning, provided in another embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0017] It should be noted that the acquisition or use of data in the embodiments of this application requires the user's consent. The relevant data can only be obtained after the user's authorization, and the acquisition or use of the data complies with the provisions of relevant laws and regulations.

[0018] To facilitate understanding of the embodiments of this application, we will first introduce the relevant content of the deep learning-based method for analyzing unspecified diabetes.

[0019] Proteins are effector molecules of gene expression, and their circulating levels are regulated by genetic variation and environmental factors, accurately reflecting the activity of biological pathways in the body. Proteomics not only provides rich information on protein post-translational modifications, protein-protein interactions, and signal transduction, but also provides a basis for disease mechanism research and the discovery of novel drug targets. With the development of proteomics technology, more and more circulating proteins have been found to be closely related to the occurrence and development of diabetes, providing new molecular markers for early disease warning. However, how to effectively extract key features from massive protein expression data to build reliable predictive models and improve the accuracy of model predictions remains a challenge in current research.

[0020] To address the aforementioned technical problems, this invention proposes a deep learning-based method for analyzing unspecified diabetes. The implementation details of this deep learning-based method for analyzing unspecified diabetes are described below. The following details are provided for ease of understanding and are not essential for implementing this solution.

[0021] Example 1: The deep learning-based method for analyzing unspecified diabetes in this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. Its specific process can be as follows: Figure 1 As shown, it includes: Step 110: Obtain the in vitro sample to be analyzed. The in vitro sample includes multiple protein expression features, including at least one important protein feature not specifically corresponding to diabetes.

[0022] In this context, "ex vivo sample" refers to isolated tissues from humans or animals, or from deceased animals or humans. "Unspecified diabetes" refers to diabetes with abnormally high blood glucose levels, but for which the type of diabetes has not yet been determined. Protein expression characteristics are used to characterize the expression level of a corresponding protein feature in isolated tissues; this can also be understood as the concentration or content of that protein feature. Higher expression levels indicate higher concentrations or content of the corresponding protein feature in the isolated sample. Important protein features refer to the expression levels of protein features that are highly correlated with unspecified diabetes. This can also be understood as the expression level of the protein feature among multiple protein features related to unspecified diabetes whose correlation with unspecified diabetes is higher than a preset correlation level or whose correlation ranking is higher than a preset ranking.

[0023] Specifically, the key protein features were determined through interpretive analysis after the deep learning model was trained.

[0024] Furthermore, such as Figure 2 As shown, interpretive analysis of a trained deep learning model is performed to identify important protein features, including the following steps: Step S01, constructing an ensemble model for interpretive analysis, and training the ensemble model using the training set used to train the deep learning model to obtain a trained ensemble model.

[0025] The ensemble model acts as a proxy model for the deep learning model, used to corroborate the importance of extracted protein features when analyzing in vitro samples to determine whether the sample is nonspecifically diabetic. The ensemble model can be a gradient boosting tree or a random forest. The training set refers to the set used to train the deep learning model. The training set consists of 80% of the samples randomly selected from the total sample set. Further, the training set is split into a new training set and a validation set in an 80%:20% ratio, with the remaining 20% ​​of the total sample set used as the test set. The new training set is used for model training, the validation set for parameter tuning, and the test set for final performance evaluation of the model.

[0026] The total sample set includes at least 53,013 training samples, each containing 2,923 protein expression features and a training sample label. The training sample label indicates whether the training sample is an ex vivo tissue sample from a person or animal with diabetes. The total sample set includes some training samples associated with unspecified diabetes and some training samples not associated with unspecified diabetes.

[0027] Furthermore, the total sample set is the sample set obtained after data preprocessing of the initial total sample set. Data preprocessing includes missing value handling and standardization. Specifically, missing value handling of the initial total sample set includes: obtaining the first total number of initial training samples in the initial total sample set; for each protein feature in the initial total sample set, obtaining multiple protein expression features corresponding to that protein feature in the initial total sample set, determining whether the multiple protein expression features are similar, and if similar, deleting the protein expression features corresponding to that protein feature in each initial training sample; obtaining a second total number of multiple protein expression features, and if the second total number is less than the first total number, determining the median value of the protein expression features; determining the target initial training sample where the protein expression feature is missing, and filling the missing protein expression feature of that protein feature in the target initial training sample using the feature median value; if no multiple protein expression features are obtained, deleting that protein feature in each initial training sample; wherein, if multiple protein expression features are all the same, or the difference between the largest and smallest protein expression features is less than a preset difference threshold, then the multiple protein expression features are determined to be similar.

[0028] The median value of protein expression features refers to the median of multiple protein expression features corresponding to the protein features in the initial total sample set.

[0029] Therefore, using the median to impute missing values ​​is more resistant to outlier interference than using the mean, especially suitable for data such as proteomics that may not follow a normal distribution. This preserves data integrity while avoiding the loss of valuable information caused by deleting samples. Secondly, by removing features that are missing or almost unchanged in all samples, "noisy features" that cannot provide discriminative information are successfully removed. This not only significantly reduces the computational complexity of training the model but also effectively prevents noise from interfering with the model's learning of true patterns, thus improving the model's learnability.

[0030] Specifically, the initial total sample set is standardized, including: for each protein expression feature in the initial total sample set, z-score standardization is used to perform numerical transformation on the protein expression features to obtain protein expression features of uniform order of magnitude. Uniform order of magnitude means that the order of magnitude of each protein expression feature is the same. Specifically, in this scheme, uniform order of magnitude means that the distribution of each protein expression feature after standardization in the initial total sample set follows a mean of 0 and a standard deviation of 1.

[0031] Therefore, by standardizing protein expression features, the problems of dimensional differences and inconsistent numerical scales between different features are effectively eliminated, ensuring that all features are compared and calculated within a uniform numerical range. This process ensures that each protein feature contributes equally to the model's decision-making during training, avoiding unnecessary bias towards certain protein features caused by excessive differences in their numerical ranges. This significantly improves the model's convergence speed and prediction accuracy.

[0032] Specifically, the number of training samples related to unspecified diabetes in the total sample set differs from the number of training samples not related to unspecified diabetes. Furthermore, the training set comprised of 80% of the randomly selected training samples from the total sample set is more likely to have an imbalance in the number of training samples related to unspecified diabetes. To balance the number of training samples related to and unrelated to unspecified diabetes in the training set, an oversampling algorithm can be used to perform data balancing on either type of training sample. This ensures that the number of training samples related to and unrelated to unspecified diabetes is balanced in the balanced training set. This reduces the bias tendency of the trained deep learning model and ensemble model, improving their ability to identify minority classes.

[0033] In this algorithm, if the number of training samples related to unspecified diabetes is less than the number of training samples not related to unspecified diabetes, then the training samples related to unspecified diabetes are considered a minority class. Conversely, if the number of training samples related to unspecified diabetes is greater than the number of training samples not related to unspecified diabetes, then the training samples not related to unspecified diabetes are considered a minority class. Oversampling algorithms involve expanding the number of training samples in the minority class to a level comparable to the number of training samples in the non-minority class. This is used to alleviate class imbalance and prevent the trained deep learning model and ensemble model from having poor prediction accuracy due to excessively large discrepancies in the number of training samples from different classes in the training set.

[0034] Furthermore, the oversampling algorithm performs data balancing on training samples that are either related to or unrelated to unspecified diabetes. This includes repeatedly performing K-nearest neighbor interpolation on the training samples that are either related to or unrelated to unspecified diabetes to synthesize new training samples. These new training samples are then used as samples in the training set until the number of training samples related to unspecified diabetes in the training set is equal to the number of training samples unrelated to unspecified diabetes. The new training samples belong to the minority class.

[0035] Furthermore, K-nearest neighbor interpolation is performed on training samples that are related to or unrelated to unspecified diabetes to synthesize new training samples. This includes: selecting k nearest neighbor training samples from the minority class that are adjacent to the target training sample based on the nearest neighbor search method, where the target training sample refers to any training sample in the minority class; and performing linear interpolation on the target training sample and the nearest training samples for each nearest training sample to obtain new training samples.

[0036] Specifically, the selection of k nearest neighbor training samples from the minority class based on the nearest neighbor search method includes: calculating the Euclidean distance between the training sample and the target training sample, and selecting the k training samples with the smallest Euclidean distance as the nearest neighbor training samples.

[0037] For example, the Euclidean distance between the training sample and the target training sample can be calculated using the following formula: Where x is the target training sample, and y is the training sample. Let be the Euclidean distance, i be the target training sample and the i-th protein feature of the training sample, and n be the target training sample or the total number of protein features contained in the training sample.

[0038] Example 2: A new training sample is obtained by interpolating the target training sample and its neighboring training samples using linear interpolation, as shown in the following formula:

[0039] Where x is the target training sample and y is the neighboring training sample. For new training samples, These are preset parameters.

[0040] Step S02: Based on the trained ensemble model, calculate the first important feature evaluation index and the second important feature evaluation index for each protein feature; the first important feature evaluation index is used to indicate the structural contribution of the protein feature in class segmentation, and the second important feature evaluation index is used to characterize the criticality of the protein feature in the prediction process.

[0041] The first importance feature evaluation index reflects the structural contribution of protein features in segmenting the boundary between diabetes and non-diabetes. The greater the structural contribution, the more important the corresponding protein feature is in distinguishing between diabetes and non-diabetes. Highly important protein features can be used for class segmentation. The second importance feature evaluation index reflects the decrease in the model's performance in distinguishing between diabetes and non-diabetes after the protein feature is occluded. The greater the decrease, the more critical the protein feature is in distinguishing between diabetes and non-diabetes.

[0042] Specifically, the trained ensemble model is validated based on the validation set of the validation deep learning model. When validating the ensemble model based on the validation set, the first important feature evaluation index and the second important feature evaluation index are calculated for each protein feature.

[0043] Specifically, when the ensemble model is a random forest model, the calculation of the first important feature evaluation index in step S02 above includes: obtaining the total number of trees in the ensemble model; for each protein feature, obtaining all nodes t that split on the protein feature in each tree structure, the sample weight of node t, and the gain of node t; determining the importance of the protein feature in the tree based on the sample weight and gain of node t; and determining the first important feature evaluation index of the protein feature based on the average importance of the protein feature in multiple trees.

[0044] The ensemble model is composed of multiple decision tree models, with each decision tree model representing a single tree. A split node is a node that divides the training set into two or more subsets. Gain refers to the degree of disorder or impurity of the subsets after the training set is partitioned using split nodes; it is used to measure the effectiveness of the split.

[0045] Furthermore, the importance of protein features in the tree is determined based on the sample weights and gains of node t, including: determining the importance of protein features in the tree based on the sample weights, gains, and total number of trees of node t.

[0046] For example, the importance of a protein feature in a tree can be determined based on the sample weights, gain, and total number of trees at node t, as shown in the following formula:

[0047] in, This refers to the importance of protein features in the tree. It refers to gain; This refers to sample weights; The total number of trees; For example, a tree in an ensemble model; , For the tree All in the features The nodes that split upwards.

[0048] Furthermore, based on the average importance of protein features across multiple trees, a first importance feature evaluation index is determined, including using the average importance as the first importance feature evaluation index.

[0049] Specifically, in step S02 above, calculating the second important feature evaluation index includes: obtaining the baseline performance index corresponding to the ensemble model validation set; performing a scrambling operation on each protein feature R times to obtain the scrambling performance index after each scrambling of the protein feature; and determining the second important feature evaluation index corresponding to the protein feature based on the baseline performance index and the scrambling performance index.

[0050] The baseline performance metric represents the best performance of the ensemble model without disrupting features. The scrambling performance metric represents the model performance after each scrambling of a protein feature (i.e., after disrupting the feature). A higher value for the second most important feature evaluation index indicates a greater impact of that protein feature on model performance. The number of repetitions of the scrambling operation can be used to reduce the randomness of the model; the more repetitions, the less randomness the model has, and the better its performance.

[0051] Furthermore, based on baseline performance metrics and scrambling performance metrics, the second most important feature evaluation index corresponding to the protein feature is determined, as shown in the following formula:

[0052] in, It is the second most important feature evaluation index; Baseline performance metrics; To scramble performance indicators; The number of scrambling operations.

[0053] Step S03: Based on the consistency between the first important feature evaluation index and the second important feature evaluation index corresponding to the protein feature, important protein features are determined.

[0054] Specifically, in the aforementioned step S03, determining important protein features based on the consistency between the first important feature evaluation index and the second important feature evaluation index corresponding to the protein features includes: calculating the important consistency score corresponding to the protein features according to the first important feature evaluation index and the second important feature evaluation index; when the important consistency score is greater than or equal to a preset consistency threshold, determining the importance ranking of the protein features according to the first important feature evaluation index and the second important feature evaluation index; and determining at least one protein feature whose importance ranking meets a preset number as an important protein feature.

[0055] The importance consistency score indicates the degree of consistency between a protein feature and its two physical meanings (the model structure expressed by the first importance feature evaluation index and the predictive performance of the two physical meanings expressed by the second importance feature evaluation index). A higher importance consistency score indicates a higher degree of consistency between the protein feature and its two physical meanings, meaning the first and second importance feature evaluation indices expressing both physical meanings are more reliable. The importance consistency score ranges from 0 to 1. If the importance consistency score is 1, it means the ranking of the first and second importance feature evaluation indices is exactly the same, indirectly verifying the mutual reliability of the first and second importance feature evaluation indices, indicating that both are reliable. If the importance consistency score is 0, it means the ranking of the first importance feature evaluation index is either first or last, and the ranking of the corresponding second importance feature evaluation index is either last or first, meaning they are completely inconsistent; if one is first, the other is last, indirectly indicating that both the first and second importance feature evaluation indices are unreliable. The preset consistency threshold can be set according to actual needs, such as 0.9, 0.8, 0.95, or other values. This is just an example and is not a practical limitation. The preset quantity can be set according to actual needs, such as 5, 10, 11, 12, 13, or other values. This is just an example and is not a practical limitation.

[0056] Furthermore, based on the first important feature evaluation index and the second important feature evaluation index, the importance consistency score corresponding to the protein feature is calculated, including: determining the first important feature evaluation ranking of each protein feature based on the first important feature evaluation index; determining the second important feature evaluation ranking of each protein feature based on the second important feature evaluation index; obtaining the number of proteins of the protein feature in the validation set; and calculating the importance consistency score corresponding to the protein feature based on the first important feature evaluation ranking, the second important feature evaluation ranking, and the number of proteins.

[0057] The ranking of the first important feature refers to the ranking of the protein feature's corresponding first important feature evaluation index among all protein features in the validation set. This ranking value ranges from 1 to less than or equal to the number of proteins. The ranking of the second important feature refers to the ranking of the protein feature's corresponding second important feature evaluation index among all protein features in the validation set. This ranking value ranges from 1 to less than or equal to the number of proteins. The ranking of the first important feature is based on the first important feature evaluation index, from largest to smallest. The smaller the ranking, the greater the structural contribution of the corresponding protein feature in segmenting the boundary between diabetes and non-diabetes. The ranking of the second important feature is based on the second important feature evaluation index, from largest to smallest. The smaller the ranking, the more critical the corresponding protein feature is in distinguishing between diabetes and non-diabetes.

[0058] For example, the importance consistency score corresponding to a protein feature is calculated based on the ranking of the first important feature, the ranking of the second important feature, and the number of proteins, as shown in the following formula:

[0059] in, This refers to the ranking of the most important features of protein feature i. This refers to the ranking of the second most important feature of protein feature i. This refers to the number of proteins with different protein characteristics in the validation set; This refers to the important consistency score of protein feature i.

[0060] Example 2: Based on the ranking of the first important feature, the ranking of the second important feature, and the number of proteins, the importance consistency score corresponding to the protein feature is calculated, as shown in the following formula:

[0061] in, This refers to the ranking of the most important features of protein feature i. This refers to the ranking of the second most important feature of protein feature i. This refers to the number of proteins with different protein characteristics in the validation set; This refers to the important consistency score of protein feature i.

[0062] Example 3: The importance consistency score corresponding to the protein feature is calculated based on the ranking of the first important feature evaluation, the ranking of the second important feature evaluation, and the number of proteins, as shown in the following formula:

[0063] in, The preset weighting factor of protein feature i This refers to the ranking of the most important features of protein feature i. This refers to the ranking of the second most important feature of protein feature i. This refers to the number of proteins with different protein characteristics in the validation set; This refers to the important consistency score of protein feature i.

[0064] Furthermore, the importance ranking of protein features is determined based on the first and second important feature evaluation indices, including: determining the first important feature evaluation ranking of each protein feature based on the first important feature evaluation index; determining the second important feature evaluation ranking of each protein feature based on the second important feature evaluation index; calculating the importance index corresponding to the protein feature based on the first and second important feature evaluation rankings; and determining the importance ranking of the protein feature based on the importance index.

[0065] For example, the importance index corresponding to a protein feature is calculated based on the ranking of the first and second most important features, as shown in the following formula:

[0066] in, This is the importance index of protein feature i. The ranking of the most important features of protein feature i; The ranking is based on the second most important feature of protein feature i.

[0067] Furthermore, the importance ranking of protein features is determined based on the importance index, including: sorting the importance index of each protein feature in ascending order to obtain the importance ranking of each protein feature. The smaller the importance index, the higher the corresponding importance ranking, and the more important the corresponding protein feature is in distinguishing between diabetes and non-diabetes.

[0068] In some examples, to corroborate the importance index and determine its accuracy and effectiveness, the validation set can be divided into a diabetes set and a non-diabetes set. First model prediction results are obtained for each validation sample in the diabetes set based on the diabetes set. Second model prediction results are obtained for each validation sample in the non-diabetes set based on the non-diabetes set. A first model prediction baseline is determined based on multiple first model prediction results. A second model prediction baseline is determined based on multiple second model prediction results. For each validation sample, a first SHAP (Significant Important Value) for each protein feature in the validation sample is determined based on the first model prediction results and the first model prediction baseline. A second SHAP for each protein feature in the validation sample is determined based on the second model prediction results and the second model prediction baseline. For each protein feature, the mean of the first SHAPs corresponding to the protein feature in the validation set is determined based on multiple first SHAPs. The mean of the second SHAPs corresponding to the protein feature in the validation set is determined based on multiple second SHAPs. A global importance index for the protein feature is determined based on the mean of the first SHAPs and the mean of the second SHAPs. Based on the global importance index, a weighted sum of the global importance index and the importance index is obtained to obtain a new importance index, which is used to determine the importance ranking of the protein features.

[0069] Here, the first SHAP and the second SHAP are both contributions of protein features to the prediction results, used to illustrate the degree and direction of the influence of protein features on the model prediction. The first SHAP is determined based on the diabetes set, and the second SHAP is determined based on the non-diabetes set.

[0070] Specifically, the first SHAP of each protein feature in the validation sample is determined based on the prediction results of the first model and the baseline of the first model prediction, as shown in the following formula:

[0071] Where g(x) is the prediction result of the first model, Predict the baseline for the first model. The first SHAP for protein characteristic i.

[0072] Furthermore, a global importance index for protein features is determined based on the first SHAP mean and the second SHAP mean, as shown in the following formula:

[0073] in, The first SHAP mean determined based on the diabetes set; The second SHAP mean was determined based on the non-diabetic set; It is an important index for the whole. A value greater than 0 indicates that the protein feature is used to push the prediction result to diabetes, while a value less than 0 indicates that the protein feature is used to push the prediction result to non-diabetes.

[0074] For example, based on the ensemble model, the top 10 protein features ranked by importance were identified as being able to effectively distinguish between diabetes and non-diabetes, such as... Figure 3 As shown in the figure, the horizontal axis represents the importance index of the corresponding protein feature. The vertical axis represents the top 10 protein features in terms of importance, specifically: PLXNB2 (Plexin-B2), GDF15 (growth differentiation factor 15), CFC1 (Cryptic protein 1), LGALS4 (galactolectin 4), SIAE (sialic acid-O-acetyl esterase), DPP6 (dipeptidyl peptidase-like protein 6), REN (renin), PAEP (glycoprotein P, also known as Glycodelin), MUC13 (mucin 13), and GIP (gastric inhibitory peptide). Among them, GDF15 is a stress-induced cytokine, and studies have shown that its elevation can suppress appetite and reduce weight; REN (renin) is the core hormone of the renin-angiotensin system, directly affecting blood pressure regulation and kidney function, and plays an important role in hypertension and kidney complications in diabetic patients; GIP belongs to the pancreaticoglucagon family of hormones, which can promote the secretion of insulin by pancreatic β cells, thereby helping to regulate blood glucose levels. These highly important protein biomarkers involve multiple pathways, including metabolic regulation, inflammatory response, and hormone signaling, which supports the role of protein characteristics as a "mediating phenotype" in the development of diabetes and helps to deepen our understanding of the molecular mechanisms of diabetes.

[0075] Therefore, by simulating and validating the deep learning model using an ensemble model, the specific contributions of each protein feature in predicting unspecified diabetes mellitus were successfully revealed, thus accurately identifying several important protein features that significantly influence the prediction results. This not only confirms that the deep learning model can automatically learn and extract complex nonlinear relationships from the training data, but also demonstrates its superiority in protein feature selection and recognition. By capturing these potential combinations of biomarkers, the accuracy and sensitivity of unspecified diabetes mellitus prediction were significantly improved.

[0076] Step 120: Input the isolated sample into the trained deep learning model to obtain the real log probability corresponding to the matching result between the isolated sample and the binary classification result. The matching probability value between the isolated sample and the binary classification result is obtained by mapping through the Sigmoid activation function.

[0077] The deep learning model, trained on the aforementioned training set, is a feedforward deep neural network model, which can be named ProteinExpressionNet. The deep learning model consists of one input layer, three hidden layers, and one output layer. The input layer has a dimension of 2923, with each dimension corresponding to a protein expression feature. The three hidden layers are fully connected networks with 128, 64, and 32 nodes respectively. The outputs of each hidden layer undergo batch normalization, ReLU activation, and a 30% random deactivation mechanism. Before activation, each hidden layer undergoes batch normalization to reduce internal covariate bias, followed by a linear rectified activation function (e.g., ReLU) to introduce non-linear modeling capabilities. After activation, a 30% random node deactivation (Dropout mechanism) is applied to enhance the model's generalization performance. By extracting high-order features layer by layer, the deep learning model can extract protein features related to diabetes status from thousands of dimensions of input, thereby determining whether the in vitro sample belongs to a human or animal with diabetes. The output layer is a fully connected node that outputs the matching probability value. The matching probability value is used to determine whether the human or animal to which the input ex vivo sample belongs has diabetes.

[0078] Specifically, during the training iterations of a deep learning model, a validation set is used to monitor the performance of the trained deep learning model on unseen data, and to adjust the training strategy accordingly, or trigger an early stopping mechanism to halt the training of the deep learning model. The validation process, based on the aforementioned validation set, verifies the deep learning model trained in each training round, including: determining the loss weights for positive and negative samples based on the weighted cross-entropy loss function and the ratio of positive to negative samples in the validation set; determining the positive and negative sample loss values ​​of the deep learning model based on each validation sample in the validation set; obtaining the first product of the positive sample loss weights and positive sample loss values; obtaining the second product of the negative sample loss weights and negative sample loss values; and using the sum of the first and second products as the loss value of the deep learning model in this training round. The learning rate of the Adam optimizer is adjusted based on the loss value, or an early stopping strategy is triggered.

[0079] Furthermore, multiple loss values ​​determined in consecutive validation rounds are obtained. If the difference between the maximum and minimum values ​​of the multiple loss values ​​is less than the loss difference threshold, the learning rate of the Adam optimizer is reduced. If the difference between the maximum and minimum values ​​among a series of preset loss values ​​is less than the loss difference threshold, an early stopping strategy is triggered to terminate the training of the deep learning model in advance.

[0080] Positive samples refer to training samples from human or animal subjects with diabetes. Negative samples refer to training samples from human or animal subjects without diabetes. The initial learning rate of the Adam optimizer is 0.001. This preset value can be set according to actual needs, for example, to 5. Furthermore, the maximum number of training epochs for the deep learning model is 200.

[0081] For example, see Figure 4 With training epochs between 15 and 25, the training and validation losses of the deep learning model change relatively smoothly. At this point, the deep learning model has basically converged. Therefore, setting the maximum training epochs to 200 provides enough training epochs to allow the analysis model to converge.

[0082] For example, the loss weights for positive samples and negative samples are determined based on the ratio of positive to negative samples in the validation set, as shown in the following formula:

[0083] in, The loss weights for positive samples, The loss weights for negative samples, To verify the number of negative samples in the set, To verify the number of positive samples in the set.

[0084] Therefore, when the proportion of positive samples is large and the proportion of negative samples is small, allocating smaller weights reduces the gradient contribution of positive samples and amplifies the gradient contribution of negative samples, thus offsetting the class bias problem caused by the imbalance of positive and negative sample types when validating the model. A learning rate of 0.001 ensures convergence speed while avoiding training oscillations. When the loss value on the validation set no longer decreases within several rounds, or the decrease is less than the loss difference threshold, the learning rate is automatically reduced by a certain factor, allowing the model to make more refined parameter adjustments near its optimal value. By introducing an early stopping mechanism, the training process is terminated early when the loss value on the validation set does not decrease for a consecutive preset number of rounds, or the decrease is very small. This effectively prevents the deep learning model from overfitting on the training set and improves the generalization ability of the deep learning model on unknown data.

[0085] In some examples, after training, the deep learning model's performance is evaluated using a test set, including: obtaining the average loss, classification accuracy, and area under the receiver operating characteristic (AUC) curve for the test set; and generating a confusion matrix and classification report based on the average loss, classification accuracy, and AUC curve. The classification report includes the deep learning model's precision, recall, and F1 score in predicting positive and negative samples.

[0086] The confusion matrix is ​​a table used to compare the predictions of the deep learning model with the true labels, clearly showing the classification performance of the deep learning model on various types of samples. Precision measures how many samples the model predicts as positive are actually positive; that is, the accuracy of the deep learning model's positive predictions. Recall refers to how many positive samples are successfully identified by the deep learning model. The F1 score is the harmonic mean of precision and recall.

[0087] Experimental results show that the deep learning model achieves a diabetes classification accuracy of approximately 0.812 and an AUC of approximately 0.732 on a large number of independent test samples, indicating that the deep learning model has good classification ability and high accuracy.

[0088] Step 130: Determine the prediction result of unspecified diabetes for the isolated sample based on the matching probability value.

[0089] Specifically, if the matching probability value is greater than 0.5, it is determined that the human or animal body to which the in vitro sample belongs has unspecified diabetes; otherwise, it is determined that the human or animal body to which the in vitro sample belongs does not have unspecified diabetes.

[0090] In summary, this approach acquires in vitro samples for analysis, which include multiple protein expression features, including at least one important protein feature corresponding to unspecified diabetes. The in vitro samples are then input into a trained deep learning model to obtain the real-valued logarithmic probability of matching the in vitro samples with the binary classification results. A sigmoid activation function is used to map the matching probability value between the in vitro samples and the binary classification results. Based on the matching probability value, the prediction result for unspecified diabetes corresponding to the in vitro samples is determined. Based on deep learning technology, this approach can automatically identify important protein features related to unspecified diabetes from multiple protein expression features of in vitro samples, resulting in a biologically interpretable deep learning model. The trained deep learning model is then used to calculate the matching probability of the in vitro samples belonging to unspecified diabetes. Compared to traditional methods, this approach can more comprehensively integrate multidimensional protein information, effectively capture complex nonlinear relationships, and significantly improve the accuracy and stability of model predictions.

[0091] Example 2: Another embodiment of this application relates to a deep learning-based unspecified diabetes analysis device. The implementation details of this deep learning-based unspecified diabetes analysis device are described below. The following details are provided for ease of understanding and are not essential for implementing this solution. A schematic diagram of the deep learning-based unspecified diabetes analysis device 50 in this embodiment can be seen as follows: Figure 5 As shown, it includes an acquisition unit 501, an input unit 502, and a determination unit 503.

[0092] The acquisition unit 501 is used to acquire an in vitro sample to be analyzed. The in vitro sample includes multiple protein expression features, including at least one important protein feature not specifically corresponding to diabetes.

[0093] The input unit 502 is used to input the isolated sample into the trained deep learning model to obtain the real log probability corresponding to the matching result between the isolated sample and the binary classification result, and to obtain the matching probability value between the isolated sample and the binary classification result through the Sigmoid activation function.

[0094] The determining unit 503 is used to determine the unspecified diabetes prediction result corresponding to the ex vivo sample based on the matching probability value.

[0095] In some examples, the key protein features are determined through interpretive analysis after the deep learning model has been trained.

[0096] In some examples, when determining important protein features by performing interpretive analysis on the trained deep model, the determining unit 503 specifically performs the following steps: constructing an ensemble model for interpretive analysis and training the ensemble model using the training set used to train the deep learning model to obtain a trained ensemble model; calculating a first important feature evaluation index and a second important feature evaluation index for each protein feature based on the trained ensemble model; the first important feature evaluation index is used to indicate the structural contribution of the protein feature in class segmentation, and the second important feature evaluation index is used to characterize the criticality of the protein feature in the prediction process; and determining the important protein feature based on the consistency between the first important feature evaluation index and the second important feature evaluation index corresponding to the protein feature.

[0097] In some examples, when the ensemble model is a random forest model, the determining unit 503, when calculating the first importance feature evaluation index, specifically performs the following steps: obtaining the total number of trees in the ensemble model; for each protein feature, obtaining all nodes t that split on the protein feature in each tree structure, the sample weights of the nodes t, and the gain of the nodes t; determining the importance of the protein feature in the tree based on the sample weights and gain of the nodes t; and determining the first importance feature evaluation index of the protein feature based on the average importance of the protein feature in multiple trees.

[0098] In some examples, when the determining unit 503 is used to calculate the second important feature evaluation index, it is specifically used to: obtain the baseline performance index corresponding to the integrated model validation set; perform a scrambling operation on each protein feature R times to obtain the scrambling performance index after each scrambling of the protein feature; and determine the second important feature evaluation index corresponding to the protein feature based on the baseline performance index and the scrambling performance index.

[0099] In some examples, when the determining unit 503 is used to determine the important protein feature based on the consistency of the first important feature evaluation index and the second important feature evaluation index corresponding to the protein feature, it is specifically used to: calculate the important consistency score corresponding to the protein feature according to the first important feature evaluation index and the second important feature evaluation index; when the important consistency score is greater than or equal to a preset consistency threshold, determine the importance ranking of the protein feature according to the first important feature evaluation index and the second important feature evaluation index; and determine at least one protein feature whose importance ranking satisfies a preset number of such protein features as the important protein feature.

[0100] It is worth mentioning that all units involved in this embodiment are logical units. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0101] Example 3: Another embodiment of this application relates to an electronic device, such as... Figure 6 As shown, it includes: at least one processor 901; and a memory 902 communicatively connected to the at least one processor 901; wherein the memory 902 stores instructions executable by the at least one processor 901, the instructions being executed by the at least one processor 901 to enable the at least one processor 901 to execute the deep learning-based non-specific diabetes analysis method in the above embodiments.

[0102] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0103] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0104] Example 4: Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0105] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A deep learning-based method for analyzing unspecified diabetes mellitus, characterized in that, include: Obtain an in vitro sample to be analyzed, the in vitro sample including multiple protein expression features, the multiple protein expression features including at least one important protein feature not specifically corresponding to diabetes; The isolated sample is input into a trained deep learning model to obtain the real log probability corresponding to the matching result between the isolated sample and the binary classification result. The matching probability value between the isolated sample and the binary classification result is obtained by mapping through the Sigmoid activation function. Based on the matching probability value, the prediction result of unspecified diabetes corresponding to the ex vivo sample is determined.

2. The method for analyzing unspecified diabetes based on deep learning according to claim 1, characterized in that, The key protein features were determined through interpretive analysis after the deep learning model was trained.

3. The method for analyzing unspecified diabetes based on deep learning according to claim 2, characterized in that, Interpretive analysis was performed on the trained deep model to identify important protein features, including: An ensemble model for interpretive analysis is constructed, and the ensemble model is trained using the training set used to train the deep learning model to obtain a trained ensemble model. Based on the trained ensemble model, a first important feature evaluation index and a second important feature evaluation index are calculated for each protein feature; the first important feature evaluation index is used to indicate the structural contribution of the protein feature in class segmentation, and the second important feature evaluation index is used to characterize the criticality of the protein feature in the prediction process. The important protein features are determined based on the consistency between the first important feature evaluation index and the second important feature evaluation index corresponding to the protein features.

4. The method for analyzing unspecified diabetes based on deep learning according to claim 3, characterized in that, When the ensemble model is a random forest model, the calculation of the first important feature evaluation index includes: Obtain the total number of trees in the ensemble model; For each protein feature, obtain all nodes t that split on the protein feature in each tree structure, the sample weight of node t, and the gain of node t; The importance of the protein feature in the tree is determined based on the sample weights and gains of node t; The first importance feature evaluation index of the protein feature is determined based on the average importance of the protein feature in multiple trees.

5. The method for analyzing unspecified diabetes based on deep learning according to claim 3, characterized in that, The calculation of the second most important feature evaluation index includes: Obtain the baseline performance metrics corresponding to the validation set of the ensemble model; The scrambling operation is repeated R times for each protein feature to obtain the scrambling performance index after each scrambling of the protein feature; Based on the baseline performance index and the scrambling performance index, a second important feature evaluation index corresponding to the protein feature is determined.

6. The method for analyzing unspecified diabetes based on deep learning according to claim 3, characterized in that, The determination of the important protein feature based on the consistency between the first important feature evaluation index and the second important feature evaluation index corresponding to the protein feature includes: Calculate the importance consistency score corresponding to the protein feature based on the first important feature evaluation index and the second important feature evaluation index; When the important consistency score is greater than or equal to a preset consistency threshold, the importance ranking of the protein feature is determined based on the first important feature evaluation index and the second important feature evaluation index. The protein feature whose importance ranking meets a preset number is identified as the important protein feature.

7. A deep learning-based device for analyzing unspecified diabetes, characterized in that, include: An acquisition unit is used to acquire an in vitro sample to be analyzed, the in vitro sample including multiple protein expression features, the multiple protein expression features including at least one important protein feature not specifically corresponding to diabetes; The input unit is used to input the isolated sample into the trained deep learning model to obtain the real log probability corresponding to the matching result between the isolated sample and the binary classification result, and to obtain the matching probability value between the isolated sample and the binary classification result through the Sigmoid activation function. A determining unit is used to determine the unspecified diabetes prediction result corresponding to the ex vivo sample based on the matching probability value.

8. The deep learning-based unspecified diabetes analysis device according to claim 7, characterized in that, The key protein features were determined through interpretive analysis after the deep learning model was trained.

9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the deep learning-based method for analyzing unspecified diabetes as described in any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the deep learning-based method for analyzing unspecified diabetes as described in any one of claims 1 to 6.

Images