Interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method, device, equipment and medium

Abstract

The application provides an interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method, device, equipment and medium. The interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method comprises the following steps: acquiring user case information, performing interpolation and enhancement processing on the user case information by using a first model to obtain user case text and user case data; performing grouping diagnosis and treatment and disease risk prediction processing on the user case text and the user case data by using a second model to obtain a prediction result, wherein the first model and the second model are obtained by training. The application has the beneficial effect of improving the timeliness and accuracy of the old-age prevention and control auxiliary diagnosis and treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical treatment and artificial intelligence, and in particular relates to a prevention and control auxiliary diagnosis and treatment method and device based on interpolation and enhancement, equipment and medium. BACKGROUND

[0002] Based on the four-group diagnosis and treatment system (non-infection non-severe, infection non-severe, non-infection severe, and infection severe) of the "infection-severe" two-dimensional classification, the clinical decision and resource allocation can be significantly optimized.

[0003] However, the existing grouping diagnosis and treatment method has the following disadvantages: the artificial grouping diagnosis and treatment can achieve a certain accuracy, but it is time-consuming and laborious; in clinical practice, there are missing fields in electronic medical records (such as laboratory tests not performed), and the subjectivity of manual filling is large and it is difficult to handle a large amount of data, but traditional interpolation methods are difficult to handle multi-modal mixed missing scenarios; traditional methods such as mean interpolation and multiple interpolation only target numerical missing data, completely ignoring the pathological information contained in text descriptions (such as "high fever with blurred consciousness"), leading to a disconnection between infection indicators and symptom descriptions, and such modality fragmentation increases the misdiagnosis rate of elderly patients.

[0004] The generated values of models such as GPT-4 often exceed the reasonable range of medicine, and require secondary verification by humans, increasing the clinical workload; the text and numerical processing processes are separated, and a quantitative mapping relationship between "pulmonary wet rales" and oxygenation index cannot be established, leading to a cross-modal semantic break; the full tuning of a large number of parameter models requires the support of supercomputing clusters, and the energy cost of a single interpolation is more than ten times that of traditional methods, which is completely inconsistent with the needs of hospital localization, and the inference delay is more than 500 ms, which is difficult to meet the real-time requirements of emergency scenarios. SUMMARY

[0005] The main purpose of the embodiments of the present application is to provide a prevention and control auxiliary diagnosis and treatment method, device, equipment and medium based on interpolation and enhancement, which improves the timeliness and accuracy of elderly prevention and control auxiliary diagnosis and treatment.

[0006] One aspect of the present application provides a prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement, comprising:

[0007] Obtaining user case information, performing interpolation and enhancement processing on the user case information using a first model to obtain user case text and user case data;

[0008] Performing grouping diagnosis and disease risk prediction processing on the user case text and the user case data using a second model to obtain a prediction result;

[0009] The training step of the first model comprises:

[0010] Adopt a model architecture pre-trained on medical corpus and fine-tune it using adversarial learning;

[0011] Obtaining a first data set, determining multimodal associations using a Transformer encoding layer and a self-attention mechanism, and performing model training and model fine-tuning using the first data set and the multimodal associations to obtain the first model;

[0012] The training step of the second model includes:

[0013] Obtain relevant features of the original features for group diagnosis and treatment and disease risk prediction;

[0014] Performing concatenation and dimensionality reduction processing on the relevant features to obtain a joint feature vector;

[0015] Model training is performed based on the second data set and the joint feature vector to obtain a second model.

[0016] According to the interpolation and augmentation-based prevention and control auxiliary diagnosis and treatment method, a model architecture pre-trained on medical corpus is used and fine-tuned using adversarial learning, including:

[0017] The self-attention matrix of NVIDIA BioMegatron-3.8B pre-trained on medical corpus is fine-tuned through LoRA and Adapter. Medical constraints are injected after the self-attention matrix is ​​inserted into each layer of Transformer. The medical constraint injection is achieved through the total loss function. for:

[0018]

[0019] in, As the basic task loss, For dynamic weight adjustment, represents the dynamic penalty weight, represents the loss of medical pathological constraints, represents adversarial training, To counter the loss weight;

[0020] Medical constraints based on value domain boundaries and pathological logic are enforced on the model architecture, where the value domain boundaries are used to characterize the indicator range and the pathological logic is used to characterize the causal relationship of the text description.

[0021] According to the interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method, a first data set is obtained, a Transformer encoding layer and a self-attention mechanism are used to determine multimodal associations, and model training and model fine-tuning are performed using the first data set and the multimodal associations to obtain the first model, including:

[0022] The PGD algorithm is used to perform perturbations along the loss gradient direction, and elderly-specific perturbations are performed according to the numerical type;

[0023] The Transformer encoding layer uses multi-head self-attention to fine-tune based on standard attention, where the standard attention for:

[0024]

[0025] Among them, Q is the Query matrix, that is, the query vector; K is the Key matrix, that is, the key vector; V is the Value matrix, that is, the value matrix; represents the dot product of the query vector and the key vector, represents the dimension of the key vector;

[0026] Among them, LoRA fine-tuning only adjusts the Query matrix, Key matrix and Value matrix;

[0027] Through the two-layer fully connected feedforward network, the GELU activated network is inserted into the adapter to perform model enhancement processing;

[0028] Using the BioWordPiece word segmenter to perform word segmentation, data cleaning, and truncation on the first dataset, wherein the data cleaning retains medical terminology;

[0029] Normalize the mean and row Z-score of the processed first dataset, add binary mask features to missing fields, and add random masked numerical fields or text fragments to the first dataset to obtain adversarial samples;

[0030] The adversarial sample and the first data set are taken as input, and adversarial training is performed according to preset training parameters to obtain a first model.

[0031] According to the interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method, the features related to group diagnosis and treatment and disease risk prediction in the original features are obtained, including:

[0032] A one-way ANOVA was performed on the original features, and the features related to group diagnosis and treatment and disease risk prediction were retained. The feature importance of the relevant features was calculated based on the Gini impurity using random forest importance, and the relevant features whose feature importance was greater than the preset value were retained.

[0033] According to the interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method, the relevant features are spliced ​​and dimensionally reduced to obtain a joint feature vector, including:

[0034] The relevant features are processed through a large model to obtain CLS vectors, and the CLS vectors are subjected to principal component analysis and dimensionality reduction to obtain text features;

[0035] The original interpolation values ​​of the related features in the dimensionality reduction process are retained, and the numerical features are enhanced by calculating the interpolation offset to obtain the numerical features;

[0036] The text features and the numerical features are concatenated to obtain the joint feature vector.

[0037] According to the interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method, model training is performed based on the second data set and the joint feature vector to obtain a second model, including:

[0038] Based on the second data set and the joint feature vector as input, a model evaluation is performed according to a five-fold cross validation of preset training parameters to obtain a model evaluation result;

[0039] Based on the model evaluation results, Bayesian search was used to optimize hyperparameters with AUC as the optimization target to obtain the second model with the optimal parameter combination;

[0040] The SHAP value is used to evaluate the feature contribution of the second model, and features exceeding the preset importance are retained;

[0041] The XGBoost model configuration of the second model is determined based on the model evaluation results, optimal parameter combination, feature contribution and feature importance, and verification processing of group diagnosis and treatment and disease risk prediction is performed on the second model.

[0042] According to the interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method, the method further includes:

[0043] Perform data normalization on user case information, obtain the group diagnosis and treatment and disease risk prediction results obtained after processing by the first model and the second model, perform interpolation time, confidence distribution and abnormal alarm monitoring on the group diagnosis and treatment and disease risk prediction results, and review if there is an abnormal alarm; perform iterative update on the first model and the second model according to the interpolation time and confidence distribution.

[0044] Another aspect of an embodiment of the present invention provides a prevention and control auxiliary diagnosis and treatment device based on interpolation and enhancement, comprising:

[0045] The first module is used to obtain user case information, and perform interpolation and enhancement processing on the user case information using a first model to obtain user case text and user case data;

[0046] The second module is used to perform group diagnosis and treatment and disease risk prediction processing on the user case text and user case data using the second model to obtain a prediction result;

[0047] The training of the first model includes:

[0048] The third module is used to fine-tune the model architecture pre-trained on medical corpus using adversarial learning;

[0049] A fourth module is configured to obtain a first data set, determine multimodal associations using a Transformer encoding layer and a self-attention mechanism, and perform model training and fine-tuning using the first data set and the multimodal associations to obtain the first model;

[0050] The training step of the second model includes:

[0051] The fifth module is used to obtain relevant features of the original features for group diagnosis and treatment and disease risk prediction;

[0052] The sixth module is used to perform concatenation and dimensionality reduction processing on the relevant features to obtain a joint feature vector;

[0053] The seventh module is used to perform model training based on the second data set and the joint feature vector to obtain a second model.

[0054] Another aspect of an embodiment of the present invention provides an electronic device, including a processor and a memory;

[0055] The memory is used to store programs;

[0056] The processor executes the program to implement the method described above.

[0057] Embodiments of the present invention further disclose a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the method described above.

[0058] The beneficial effects of the present invention are: learning through the efficient fine-tuning strategy of LoRA+Adapter parameters and medical rule constraints, which improves the accuracy and clinical rationality of data interpolation; the innovative advantages of downstream model access and training, the feature fusion and XGBoost optimization process designed by the present invention, achieve a balance between high-precision graded prediction and clinical interpretability; use multimodal joint decision-making, fuse text semantic features and numerical offset features (Δ value), overcome the limitations of a single data source, and significantly improve the AUC of severe infection grouping compared with traditional models; at the same time, realize real-time response and fault-tolerant mechanism, such as when the interpolation value exceeds the medical range, automatically review and synchronously return the traditional and large model results, and significantly reduce the risk of misoperation; improve the timeliness and accuracy of auxiliary diagnosis and treatment for elderly prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0060] Figure 1 It is a flowchart of the prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement according to an embodiment of the present invention.

[0061] Figure 2 This is a flowchart of the use of the online auxiliary diagnosis and treatment platform according to an embodiment of the present invention.

[0062] Figure 3 Reference to the embodiments of the present invention Figure 3 The reference interpolation and enhancement neural network structure diagram is shown.

[0063] Figure 4 Schematic diagram of a multimodal data intelligent interpolation and enhancement system according to an embodiment of the present invention.

[0064] Figure 5 This is a flowchart of multi-task model training and use according to an embodiment of the present invention.

[0065] Figure 6 Schematic diagram of a prevention and control auxiliary diagnosis and treatment device based on interpolation and enhancement according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. In the subsequent description, suffixes such as "module," "component," or "unit" used to represent elements are used solely to facilitate the description of the present invention and have no specific meaning in themselves. Therefore, "module," "component," or "unit" may be used interchangeably. "First," "second," and the like are used solely to distinguish technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features. In this subsequent description, the consecutive numbering of method steps is for ease of review and understanding. In conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, adjusting the order of implementation of the steps does not affect the technical effects achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and should not be construed as limiting the present invention.

[0067] refer to Figure 1 ,in Figure 1 This is a flow chart of the interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method according to an embodiment of the present invention. It includes but is not limited to steps S100 to S700:

[0068] S100, obtaining user case information, performing interpolation and enhancement processing on the user case information using a first model, and obtaining user case text and user case data.

[0069] S200, performing group diagnosis and treatment and disease risk prediction processing on the user case text and user case data using the second model to obtain a prediction result.

[0070] In some embodiments, reference Figure 2 The online auxiliary diagnosis and treatment platform usage flow chart shown in the figure has the following steps:

[0071] The third part of this invention is an online auxiliary diagnosis and treatment platform based on the above two inventions. Through the excellent predictive ability of the model and simple interactive operation, after entering the patient information, doctors can directly obtain the group diagnosis and treatment results of elderly patients and the mortality risk of patients with severe infections. This greatly reduces the workload of doctors and also improves the accuracy and efficiency of diagnosis and treatment to a certain extent. It is divided into the following steps:

[0072] (1) Input data normalization

[0073] Formatted input, using a table or manual input, for example: {"Diagnosis Information": "82-year-old male, cough and dyspnea for 2 days, CT showed diffuse ground-glass opacities in both lungs", "ID": {378412434 "WBC": null, "PCT": 3.8, "PaO2 / FiO2": 180...}};

[0074] Data verification checks field integrity. If key fields (such as age and chief complaint) are missing, an exception request response is triggered.

[0075] (2) Joint interpolation of multimodal data

[0076] Perform interpolation based on traditional methods. For example, for the above missing data, random forest predicts the missing WBC=13.2×10 9 / L; large model correction, combined with the "dyspnea" and "ground glass shadow" in the text, output WBC=18.5×10 9 / L (confidence level 0.91); medical logic verification: if the interpolated value exceeds the preset range (such as WBC>50), the manual review process is triggered.

[0077] (3) Feature generation and encoding text encoding

[0078] Text encoding extracts the [CLS] vector of the text and generates 64-dimensional semantic features through PCA dimensionality reduction. Numerical encoding concatenates the interpolated value, original value, and offset into a joint numerical feature (e.g., PaO2 / FiO2=180, ΔWBC=5.3).

[0079] (4) Grading prediction and risk assessment

[0080] Group prediction: Based on the test data input into the model, predictions are made and prediction results are output. For example, for the sample data, XGBoost outputs a 94% probability of severe infection, triggering the ICU priority process. Mortality risk calculation: If the infection is severe, the mortality risk prediction is triggered, and further prediction of hospital mortality risk is made. For example, for the sample data, the model group prediction result is severe infection, triggering the mortality risk prediction, and further predicting the probability of death is 68%.

[0081] (5) Result output and interpretation

[0082] Structured output: Based on the model prediction results, the output is structured data. For the example data, the output results are: "Grouping result": "Severe infection", "Grouping probability": 0.94, "Death risk": 0.68, "Key features" (PCT>2, PaO2 / FiO2<200).

[0083] The explanatory information is output simultaneously, including interpolation information, data quality information, and abnormal information alarm. For example, for example data, the output explanatory result is: "data quality": {"interpolation data": ["WBC"], "confidence": 0.91, "whether to trigger an abnormal alarm": no}. Abnormal processing, when "whether to trigger an abnormal alarm" is "yes", the traditional interpolation value and the large model result are returned synchronously for the clinician to review;

[0084] (6) System monitoring and feedback

[0085] Real-time monitoring, recording interpolation time consumption (average 148 ms), prediction confidence distribution; feedback loop, returning actual grouping results and clinical outcomes (such as survival / death) for monthly model iteration update.

[0086] In some embodiments, the training step of the first model comprises:

[0087] S300, using a model architecture pre-trained on a medical corpus, fine-tuning using adversarial learning.

[0088] S400, obtaining a first data set, determining multi-modal association using a Transformer encoding layer and a self-attention mechanism, performing model training and model fine-tuning through the first data set and the multi-modal association, and obtaining a first model.

[0089] In some embodiments, the self-attention matrix of the NVIDIA BioMegatron-3.8B pre-trained on the medical corpus is mixed fine-tuned through LoRA and Adapter, and the self-attention matrix is inserted after each layer of the Transformer to perform medical constraint injection, wherein the medical constraint injection is realized through a total loss function, and the total loss function is:

[0090]

[0091] wherein, is the basic task loss, is the dynamic weight adjustment, represents a dynamic penalty weight, represents the loss of medical pathology constraints, represents adversarial training, is the adversarial loss weight;

[0092] The model architecture is subjected to medical constraints based on value range boundaries and medical constraints based on pathological logic, wherein the value range boundaries are used to represent index ranges, and the pathological logic is used to represent causal relationships described in text.

[0093] In some embodiments, the PGD algorithm is used to perform perturbations along the loss gradient direction, while performing elderly-specific perturbations according to the numerical type;

[0094] The Transformer encoding layer uses multi-head self-attention to fine-tune based on standard attention, where the standard attention for:

[0095]

[0096] Among them, Q is the Query matrix, that is, the query vector; K is the Key matrix, that is, the key vector; V is the Value matrix, that is, the value matrix; represents the dot product of the query vector and the key vector, represents the dimension of the key vector;

[0097] Among them, LoRA fine-tuning only adjusts the Query matrix, Key matrix, and Value matrix; through the two-layer fully connected feedforward network, the GELU activated network is used to insert the adapter to perform model enhancement processing;

[0098] The BioWordPiece word segmenter is used to perform word segmentation, data cleaning, and truncation settings on the first dataset, wherein the data cleaning retains medical terms;

[0099] Normalize the mean and row Z-score of the processed first dataset, add binary mask features to missing fields, and add random masked numerical fields or text fragments to the first dataset to obtain adversarial samples;

[0100] The adversarial sample and the first data set are taken as input, and adversarial training is performed according to preset training parameters to obtain a first model.

[0101] refer to Figure 3 The reference interpolation and enhancement neural network structure diagram shown in the figure, and the reference Figure 4 The diagram below shows a multimodal data intelligent interpolation and enhancement system. The training process of interpolation and enhancement is as follows:

[0102] (1) Select the model architecture: Use NVIDIA Bio Megatron-3.8B (pre-trained on medical corpora such as PubMed and MIMIC-III).

[0103] (2) Determine the model fine-tuning strategy: Use LoRA (Low-Rank Adaptation) + Adapter hybrid fine-tuning, only fine-tune the query_key_value matrix, rank r = 64, scaling factor α = 32, bottleneck dimension d = 128, and insert each layer of Transformer. At the same time, perform medical constraint injection, that is, add rule penalty terms to the loss function.

[0104] Among them, the total loss function is:

[0105] ;

[0106] : Basic task loss (text generation cross entropy + numerical interpolation MSE);

[0107] : Dynamic penalty weight (initial 0.1, linearly increased to 0.5);

[0108] : represents the loss of medical pathology constraints,

[0109] : Adversarial loss weight (fixed at 0.1);

[0110] : Dynamic weight adjustment;

[0111] : Adversarial training.

[0112] Based on the medical constraints of the value range boundary constraints, the key indicator range is defined based on the "Guidelines for the Diagnosis and Treatment of Infections in the Elderly" (such as WBC: 2-50×10 9 / L).

[0113] Medical constraints based on pathological logic mainly include causality: if the text description contains "sepsis", the lactate value must be ≥2.0mmol / L, and correlation: if CRP increases, procalcitonin (PCT) should also increase (ΔPCT / ΔCRP≥0.1ΔPCT / ΔCRP≥0.1).

[0114] Adversarial training, based on the PGD (Projected Gradient Descent) algorithm, perturbation direction: along the loss gradient direction, perturbation amplitude: =0.1 (numeric type), =0.05 (text word vector); based on elderly-specific perturbations, for numerical values: creatinine value ±15%, albumin ±10% (simulating physiological fluctuations in the elderly); for text types: replace "fever" → "normal body temperature" (simulating atypical symptoms).

[0115] (3) Transformer encoding layer: Models multimodal associations through the self-attention mechanism to achieve data interpolation and feature extraction. Hierarchical structure: 24-layer Transformer, each layer contains the following components: multi-head self-attention, feedforward network, layer normalization and residual connection.

[0116] The multi-head self-attention is fine-tuned based on the standard attention, and the annotated attention is:

[0117]

[0118] In some embodiments, the fine-tuning method adopts LoRA fine-tuning, and only the Query, Key, and Value matrices are adjusted.

[0119] The feedforward network is based on the standard structure of two fully connected layers (768→3072→768), and the network with GELU activation is enhanced by adapter: down projection, 768→128 dimensions, ReLU activation; up projection, 128→768 dimensions, no activation; the residual connection of the feedforward network is: .

[0120] Layer normalization and residual connection, including: Pre-LN structure, normalization before entering the attention / feedforward layer; residual connection, .

[0121] (4) Data processing: Word segmentation was performed based on the Bio Word Piece tokenizer (the vocabulary size was 30,522); data cleaning was performed to retain medical terms (regular expressions matched ICD-11 entities); truncation was set to a fixed length of 512 tokens (if insufficient, padding was performed with PAD).

[0122] (5) Model training: numerical processing, using the mean and variance from the training set for Z-score normalization, and adding binary mask features to missing fields to handle missing data; training data construction: artificially creating a 20% missing rate (randomly masking numerical fields or text fragments), and generating adversarial samples, that is, adding elderly-specific perturbations to 10% of the data (such as creatinine values ​​±15%).

[0123] (6) Determine the training parameters and fine-tune them: learning rate, set to 3e-5 (AdamW optimizer, linear warmup 1000 steps); batch size: 16 (gradient accumulation 4 steps, equivalent to batch_size=64); training rounds, 10 epochs (early stopping mechanism: termination if the validation loss does not decrease after 3 rounds); accuracy, mixed precision (FP16) + gradient clipping (max_norm=1.0).

[0124] The training steps for the second model include:

[0125] S500, obtain the features related to grouping diagnosis and treatment and disease risk prediction from the original features.

[0126] In some embodiments, one-way ANOVA is performed on the original features, features related to grouping diagnosis and treatment and disease risk prediction are retained, the importance of the related features is calculated based on Gini impurity using random forest importance, and the related features with feature importance greater than a preset value are retained.

[0127] S600, the related features are spliced and dimensionally reduced to obtain a joint feature vector.

[0128] In some embodiments, the related features are processed by a large model to obtain a CLS vector, principal component analysis is performed on the CLS vector for dimensionality reduction to obtain a text feature, the original feature interpolation value is retained for the dimensionally reduced related features, numerical feature enhancement is performed by calculating the interpolation offset to obtain a numerical feature, and the text feature and the numerical feature are spliced to obtain a joint feature vector.

[0129] S700, model training is performed according to the second data set and the joint feature vector to obtain a second model.

[0130] In some embodiments, the second data set and the joint feature vector are used as inputs, model evaluation is performed according to a preset training parameter five-fold cross-validation to obtain a model evaluation result, a Bayesian search is performed according to the model evaluation result, hyperparameter optimization is performed with AUC as the optimization target to obtain a second model with optimal parameter combination, SHAP value is used to evaluate feature contribution, and features with importance greater than a preset value are retained; the XGBoost model configuration of the second model is determined according to the model evaluation result, the optimal parameter combination, the feature contribution, and the feature importance, and verification processing of grouping diagnosis and disease risk prediction is performed on the second model.

[0131] In some embodiments, reference is made to the multi-task model training and use flowchart shown in FIG. 8, and the steps include: Figure 5

[0132] (1) Feature engineering and selection: variance analysis screening, one-way ANOVA (ANOVA) is performed on the original features, features significantly related to the target variable (grouping result and death risk) (p<0.05) are retained, random forest importance evaluation is used, feature importance is calculated based on Gini impurity, and key features with a score >0.05 are retained in the previous step.

[0133] ​(2) Feature concatenation and dimension reduction: Text feature processing, extract the [CLS] vector (768 dimensions) output by the large model, and reduce the dimension through principal component analysis (PCA); numerical feature enhancement, retain the original interpolation value, and calculate the interpolation offset (Delta = | large model interpolation value - traditional interpolation value |) as an additional feature; feature fusion, concatenate the reduced text features and numerical features to form a joint feature vector.

[0134] (3) Model training and optimization: Cross-validation, 5-fold cross-validation is used, with early stopping rounds set to 50 and maximum iteration number set to 1000; hyperparameter optimization, based on Bayesian search (50 iterations), with AUC as the optimization target, to select the optimal parameter combination; feature selection, using SHAP value to evaluate feature contribution, retaining features with importance > 0.001.

[0135] (4) XGBoost model configuration, the final XGBoost model configuration is as follows: 'objective': 'binary:logistic';'max_depth': 5, # tuning range 3-7; 'learning_rate': 0.1, # grid search 0.01-0.3;'subsample': 0.8, # row sampling ratio; 'colsample_bytree': 0.7, # column sampling ratio; 'lambda': 1.5, # L2 regularization coefficient; 'gamma': 0.2 # minimum gain for splitting.

[0136] (5) Model testing and verification: Evaluation indicators include severe infection group AUC and death risk prediction F1-score (severe group), reference Table 1 for model performance comparison, comparison experiments show that compared with traditional logistic regression and random forest model, the validation AUC is significantly improved (p < 0.05).

[0137] Table 1 Model performance comparison (wherein, test set N = 307, severe infection group accounts for 30%)

[0138]

[0139] Figure 6 is the schematic diagram of the auxiliary diagnosis and treatment method and device based on interpolation and enhancement of the embodiment of the application. The device includes a first module 610, a second module 620, a third module 630, a fourth module 640, a fifth module 650, a sixth module 660, and a seventh module 670.

[0140] Among them, the first module is used to obtain user case information, and use the first model to perform interpolation and enhancement processing on the user case information to obtain user case text and user case data; the second module is used to use the second model to perform group diagnosis and treatment and disease risk prediction processing on the user case text and user case data to obtain prediction results; the training of the first model includes: the third module is used to use the model architecture pre-trained on medical corpus, and use adversarial learning for fine-tuning; the fourth module is used to obtain the first data set, use the Transformer encoding layer and self-attention mechanism to determine the multimodal association, and perform model training and model fine-tuning through the first data set and multimodal association to obtain the first model; the training of the second model includes: the fifth module is used to obtain relevant features of the original features that are related to group diagnosis and treatment and disease risk prediction; the sixth module is used to splice and reduce the dimensionality of the relevant features to obtain a joint feature vector; the seventh module is used to perform model training based on the second data set and the joint feature vector to obtain the second model.

[0141] Illustratively, with the cooperation of the first to seventh modules in the device, the embodiment device can implement any of the aforementioned prevention and control auxiliary diagnosis and treatment methods based on interpolation and enhancement, that is, obtaining user case information, using the first model to perform interpolation and enhancement processing on the user case information to obtain user case text and user case data; using the second model to perform group diagnosis and treatment and disease risk prediction processing on the user case text and user case data to obtain prediction results, wherein the first model and the second model are both obtained through training. The beneficial effects of the present invention are: through the efficient fine-tuning strategy of LoRA+Adapter parameters and the medical rule constraint mechanism, the accuracy and clinical rationality of data interpolation are improved; the innovative advantages of downstream model access and training, the feature fusion and XGBoost optimization process designed by the present invention, achieve a balance between high-precision graded prediction and clinical interpretability; use multimodal joint decision-making, fuse text semantic features and numerical offset features (Δ value), overcome the limitations of a single data source, and significantly improve the AUC of severe infection grouping compared with traditional models; at the same time, realize real-time response and fault-tolerant mechanism, such as when the interpolation value exceeds the medical range, automatically review and synchronously return the traditional and large model results, and significantly reduce the risk of misoperation; improve the timeliness and accuracy of auxiliary diagnosis and treatment for elderly prevention and control.

[0142] An embodiment of the present invention further provides an electronic device, the electronic device including a processor and a memory;

[0143] The memory stores a program;

[0144] The processor executes a program to perform the aforementioned interpolation and enhancement-based prevention, control, auxiliary diagnosis and treatment method; the electronic device has the function of carrying and running the software system based on interpolation and enhancement of prevention, control, auxiliary diagnosis and treatment provided by an embodiment of the present invention, such as a personal computer, a minicomputer, a main frame, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or communicating with a charged particle tool or other imaging device, etc.

[0145] An embodiment of the present invention also provides a computer-readable storage medium, which stores a program, and the program is executed by a processor to implement the prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement as described above.

[0146] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0147] Embodiments of the present invention further disclose a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned interpolation- and enhancement-based prevention and control auxiliary diagnosis and treatment method.

[0148] Furthermore, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise indicated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It will also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art using ordinary skill will be able to implement the present invention set forth in the claims without undue experimentation. It will also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0149] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0150] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0151] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting, or processing it in another suitable manner as necessary, and then storing it in a computer memory.

[0152] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0153] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0154] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0155] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement, characterized in that: include: Acquire user case information, and perform interpolation and enhancement processing on the user case information using a first model to obtain user case text and user case data; The second model is used to perform group diagnosis and treatment and disease risk prediction processing on the user case text and user case data to obtain a prediction result; The training steps of the first model include: Adopt a model architecture pre-trained on medical corpus and fine-tune it using adversarial learning; Obtaining a first data set, determining multimodal associations using a Transformer encoding layer and a self-attention mechanism, and performing model training and model fine-tuning using the first data set and the multimodal associations to obtain the first model; The training step of the second model includes: Obtain relevant features of the original features for group diagnosis and treatment and disease risk prediction; Performing concatenation and dimensionality reduction processing on the relevant features to obtain a joint feature vector; Model training is performed based on the second data set and the joint feature vector to obtain a second model.

2. The prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement according to claim 1 is characterized in that: The model architecture pre-trained on medical corpus is fine-tuned using adversarial learning, including: The self-attention matrix of NVIDIA BioMegatron-3.8B pre-trained on medical corpus is fine-tuned through LoRA and Adapter. The self-attention matrix is ​​inserted into each layer of Transformer and medical constraints are injected. The medical constraint injection is achieved through the total loss function. for: in, As the basic task loss, For dynamic weight adjustment, represents the dynamic penalty weight, represents the loss of medical pathological constraints, represents adversarial training, To counter the loss weight; Medical constraints based on value domain boundaries and pathological logic are enforced on the model architecture, where the value domain boundaries are used to characterize the indicator range and the pathological logic is used to characterize the causal relationship of the text description.

3. The interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method according to claim 2, characterized in that: The first data set is obtained, a Transformer encoding layer and a self-attention mechanism are used to determine multimodal associations, and model training and model fine-tuning are performed using the first data set and the multimodal associations to obtain the first model, including: The PGD algorithm is used to perform perturbations along the loss gradient direction, and elderly-specific perturbations are performed according to the numerical type; The Transformer encoding layer uses multi-head self-attention to fine-tune based on standard attention, where the standard attention for: Among them, Q is the Query matrix, that is, the query vector; K is the Key matrix, that is, the key vector; V is the Value matrix, that is, the value matrix; represents the dot product of the query vector and the key vector, represents the dimension of the key vector; Among them, LoRA fine-tuning only adjusts the Query matrix, Key matrix and Value matrix; Through the two-layer fully connected feedforward network, the GELU activated network is inserted into the adapter to perform model enhancement processing; Using the BioWordPiece word segmenter to perform word segmentation, data cleaning, and truncation settings on the first dataset, wherein data cleaning retains medical terms; Normalize the mean and row Z-score of the processed first dataset, add binary mask features to missing fields, and add random masked numerical fields or text fragments to the first dataset to obtain adversarial samples; The adversarial sample and the first data set are taken as input, and adversarial training is performed according to preset training parameters to obtain a first model.

4. The prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement according to claim 1 is characterized in that: The features related to group diagnosis and treatment and disease risk prediction obtained from the original features include: A one-way ANOVA was performed on the original features, and the features related to group diagnosis and treatment and disease risk prediction were retained. The feature importance of the relevant features was calculated based on the Gini impurity using random forest importance, and the relevant features whose feature importance was greater than the preset value were retained.

5. The prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement according to claim 4 is characterized in that: The concatenation and dimensionality reduction processing of the related features to obtain a joint feature vector includes: The relevant features are processed through a large model to obtain CLS vectors, and the CLS vectors are subjected to principal component analysis and dimensionality reduction to obtain text features; The original interpolation values ​​of the related features in the dimensionality reduction process are retained, and the numerical features are enhanced by calculating the interpolation offset to obtain the numerical features; The text features and the numerical features are concatenated to obtain the joint feature vector.

6. The method for auxiliary diagnosis and treatment based on interpolation and enhancement according to claim 1 is characterized in that: The performing model training according to the second data set and the joint feature vector to obtain a second model includes: Based on the second data set and the joint feature vector as input, a model evaluation is performed according to a five-fold cross validation of preset training parameters to obtain a model evaluation result; Based on the model evaluation results, Bayesian search was used to optimize hyperparameters with AUC as the optimization target to obtain the second model with the optimal parameter combination; The SHAP value is used to evaluate the feature contribution of the second model, and features exceeding the preset importance are retained; The XGBoost model configuration of the second model is determined based on the model evaluation results, optimal parameter combination, feature contribution and feature importance, and verification processing of group diagnosis and treatment and disease risk prediction is performed on the second model.

7. The interpolation and enhancement-based prevention and control auxiliary diagnosis and treatment method according to claim 1, characterized in that: The method further comprises: Perform data normalization on user case information, obtain the group diagnosis and treatment and disease risk prediction results obtained after processing by the first model and the second model, perform interpolation time, confidence distribution and abnormal alarm monitoring on the group diagnosis and treatment and disease risk prediction results, and review if there is an abnormal alarm; perform iterative update on the first model and the second model according to the interpolation time and confidence distribution.

8. A prevention and control auxiliary diagnosis and treatment device based on interpolation and enhancement, characterized in that: include: The first module is used to obtain user case information, and perform interpolation and enhancement processing on the user case information using a first model to obtain user case text and user case data; The second module is used to perform group diagnosis and treatment and disease risk prediction processing on the user case text and user case data using the second model to obtain a prediction result; The training of the first model includes: The third module is used to fine-tune the model architecture pre-trained on medical corpus using adversarial learning; A fourth module is configured to obtain a first data set, determine multimodal associations using a Transformer encoding layer and a self-attention mechanism, and perform model training and fine-tuning using the first data set and the multimodal associations to obtain the first model; The training step of the second model includes: The fifth module is used to obtain relevant features of the original features for group diagnosis and treatment and disease risk prediction; The sixth module is used to perform concatenation and dimensionality reduction processing on the relevant features to obtain a joint feature vector; The seventh module is used to perform model training based on the second data set and the joint feature vector to obtain a second model.

9. An electronic device, characterized in that: including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The storage medium stores a program, and the program is executed by a processor to implement the prevention and control auxiliary diagnosis and treatment method based on interpolation and enhancement as described in any one of claims 1 to 7.

Images