Deep learning-based epidemic patient number prediction method and system

Abstract

The invention discloses a deep learning-based epidemic patient number prediction method and system. The method comprises the steps of constructing a basic module comprising an LSTM and a full connection layer; an epidemic patient number prediction model is constructed based on the basic module and is trained, and prediction of the number of epidemic patients is realized; the epidemic patient number prediction model comprises a trend module used for capturing the long-term trend of input data, and outputting a trend part and a first intermediate residual part; the season module is used for capturing a season rule of the first residual error part and outputting a season part and a second middle residual error part; the residual module receives the second middle residual part and outputs a first residual part and a second residual part; and the multi-head attention layer receives the trend part, the season part, the first residual part and the second residual part, and outputs a prediction result. According to the method, information is decomposed into three levels to capture trend features, seasonal features and abrupt change point features respectively, the prediction precision is improved through multi-mode combination, and the interpretability of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the intersection of medicine and engineering, and in particular to a method and system for predicting the number of epidemic patients based on deep learning. Background Art

[0002] In the current healthcare environment, the outbreak and spread of epidemic diseases pose great challenges to public health systems, especially in the context of limited resources. The prediction and prevention of epidemic diseases are crucial in many countries because these diseases can spread rapidly, affect a large range of populations, and have a significant impact on society and the economy. Over time, the development of medical information has brought increasing complexity to epidemiological research and unpredictability in resource allocation. In the current healthcare environment, diverse clinical information represented by a large amount of electronic medical records and public health data forces public health experts and clinicians to make decisions based on the evaluation of large amounts of data. It has been proven that deep learning can successfully solve complex intelligent tasks by effectively processing complex input data and intelligently learning different knowledge. In practice, deep learning is an efficient feature extraction method that can capture more abstract features and achieve a more essential description of big data. In this case, deep learning can efficiently capture characteristics such as data features, while capturing structural changes in previous data and the occurrence of mutation points, which is helpful for the prediction of epidemic diseases, the formulation of preventive measures, and the rational allocation of public health resources. The Chinese invention application with publication number CN117540336A discloses a time series prediction method, which extracts abnormal variable data and seasonal variable data, trend variable data, and residual variable data from the time series data by performing multivariate decomposition on historical time series data, and then inputs the extracted abnormal variable data and seasonal variable data, trend variable data, and residual variable data together with the historical time series data into a pre-trained time series prediction model, and uses the time series prediction model to predict the time series prediction result of the target prediction time. However, the multiplicative decomposition method used in this method will face complex and huge computational complexity when applied to the prediction of the number of epidemic patients, and the attention mechanism used in this method is mainly used to process the data of abnormal time steps, focusing on the changes of abnormal time points. This localized processing method may cause the following problems in the model's ability to capture global trends and long-term dependencies when applied to the prediction of the number of epidemic patients: the model focuses too much on abnormal time points and may ignore the information in normal time steps, resulting in a decrease in the accuracy of the overall prediction; the attention mechanism fails to specifically handle seasonal characteristics and cannot fully utilize the periodic changes in the data, which is an important defect for the seasonal fluctuations common in epidemic prediction; when the data has long-term characteristics, the model may be biased towards short-term abnormal changes and ignore long-term stable trends and patterns. This limitation affects the model's ability to capture global trends and long-term dependencies, thereby limiting its application in epidemic data analysis. Summary of the invention

[0003] The purpose of the present invention is to solve the problems in the prior art.

[0004] The technical solution adopted by the present invention to solve the technical problem is: to provide a method for predicting the number of epidemic patients based on deep learning, comprising the following steps:

[0005] Build a basic module including a long short-term memory network and two fully connected layers;

[0006] Build and train an epidemic patient quantity prediction model based on the basic module;

[0007] Use the trained epidemic patient number prediction model to predict the number of epidemic patients;

[0008] The epidemic patient quantity prediction model includes a trend module, a season module, a residual module and a multi-head attention layer which are connected in sequence; the trend module includes a basic module and a polynomial function, which is used to capture the long-term trend of the input data, output the trend part, and use the difference between the trend part and the input data as the first intermediate residual part to capture the seasonal cycle characteristics in the next step; the season module includes a basic module and a periodic function, which is used to capture the seasonal law of the first intermediate residual part, output the season part, and use the difference between the season part and the first intermediate residual part as the second intermediate residual part to capture the mutation point characteristics in the next step; the residual module includes two basic modules, which receive the second intermediate residual part output by the season module, capture the features of different levels such as mutation points, and output the first residual part and the second residual part; the multi-head attention layer receives the trend part, the season part, the first residual part and the second residual part, strengthens the processing of important information, and suppresses the influence of irrelevant or secondary information, and the output result is used as the prediction result after passing through a fully connected layer.

[0009] Preferably, the basic module comprises a long short-term memory network with 256 units, followed by two fully connected layers with 128 and 64 neurons respectively.

[0010] Preferably, the basic module also includes a dropout layer.

[0011] Preferably, the polynomial function of the trend module It is expressed as:

[0012]

[0013] in, Represents the forward prediction coefficient of the network layer contained in the basic module, T = [1, t, ..., t p ] is the power matrix of t, t is the time vector, representing the prediction step; p is the preset value.

[0014] Preferably, the periodic function of the seasonal module It is expressed as:

[0015]

[0016] in, Represents the forward prediction coefficient of the network layer contained in the basic module, S = [1, cos(2πt), ..., cos(2πit)), sin(2πt), ..., sin(2πit))].

[0017] Preferably, the method of using the trained epidemic patient number prediction model to predict the number of epidemic patients comprises the following steps:

[0018] The observed data is used as input data, the trend module captures the trend result of the input data as the trend part output, and the residual result between the input data and the trend result is output as the first intermediate residual part;

[0019] The season module receives the first intermediate residual part, captures the seasonal regularity result, and outputs the seasonal part and the second intermediate residual part;

[0020] The residual module receives the second intermediate residual part and outputs the first residual part and the second residual part;

[0021] The multi-head attention layer receives the trend part, the season part, the first residual part and the second residual part, and the multi-head self-attention mechanism assigns different weights. The output result passes through a fully connected layer and then outputs the prediction result.

[0022] Preferably, the observed data include data on the number of patients with different variables during the observation period; the variables include patient gender, patient age and epidemic situation;

[0023] The data on the number of patients with different variables during the observation period are used as input data for the epidemic patient number prediction model, and the prediction results obtained include the predicted data on the number of patients with different variables in the future.

[0024] The present invention also provides a method for predicting the number of epidemic patients based on deep learning, comprising:

[0025] Basic building blocks, which build basic modules including long short-term memory networks and two fully connected layers;

[0026] Model building module, which builds and trains the epidemic patient number prediction model based on the basic module;

[0027] The prediction module uses the trained epidemic patient number prediction model to predict the number of epidemic patients;

[0028] The epidemic patient quantity prediction model includes a trend module, a season module, a residual module and a multi-head attention layer connected in sequence; the trend module includes a basic module and a polynomial function, which is used to capture the long-term trend of the input data, and outputs the trend part and the first intermediate residual part; the season module includes a basic module and a periodic function, which is used to capture the seasonal law of the first residual part, and outputs the season part and the second intermediate residual part; the residual module includes two basic modules, which receive the second intermediate residual part and output the first residual part and the second residual part; the multi-head attention layer receives the trend part, the season part, the first residual part and the second residual part, strengthens the processing of important information, and suppresses the influence of irrelevant or secondary information, and the output result is used as the model prediction result after passing through a fully connected layer. The present invention has the following beneficial effects:

[0029] (1) The present invention supports inputting observation data of multiple variables such as the patient's gender, age, and disease system into the prediction model, and simultaneously predicts the number of patients with multiple different variables in the future by fitting the potential patterns of the observation data. It can simultaneously consider the interrelationships and dependencies between multiple variables, explore the influenza infection patterns of different age groups, regions, and populations, and provide a more comprehensive analytical perspective for formulating targeted prevention and control strategies and medical measures;

[0030] (2) The prediction model of the present invention includes a trend module, a season module and a residual module, taking into account factors such as trend, seasonality and mutation points, thereby enhancing the accuracy and interpretability of the prediction. This means that not only can the number of influenza cases be predicted, but also the changing trends and influencing factors therein can be analyzed, providing a deeper understanding for the medical clinical and public health fields;

[0031] (3) The prediction model of the present invention introduces a multi-head attention mechanism to strengthen the processing of important information, while suppressing the influence of irrelevant or minor information. It supports viewing the weight information of the three modules, thereby analyzing the influence of the three modules on the prediction results, enhancing the interpretability of the model and providing a clearer analysis method for the prediction results.

[0032] (4) The basic module of the present invention includes an LSTM and two fully connected layers. The LSTM captures context information and long-term dependencies to alleviate the problem of gradient vanishing. The fully connected layer can learn the complex nonlinear relationship of the input data, thereby adapting to various types of data.

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

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

[0035] Figure 2 A schematic diagram of a model for predicting the number of epidemic patients according to an embodiment of the present invention;

[0036] Figure 3 is an application schematic diagram of an embodiment of the present invention;

[0037] Figure 4 is a system structure diagram of an embodiment of the present invention;

[0038] Figure 5 This is a comparison and evaluation diagram of multi-feature prediction results of an embodiment of the present invention;

[0039] Figure 6 This is a feature correlation analysis between the embodiments of the present invention and different prediction models. DETAILED DESCRIPTION

[0040] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0041] S101, builds a basic module including a long short-term memory network and two fully connected layers;

[0042] S102, constructing an epidemic patient quantity prediction model based on the basic module and performing training;

[0043] S103, using the trained epidemic patient number prediction model to predict the number of epidemic patients.

[0044] For details, see Figure 2 As shown, the epidemic patient number prediction model includes a trend module, a season module, a residual module and a multi-head attention layer connected in sequence; the trend module includes a basic module and a polynomial function, which is used to capture the long-term trend of the input data, and outputs the trend part and the first intermediate residual part; the season module includes a basic module and a periodic function, which is used to capture the seasonal law of the first intermediate residual part, and outputs the season part and the second intermediate residual part; the residual module includes two basic modules, which receive the second intermediate residual part and output the first residual part and the second residual part; the multi-head attention layer receives the trend part, the season part, the first residual part and the second residual part, strengthens the processing of important information, and suppresses the influence of irrelevant or minor information. The output result is used as the model prediction result after passing through a fully connected layer. The basic module captures contextual information and long-term dependencies through LSTM (Long Short-Term Memory Network) to alleviate the problem of gradient disappearance. Backcast represents the part that is different from the input data, that is, the first intermediate residual part; forecast represents the forward propagation prediction result, that is, the component that correctly captures the trend.

[0045] The multi-head attention layer of the embodiment of the present invention is intended to comprehensively analyze the trends, seasonal changes, and mutation points in the time series data, and assign different weights to each part in order to more accurately capture the complex patterns and long-term dependencies of the data; through this structured attention mechanism, different patterns in the time series data can be effectively separated and analyzed, and the weights of each part can be comprehensively considered in the prediction, thereby improving the accuracy of the overall prediction. This method has significant advantages in complex data analysis scenarios such as epidemic prediction, and can help relevant departments better understand and respond to influencing factors at different time points. For example, when the weight of the seasonal part is high, it means that the impact of climate change is significant; when the weight of the mutation point part is high, it indicates that there may be a new epidemic and it needs to be paid special attention. The multi-head attention mechanism focuses on the weights of trends, seasons, and mutation points, and analyzes the degree of influence of the three parts at each time point.

[0046] For details, see Figure 2 As shown, it is a schematic diagram of the epidemic patient quantity prediction model according to an embodiment of the present invention, including a trend module, a season module, a residual module, and a multi-head attention layer.

[0047] The basic module consists of an LSTM with 256 units, followed by two fully connected layers with 128 and 64 neurons respectively. LSTM can store sequence data for a long time, alleviate the problem of gradient disappearance, and capture long-term dependencies through a gating mechanism.

[0048] The trend module only contains a basic module and a polynomial function to fit the trend changes. The matrix form can be expressed as:

[0049]

[0050] Among them, tr represents the trend module, l represents the output of the basic module; θ f Represents the forward prediction coefficient of the network layer contained in the basic module, where T = [1, t, ..., t p ] is the power matrix of t, t is the time vector, representing the prediction step, p can be set to a smaller value, and this polynomial fits the trend changes through a higher-order function.

[0051] The seasonal module consists of a basic module and a periodic function to capture the law of seasonal changes. The matrix form used can be expressed as:

[0052]

[0053] Among them, S = [1, cos(2πt), ..., cos(2πit)), sin(2πt), ..., sin(2πit))], and i is set to 3 in this study. In addition, the input of the seasonal module is the residual part of the trend module, that is, the trend module retains the correct part of the learning as the trend result output, and the part that is different from the original input and the trend result is output as the residual as the input of the next module, which is used as the input of the seasonal module to learn the periodic law. The periodic function is used to automatically fit the periodic characteristics in the time series data. This method avoids the limitations of the traditional method of presetting a fixed period (the preset time period may not always accurately reflect the actual law in the data, for example, epidemic data may have multiple seasonal patterns, or the seasonal pattern changes over time, so the seasonal decomposition method based on the fixed period has significant limitations in capturing complex seasonal changes). Specifically, the model captures the periodic pattern in the data by constructing a periodic function. The model determines the optimal periodic coefficient, that is, the minimum error point, by minimizing the loss function value, so as to fit the data in the best way and capture the potential pattern in the data more accurately.

[0054] In the epidemic patient number prediction model of the embodiment of the present invention, the observation data is used as input data, and the trend result captured by the trend module is used as an output. Another output part is the residual result of the input data and the trend result, which is used as the input of the season module. After the season module captures the seasonal regularity result, the residual part is used as the input of the next module. The last module is implemented with the basic module to capture the impact of the mutation point. The outputs of the three modules are passed to the multi-head self-attention mechanism to assign different weights, and finally the prediction results of each variable in the future are output. The multi-head self-attention mechanism can help the model adapt to the degree of influence of different factors, assign different weights to each module, and make interpretable decisions.

[0055] For details, see Figure 3As shown, it is a schematic diagram of the application of the epidemic patient number prediction model of the embodiment of the present invention, and the process is: the pediatric medical treatment data of the past few years is input as observation data into the epidemic patient number prediction model, and the model analyzes the data and captures the seasonal cycle changes, trend changes, mutation points and other uncontrollable factors of the data. The health department can make effective resource planning for the pediatric department by predicting the number of patients and distribution patterns of the epidemic system, and take effective preventive measures to reduce the transmission rate before the arrival of the flu season. For example, the 6-year observation data [male, female, age, respiratory system disease, ..., mental and psychological system disease] is used as input, where male represents the number of male patients with diseases per day during the observation period, female represents the number of female patients with diseases per day during the observation period, age represents the number of patients with diseases of different ages from 2 to 14 years old per day during the observation period, respiratory system disease represents the number of patients with respiratory system disease per day during the observation period, and mental and psychological system disease represents the number of patients with mental and psychological system disease per day during the observation period. The prediction result [male, female, age, respiratory system disease, ..., mental and psychological system disease] is the number of patients with different variables per day in the next year. According to the results, the incidence of respiratory diseases is higher in spring and autumn. Based on these data, parents should pay attention to children's respiratory protection during the special period of transition between winter and spring and autumn and winter, and avoid allergic coughs caused by decreased resistance in children due to alternating hot and cold weather in the morning and evening.

[0056] For details, see Figure 4 FIG. 1 is a system structure diagram of an embodiment of the present invention, including:

[0057] Basic building block 401, building a basic module including a long short-term memory network and two fully connected layers;

[0058] A model building module 402 builds an epidemic patient quantity prediction model based on the basic module and performs training;

[0059] The prediction module 403 uses the trained epidemic patient number prediction model to predict the number of epidemic patients.

[0060] Relevant experiments were conducted on the embodiments of the present invention to explore the detection effect thereof.

[0061] (1) Evaluation indicators

[0062] In order to evaluate the performance of the model, RMSE (root mean square error) and MASE (mean absolute proportional error) are used as evaluation indicators, combined with the Pearson correlation coefficient for correlation verification; Root mean square error (RMSE) is one of the commonly used indicators for evaluating the error of the forecast model. It measures the degree of difference between the model's predicted value and the actual observed value, and can intuitively understand the model's forecast accuracy; MASE and other common error metrics (such as mean square error), MASE takes into account the seasonality of the data and compares the forecast error with the error of a simple benchmark model. MASE does not depend on the scale of the data, so the model performance on different data sets can be compared; Pearson correlation coefficient is a statistic that measures the strength and direction of the linear relationship between two variables. It represents the degree of linear correlation between two variables, and the value range is between -1 and 1, where 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation. The definitions of these three indicators are as follows:

[0063]

[0064]

[0065]

[0066] Where n is the number of samples, y i is the true value, is the model prediction value; T is the length of the time series, m is the length of the seasonal cycle, e t is the actual observed value y t and predicted values The absolute error between t | is the absolute value of the absolute error; X represents the predicted value, Y represents the true value, represents the mean of the predicted values, Represents the true mean value.

[0067] (2) Experimental results

[0068] The experiments compared different models and trained them with the same parameters, including using the Adam optimizer, mean squared error (MSE) as the loss function, the learning rate set to 0.0001, and the batch size of 64.

[0069] The experimental results are as follows Figure 5 , Figure 6 And as shown in Table 1.

[0070] Table 1: Model performance indicators of different models - RMSE, MASE and correlation coefficient

[0071]

[0072] The experimental results show that the epidemic patient number prediction model proposed by the present invention shows better performance than other models, among which the RMSE is 0.042 lower than the traditional model ARIMA and 0.022 lower than other advanced deep learning models TCN. It is worth noting that among the multiple models, only the model of the present invention embodiment is less than 1 in the MASE evaluation, and reaches 0.610, which can be explained that the model of the present invention embodiment is better than the benchmark model in terms of prediction accuracy. At the same time, the average correlation coefficient is calculated, and the correlation coefficient between the prediction results of the model of the present invention embodiment and the true observation value is higher.

[0073] See also Figure 5 As shown, it is a comparison and evaluation diagram of the multi-feature prediction results of an embodiment of the present invention, in which MPD, DSD, NMSD and RSD represent mental and psychological system diseases, digestive system diseases, neuromuscular system diseases and respiratory system diseases, respectively. In the displayed prediction results, there is a correlation between different genders, ages and diseases and real data. Among them, the correlation coefficient of the prediction of the two variables of gender is above 0.8, and among the six variables of disease, the number of neuromuscular system diseases is more accurately predicted, the prediction effect of mental and psychological system diseases is weaker, and the prediction correlation coefficient of age groups remains at around 0.65.

[0074] See also Figure 6 The figure shows the feature correlation analysis between the embodiment of the present invention and different prediction models. Multiple models have different prediction effects on different variables. By comparing the depth of colors, we can intuitively see the performance of each model in different categories. The performance of the two versions of N-BEATS is similar to that of the model of the embodiment of the present invention, but the model of the embodiment of the present invention is slightly better than the NBEATS series overall, and is significantly better than the traditional models ARIMA and TCN.

[0075] It can be seen that the epidemic patient number prediction model proposed by the present invention includes a trend module, a season module and a residual module, which can not only capture trends and seasonal changes, but also capture the characteristics of mutation points and other uncontrollable factors, and support multivariate factors for simultaneous prediction. The prediction results are highly interpretable, and the weights of different modules on the prediction results can be observed. The model predicts the law of the number of future patients by observing past data and maintaining a low error value, which is of great significance to the allocation of hospital resources and the health department's measures to prevent and control influenza.

[0076] The above are only preferred embodiments of the present invention and are 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 in the protection scope of the present invention.

Claims

1. A method for predicting the number of epidemic patients based on deep learning, characterized in that: The following steps are involved: Build a basic module including a long short-term memory network and two fully connected layers; Build and train an epidemic patient quantity prediction model based on the basic module; Use the trained epidemic patient number prediction model to predict the number of epidemic patients; The epidemic patient quantity prediction model comprises a trend module, a season module, a residual module and a multi-head attention layer connected in sequence; The trend module includes a basic module and a polynomial function, which is used to capture the long-term trend of the input data, output the trend part, and use the difference between the trend part and the input data as the first intermediate residual part to capture the seasonal cycle characteristics in the next step; The seasonal module includes a basic module and a periodic function, which are used to capture the seasonal regularity of the first intermediate residual part, output the seasonal part, and use the difference between the seasonal part and the first intermediate residual part as the second intermediate residual part for the next step of capturing the mutation point features; the residual module includes two basic modules, which receive the second intermediate residual part output by the seasonal module, capture the features of different levels such as the mutation point, and output the first residual part and the second residual part; the multi-head attention layer receives the trend part, the seasonal part, the first residual part and the second residual part, strengthens the processing of important information, and suppresses the influence of irrelevant or minor information. The output result is passed through a fully connected layer as the prediction result.

2. The method for predicting the number of epidemic patients based on deep learning according to claim 1, characterized in that: The basic module consists of a LSTM network with 256 units, followed by two fully connected layers with 128 and 64 neurons respectively.

3. The method for predicting the number of epidemic patients based on deep learning according to claim 2, characterized in that: The base module also includes a dropout layer.

4. The method for predicting the number of epidemic patients based on deep learning according to claim 1, characterized in that: Polynomial functions of the Trend module It is expressed as: in, Represents the forward prediction coefficient of the network layer contained in the basic module, T = [1, t, ..., t p ] is the power matrix of t, t is the time vector, representing the prediction step; p is the preset value.

5. The method for predicting the number of epidemic patients based on deep learning according to claim 1, characterized in that: Periodic functions of the seasonal module It is expressed as: in, Represents the forward prediction coefficient of the network layer contained in the basic module, S = [1, cos(2πt), ..., cos(2πit)), sin(2πt), ..., sin(2πit))].

6. The method for predicting the number of epidemic patients based on deep learning according to claim 1, characterized in that: The method of using the trained epidemic patient number prediction model to predict the number of epidemic patients comprises the following steps: The observed data is used as input data, the trend module captures the trend result of the input data as the trend part output, and the residual result between the input data and the trend result is output as the first intermediate residual part; The season module receives the first intermediate residual part, captures the seasonal regularity result, and outputs the seasonal part and the second intermediate residual part; The residual module receives the second intermediate residual part and outputs the first residual part and the second residual part; The multi-head attention layer receives the trend part, the season part, the first residual part and the second residual part, and the multi-head self-attention mechanism assigns different weights. The output result passes through a fully connected layer and then outputs the prediction result.

7. The method for predicting the number of epidemic patients based on deep learning according to claim 6, characterized in that: The observed data include the number of diseased persons of different variables during the observation period; the variables include patient gender, patient age and epidemic situation; The data on the number of patients with different variables during the observation period are used as input data for the epidemic patient number prediction model, and the prediction results obtained include the predicted data on the number of patients with different variables in the future.

8. A method for predicting the number of epidemic patients based on deep learning, characterized in that: include: Basic building blocks, which build basic modules including long short-term memory networks and two fully connected layers; Model building module, which builds and trains the epidemic patient number prediction model based on the basic module; The prediction module uses the trained epidemic patient number prediction model to predict the number of epidemic patients; The epidemic patient quantity prediction model comprises a trend module, a season module, a residual module and a multi-head attention layer connected in sequence; The trend module includes a basic module and a polynomial function, which is used to capture the long-term trend of the input data and output the trend part and the first intermediate residual part; The seasonal module includes a basic module and a periodic function, which are used to capture the seasonal regularity of the first residual part and output the seasonal part and the second intermediate residual part; the residual module includes two basic modules, which receive the second intermediate residual part and output the first residual part and the second residual part; the multi-head attention layer receives the trend part, the seasonal part, the first residual part and the second residual part, strengthens the processing of important information, and suppresses the influence of irrelevant or minor information. The output result is passed through a fully connected layer as the model prediction result.

Images