Disease prognosis evaluation and dynamic prediction method based on time sequence analysis

Abstract

The invention discloses a disease prognosis evaluation and dynamic prediction method based on time sequence analysis, which comprises the following steps: acquiring disease prognosis medical data of a patient to construct a medical data time sequence set, extracting medical data time sequence characteristics and constructing a prognosis evaluation model, updating the prognosis evaluation model, and performing dynamic prediction. Inputting the to-be-evaluated patient disease prognosis medical data into the updated prognosis evaluation model to obtain a disease prognosis evaluation result, constructing a prognosis prediction model, inputting the to-be-evaluated patient disease prognosis medical data into the prognosis prediction model to obtain a disease prognosis prediction result, and updating the disease prognosis prediction result. And generating evaluation interpretation, prediction interpretation, a disease prognosis evaluation chart and a disease prediction trend chart. According to the method, the efficiency and accuracy of disease prognosis evaluation and prediction can be improved, timely and effective decision support is provided for clinic, the requirements of personalized medical treatment are met, and more accurate treatment schemes and rehabilitation guidance are provided for patients.

Description

Technical Field

[0001] This invention relates to the fields of medical data analysis and artificial intelligence technology, and in particular to a method for disease prognosis assessment and dynamic prediction based on time series analysis. Background Technology

[0002] In the medical field, disease prognostic assessment and dynamic prediction are crucial for developing personalized treatment plans, rationally allocating medical resources, and managing patient rehabilitation. Traditional methods are mostly based on static clinical data and rely on statistical models or expert experience to establish assessment criteria. Therefore, these methods have significant limitations: First, disease development is a dynamic process, and traditional static methods cannot capture this dynamic information, making it difficult to accurately reflect the true development trend and prognosis of the disease. Second, traditional methods lack in-depth analysis of the time-series characteristics of the data, resulting in insufficient predictive accuracy and significant bias, failing to provide a reliable basis for clinical decision-making. Finally, the disease progression process varies among different patients, and uniform assessment criteria cannot meet the needs of personalized medicine, failing to dynamically adjust and accurately assess the characteristics of different patients.

[0003] In recent years, with the development of medical informatization, a large amount of medical data has been recorded and stored, making data-driven disease prognosis assessment possible. Therefore, this invention proposes a disease prognosis assessment and dynamic prediction method based on time series analysis. This method collects patient disease prognosis medical data from multiple sources, extracts the time-series features of the medical data using a hybrid feature extraction method, constructs a prognosis assessment model and a prognosis prediction model, and updates the models. This achieves accurate dynamic assessment and prediction of disease prognosis, making it easier for doctors to understand and apply, meeting the needs of personalized medicine, and providing patients with more precise treatment plans and rehabilitation guidance. Summary of the Invention

[0004] The purpose of this invention is to provide a method for disease prognosis assessment and dynamic prediction based on time series analysis.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention provides a disease prognostic assessment method based on time series analysis, comprising the following steps:

[0007] Medical data on patient disease prognosis is collected from multiple sources, preprocessed, and a time series set of medical data is constructed according to timestamps; the multiple sources include hospital electronic medical record databases, laboratory examination databases, and monitoring equipment data platforms.

[0008] A hybrid feature extraction method is used to extract time-series features from medical data, and a prognostic assessment model is constructed based on the time-series features of historical medical data. The hybrid feature extraction method includes statistical feature extraction, deep learning feature extraction, and time-frequency feature analysis. The time-series features of the medical data include statistical time-series features, deep time-series features, and time-frequency features.

[0009] The model update index is calculated based on the newly collected patient disease prognosis medical data. The update trigger conditions are determined based on the model update index to update the model. The disease prognosis medical data of the patient to be evaluated is input into the updated prognosis evaluation model to obtain the disease prognosis evaluation result. The trigger conditions include quantitative triggering and performance triggering.

[0010] An assessment interpretation and a disease prognosis assessment graph are generated based on the disease prognosis assessment results; the assessment interpretation reflects the contribution of the time-series characteristics of the medical data; the disease prognosis assessment graph reflects the health status and reliability of each assessment indicator.

[0011] Furthermore, the method for extracting time-series features from medical data includes:

[0012] The time-series characteristics of the medical data include statistical time-series characteristics, deep time-series characteristics, and time-frequency characteristics;

[0013] Statistical time-series characteristics are obtained by statistically analyzing medical data on patient disease prognosis; the statistical time-series characteristics include the mean, variance, standard deviation, autocorrelation coefficient, and partial autocorrelation coefficient of the data sequence;

[0014] Wavelet transform was used to perform time-frequency feature analysis on patient disease prognosis medical data to obtain time-frequency features; the time-frequency features are feature representations under different time scales and frequency domains.

[0015] A deep learning dataset was constructed using historical patient disease prognosis medical data and corresponding deep temporal features, and randomly divided into a first training set and a first test set in a 7:3 ratio. The first training set was used to train a deep learning-based temporal feature extraction model, and the first test set was used to evaluate the performance of the deep learning-based temporal feature extraction model.

[0016] The deep learning-based temporal feature extraction model includes an input layer, a long and short extraction layer, a feature fusion layer, and an output layer.

[0017] The long and short extraction layers include parallel LSTM and GRU branches. The LSTM branch is used to capture long-term dependencies in patient disease prognosis medical data to generate long-term time-series features, while the GRU branch is used to learn short-term patterns in patient disease prognosis medical data to generate short-term time-series features.

[0018] The feature fusion layer concatenates long-term and short-term temporal features to generate deep temporal features;

[0019] The deep learning-based temporal feature extraction model employs a multi-scale temporal reconstruction loss optimization model, the expression of which is:

[0020]

[0021] in For multi-scale temporal reconstruction loss, α1, α2, and α3 are the local reconstruction weight, frequency domain weight, and trend distribution weight, respectively. X t:t+k Given the original input sequence X from time t to t+k, The reconstructed sequence from time t to t+k That is, deep temporal features, DWT(·) is the discrete wavelet transform, KL(p trend ‖q trend To measure the trend distribution of real data p trend With the model predicting the trend distribution q trend KL divergence of the difference.

[0022] Furthermore, the method for constructing the prognostic assessment model includes:

[0023] An evaluation dataset is formed based on the time-series characteristics of historical medical data and the corresponding disease prognosis assessment results. This dataset is then randomly divided into a second training set and a second test set in a 6:4 ratio. The prognosis assessment model is trained using the second training set, and its performance is evaluated using the second test set. The disease prognosis assessment results include prognosis category and event probability.

[0024] The prognostic assessment model includes an input layer, a base model layer, a feature fusion layer, and an output layer;

[0025] The base model layer includes parallel LSTM base models, random forest base models, and linear regression base models, which output disease prognosis branch assessment results through the base models;

[0026] The feature fusion layer uses a temporal attention mechanism to fuse the disease prognostic branch assessment results output by the base model layer to obtain the disease prognostic assessment result, expressed as:

[0027]

[0028] Where A t Let W be the attention weight matrix at time step t. q For querying the matrix, W k The key matrix, d represents the predicted probability output of the LSTM at time step t. k The scaling dimension of the key vector. Let be the original predicted output of base model i, and MLP(·) be a multilayer perceptron. Let be the learnable weight matrix of base model i, where i∈{lstm,rf,lr} represents the index set of the LSTM base model, random forest base model, and linear regression base model, and o1(t)∈[0.1]. c The result represents the prognostic assessment of the disease, where [0.1] represents the probability range and c represents the prognostic category, including deterioration, stability, and recovery.

[0029] Furthermore, the method for updating the model includes:

[0030] The model update metrics include the amount of new data and the prediction confidence level;

[0031] The amount of new data is determined directly by statistics, and quantitative triggering is determined when the amount of new data exceeds the threshold of data increase.

[0032] The current prediction error is calculated using cross-entropy loss. If the current prediction error exceeds the error threshold, it is determined to be a performance trigger.

[0033] The model is updated when any update trigger condition is met, as expressed in the following expression:

[0034]

[0035] Where Θ new For the updated model parameters, Θ old These are the model parameters before the update, and η is the learning rate. For the gradient operator with respect to parameter Θ, For the new dataset The loss function on, μ is the regularization coefficient, ||Θ-Θ old || 2 The L2 distance between the old and new parameters.

[0036] On the other hand, a dynamic prediction method for disease prognosis based on time series analysis includes the following steps:

[0037] A prognostic prediction model is constructed based on historical time series characteristics, historical assessment results, and historical development results.

[0038] The medical data of the patient's disease prognosis to be evaluated are input into the prognosis prediction model to obtain the disease prognosis prediction result; the disease prognosis prediction result includes the disease prognosis development category and development probability at different future time points;

[0039] Define a data sliding window and dynamically update the disease prognosis prediction results based on the time series set of disease prognosis prediction results within the window;

[0040] Based on the disease prognosis prediction results, a prediction interpretation and a disease prediction trend map are generated; the prediction interpretation reflects the contribution of the time-series characteristics of the medical data; the disease prediction trend map reflects the development trend and reliability of the disease prognosis.

[0041] Furthermore, the method for constructing the prognostic prediction model includes:

[0042] The prediction dataset is composed of historical time-series features, historical assessment results, and historical development results, and is randomly divided into a third training set and a third test set in a 6:4 ratio. The prognostic prediction model is trained using the third training set, and the performance of the prognostic assessment model is evaluated using the third test set. The historical development results include the disease prognostic development categories and development probabilities at different future time points.

[0043] The prognostic prediction model includes an input layer, a base model layer, a feature fusion layer, and an output layer;

[0044] The base model layer includes a parallel BiLSTM base model, a Cox proportional hazards model, and an XGBoost base model. The BiLSTM base model outputs disease prognosis branch prediction results through the base model.

[0045] The feature fusion layer embeds temporal attention units and model attention units. The temporal attention units generate weights for each time step, and the model attention units generate weights for each base model. Based on the weights for each time step and the weights for each base model, the disease prognosis branch prediction results are cross-fused to obtain the disease prognosis prediction results.

[0046] Furthermore, the method for updating disease prognostic prediction results includes:

[0047] Acquire new time-series patient disease prognosis medical data, input it into the prognosis prediction model to obtain a time-series set of disease prognosis prediction results, determine the length of the data sliding window, and dynamically update the disease prognosis prediction results based on the historical time-series set of disease prognosis prediction results within the window. The expression is:

[0048]

[0049] Where o'2(t) is the update of the disease prognosis prediction result o2(t) at time t, and o2(t)∈[0.1] c [0.1] represents the probability range, c represents the prognostic category, including worsening and recovery, β represents the smoothing factor, and N represents the probability range. t λ is the length of the data sliding window. j Let j be the attention weight. For time point j, the local length is N. t The mean of historical disease prognostic prediction results, W a Let b be the weight matrix. a is the bias vector, and v is the attention score vector.

[0050] The beneficial effects of this invention are:

[0051] This invention is a method for disease prognosis assessment and dynamic prediction based on time series analysis. Compared with existing technologies, this invention has the following technical advantages:

[0052] Improving prediction accuracy: By deeply mining the time-series characteristics of medical data and combining multiple analysis methods and ensemble learning models, it is possible to capture the development patterns and trends of diseases more comprehensively and accurately, which significantly improves the accuracy of disease prognosis prediction compared with traditional methods.

[0053] Achieving dynamic prediction: The dynamic update module can incorporate newly generated medical data in real time, adjust and optimize the model in a timely manner, enabling the system to adapt to the dynamic changes of the disease, achieve dynamic prediction of disease prognosis, and provide more timely and effective decision support for clinical practice;

[0054] Enhanced interpretability: The results interpretation and visualization module provides detailed analysis and interpretation of the prediction results, clarifies the degree of influence of each feature on the prognosis, and presents them in an intuitive visualization manner, which improves the interpretability of the prediction results, makes it easier for doctors to understand and apply them, and promotes trust and adoption in clinical practice;

[0055] Adapting to Personalized Medicine: The system can build suitable prognostic assessment models based on the individual differences of different patients through personalized analysis of their medical data time series, thereby meeting the needs of personalized medicine and providing patients with more precise treatment plans and rehabilitation guidance. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the steps of the disease prognosis assessment and dynamic prediction method based on time series analysis of the present invention. Detailed Implementation

[0057] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0058] The disease prognosis assessment and dynamic prediction method based on time series analysis of this invention includes the following steps:

[0059] like Figure 1 As shown, this embodiment includes the following steps:

[0060] Taking the prognostic assessment and prediction of heart failure patients as an example, medical data on the prognostic status of patients are collected from the hospital's electronic medical record system, cardiovascular monitoring equipment, and laboratory testing system. This includes basic information upon admission (age, gender, medical history, etc.), vital signs data (such as daily blood pressure, heart rate, respiratory rate, blood oxygen saturation, etc.), laboratory indicator data (such as weekly brain natriuretic peptide (BNP) levels, ejection fraction (EF), complete blood count, blood biochemical indicators, etc.), symptom record data (such as the degree of dyspnea, edema, etc.), and treatment intervention data (such as medication, treatment measures, etc.).

[0061] The collected data was cleaned to remove obvious errors and duplicate records; the Z-score standardization method was used to standardize the data to make the data of different indicators comparable; for missing data, a time series-based linear interpolation method was used to imputate the missing data and construct a complete time series set of patient medical data.

[0062] For numerical time series such as vital signs data and test index data, calculate their mean, variance, standard deviation, autocorrelation coefficient, and partial autocorrelation coefficient to obtain statistical time series characteristics;

[0063] Wavelet transform is used to perform time-frequency analysis on the data to obtain time-frequency characteristics at different time scales and in the frequency domain;

[0064] A deep learning-based temporal feature extraction model was used to process patient disease prognosis medical data to obtain deep learning features. In the deep learning-based temporal feature extraction model, the LSTM branch layer had 3 layers, 256 hidden units, and dropout = 0.2; the GRU branch layer had 2 layers, 128 hidden units, and dropout = 0.1; the fully connected layer of the feature fusion layer had a dimension of 192; the local reconstruction weight α1 / frequency domain weight α2 / trend distribution weight α3 were 0.4 / 0.4 / 0.2; and the DWT used the db4 wavelet basis with 3 decomposition layers.

[0065] A prognostic assessment model is constructed based on the temporal characteristics of historical medical data and the corresponding disease prognostic evaluation results. In the prognostic assessment model, the LSTM base model has a time step of 24, 128 hidden units, a weighted cross-entropy loss function, and the Adam optimizer. The random forest base model has 200 trees and a maximum depth of 10. The linear regression base model has an L2 regularization coefficient of 0.01. The scaling dimension d of the key vector in the feature fusion layer is... k The value is 64, and the MLP hidden layer is [32, 16].

[0066] Taking the prognostic assessment model update determination at a certain moment as an example, the new data volume is 62 cases / historical data is 1000 cases, and the new data volume ratio is 6.2%, which is greater than the data increase threshold of 5%, and it is determined to be quantitative trigger; the prediction error (MAE) is 0.18, which exceeds the error threshold of 0.15, and it is determined to be performance trigger; both types of update trigger conditions are triggered at the same time to update the prognostic assessment model, where the learning rate η is 0.0001 and the regularization coefficient μ is 0.2.

[0067] The disease prognosis medical data of the patients to be evaluated were input into the updated prognosis assessment model to obtain the disease prognosis assessment results: recovery probability 85%, deterioration probability 43%, confidence level 0.82 (high confidence level), and in-window prediction standard deviation 0.05 (recovery probability 85%, deterioration probability 43%). The recovery probability 85% - deterioration probability 43% = 42% > 30%, which is a good prognosis.

[0068] The prognostic assessment model's evaluation results are analyzed. By calculating the Shapley value of each feature (AUC-ROC / deterioration class, C-index / survival analysis, MAE / critical events, physician approval), the contribution of each feature to the prediction results is determined, and a detailed explanatory report is generated. Charting tools are used to plot the curve of the patient's heart failure prognosis probability over time, as well as trend graphs of key indicators (such as BNP level, EF value, etc.), to present them to doctors and patients in an intuitive way.

[0069] A prognostic prediction model is constructed based on historical time-series characteristics, historical assessment results, and historical development results. In the prognostic prediction model, the BiLSTM base model has 3 layers, 128 hidden units, dropout = 0.3, and uses time-weighted cross-entropy and Adam optimizers. The Cox model has a regularization coefficient of 0.5, a maximum number of iterations of 500, and uses partial likelihood loss and L-BFGS optimizers. The XGBoost model has 300 trees, a maximum depth of 7, and a learning rate of 0.1. In the feature fusion layer, the temporal attention unit generates weights for each time step and weights the disease prognostic branch prediction results output by the BiLSTM base model to obtain the time-weighted disease prognostic prediction result. The model attention unit generates base model weights for Cox and XGBoost and weights the corresponding disease prognostic branch prediction results output by the base models to obtain the model-weighted disease prognostic prediction result. The time-weighted disease prognostic prediction result and the model-weighted disease prognostic prediction result are cross-fused to obtain the final disease prognostic prediction result.

[0070] Acquire new time-series patient disease prognosis medical data, input them into the prognostic prediction model to obtain a time-series set of disease prognosis prediction results, determine the data sliding window length as 5, update the disease prognosis prediction results, where the smoothing factor β is 0.4, and the weight matrix W... a Bias vector b aThe attention score vector v is fixed through pre-training;

[0071] The medical data of the patient's disease prognosis to be evaluated are input into the prognosis prediction model to obtain the disease prognosis prediction results, and the disease prognosis prediction results are updated (predicting the disease prognosis after 1 / 3 / 6 months): 28% probability of deterioration / 68% probability of recovery after 1 month, 15% probability of deterioration / 79% probability of recovery after 3 months, and 28% probability of deterioration / 68% probability of recovery after 6 months.

[0072] The evaluation results of the prognostic prediction model were analyzed. By calculating the Shapley value of each feature (AUC-ROC / deterioration class, C-index / survival analysis, MAE / critical events, physician approval), the contribution of each feature to the prediction results was determined, and a detailed explanatory report was generated. Charting tools were used to draw the trend of the patient's prognostic probability over time and the confidence trend.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A disease prognostic assessment method based on time series analysis, characterized in that, Includes the following steps: Medical data on patient disease prognosis is collected from multiple sources, preprocessed, and a time series set of medical data is constructed according to timestamps; the multiple sources include hospital electronic medical record databases, laboratory examination databases, and monitoring equipment data platforms. A hybrid feature extraction method is used to extract time-series features from medical data, and a prognostic assessment model is constructed based on the time-series features of historical medical data. The hybrid feature extraction method includes statistical feature extraction, deep learning feature extraction, and time-frequency feature analysis. The time-series features of the medical data include statistical time-series features, deep time-series features, and time-frequency features. The model update index is calculated based on the newly collected patient disease prognosis medical data. The update trigger conditions are determined based on the model update index to update the model. The disease prognosis medical data of the patient to be evaluated is input into the updated prognosis evaluation model to obtain the disease prognosis evaluation result. The trigger conditions include quantitative triggering and performance triggering. An assessment interpretation and a disease prognosis assessment graph are generated based on the disease prognosis assessment results; the assessment interpretation reflects the contribution of the time-series characteristics of the medical data; the disease prognosis assessment graph reflects the health status and reliability of each assessment indicator.

2. The disease prognostic assessment method based on time series analysis according to claim 1, characterized in that, The method for extracting time-series features from medical data includes: The time-series characteristics of the medical data include statistical time-series characteristics, deep time-series characteristics, and time-frequency characteristics; Statistical time-series characteristics are obtained by statistically analyzing medical data on patient disease prognosis; the statistical time-series characteristics include the mean, variance, standard deviation, autocorrelation coefficient, and partial autocorrelation coefficient of the data sequence; Wavelet transform was used to perform time-frequency feature analysis on patient disease prognosis medical data to obtain time-frequency features; the time-frequency features are feature representations under different time scales and frequency domains. A deep learning dataset was constructed using historical patient disease prognosis medical data and corresponding deep temporal features, and randomly divided into a first training set and a first test set in a 7:3 ratio. The first training set was used to train a deep learning-based temporal feature extraction model, and the first test set was used to evaluate the performance of the deep learning-based temporal feature extraction model. The deep learning-based temporal feature extraction model includes an input layer, a long and short extraction layer, a feature fusion layer, and an output layer. The long and short extraction layers include parallel LSTM and GRU branches. The LSTM branch is used to capture long-term dependencies in patient disease prognosis medical data to generate long-term time-series features, while the GRU branch is used to learn short-term patterns in patient disease prognosis medical data to generate short-term time-series features. The feature fusion layer concatenates long-term and short-term temporal features to generate deep temporal features; The deep learning-based temporal feature extraction model employs a multi-scale temporal reconstruction loss optimization model, the expression of which is: in For multi-scale temporal reconstruction loss, α1, α2, and α3 are the local reconstruction weight, frequency domain weight, and trend distribution weight, respectively. X t:t+k Given the original input sequence X from time t to t+k, The reconstructed sequence from time t to t+k That is, deep temporal features, DWT(·) is the discrete wavelet transform, KL(p trend ‖q trend To measure the trend distribution of real data p trend With the model predicting the trend distribution q trend KL divergence of the difference.

3. The disease prognostic assessment method based on time series analysis according to claim 1, characterized in that, The method for constructing the prognostic assessment model includes: An evaluation dataset is formed based on the time-series characteristics of historical medical data and the corresponding disease prognosis assessment results. This dataset is then randomly divided into a second training set and a second test set in a 6:4 ratio. The prognosis assessment model is trained using the second training set, and its performance is evaluated using the second test set. The disease prognosis assessment results include prognosis category and event probability. The prognostic assessment model includes an input layer, a base model layer, a feature fusion layer, and an output layer; The base model layer includes parallel LSTM base models, random forest base models, and linear regression base models, which output disease prognosis branch assessment results through the base models; The feature fusion layer uses a temporal attention mechanism to fuse the disease prognostic branch assessment results output by the base model layer to obtain the disease prognostic assessment result, expressed as: Where A t Let W be the attention weight matrix at time step t. q For querying the matrix, W k The key matrix, d represents the predicted probability output of the LSTM at time step t. k The scaling dimension of the key vector. Let be the original predicted output of base model i, and MLP(·) be a multilayer perceptron. Let be the learnable weight matrix of base model i, where i∈{lstm,rf,lr} represents the index set of the LSTM base model, random forest base model, and linear regression base model, and o1(t)∈[0.1]. c The result represents the prognostic assessment of the disease, where [0.1] represents the probability range and c represents the prognostic category, including deterioration, stability, and recovery.

4. The disease prognostic assessment method based on time series analysis according to claim 1, characterized in that, The method for updating the model includes: The model update metrics include the amount of new data and the prediction confidence level; The amount of new data is determined directly by statistics, and quantitative triggering is determined when the amount of new data exceeds the threshold of data increase. The current prediction error is calculated using cross-entropy loss. If the current prediction error exceeds the error threshold, it is determined to be a performance trigger. The model is updated when any update trigger condition is met, as expressed in the following expression: Where Θ new For the updated model parameters, Θ old These are the model parameters before the update, and η is the learning rate. For the gradient operator with respect to parameter Θ, For the new dataset The loss function on, μ is the regularization coefficient, ||Θ-Θ old || 2 The L2 distance between the old and new parameters.

5. A method for dynamic prediction of disease prognosis based on time series analysis, characterized in that, Includes the following steps: A prognostic prediction model is constructed based on historical time series characteristics, historical assessment results, and historical development results. The medical data of the patient's disease prognosis to be evaluated are input into the updated prognostic prediction model to obtain the disease prognosis prediction result; the disease prognosis prediction result includes the disease prognosis development category and development probability at different future time points; Define a data sliding window and dynamically update the disease prognosis prediction results based on the time series set of disease prognosis prediction results within the window; Based on the disease prognosis prediction results, a prediction interpretation and a disease prediction trend map are generated; the prediction interpretation reflects the contribution of the time-series characteristics of the medical data. The disease prediction trend chart reflects the development trend and reliability of disease prognosis.

6. The method for dynamic prediction of disease prognosis based on time series analysis according to claim 5, characterized in that, The method for constructing a prognostic prediction model includes: The prediction dataset is composed of historical time-series features, historical assessment results, and historical development results, and is randomly divided into a third training set and a third test set in a 6:4 ratio. The prognostic prediction model is trained using the third training set, and the performance of the prognostic assessment model is evaluated using the third test set. The historical development results include the disease prognostic development categories and development probabilities at different future time points. The prognostic prediction model includes an input layer, a base model layer, a feature fusion layer, and an output layer; The base model layer includes a parallel BiLSTM base model, a Cox proportional hazards model, and an XGBoost base model. The BiLSTM base model outputs disease prognosis branch prediction results through the base model. The feature fusion layer embeds temporal attention units and model attention units. The temporal attention units generate weights for each time step, and the model attention units generate weights for each base model. Based on the weights for each time step and the weights for each base model, the disease prognosis branch prediction results are cross-fused to obtain the disease prognosis prediction results.

7. The method for dynamic prediction of disease prognosis based on time series analysis according to claim 5, characterized in that, The method for updating disease prognostic prediction results includes: Acquire new time-series patient disease prognosis medical data, input it into the prognosis prediction model to obtain a time-series set of disease prognosis prediction results, determine the length of the data sliding window, and dynamically update the disease prognosis prediction results based on the historical time-series set of disease prognosis prediction results within the window. The expression is: Where o'2(t) is the update of the disease prognosis prediction result o2(t) at time t, and o2(t)∈[0.1] c [0.1] represents the probability range, c represents the prognostic category, including worsening and recovery, β represents the smoothing factor, and N represents the probability range. t λ is the length of the data sliding window. j Let j be the attention weight. For time point j, the local length is N. t The mean of historical disease prognostic prediction results, W a Let b be the weight matrix. a is the bias vector, and v is the attention score vector.

Images