Immune state dynamic trajectory-based elderly infection auxiliary diagnosis and treatment method and system

Abstract

The invention discloses an immune state dynamic trajectory-based elderly infection auxiliary diagnosis and treatment method and system. The method comprises the following steps: constructing a multi-time scale immune trajectory of each patient; calculating an immune feature matrix according to the multi-time-scale immune trajectory of each patient; carrying out dimension reduction processing on the immune feature matrix of each patient, calculating a severity score, a survival risk index and a treatment effect index, and further constructing a sample data set; constructing a long and short-term memory network based on an attention mechanism; training and verifying an attention mechanism-based long-short-term memory network by utilizing the sample data set to obtain an infection auxiliary diagnosis and treatment model; and predicting the infection severity, survival risk and treatment effect by using the infection auxiliary diagnosis and treatment model to realize the auxiliary diagnosis and treatment of the elderly infection. According to the method, the accuracy and robustness of auxiliary diagnosis and treatment are improved, and a new technical means is provided for precise diagnosis and treatment of elderly infection.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent medical treatment, and particularly relates to an old-age infection auxiliary diagnosis and treatment method and system based on an immune state dynamic trajectory. BACKGROUND

[0002] Traditional infection diagnosis and treatment methods mainly rely on clinical symptoms, imaging examinations and immune index detection at a single time point, such as blood routine, C-reactive protein (CRP), procalcitonin (PCT) and the like. However, these methods are difficult to comprehensively reflect the dynamic change process of the immune system of an old-age patient. Taking CRP as an example, the increase amplitude and duration thereof at the initial stage of infection are closely related to the prognosis of the old-age patient, but the traditional infection diagnosis and treatment only focuses on the value at a single time point, and ignores the dynamic change trend. Studies have shown that the mortality rate of the old-age patient whose CRP does not effectively decrease within 48-72 hours after infection is 2.8 times higher than that of the old-age patient whose CRP obviously decreases. In addition, the old-age patient often has multiple underlying diseases, such as diabetes, chronic obstructive pulmonary disease and the like, which further interfere with the interpretation of the immune index, and cause the traditional infection diagnosis and treatment method to be difficult to timely and accurately evaluate the infection severity and treatment effect.

[0003] In recent years, with the development of immunology and molecular biology technology, more and more immune indexes have been found to be closely related to the occurrence, development and prognosis of infection. Peripheral blood lymphocyte subset analysis shows that the decrease amplitude of CD4+T cell count within 24 hours after infection of the old-age infection patient is positively correlated with the incidence of sepsis, and the activation degree of CD8+T cell is directly related to the virus clearance efficiency. The research on the cytokine network finds that the peak value of pro-inflammatory factors such as IL-6 and TNF-α and the dynamic balance of anti-inflammatory factors such as IL-10 have important predictive value for the prognosis of old-age infection. A study of 1200 old-age sepsis patients shows that the 28-day mortality rate of the patients whose IL-6 peak value within 12 hours after infection exceeds 500 pg / ml is as high as 67%, and the mortality rate of the patients whose IL-6 peak value is lower than 100 pg / ml is only 19%.

[0004] Meanwhile, with the progress of bioinformatics and machine learning technology, it is possible to use multi-dimensional immune data for infection diagnosis and treatment. By integrating lymphocyte subsets, cytokines, immunoglobulins and other multi-dimensional indexes, a machine learning model can be constructed to more comprehensively reflect the overall state of the immune system. However, current researches mainly focus on the analysis of immune indexes at a single time point, and lack in-depth research on the dynamic change of immune state. Existing researches show that the dynamic change pattern of immune indexes has higher clinical value than the value at a single time point, for example, the recovery speed of NK cell activity within 3 days after infection is significantly positively correlated with the survival rate of the old-age patient, but most of the current models have not yet included such dynamic characteristics.

[0005] In addition, the immune system of elderly patients has unique age-related characteristics, and the traditional diagnosis and treatment model is difficult to adapt to the special needs of elderly infections. The immune response of the elderly population presents the characteristics of coexistence of "low reactivity" and "excessive inflammation". In the early stage of infection, it often shows insufficient immune activation, while in the middle and late stages of infection, it is prone to inflammatory factor storm. This unique immune dynamic change law makes the generalization ability of the traditional model poor. A comparative study showed that the accuracy of the infection prediction model trained based on young patient data decreased by 15-20% in the elderly population, especially when predicting immune suppression-related complications, the error was greater.

[0006] In the prior art, the infection diagnosis and treatment system based on immune indicators has the following shortcomings:

[0007] Firstly, the multi-time scale dynamic change of immune state is not modeled, and the time sequence characteristics of immune response cannot be captured. Immune response is a dynamic process, from natural immune activation in the early stage of infection (hour level) to adaptive immune response (day level), and then to immune regulation and recovery (week level). The changes of immune indicators at different time scales have different clinical significance. However, the existing system mainly analyzes the indicators at a single time point or a short time window, and cannot fully reflect this dynamic process.

[0008] Secondly, the special nature of the immune system of elderly patients is not considered, and the generalization ability of the model is poor. Immune aging in the elderly leads to significant differences in baseline levels of immune indicators compared with young people, for example, CD4+ T cell counts in healthy elderly people are 20-30% lower than in young people, and the immune activation threshold after infection is higher. The traditional model is not optimized for these characteristics, resulting in poor application effect in the elderly population.

[0009] Thirdly, the diagnosis and treatment process relies on single-time-point detection indicators, and it is difficult to dynamically evaluate treatment effect and predict prognosis. The dynamic change of immune indicators after treatment is the key to evaluate the efficacy, such as the IL-6 level of patients with effective antibiotic treatment usually begins to decline within 24-48 hours, while the IL-6 level of patients with ineffective antibiotic treatment continues to rise. However, the existing system mainly evaluates based on single-time-point detection results, and cannot capture this dynamic change in time, resulting in lag in treatment adjustment.

[0010] Fourthly, there is a lack of in-depth analysis of the complex relationship between immune indicators and clinical outcomes, and the model has poor interpretability. There are complex interactions between immune indicators, such as the synergistic pro-inflammatory effect between IL-6 and TNF-a, and the inhibitory effect of IL-10 on both. These mutual relationships have an important influence on clinical outcomes. However, the existing model mainly uses black box algorithm, which cannot explain the contribution degree and interaction mechanism of each indicator, limiting its popularization and application in clinical practice. SUMMARY

[0011] The application aims to provide an elderly infection auxiliary diagnosis and treatment method and system based on immune state dynamic trajectory, so as to solve the problems of poor diagnosis and treatment accuracy and clinical practicability caused by the lack of modeling of multi-time scale dynamic changes of immune state, the lack of consideration of the particularity of the immune system of elderly patients, the dependence on single-time point detection indexes and the lack of in-depth analysis of the complex relationship between immune indexes and clinical outcomes in traditional diagnosis and treatment methods.

[0012] The application solves the above technical problems through the following technical scheme: an elderly infection auxiliary diagnosis and treatment method based on immune state dynamic trajectory, comprising:

[0013] constructing a multi-time scale immune trajectory of each patient;

[0014] calculating the time domain feature, the frequency domain feature, the time-frequency domain feature and the trajectory morphological feature of each immune index of each patient according to the multi-time scale immune trajectory of the patient, and then obtaining an immune feature matrix composed of the time domain feature, the frequency domain feature, the time-frequency domain feature and the trajectory morphological feature of all immune indexes of the patient;

[0015] performing dimension reduction processing on the immune feature matrix of each patient, and calculating the severity score, the survival risk index and the treatment effect index;

[0016] constructing a sample data set according to the immune feature matrix after dimension reduction processing and the corresponding severity score, survival risk index and treatment effect index;

[0017] constructing a long short-term memory network based on an attention mechanism; wherein the long short-term memory network based on the attention mechanism comprises an input layer, an LSTM layer, an attention layer and an output layer connected in sequence; the LSTM layer is used for feature extraction on the immune feature matrix after dimension reduction, to obtain the hidden state of each time step; the attention layer is used for attention weighted summation on the hidden state of each time step output by the LSTM layer, to obtain a context vector; and the output layer is used for infection severity, survival risk and treatment effect prediction according to the context vector, to obtain the prediction results of the severity score, the survival risk index and the treatment effect index;

[0018] training and verifying the long short-term memory network based on the attention mechanism by using the sample data set, to obtain an infection auxiliary diagnosis and treatment model;

[0019] performing infection severity, survival risk and treatment effect prediction by using the infection auxiliary diagnosis and treatment model, to realize the auxiliary diagnosis and treatment of elderly infection.

[0020] Further, the specific construction process of the multi-time scale immune trajectory of each patient comprises:

[0021] Obtaining immune index data of a patient at different time points; wherein the immune indexes include peripheral blood lymphocyte subsets, cytokines, immunoglobulins and complement components;

[0022] Preprocessing the immune index data at different time points;

[0023] Based on the immune index data at different time points after preprocessing, the detection values of each immune index at different time points are extracted to form corresponding immune index time series;

[0024] According to the sampling frequency, the immune indexes are divided into short-term high-frequency sampling immune indexes and long-term low-frequency sampling immune indexes; the short-term high-frequency sampling refers to a sampling time period of one week and a sampling frequency of hours, and the long-term low-frequency sampling refers to a sampling time period of more than one week and a sampling frequency of weeks;

[0025] All immune index time series are unified to the same time axis, and all short-term high-frequency sampling immune index time series and all long-term low-frequency sampling immune index time series are aligned respectively to form each immune index track, and then form multi-time scale immune tracks.

[0026] Further, the immune index data at different time points are preprocessed, including:

[0027] The immune index data at different time points are cleaned to remove abnormal values and noises;

[0028] The immune index data at different time points after cleaning are standardized;

[0029] The immune index data at different time points after standardization are processed for missing values;

[0030] For short-term high-frequency sampling immune indexes, forward or backward filling method is used for missing value filling; for long-term low-frequency sampling immune indexes, random forest regression multiple imputation method is used for missing value filling.

[0031] Further, the cubic spline interpolation method is used to align all short-term high-frequency sampling immune index time series and all long-term low-frequency sampling immune index time series respectively, and the specific formula is:

[0032] S i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 ,x∈[xi , i+1 ]

[0033] wherein S i (x) represents a piecewise cubic polynomial function on the i-th interval [x i , x i+1 ], a i , b i , c i , d i all represent undetermined coefficients, x represents a detection value of the to-be-interpolated point, x i , x i+1 respectively represent detection values of the i-th time point and the i+1-th time point of the immune index.

[0034] Further, the time domain features include mean, standard deviation and peak value, the frequency domain features include power spectral density, the time-frequency domain features include wavelet transform coefficients, and the trajectory morphological features include trajectory slope and fluctuation amplitude.

[0035] Further, the immune feature matrix is processed by dimension reduction using a linear discriminant analysis method.

[0036] Further, the severity score is calculated using a SOFA score model, and the SOFA score model is:

[0037] SOFA = λ1·max(IL-6) + λ2·(CD4+T cell count) + λ3·APACHE;

[0038] wherein SOFA represents the severity score; λ1, λ2, λ3 all represent regression coefficients or weights of the SOFA score model; max(IL-6) represents the peak concentration of IL-6, reflecting the severity of inflammatory response; CD4+T cell count represents the absolute number of immune cells CD+T, evaluating the immune function status; and APACHE represents Apache II score.

[0039] The survival risk indicator is calculated using a COX proportional risk model, and the COX proportional risk model is:

[0040] ln(h(t,X)) = ln(h0(t)) + ∑β i X i i=1

[0041] wherein h(t,X) represents a risk function under the given immune index trajectory X at time t; h0(t) represents a baseline risk function; β i represents a regression coefficient; and X i represents the i-th immune index trajectory.

[0042] The calculation formula of the treatment effect index is:

[0043]

[0044] Wherein, TEI represents the treatment effect index; X pre (t) represents the immune index trajectory before treatment; X post (t) represents the immune index trajectory after treatment; T represents the evaluation period.

[0045] Based on the same concept, the application also provides an elderly infection auxiliary diagnosis and treatment system based on immune state dynamic trajectory, comprising:

[0046] The first construction unit is used for constructing the multi-time scale immune trajectory of each patient;

[0047] The calculation unit is used for calculating the time domain feature, frequency domain feature, time-frequency domain feature and trajectory morphological feature of each immune index of each patient according to the multi-time scale immune trajectory of each patient, and then obtaining the immune feature matrix composed of the time domain feature, frequency domain feature, time-frequency domain feature and trajectory morphological feature of all immune indexes of the patient;

[0048] The dimension reduction and calculation unit is used for dimension reduction processing of the immune feature matrix of each patient, and calculating the severity score, survival risk index and treatment effect index;

[0049] The second construction unit is used for constructing the sample data set according to the immune feature matrix after dimension reduction processing and the corresponding severity score, survival risk index and treatment effect index;

[0050] The third construction unit is used for constructing the long short-term memory network based on attention mechanism; wherein the long short-term memory network based on attention mechanism comprises an input layer, an LSTM layer, an attention layer and an output layer connected in turn; the LSTM layer is used for feature extraction of the immune feature matrix after dimension reduction, and the hidden state of each time step is obtained; the attention layer is used for attention weighted summation of the hidden state of each time step output by the LSTM layer, and the context vector is obtained; the output layer is used for infection severity, survival risk and treatment effect prediction according to the context vector, and the prediction result of the severity score, survival risk index and treatment effect index is obtained;

[0051] The training and verification unit is used for training and verifying the long short-term memory network based on attention mechanism by using the sample data set, and obtaining the infection auxiliary diagnosis and treatment model;

[0052] The prediction unit is used for infection severity, survival risk and treatment effect prediction by using the infection auxiliary diagnosis and treatment model, and realizing the auxiliary diagnosis and treatment of elderly infection.

[0053] Compared with the prior art, the present application has the beneficial effects that:

[0054] The multi-time scale immune trajectory constructed by the present application can comprehensively reflect the dynamic change process of the immune state of the elderly infected patient, capture the immune response characteristics at different time scales, and improve the accuracy and comprehensiveness of auxiliary diagnosis and treatment; the attention mechanism and long short-term memory network can effectively capture the time sequence characteristics and key information of the immune trajectory, improve the understanding ability of the model to the immune-infection relationship, discover complex relationships and patterns that are difficult to capture by traditional methods, and improve the prediction ability and generalization ability of the model. Through the multi-time scale immune trajectory and the long short-term memory network based on the attention mechanism, the present application not only improves the accuracy and robustness of auxiliary diagnosis and treatment, but also provides a new technical means for precise diagnosis and treatment of elderly infections. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the following embodiment description will be briefly introduced. Obviously, the drawings in the following description are only one embodiment of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0056] Figure 1 is a flow chart of the elderly infection auxiliary diagnosis and treatment method based on the immune state dynamic trajectory in the embodiment of the present application;

[0057] Figure 2 is a long short-term memory network architecture diagram based on an attention mechanism in the embodiment of the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the present application will be described clearly and completely in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0059] The technical solutions of the present application will be described in detail in combination with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0060] Embodiment one

[0061] Figure 1 The flow chart of the elderly infection auxiliary diagnosis and treatment method based on the immune state dynamic trajectory provided by the present application is shown. As shown in Figure 1 The elderly infection auxiliary diagnosis and treatment method includes the following steps:

[0062] Step 1: Data collection and preprocessing.

[0063] The immune index data of the elderly infected patients at different time points are collected from the data interface of each hospital, wherein the immune indexes include peripheral blood lymphocyte subsets, cytokines, immunoglobulins and complement components, the peripheral blood lymphocyte subsets include CD4+, CD8+, NK cell count, the cytokines include IL-6, TNF-a, IFN-g, the immunoglobulins include IgG, IgM, and the complement components include C3, C4. At the same time, the clinical data of the elderly infected patients can also be collected, including the infection type (such as bacterial, viral, fungal infection), severity score (such as SOFA, APACHEII), treatment plan (such as antibiotic type, dose, medication time, immune modulator use) and survival risk index (such as survival / death, complication occurrence).

[0064] The immune indexes can be obtained from the LIS system (Laboratory Information System) of the hospital, and according to the sampling frequency of the immune indexes, the immune indexes are divided into short-term high-frequency sampling immune indexes and long-term low-frequency sampling immune indexes. The short-term high-frequency sampling refers to the sampling time period within one week and the sampling frequency is counted by hours, for example, collecting once every 12 hours or 24 hours within 72 hours; the long-term low-frequency sampling refers to the sampling time period more than one week and the sampling frequency is counted by weeks, for example, collecting once a week within one month. The short-term high-frequency sampling immune indexes include IL-6, TNF-a, CD4+, CD8+, NK cell count, and the long-term low-frequency sampling immune indexes include immunoglobulins and complement components. Part of the clinical data can be obtained from the HIS system (Hospital Information System), such as infection type and treatment plan; the rest of the clinical data can be calculated, such as severity score and survival risk index.

[0065] The short-term high-frequency sampling immune indexes can capture the rapid change characteristics of the early immune activation stage of infection, for example, for acute infection patients, IL-6 sharply increases in a short time after infection; the long-term low-frequency sampling immune indexes can track the slow change characteristics of the immune regulation and recovery stage, for example, for subacute and chronic infection patients, the immunoglobulin level gradually recovers in a long time. The immune state dynamic trajectory is constructed by the short-term high-frequency sampling immune indexes and the long-term low-frequency sampling immune indexes. By fully mining the dynamic change characteristics of the immune state, not only the accuracy and timeliness of diagnosis and treatment are improved, but also high-risk patients can be identified early, and more targeted treatment decision support is provided for clinicians.

[0066] The data collected from the hospital data interface often has many quality problems, such as missing values, abnormal values and noise data. Therefore, in order to improve the data quality, the multi-dimensional immune index data at different time points needs to be preprocessed. In the specific embodiments of the present application, the immune index data at different time points is preprocessed, including:

[0067] Step 1.1: cleaning the immune index data at different time points to remove abnormal values and noise;

[0068] Step 1.2: standardizing the immune index data at different time points after cleaning;

[0069] Step 1.3: missing value processing of the immune index data at different time points after standardization.

[0070] In the missing value processing, for the immune indicators with a missing proportion exceeding a certain threshold (for example, 50%), considering that they may not provide effective information, a deletion strategy is adopted; for the remaining immune indicators with missing values, different missing value filling methods are adopted according to the type and clinical significance of the immune indicators.

[0071] In this embodiment, for long-term low-frequency sampling immune indicators, a random forest regression multiple imputation method is used for missing value filling. The random forest regression multiple imputation method estimates missing values by constructing multiple random forest models. This method combines the powerful prediction ability of the random forest algorithm and the idea of multiple imputation, can effectively handle missing values, and ensures the data quality, providing a reliable data basis for the subsequent steps. For short-term high-frequency sampling immune indicators, forward or backward filling method is used for missing value filling. The method of identifying outliers based on interquartile range is used to identify outliers, and the adjacent normal values are replaced to ensure data reliability.

[0072] Step 2: constructing multi-time scale immune trajectory.

[0073] Immune response is a complex process that changes dynamically at multiple time scales, from rapid immune activation in the early stage of infection (hour level) to the establishment of adaptive immune response (day level), and then to immune regulation and recovery (week level), each time scale contains different clinical information. In order to be able to fully obtain the dynamic change process of the immune state of the elderly infected patients, the immune index time series is constructed based on the preprocessed immune index data at different time points. Specifically, for each immune indicator, the detection value at different time points is extracted to form an immune index time series. In order to remove noise interference and highlight the key features of the immune trajectory, smoothing processing, feature point detection and other methods are used to remove the noise of the immune index time series.

[0074] In order to construct the multi-time scale immune trajectory, all the time series of the immune indicators of the same patient are unified to the same time axis, and all the short-term high-frequency sampling immune indicator time series are aligned to form the short-term high-frequency sampling immune indicator trajectory; all the long-term low-frequency sampling immune indicator time series are aligned to form the long-term low-frequency sampling immune indicator trajectory.

[0075] Since the sampling periods or sampling frequencies of the short-term high-frequency sampling immune indicators (or the long-term low-frequency sampling immune indicators) are inconsistent, an interpolation method is needed to align the short-term high-frequency sampling immune indicator time series (or the long-term low-frequency sampling immune indicator time series) with different sampling frequencies, so that all the short-term high-frequency sampling immune indicator time series (or the long-term low-frequency sampling immune indicator time series) are aligned to form the short-term high-frequency sampling immune indicator trajectory (or the long-term low-frequency sampling immune indicator trajectory). The multi-time scale immune trajectory is composed of all the short-term high-frequency sampling immune indicator trajectories and all the long-term low-frequency sampling immune indicator trajectories.

[0076] In this embodiment, the cubic spline interpolation method is used to align all the short-term high-frequency sampling immune indicator time series and all the long-term low-frequency sampling immune indicator time series, and the specific formula is:

[0077] S i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 ,x∈[x i ,x i+1 ](1)

[0078] Where S i (x) represents the piecewise cubic polynomial function in the i-th interval [x i ,x i+1 ], a i , b i , c i , d i all represent undetermined coefficients, x represents the value of the point to be interpolated, x i , x i+1 represent the measurement values of the immune indicators at the i-th time point and the i+1-th time point, respectively. The cubic spline interpolation method not only ensures the smoothness of the interpolation, but also better preserves the characteristics of the original data.

[0079] According to the dynamic characteristics of immune response, multi-time scale immune trajectories are divided: the IL-6 / TNF-α acute inflammation trajectory is constructed at the hourly level in the early stage of infection (0-72 hours), and the CD4+ / IgG immune recovery trajectory is constructed at the weekly level in the recovery period (>168 hours), which fully covers the whole process from rapid immune activation to long-term regulation.

[0080] By constructing multi-time scale immune trajectories, the dynamic changes of immune response can be comprehensively captured, including rapid immune activation and slow immune regulation process; the immune trajectory representation is improved by adapting to the immune response characteristics of different infection types and individual patients; and uniform time series data is provided for subsequent feature extraction and modeling, which is convenient for analysis and comparison.

[0081] Step 3: Extracting features of multi-time scale immune trajectories.

[0082] Feature extraction of multi-time scale immune trajectories is a key step to convert complex trajectory data into feature vectors with clinical significance. In this embodiment, the features of each immune indicator trajectory in the multi-time scale immune trajectory include time domain features, frequency domain features, time-frequency domain features and trajectory morphology features. The time domain features include mean, standard deviation and peak value, which reflect the overall level and fluctuation of the immune indicators in the time series, for example, the mean of CD4+ T cell count can reflect the overall immune status, and the standard deviation can reflect its stability. The frequency domain features include power spectral density, which is obtained by Fourier transform, and the specific formula is:

[0083] P(f) = |F(x(t))| 2 (2)

[0084] Where P represents the power spectral density, f represents the frequency; F represents the Fourier transform, which helps to discover the potential rules of immune response; x(t) represents the immune indicator trajectory, and t represents the time. The power spectral density can reveal the periodicity and frequency characteristics of the immune indicator changes.

[0085] The time-frequency domain features include wavelet transform coefficients, which can analyze the time and frequency characteristics of the immune indicators, and are suitable for capturing the dynamic characteristics of non-stationary signals. The trajectory morphology features include trajectory slope (rising slope and falling slope) and fluctuation amplitude, which can describe the speed and amplitude of the immune indicator changes, for example, the rising slope of IL-6 can reflect the severity of the inflammatory response.

[0086] The time domain features, frequency domain features, time-frequency domain features and trajectory morphology features of all immune indicators of each patient form the immune feature matrix of the patient, that is, each patient corresponds to an immune feature matrix.

[0087] Step 4: Dimensionality reduction processing of the immune feature matrix to calculate the severity score, survival risk index and treatment effect index.

[0088] Since the extracted features are often high-dimensional and redundant, to address this issue, the present invention uses linear discriminant analysis (LDA) to reduce the dimensionality of the immune signature matrix. LDA is a supervised learning method that aims to find projection directions that maximize inter-class differences and minimize intra-class differences. For example, by maximizing and minimizing inter-class divergence, a low-dimensional vector that retains 95% of the discriminant information is obtained, such as a combined feature that fuses the CD4+ mean and IL-6 slope.

[0089] LDA is used to perform dimensionality reduction processing, remove redundant information, improve the efficiency of subsequent modeling, and avoid the "dimensionality disaster" problem caused by high-dimensional data.

[0090] In a specific embodiment of the present invention, the severity score is calculated using the SOFA scoring model, which is:

[0091] SOFA=λ1·max(IL-6)+λ2·(CD4+ T cell count)+λ3·APACHE (3)

[0092] Among them, SOFA represents severity score; λ1, λ2, and λ3 represent the regression coefficients or weights of the SOFA scoring model; max(IL-6) represents the peak concentration of IL-6, which reflects the severity of the inflammatory response; CD4+ T cell count represents the absolute number of immune cells CD+T, which assesses the immune function status; APACHE represents Apache II score.

[0093] The survival risk index can predict the patient's survival probability, complication risk, etc. within a certain period of time, helping doctors to formulate personalized treatment plans and nursing plans. In this embodiment, the survival risk index is calculated using the COX proportional hazard model, which is:

[0094] ln(h(t,X))=ln(h0(t))+∑β i X i (4)

[0095] Where h(t,X) represents the risk function under the given immune indicator trajectory X at time t, that is, the instantaneous risk of an individual experiencing an event (such as death or recurrence) at time t; h0(t) represents the baseline risk function, that is, the basic risk level when the immune indicator trajectory is not considered. It can be estimated by analyzing the historical data of a large number of patients and using the Kaplan-Meier estimate to estimate the survival function S(t), and then calculate h0(t); β i represents the regression coefficient, which is obtained by the maximum likelihood estimation method and reflects the contribution of each immune indicator trajectory to the risk; X i represents the i-th immune indicator trajectory.

[0096] The treatment effect index is used to dynamically evaluate the effectiveness of the treatment plan and provide a basis for treatment adjustment. By comparing the changes in immune index trajectories before and after treatment, the regulatory effect of treatment on the immune system is analyzed, such as the decline in inflammatory factor levels, the recovery of immune cell function, etc., to quantify the treatment effect. When the treatment effect is poor, the doctor is prompted to adjust the treatment plan in a timely manner. In this embodiment, the calculation formula of the treatment effect index is:

[0097]

[0098] wherein TEI represents the treatment effect index; X pre (t) represents the immune index trajectory before treatment; X post (t) represents the immune index trajectory after treatment; T represents the evaluation period. The larger the TEI, the better the treatment effect.

[0099] Step 5: Construct a sample data set according to the immune feature matrix after dimensionality reduction and the corresponding severity score, survival risk index and treatment effect index.

[0100] Each sample in the sample data set includes the immune feature matrix after dimensionality reduction and the corresponding severity score, survival risk index and treatment effect index, wherein the input quantity is the immune feature matrix after dimensionality reduction, and the output quantity is the corresponding severity score, survival risk index and treatment effect index.

[0101] Step 6: Construct a long short-term memory network (LSTM) based on attention mechanism.

[0102] As shown in Figure 2 The long short-term memory network based on attention mechanism includes an input layer, an LSTM layer, an attention layer and an output layer connected in turn. The LSTM layer includes a forgetting gate, an input gate and an output gate. Through the control of these gates, past information can be selectively forgotten, current state can be updated and useful information can be output. The forgetting gate determines which information to discard, the input gate determines which information to update, and the output gate determines which information to output. The LSTM layer is used to extract features from the immune feature matrix after dimensionality reduction to obtain the hidden state of each time step. The attention layer is used to perform attention weighted summation on the hidden state of each time step output by the LSTM layer to obtain a context vector. The output layer is used to predict the severity of infection, survival risk and treatment effect according to the context vector to obtain the prediction results of the severity score, survival risk index and treatment effect index.

[0103] In the attention layer, the attention weight coefficient is first calculated according to the hidden state of each time step, and then the weighted sum of the hidden state of each time step is calculated according to the attention weight coefficient. The specific formula is:

[0104]

[0105] c =∑ i a i h i (7)

[0106] where a i denotes the attention weight coefficient of the i-th time step; w and b denote the weight matrix and bias of the attention layer, respectively; h i denotes the hidden state of the i-th time step; n denotes the number of time steps; the superscript T denotes matrix transposition; and c denotes the context vector.

[0107] The output layer includes a fully connected layer, which generates three outputs directly related to clinical diagnosis and treatment of the domain: severity score, survival risk indicator, and prediction of treatment effect index, after feature transformation of the context vector output by the attention layer.

[0108] The LSTM layer solves the gradient vanishing problem in traditional RNN through the gating mechanism, and can better handle long-distance dependencies; the context vector output by the attention layer integrates the hidden states of all time steps, and gives greater weight to the hidden states of key time steps according to the attention weight coefficient; the output layer performs infection severity assessment, survival risk and treatment effect prediction based on the context vector, and inputs the context vector into the output layer to obtain the final prediction result through structures such as fully connected layers.

[0109] Step 7: Train and verify the long short-term memory network based on the attention mechanism using the sample data set to obtain an infection auxiliary diagnosis and treatment model.

[0110] In the training phase, the reduced immune feature matrix is input into the input layer of the long short-term memory network based on the attention mechanism, and the long short-term memory network based on the attention mechanism outputs the predicted severity score, survival risk indicator, and treatment effect index. By calculating the loss value between the predicted severity score, survival risk indicator, and treatment effect index and the severity score, survival risk indicator, and treatment effect index in the sample data set, the parameters of the long short-term memory network based on the attention mechanism are adjusted in reverse to realize model training.

[0111] Step 8: Use the infection auxiliary diagnosis and treatment model to predict the severity of infection, survival risk, and treatment effect, and realize the auxiliary diagnosis and treatment of senile infection.

[0112] Step 9: Diagnosis and treatment decision support and visualization.

[0113] The output results of the infection auxiliary diagnosis and treatment model are visualized, which is an important means to convert complex model results into information that is easy for clinicians to understand. The visualization content includes a multi-time scale immune trajectory dynamic change graph, which intuitively displays the change trend of the immune index at different time points; an infection severity grading graph, which displays the infection severity of the patient in a graphical manner; a survival risk distribution graph, which displays the probability of different prognosis outcomes of the patient; and a treatment effect trend graph, which reflects the recovery of immune function and the change of treatment effect during the treatment process.

[0114] At the same time, combined with clinical guidelines and expert experience, personalized diagnosis and treatment decision suggestions are provided for clinicians. For example, according to the multi-time scale immune trajectory and the severity score, the appropriate type and dose of antibiotic is suggested; according to the survival risk index, it is suggested whether to need to strengthen the support treatment or consider other treatment methods; according to the treatment effect index, it is suggested whether to need to adjust the treatment plan or extend the treatment course, etc. The decision suggestion model is constructed based on a rule engine and a machine learning model, and the mathematical expression is:

[0115] R=f(I,C,G)

[0116] Wherein, R is the decision suggestion, I is the multi-time scale immune trajectory, C is the clinical data, G is the clinical guideline rule set, f is the decision function, which is determined by expert knowledge and data training.

[0117] For example, when the IL-6 decline slope k <-0.5 and creatinine >2mg / dL, combined with the IDSA guideline recommendation for reducing nephrotoxic antibiotics, data-driven personalized intervention is realized.

[0118] The present application captures the rapid immune activation characteristics in the early stage of infection through short-term high-frequency sampling (hour level), monitors immune regulation and recovery through long-term low-frequency sampling (week level), and comprehensively constructs the dynamic trajectory of immune state. By fully mining the dynamic change characteristics of the immune state, not only the accuracy and timeliness of diagnosis and treatment are improved, but also high-risk patients are identified early, and more targeted treatment decision support is provided for clinicians, such as timely adjustment of antibiotic regimen and initiation of immune regulation therapy. Preliminary clinical test verification shows that the present application can reduce the misdiagnosis rate of elderly infection patients by 25%, and improve the timeliness of treatment adjustment by 40%, which has important clinical application value and broad popularization prospect.

[0119] Embodiment two

[0120] The elderly infection auxiliary diagnosis and treatment system based on the dynamic trajectory of immune state provided by the embodiment of the present application comprises a first construction unit, a calculation unit, a dimension reduction and calculation unit, a second construction unit, a third construction unit, a training and verification unit and a prediction unit.

[0121] a first construction unit configured to construct a multi-time scale immune trajectory of each patient;

[0122] a calculation unit configured to calculate time domain features, frequency domain features, time-frequency domain features and trajectory morphology features of each immune indicator of the patient according to the multi-time scale immune trajectory of each patient, and further obtain an immune feature matrix composed of the time domain features, the frequency domain features, the time-frequency domain features and the trajectory morphology features of all immune indicators of the patient;

[0123] a dimension reduction and calculation unit configured to perform dimension reduction processing on the immune feature matrix of each patient, and calculate a severity score, a survival risk indicator and a treatment effect index;

[0124] a second construction unit configured to construct a sample data set according to the immune feature matrix after the dimension reduction processing and the corresponding severity score, the survival risk indicator and the treatment effect index;

[0125] a third construction unit configured to construct a long short-term memory network based on an attention mechanism; wherein the long short-term memory network based on the attention mechanism comprises an input layer, an LSTM layer, an attention layer and an output layer connected in sequence; the LSTM layer is configured to perform feature extraction on the immune feature matrix after the dimension reduction, and obtain a hidden state of each time step; the attention layer is configured to perform attention weighted summation on the hidden state of each time step output by the LSTM layer, and obtain a context vector; and the output layer is configured to perform infection severity, survival risk and treatment effect prediction according to the context vector, and obtain a prediction result of the severity score, the survival risk indicator and the treatment effect index;

[0126] a training and verification unit configured to train and verify the long short-term memory network based on the attention mechanism by using the sample data set, and obtain an infection auxiliary diagnosis and treatment model;

[0127] a prediction unit configured to perform infection severity, survival risk and treatment effect prediction by using the infection auxiliary diagnosis and treatment model, and realize the auxiliary diagnosis and treatment of the senile infection.

[0128] In some embodiments of the present application, the senile infection auxiliary diagnosis and treatment system can combine the features of the senile infection auxiliary diagnosis and treatment method in the first embodiment of the present application, and vice versa.

[0129] The above disclosure is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or modifications within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. An immune status dynamic trajectory-based auxiliary diagnosis and treatment method for elderly infections, characterized in that, The auxiliary diagnosis and treatment method comprises: constructing a multi-time scale immune trajectory for each patient; calculating the time domain feature, frequency domain feature, time-frequency domain feature and trajectory morphological feature of each immune index of each patient according to the multi-time scale immune trajectory of each patient, and then obtaining an immune feature matrix composed of the time domain feature, frequency domain feature, time-frequency domain feature and trajectory morphological feature of all immune indexes of the patient; dimension reduction processing is performed on the immune feature matrix of each patient, and a severity score, a survival risk index and a treatment effect index are calculated; a sample data set is constructed according to the immune feature matrix after dimension reduction processing and the corresponding severity score, survival risk index and treatment effect index; a long short-term memory network based on an attention mechanism is constructed; wherein the long short-term memory network based on the attention mechanism comprises an input layer, an LSTM layer, an attention layer and an output layer connected in turn; the LSTM layer is used for feature extraction on the immune feature matrix after dimension reduction, to obtain a hidden state of each time step; the attention layer is used for attention weighted summation on the hidden state of each time step output by the LSTM layer, to obtain a context vector; and the output layer is used for infection severity, survival risk and treatment effect prediction according to the context vector, to obtain a prediction result of the severity score, survival risk index and treatment effect index; the long short-term memory network based on the attention mechanism is trained and verified by using the sample data set, to obtain an infection auxiliary diagnosis and treatment model; infection severity, survival risk and treatment effect prediction is performed by using the infection auxiliary diagnosis and treatment model, to realize the auxiliary diagnosis and treatment of the infection of the elderly.

2. The immune status dynamic trajectory-based senile infection auxiliary diagnosis and treatment method according to claim 1, characterized in that, The specific construction process of the multi-time scale immune trajectory of each patient comprises: obtaining immune index data of the patient at different time points; wherein the immune indexes include peripheral blood lymphocyte subsets, cytokines, immunoglobulins and complement components; preprocessing the immune index data at different time points; based on the immune index data at different time points after preprocessing, extracting the detection values of each immune index at different time points to form corresponding immune index time series; dividing the immune indexes into short-term high-frequency sampling immune indexes and long-term low-frequency sampling immune indexes according to the sampling frequency; the short-term high-frequency sampling refers to a sampling time period of one week and a sampling frequency of hours, and the long-term low-frequency sampling refers to a sampling time period of more than one week and a sampling frequency of weeks; aligning all short-term high-frequency sampling immune index time series and all long-term low-frequency sampling immune index time series on the same time axis to form an immune index trajectory of each immune index, and then forming a multi-time scale immune trajectory.

3. The method according to claim 2, wherein the method is characterized by, The preprocessing of the immune index data at different time points comprises: cleaning the immune index data at different time points to remove abnormal values and noises; standardizing the immune index data at different time points after cleaning; processing the immune index data at different time points after standardization for missing values; Wherein, for short-term high-frequency sampling immune indicators, forward or backward padding method is used for missing value filling; for long-term low-frequency sampling immune indicators, random forest regression multiple imputation method is used for missing value filling.

4. The immune status dynamic trajectory-based senile infection auxiliary diagnosis and treatment method according to claim 2, characterized in that, The time series of all short-term high-frequency sampling immune indicators and the time series of all long-term low-frequency sampling immune indicators are aligned by using cubic spline interpolation method, and the specific formula is: S i (x) = a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 ,x∈[x i ,x i+1 ] wherein S i (x) represents a piecewise cubic polynomial function on the ith interval [x i , x i+1 ], a i , b i , c i , d i all represent undetermined coefficients, x represents a detection value of the point to be interpolated, x i , x i+1 respectively represent detection values of the immune index at the ith time point and the i+1th time point.

5. The immune status dynamic trajectory-based senile infection auxiliary diagnosis and treatment method according to claim 1, characterized in that, The time domain features include mean, standard deviation and peak value, the frequency domain features include power spectral density, the time-frequency domain features include wavelet transform coefficients, and the trajectory shape features include trajectory slope and fluctuation amplitude. 6.The method of claim 1, wherein the method further comprises: determining a dynamic trajectory of the immune status of the elderly patient; and determining a dynamic trajectory of the infection status of the elderly patient. The immune feature matrix is processed by dimension reduction by using linear discriminant analysis method.

7. The immune status dynamic trajectory-based senile infection auxiliary diagnosis and treatment method according to any one of claims 1-6, characterized in that, The severity score is calculated by using SOFA score model, and the SOFA score model is: SOFA = λ1 max (IL-6) + λ2 (CD4+T cell count) + λ3 APACHE; Wherein, SOFA represents the severity score; λ1, λ2 and λ3 all represent the regression coefficients or weights of the SOFA score model; max (IL-6) represents the peak concentration of IL-6, reflecting the severity of inflammatory response; CD4+T cell count represents the absolute number of immune cells CD+T, evaluating the immune function state; APACHE represents Apache II score; The survival risk index is calculated by using COX proportional risk model, and the COX proportional risk model is: ln(h(t,X)) = ln(h0(t)) +∑β i X i ; wherein h(t, X) represents the risk function at time t given the trajectory of immune indicators X; h0(t) represents the baseline risk function; β i represents the regression coefficient; X i represents the i-th trajectory of immune indicators; The formula for calculating the treatment effect index is: wherein TEI represents a treatment effect index; X pre (t) represents the immune index trajectory before treatment; X post (t) represents the immune index trajectory after treatment; T represents the evaluation period.

8. An elderly infection auxiliary diagnosis and treatment system based on immune state dynamic trajectory, comprising: A first construction unit for constructing a multi-time scale immune trajectory of each patient; A calculation unit for calculating the time domain features, frequency domain features, time-frequency domain features and trajectory shape features of each immune indicator of each patient according to the multi-time scale immune trajectory of each patient, and then obtaining an immune feature matrix composed of the time domain features, frequency domain features, time-frequency domain features and trajectory shape features of all immune indicators of the patient; A dimension reduction and calculation unit for performing dimension reduction processing on the immune feature matrix of each patient, and calculating the severity score, survival risk index and treatment effect index; A second construction unit for constructing a sample data set according to the dimension-reduced immune feature matrix and the corresponding severity score, survival risk index and treatment effect index; A third construction unit for constructing an attention mechanism-based long short-term memory network; wherein the attention mechanism-based long short-term memory network comprises an input layer, an LSTM layer, an attention layer and an output layer connected in turn; the LSTM layer is used for feature extraction on the dimension-reduced immune feature matrix to obtain the hidden state of each time step; the attention layer is used for attention weighted summation on the hidden state of each time step output by the LSTM layer to obtain a context vector; and the output layer is used for infection severity, survival risk and treatment effect prediction according to the context vector to obtain the prediction results of the severity score, survival risk index and treatment effect index. A training and verification unit is configured to train and verify the long short-term memory network based on the attention mechanism by using the sample data set, so as to obtain an infection auxiliary diagnosis and treatment model. A prediction unit is configured to perform infection severity, survival risk and treatment effect prediction by using the infection auxiliary diagnosis and treatment model, so as to realize the auxiliary diagnosis and treatment of the senile infection.

Images