Chronic disease progression prediction method and system based on feature progressive fusion

Abstract

The present invention belongs to the field of medical data processing technology, and discloses a method and system for predicting the progression of chronic diseases based on progressive feature fusion. In the process of progressive feature fusion, the present invention dynamically selects modal features with better prediction performance in the patient's demographic representation, disease representation, and time series clinical test representation, and then fuses them with the remaining two modal features to achieve gradual multiple fusions, which can make full use of the complementary information between multimodal features and improve the accuracy of chronic disease progression prediction. In the process of filling missing time series clinical test features based on LSTM, in addition to directly using LSTM to predict the missing features, the present invention also considers the similarity between the missing features and the complete features, and fills the missing time series clinical test features by the mean of the two, thereby improving the accuracy of data filling. The present invention takes into account the size of the intervals between patient medical records to more accurately extract time series clinical test features.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a method and system for predicting the progression of chronic diseases based on progressive feature fusion. Background Art

[0002] Chronic diseases are a type of disease that develops slowly and lasts for a long time. Common chronic diseases include cardiovascular disease, diabetes, chronic respiratory disease and cancer. The long-term and complex nature of these diseases means that treatment should not only focus on symptom relief, but also on prevention and risk management to prolong patients' lives and improve their quality of life. Accurate prediction of the progression of chronic diseases can help governments and health organizations develop more effective public health policies and interventions to improve the quality of life of patients with chronic diseases.

[0003] Chronic disease progression prediction methods can be divided into two categories: statistical methods and artificial intelligence methods. The most famous prediction model in statistical methods is the Cox proportional hazard model, which does not make any assumptions about the distribution of progression time and uses a linear regression model to model the relationship between the characteristics of chronic disease patients and logarithmic risk. Although the Cox model is very powerful, it is difficult to handle nonlinearity and interactivity in high-dimensional data due to its proportional analysis risk assumption and linear risk assumption.

[0004] The current methods for predicting the progression of chronic diseases have the following shortcomings. First: Existing methods often use simple methods (such as concatenation, addition, vector inner product, etc.) to fuse information from different modalities such as demographics, diseases, and time series tests of chronic disease patients, without fully exploring the correlation and heterogeneity between different modal data. Second: Since chronic diseases are long-term and difficult to eradicate, chronic disease patients often have multiple medical records. However, most of the current methods for predicting the progression of chronic diseases only use the data from the last visit of chronic disease patients, ignoring medical records at other times. Third: Current methods for predicting the progression of chronic diseases use models based on recurrent neural networks or Transformers to extract time series features, but these models all assume that the time intervals between medical records are regular; however, the irregular time intervals between medical records of chronic disease patients are very common in real medical scenarios, and frequent hospitalizations indicate that the patient may have health problems that are difficult to eradicate. Summary of the invention

[0005] The present invention aims to solve at least one of the above technical problems and provides a method and system for predicting the progression of chronic diseases based on progressive feature fusion.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting the progression of chronic diseases based on progressive feature fusion, the method comprising:

[0008] Acquiring medical data of patients with chronic diseases, wherein the medical data includes demographic characteristics, comorbidity characteristics, time-series clinical test characteristics, whether the chronic disease of interest has progressed, and the time when the progression occurred for each patient;

[0009] Feature extraction was performed on the time series clinical test features, demographic features and comorbidity features respectively to obtain the time series clinical test feature representation, demographic feature representation and comorbidity feature representation;

[0010] The temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation are integrated into the initial integrated representation;

[0011] The representation with the best predictive performance among the temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation is selected as the intermediate representation;

[0012] The intermediate representation is concatenated with the initial fusion representation to replace the representation with the best prediction performance;

[0013] The representation with the best prediction performance after replacement is fused with the representations of the other two modalities, and the initial fused representation is replaced with the fused representation, the intermediate representation is selected once after each fusion, and the splicing and fusion processes are repeated in sequence until a preset number of fusions are completed to obtain a final fused representation;

[0014] The probability that a patient will progress to the chronic disease of interest at a certain moment is calculated based on the final fused representation.

[0015] Preferably, the method of selecting the representation with the best predictive performance among the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation as the intermediate representation is:

[0016] The temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation were input into the DeepSurv model respectively. The consistency index of each modality representation was calculated based on whether the chronic disease of interest progressed and the time when the progression occurred. The one with the highest consistency index was taken as the intermediate representation.

[0017] Preferably, before extracting features from the time series clinical test features, data filling of missing time series clinical test features is performed based on LSTM.

[0018] Preferably, the method of data filling is:

[0019] Use the trained LSTM model to predict the time series clinical test characteristics of patients with missing data as the first imputation value;

[0020] Using the cosine similarity between patients with missing data and patients with complete data as the weight, weighted summation is performed on patients with complete data, and the weighted summation value at the missing moment is used as the second filling value;

[0021] The missing time-series clinical test characteristics were imputed using the average of the first and second imputed values.

[0022] Preferably, a training LSTM model is utilized using time-series clinical test features of patients with complete data.

[0023] Preferably, the time series clinical test features are extracted based on a hybrid attention mechanism, where the hybrid attention includes attention to the time interval between two adjacent visits and attention to the time interval between any visit and the present, and the attention weight formulas are:

[0024]

[0025] in, T is the time interval between two consecutive visits I Inserted T at the first position I The vector obtained after the average value of IN T I Length, T I The cumulative sum T cum The vector obtained by inserting 0 at the last position, i is the i-th visit, and the expression of sigmoid is: The expression of tanh is: Where e is a natural constant.

[0026] Preferably, the method for calculating the probability of a patient developing a chronic disease of interest at a certain time according to the final fused representation is:

[0027] Using the final fusion representation as input, the DeepSurv model is used to predict the risk value;

[0028] Estimate the patient's baseline risk function based on the Breslow method;

[0029] The probability of a patient developing a chronic disease of interest to the user at a certain moment is calculated based on the risk value and the baseline risk function.

[0030] Preferably, based on the loss function L 1 and L 2 Training model, L 1 The formula is:

[0031] Among them, h is the risk value, R(ti ) indicates that at t i The set of patients who have not yet experienced the chronic disease progression event of interest to the user, e i Whether the i-th patient has the chronic disease of interest;

[0032] L 2 The formula is:

[0033] Where t is the time when chronic disease progression occurs, and p(v) is the probability that the patient progresses to the chronic disease of interest at time v.

[0034] In a second aspect, the present invention provides a chronic disease progression prediction system based on progressive feature fusion, the system comprising:

[0035] An acquisition module, for acquiring medical data of patients with chronic diseases, wherein the medical data includes demographic characteristics, comorbidity characteristics, time series clinical test characteristics, whether the chronic disease of interest has progressed, and the time when the progression occurred for each patient;

[0036] A feature extraction module is used to extract features of time series clinical test features, demographic features and comorbidity features respectively, to obtain time series clinical test feature representation, demographic feature representation and comorbidity feature representation;

[0037] The first feature fusion module is used to fuse the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation into an initial fusion representation;

[0038] A selection module, used to select the representation with the best predictive performance among the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation as the intermediate representation;

[0039] A replacement module, used to replace the representation with the best prediction performance after splicing the intermediate representation with the initial fusion representation;

[0040] The second feature fusion module is used to fuse the replaced representation with the best prediction performance with the representations of the other two modalities, and replace the initial fused representation with the fused representation, select the intermediate representation once after each fusion, and repeat the splicing and fusion process in sequence until a preset number of fusions are completed to obtain a final fused representation;

[0041] A calculation module is used to calculate the probability of a patient progressing to a chronic disease of interest at a certain moment according to the final fused representation.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] In the process of progressive feature fusion, the present invention dynamically selects modal features with better predictive performance from the patient's demographic representation, disease representation and time-series clinical test representation, and then fuses them with other modal features to achieve gradual and multiple fusions. This can make full use of the complementary information between multi-modal features and improve the accuracy of predicting the progression of chronic diseases.

[0044] In the process of filling the missing time series clinical test features based on LSTM, in addition to directly using LSTM to predict the missing features, the present invention also considers the similarity between the missing features and the complete features, and fills the missing time series clinical test features by the mean of the two, thereby improving the accuracy of data filling.

[0045] The present invention takes into account the interval size between patient visit records, including the time interval between two adjacent visits and the time interval between any visit and the present, and more accurately extracts temporal clinical examination features through a hybrid attention mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of a process flow of a method for predicting chronic disease progression based on progressive feature fusion provided by an embodiment of the present invention;

[0047] Figure 2 A schematic block diagram of a feature progressive fusion method provided by an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of the structure of a chronic disease progression prediction system based on progressive feature fusion provided by an embodiment of the present invention;

[0049] Figure 4 FIG. 1 is a schematic block diagram of an exemplary electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] The first embodiment of the present invention provides a method for predicting the progression of chronic diseases based on progressive feature fusion. Figure 1 A schematic flow chart of a method for predicting chronic disease progression based on progressive feature fusion provided in an embodiment of the present invention.

[0052] like Figure 1 As shown in FIG. 1 , the chronic disease progression prediction method based on progressive feature fusion specifically includes the following steps:

[0053] S101, obtaining medical data of patients with chronic diseases, wherein the medical data includes demographic characteristics, comorbidity characteristics, time series clinical test characteristics, whether the chronic disease of interest has progressed, and the time when the progression occurred for each patient.

[0054] Patients can be all chronic disease patients in a certain area, city, province, country, etc. The larger the number of patients, the more accurate the final prediction result. The medical data of each chronic disease patient includes: demographic characteristics (x d ), comorbidity characteristics (x c ), time series clinical test characteristics (x t ), whether the chronic disease of interest has progressed (e) and the time when the progression occurred (t). Demographic characteristics generally include information such as the patient's age, gender, marital status, etc. The patient's comorbidity characteristics include the patient's disease history information; the patient's time-series clinical test characteristics include clinical test information (such as blood cell count, urine pH, etc.) during multiple visits. Chronic disease progression is defined as a patient's contraction of a chronic disease that he has never had before. The present invention requires the patient to define the chronic disease progression events of interest, such as hypertension progression, diabetes progression, etc.; if e=1, it means that the patient has the chronic disease of interest at time t, that is, a chronic disease progression event has occurred; if e=0, it means that the patient has not yet suffered from the chronic disease at time t. The symbol N represents the number of chronic disease patients collected, and the entire medical data set D is expressed as:

[0055] S102, extracting features of the time series clinical test features, demographic features and comorbidity features respectively to obtain a time series clinical test feature representation, a demographic feature representation and a comorbidity feature representation.

[0056] For demographic characteristics, input them into the fully connected network to obtain the demographic representation R d .

[0057] For the comorbidity features, after one-hot encoding, the one-hot encoded data is input into the fully connected network to obtain the comorbidity feature representation R c .

[0058] For time series clinical test features, since some data are inevitably missing in actual situations, in some preferred implementations, in order to more accurately extract the time series clinical test feature representation, the missing data is first filled.

[0059] In these preferred embodiments, the time series clinical test features are filled based on LSTM, including:

[0060] S102-1, using the trained LSTM model to predict the time series clinical test characteristics of patients with missing data as the first filling value.

[0061] It can be understood that the LSTM model is trained using the time series clinical test features of patients with complete data. For example, all the consultation times of all patients in the entire medical data set are traversed, and the consultation times with a missing rate of less than 20% of the time series clinical test features are screened out, denoted as T; from the time series clinical test features x t The columns with the consultation time T are selected from the dataset, and the selected time series clinical test characteristics are recorded as x ts ; in x ts Patients with complete temporal clinical test characteristics were selected and marked with Indicates; in x ts Patients with missing temporal clinical test characteristics were selected from the express; for For each patient i in , count the visit time with missing values, recorded as Will The values ​​in are randomly set to 0, denoted as Will Input into the LSTM model and based on L 3 Loss function training, L 3 The mathematical expression is: in, is the output of the LSTM model; For each patient i in , first fill its missing value with 0, then input it into the above trained LSTM, and finally save the output of LSTM in The value at the moment is taken as the first filling value and is recorded as

[0062] S102-2, using the cosine similarity between patients with missing data and patients with complete data as a weight, performing a weighted sum on the patients with complete data, and using the weighted sum value at the missing time as the second filling value.

[0063] For example, for For each patient i in The cosine similarity SIM between each patient and the patient cm , The formula for calculating the cosine similarity between the jth patient and the patient is:

[0064] Will The weighted sum of the patients in is recorded as x weight , and its calculation formula is Among them, N c for The number of patients in; save x weight exist The value at time is used as the second filling value and is recorded as

[0065] S102-3, use the average of the first filling value and the second filling value to fill in the data of the missing time series clinical test characteristics.

[0066] Specifically, The missing value for each patient i in and The average value is used to fill in the patient's time series clinical test data, using the symbol represents the time-series clinical test characteristics after imputation.

[0067] Since chronic disease patients have more than one medical record in actual medical scenarios, the disease conditions corresponding to each medical record are different, and the time intervals between medical records are irregular, frequent hospitalizations indicate that the patient may have health problems that are difficult to eradicate. Therefore, the intervals between medical records can provide information support for the prediction of chronic disease progression. In order to more accurately extract the time series clinical test feature representation, in some preferred embodiments, the time series clinical test features are extracted based on a hybrid attention mechanism. Hybrid attention includes attention to the time interval between two adjacent medical visits, and attention to the time interval between any medical visit and the present.

[0068] Specifically, based on the visit time list T, the time interval T between two consecutive visits can be calculated. I , using the symbol T IN Indicates T I length, and T I Cumulatively calculate from right to left and get T 1 The cumulative sum T cum It is not difficult to know that T cum The length of T IN , T cum [i] can represent the time from the i-th visit to the present.

[0069] If the time interval between two consecutive visits is T I If the value is small, it means that patients with chronic diseases may have major health problems and should be given more attention. The attention weight Attn is calculated based on the following formula 1 :

[0070] in, T is the time interval between two consecutive visitsI Inserted T at the first position I The vector obtained after the average value of , the expression of sigmoid is: Where e is a natural constant.

[0071] According to clinical experience, we should pay more attention to clinical test features that are closer to the present. For example, the clinical test information of a chronic disease patient 5 years ago is unlikely to contribute as much to predicting the progression of the chronic disease as the clinical test information 1 month ago. The attention weight Attn is calculated based on the following formula 2 :

[0072] in, T cum After inserting 0 at the last position, the expression of tanh is: Where e is a natural constant.

[0073] In some preferred embodiments, in order to capture the association between different time series clinical test features, the self-attention weight of the filled time series clinical test features can also be calculated. Input the self-attention layer and get the attention weight Attn 3 , and its calculation formula is:

[0074]

[0075] in, represents the temporal clinical test characteristics at the ith visit, for The matrix transpose of .

[0076] It is understandable that, based on the attention to the time interval between two adjacent visits and the attention to the time interval between any visit and the present, those skilled in the art may also selectively mix other commonly used attention mechanisms according to actual circumstances.

[0077] When mixing the above three attention mechanisms, it is not difficult to find that Attn 1 、Attn 2 Both with Attn 3 The number of columns is the same as that of the row, but the number of rows is 1. You need to first change Attn 1 、Attn 2 The matrix size is converted to Attn 3 Similarly, three attention mechanisms are mixed to extract features.

[0078] For example, Attn 1 and Attn 2 Copy TIN +1 copy, and splice according to the row to get and Will and Attn 3 Perform matrix addition to obtain the final attention weight Attn 4 , the calculation formula is as follows:

[0079]

[0080] Attn 4 and Perform matrix multiplication and record the result as Its mathematical expression is:

[0081] in, Represents matrix multiplication.

[0082] Will Input into Transformer, and record the output of Transformer as h q ; will h q After inputting into the fully connected network, a prediction is constructed auxiliary tasks, and based on L 4 Loss function training, L 4 The mathematical expression is:

[0083] in, is the predicted value of the fully connected network.

[0084] h q After inputting into the Flatten layer, we get the time series clinical test feature representation R t .

[0085] S103, fusing the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation into an initial fused representation.

[0086] Understandably, using the time series clinical test features after data filling in S102 and the time series clinical test feature representation extracted by the hybrid attention mechanism can make the final prediction effect more accurate.

[0087] Characterizing comorbidity characteristics c , demographic representation d and time series clinical test characterization R t Input into the fusion module to obtain the initial fusion representation R f .

[0088] S104, selecting the representation with the best predictive performance among the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation as the intermediate representation.

[0089] Specifically, the temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation were respectively input into the DeepSurv model, and the consistency index (c-index value) of each modality representation was calculated based on whether the chronic disease of interest progressed and the time when the progression occurred. The one with the highest consistency index was taken as the intermediate representation.

[0090] S105, concatenating the intermediate representation with the initial fused representation to replace the representation with the best prediction performance.

[0091] Specifically, the intermediate representation is concatenated with the initial fusion as the new representation with the best prediction performance selected in S104. For example, if the one with the highest c-index value is the comorbidity feature representation R c , then R c With the initial fusion representation R f The spliced ​​result is used as the new R c .

[0092] S106, the representation with the best prediction performance after replacement is fused with the representations of the other two modalities, and the initial fused representation is replaced with the fused representation. The intermediate representation is selected once after each fusion, and the splicing and fusion processes are repeated in sequence until a preset number of fusions are completed to obtain a final fused representation.

[0093] Understandably, in the progressive fusion process, the representation with the best prediction performance is dynamically determined and spliced ​​with the fused representation obtained last time to replace the representation of the corresponding modality, and then fused with the representations of the remaining two modalities. Repeat the aforementioned splicing, fusion, and selection process, that is, repeat steps S104-S106 in sequence. The specific number of fusions can be freely adjusted.

[0094] Figure 2 This is a schematic diagram of the feature progressive fusion method, which is only used as an example. Figure 2 The methods of S104-S106 can be described in detail as follows:

[0095] Step 1: Set the number of progressive fusions to M f ,M f It is a hyperparameter and can be adjusted freely;

[0096] Step 2: Initialize loop marker variable I f is 1;

[0097] Step 3: Judgment I f Is it greater than M? f, if yes, end the progressive integration;

[0098] Step 4: R d , R c and R t Input into the DeepSurv model and calculate the c-index of each modality representation based on e and t;

[0099] Step 5: Find the representation with the highest c-index value, assuming it is R c If I f =1, then R c , R d and R t according to Figure 2 Sequentially input into the fusion module to obtain the initial fusion representation R f If I f ≠1, R c With the initial fusion representation R f The spliced ​​result is used as the new R c , and the new R c , R d and R t Input into the fusion module to obtain the intermediate fusion representation, and replace the initial fusion representation with the fused representation. f After incrementing by 1, jump to step 3.

[0100] S107, calculating the probability that the patient will develop a chronic disease of interest at a certain moment based on the final fused representation.

[0101] Specifically, the final fusion representation is used as input, and the risk value h is predicted using the DeepSurv model; the patient's baseline risk function h is estimated based on the Breslow method. 0 ; According to the risk value h and the baseline risk function h 0 Calculate the probability that a patient will develop a chronic disease of interest to the user at a certain moment.

[0102] The probability P(v) that a patient develops a chronic disease of interest to the user at time v is calculated by the following formula:

[0103]

[0104] In some preferred embodiments, based on the loss function L 1 and L 2 Training the chronic disease progression prediction model according to the embodiment of the present invention, L 1 The formula is:

[0105] Among them, h is the risk value, R(t i ) indicates that at ti The set of patients who have not yet experienced the chronic disease progression event of interest to the user, e i Whether the i-th patient has the chronic disease of interest;

[0106] L 2 The formula is:

[0107] Where t is the time when chronic disease progression occurs, and p(v) is the probability that the patient progresses to the chronic disease of interest at time v.

[0108] The second embodiment of the present invention provides a chronic disease progression prediction system based on progressive feature fusion.

[0109] Figure 3 A schematic diagram of the structure of a chronic disease progression prediction system based on progressive feature fusion provided in an embodiment of the present invention.

[0110] like Figure 3 As shown, the chronic disease progression prediction system 300 based on feature progressive fusion includes an acquisition module 301, a feature extraction module 302, a first feature fusion module 303, a selection module 304, a replacement module 305, a second feature fusion module 306, and a calculation module 307. The functions of each module are described in detail as follows:

[0111] An acquisition module 301 is used to acquire medical data of patients with chronic diseases, wherein the medical data includes demographic characteristics, comorbidity characteristics, time series clinical test characteristics, whether the chronic disease of interest has progressed, and the time when the progression occurred for each patient;

[0112] A feature extraction module 302 is used to extract the representations of the temporal clinical test features, the demographic features and the comorbidity features, respectively, to obtain the temporal clinical test feature representation, the demographic feature representation and the comorbidity feature representation;

[0113] The first feature fusion module 303 is used to fuse the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation into an initial fusion representation;

[0114] A selection module 304 is used to select the representation with the best predictive performance among the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation as the intermediate representation;

[0115] A replacement module 305, configured to replace the representation with the best prediction performance by splicing the intermediate representation with the initial fused representation;

[0116] The second feature fusion module 306 is used to fuse the representation with the best prediction performance after replacement with the representations of the other two modalities, select the intermediate representation once after each fusion, and repeat the splicing, fusion, and selection processes in sequence until a preset number of fusions are completed to obtain a final fused representation;

[0117] The calculation module 307 is used to calculate the probability of the patient developing a chronic disease of interest at a certain moment according to the final fused representation.

[0118] Based on the above embodiments, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the steps in the method for predicting chronic disease progression based on progressive feature fusion in the embodiment of the first aspect are implemented.

[0119] Figure 4 A schematic block diagram of an example electronic device 400 that can be used to implement an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0120] like Figure 4 As shown, the electronic device 400 may include a computing unit 401, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. In the RAM 403, various programs and data required for the operation of the device 400 can also be stored. The computing unit 401, the ROM 402, and the RAM 403 are connected to each other via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0121] Multiple components in the device 400 are connected to the I / O interface 405, including: an input unit 406, such as a keyboard, a mouse, etc.; an output unit 407, such as various types of displays, speakers, etc.; a storage unit 408, such as a disk, an optical disk, etc.; and a communication unit 409, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 409 allows the device 400 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0122] The computing unit 401 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 401 performs the various methods and processes described above, such as omics data processing methods or model training methods for omics data processing. For example, in some embodiments, the omics data processing method or the model training method for omics data processing may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the omics data processing method or the model training method for omics data processing described above may be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to execute the omics data processing method or the model training method for omics data processing in any other suitable manner (eg, by means of firmware).

[0123] Based on the above embodiments, the present invention further provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute the cancer dynamic survival analysis method based on hierarchical attention disclosed in the embodiments of the present invention.

[0124] Based on the above embodiments, the present invention further provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the cancer dynamic survival analysis method based on hierarchical attention disclosed in the embodiments of the present invention.

[0125] The computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer readable media do not include electric carrier signals and telecommunication signals.

[0126] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0127] In the several embodiments provided in the present application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0128] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0129] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0130] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of each embodiment of the present application.

[0131] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting the progression of chronic diseases based on progressive feature fusion, characterized in that: The method includes: Acquiring medical data of patients with chronic diseases, wherein the medical data includes demographic characteristics, comorbidity characteristics, time-series clinical test characteristics, whether the chronic disease of interest has progressed, and the time when the progression occurred for each patient; Feature extraction was performed on the time series clinical test features, demographic features and comorbidity features respectively to obtain the time series clinical test feature representation, demographic feature representation and comorbidity feature representation; The temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation are integrated into the initial integrated representation; The representation with the best predictive performance among the temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation is selected as the intermediate representation; The intermediate representation is concatenated with the initial fusion representation to replace the representation with the best prediction performance; The representation with the best prediction performance after replacement is fused with the representations of the other two modalities, and the initial fused representation is replaced with the fused representation, the intermediate representation is selected once after each fusion, and the splicing and fusion processes are repeated in sequence until a preset number of fusions are completed to obtain a final fused representation; Using the final fusion representation as input, the DeepSurv model is used to predict the risk value; Estimate the patient's baseline risk function based on the Breslow method; The probability of a patient developing a chronic disease of interest to the user at a certain moment is calculated based on the risk value and the baseline risk function.

2. The method for predicting chronic disease progression based on progressive feature fusion according to claim 1, characterized in that: The method of selecting the best predictive performance among the temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation as the intermediate representation is: The temporal clinical test feature representation, demographic feature representation, and comorbidity feature representation were input into the DeepSurv model respectively. The consistency index of each modality representation was calculated based on whether the chronic disease of interest progressed and the time when the progression occurred. The one with the highest consistency index was taken as the intermediate representation.

3. The method for predicting chronic disease progression based on progressive feature fusion according to claim 1, characterized in that: Before extracting the time series clinical test features, data filling of missing time series clinical test features is performed based on LSTM.

4. The method for predicting chronic disease progression based on progressive feature fusion according to claim 3, characterized in that: The method of data filling is: Use the trained LSTM model to predict the time series clinical test characteristics of patients with missing data as the first imputation value; Using the cosine similarity between patients with missing data and patients with complete data as the weight, weighted summation is performed on patients with complete data, and the weighted summation value at the missing moment is used as the second filling value; The missing time-series clinical test characteristics were imputed using the average of the first and second imputed values.

5. The method for predicting chronic disease progression based on progressive feature fusion according to claim 4, characterized in that: The LSTM model is trained using the time-series clinical test features of patients with complete data.

6. The method for predicting chronic disease progression based on progressive feature fusion according to claim 1, characterized in that: The time series clinical test features are extracted based on the hybrid attention mechanism. The hybrid attention includes the attention to the time interval between two adjacent visits and the attention to the time interval between any visit and the present. The attention weight formulas are: ， ， in, The time interval between two consecutive visits Inserted at the first position The vector obtained after the average value of for Length, for The cumulative sum The vector obtained by inserting 0 at the last position, i is the i-th visit, and the expression of sigmoid is: , the expression of tanh is: ,in is a natural constant.

7. The method for predicting chronic disease progression based on progressive feature fusion according to claim 6, characterized in that: The model is trained based on the loss functions L1 and L2. The formula for L1 is: , where h is the risk value, Indicated in The set of patients who have not yet experienced the chronic disease progression event of interest to the user, e i Whether the i-th patient has the chronic disease of interest; The formula for L2 is: , where t is the time when chronic disease progression occurs, is the probability that the patient develops the chronic disease of interest at time v.

8. A chronic disease progression prediction system based on progressive feature fusion, characterized in that: The system includes: An acquisition module, for acquiring medical data of patients with chronic diseases, wherein the medical data includes demographic characteristics, comorbidity characteristics, time series clinical test characteristics, whether the chronic disease of interest has progressed, and the time when the progression occurred for each patient; A feature extraction module is used to extract features of time series clinical test features, demographic features and comorbidity features respectively, to obtain time series clinical test feature representation, demographic feature representation and comorbidity feature representation; The first feature fusion module is used to fuse the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation into an initial fusion representation; A selection module, used to select the representation with the best predictive performance among the temporal clinical test feature representation, the demographic feature representation, and the comorbidity feature representation as the intermediate representation; A replacement module, used to replace the representation with the best prediction performance after splicing the intermediate representation with the initial fusion representation; The second feature fusion module is used to fuse the replaced representation with the best prediction performance with the representations of the other two modalities, and replace the initial fused representation with the fused representation, select the intermediate representation once after each fusion, and repeat the splicing and fusion process in sequence until a preset number of fusions are completed to obtain a final fused representation; The calculation module is used to use the final fusion representation as input and use the DeepSurv model to predict the risk value; Estimate the patient's baseline risk function based on the Breslow method; The probability of a patient developing a chronic disease of interest to the user at a certain moment is calculated based on the risk value and the baseline risk function.

Images