Disease progress prediction method and system based on frequency domain characteristics

Abstract

The invention discloses a disease progress prediction method and system based on frequency domain characteristics, and the method comprises the steps: carrying out the modeling of multi-dimensional frequency domain characteristics, and converting an original physiological signal into a global frequency domain characteristic; self-adaptive frequency band selection: selecting an optimal frequency band combination in the global frequency domain characteristics by using a genetics algorithm; extracting a frequency domain template of the disease, constructing a disease specific frequency domain feature library (DSFP), and calculating a matching score; training a time sequence prediction model according to the optimal frequency band combination and the matching score, and performing disease progress prediction by using the trained prediction model; and using the trained prediction model to predict the disease process. The system comprises a time sequence prediction model training module and a real-time prediction model application module. According to the method, the frequency domain and multi-scale time-frequency structure features are fully mined, and dynamic matching with disease pertinence is realized through the constructed disease specific frequency domain feature library (DSFP).

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical health and artificial intelligence, and in particular to a disease progression prediction method and system based on frequency domain features. BACKGROUND

[0002] With the rapid development of artificial intelligence and big data technology, deep learning models have made significant progress in medical image analysis, physiological signal monitoring, disease diagnosis and risk assessment, etc. In particular, in the aspect of disease progression prediction, by modeling the physiological signals, medical images, electronic pathology and multi-modal data of patients, the transformation from disease detection to disease prediction and intervention is realized, which is of great significance to improve clinical precision medicine and individualized decision-making.

[0003] However, the existing disease progression prediction methods mostly rely on time domain signal modeling, focusing on directly processing time series data such as electrocardiogram (ECG), electroencephalogram (EEG), blood oxygen, and motion signals, which cannot effectively capture the implicit periodicity, lack disease spectrum structure, and are difficult to realize knowledge reuse. At the same time, time domain features are sensitive to noise and have limited feature expression ability. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a disease progression prediction method and system based on frequency domain features, which realizes high-precision disease progression prediction through frequency domain feature modeling, disease-specific frequency domain feature library (DSFP) construction, and dynamic optimization technology, to solve the problems raised in the background art.

[0005] To achieve the above purpose, the present application provides the following technical scheme: A disease progression prediction method based on frequency domain features is provided, comprising: Step 1: Multi-dimensional frequency domain feature modeling, converting the original physiological signals into global frequency domain features: Step 11: Collecting multi-modal physiological signals of patients, after data preprocessing, segmenting the preprocessed signals, and simultaneously smoothing the segmented signal boundaries through Hanning window to reduce frequency spectrum leakage caused by signal stages. Finally, the segmented signals are converted into local frequency domain features through fast Fourier transform. At the same time, the mapping relationship between the phase change direction time sequence and the disease stage is established through a random forest model.

[0006] Among them, the local frequency domain feature can be represented as: Among them, represents the first segmented signal, FFT represents fast Fourier transform, represents the amplitude spectrum, represents the phase spectrum, represents the phase change direction.

[0007] It should be noted that the amplitude spectrum is used to identify the main frequency components, and the phase spectrum reflects the time delay and harmonic relationship of the signal. The frequency domain feature modeling method combining the introduced amplitude spectrum and phase spectrum solves the problem of insufficient single feature dimension.

[0008] It should be further pointed out that, The sampling frequency is set higher than the Nyquist frequency to avoid frequency aliasing into false low-frequency signals in the sampling results, causing spectral distortion, and the Nyquist frequency is equal to half of the sampling rate.

[0009] In addition, the mapping relationship is specifically: through a data mapping mechanism, the time sequence of the phase change direction is taken as an input feature, a random forest model is used to associate and model the corresponding disease staging labels, and the correlation coefficient between the time sequence feature of the phase change direction and each disease stage is calculated in the form of feature importance evaluation or node division contribution, which is used to quantify the degree of correlation.

[0010] Step 12: Calculate the mutual information between each local frequency domain feature and the label: Among them, represents the first The entropy of the local frequency domain feature, represents the entropy of the label, represents the first The joint entropy of the local frequency domain feature and the label, represents the correlation coefficient between the phase change and the disease stage.

[0011] Among them, the correlation coefficient ranges from 0 to 1, when the correlation coefficient is close to 1, it means that the time sequence feature has a high correlation with the current disease stage, and when the correlation coefficient is close to 0, it means that the time sequence feature has a weak correlation with the current disease stage, and has no significant contribution.

[0012] Step 13: Dynamically adjust the initial weight of each local frequency domain feature according to the mutual information: Among them, represents the number of segments in step 11.

[0013] Step 14: Merge the weighted local frequency domain features to obtain the global frequency domain feature: Among them, represents the initial weight proportional to the different value.

[0014] Step 2: Adaptive frequency band selection, using genetic algorithm, select the optimal frequency band combination in the global frequency domain feature: Step 21: Define the initial population, construct the fitness function: Wherein, AUC represents the AUC of the ROC curve, F1 represents the F1 score, C represents the stability coefficient of the disease.

[0015] It should be noted that the initial frequency band is the global frequency domain feature.

[0016] In addition, it should be noted that the stability coefficient is obtained using a correlation method based on frequency band consistency, focusing on the similarity of trends and shapes shown between the same frequency bands of different patients.

[0017] Step 22: Define a frequency band combination with chromosomes, then perform genetic operations to generate new frequency band combinations.

[0018] Wherein, the new frequency band combination is found by automatic search of the algorithm to find frequency bands containing stronger prediction information, discover features with optimal value for prediction, and optimize the prediction performance of the model.

[0019] Step 23: For each frequency band combination, calculate the classification performance based on the training set to obtain And , the fitness score is obtained by weighting And Wherein, W represents the weight of W represents the weight of W represents the weight of C represents the stability coefficient of the disease. Step 24: Optimize to fitness convergence to obtain the optimal frequency band combination.

[0020] Step 3: Extract the frequency domain template of the disease, construct the disease-specific frequency domain feature library (DSFP), and calculate the matching score:

[0021] Step 31: Using large-scale clinical data, based on the signals of patients with different diseases, extract the frequency domain template of each disease: Wherein, C represents the stability coefficient of the disease.​ frequency domain templates of different diseases, dominant frequency representing spectral feature, amplitude spectrum representing spectral feature in frequency domain feature, phase spectrum representing spectral feature in frequency domain feature.

[0022] Step 32: frequency domain templates of different diseases are collectively constructed into a disease-specific frequency domain feature library (DSFP).

[0023] Step 33: the single frequency band and the frequency domain template of the optimal frequency band combination input are normalized: wherein, frequency domain feature of single frequency band, frequency domain template.

[0024] Step 34: calculate the matching score: wherein, frequency domain feature of single frequency band, frequency domain template.

[0025] The matching score is used to determine the possibility of the current disease.

[0026] Step 4: train the time series prediction model according to the optimal frequency band combination obtained in step 2 and the matching score obtained in step 3, use the trained prediction model to predict the disease progression, and learn the law of disease change over time.

[0027] The law of disease change over time is specifically to combine short-term, medium-term and long-term frequency domain features to dynamically capture the multidimensional changes of disease progression.

[0028] Step 5: use the trained prediction model to predict the disease progression, specifically: collect multi-modal physiological signals of a new patient, convert the physiological data into time-frequency domain features through Morlet wavelet transform, perform adaptive frequency band selection, obtain frequency domain features of the optimal frequency band, and at the same time divide the time scale of the time domain features, calculate the matching score of the frequency domain features at different time scales, input the matching score and the frequency domain features of the optimal frequency band into the trained time series prediction model, and obtain the future disease progression prediction result.

[0029] wherein, the different time scales are short-term window, medium-term window and long-term window.

[0030] It should be noted that in the process of obtaining the future disease progression prediction result, the mutual information calculated in the matching score is also used to calculate the weight value, further dynamically weighted fusion of multi-scale features, and then optimization of time correlation.

[0031] In addition, it should be noted that when making predictions, the Morlet wavelet transform is used to obtain frequency domain features while retaining time domain features, i.e., time dimension information. Specifically, the Morlet wavelet transform has good time-frequency localization characteristics, and the combination of the complex exponential and the Gaussian window function enables it to accurately locate the time position of the feature occurrence while performing frequency spectrum analysis.

[0032] A disease progression prediction system based on frequency domain features is provided, comprising a time series prediction model training module and a real-time prediction model application module. The time series prediction model application module comprises multi-dimensional frequency domain feature modeling, adaptive frequency band selection, construction of a disease-specific frequency domain feature library (DSFP), and training of the time series prediction module. The real-time prediction model application module comprises data acquisition, optimal frequency band acquisition, calculation of a matching score with the disease-specific frequency domain feature library (DSFP), and future disease progression prediction.

[0033] The disease progression prediction system based on frequency domain features has the following technical effects and advantages: By fusing frequency domain analysis, signal processing, and dynamic modeling techniques, the following is achieved: 1. A multi-dimensional frequency domain feature model is constructed, which can more comprehensively depict the potential laws of disease signals and improve feature expression capability compared to traditional time domain or single frequency domain methods. 2. A disease-specific frequency domain feature library (DSFP) is established, forming a standardized disease frequency domain template and enabling high-precision matching and recognition of different disease types or stages. 3. A genetic algorithm is introduced for optimal frequency band selection, allowing the model to automatically adjust the frequency domain range based on patient individual signal characteristics, improving the system's generalization ability and individual adaptability. 4. Multi-time scale frequency domain features are fused to construct a disease progression dynamic atlas, enabling continuous monitoring and trend prediction of diseases from early stages to evolution stages. 5. An offline training and online prediction mode is adopted to ensure high precision of the model while meeting real-time application requirements, improving the practicality and reliability in clinical monitoring and auxiliary diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0034] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0035] Figure 1 The method flowchart of the present application.

[0036] Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1: Figure 1 This invention provides a disease progression prediction method based on frequency domain features, comprising: Step 1: Multidimensional frequency domain feature modeling, converting the original physiological signal into global frequency domain features: Step 11: Acquire the patient's multimodal physiological signals, perform data preprocessing, segment the preprocessed signals, and smooth the signal boundaries of the segments using a Hanning window to reduce spectral leakage caused by signal segmentation. Finally, convert the segmented signals into local frequency domain features using a Fast Fourier Transform. Meanwhile, a mapping relationship between the phase change direction time series and disease staging was established using a random forest model.

[0039] The local frequency domain features can be expressed as: in, Indicates the first Segmented signals, FFT stands for Fast Fourier Transform. Represents the amplitude spectrum. Represents the phase spectrum. Indicates the direction of phase change.

[0040] It should be noted that the amplitude spectrum is used to identify the main frequency components, and the phase spectrum reflects the time delay and harmonic relationship of the signal. The frequency domain feature modeling method that combines the amplitude spectrum and the phase spectrum solves the problem of insufficient single feature dimension.

[0041] It should be further explained that, The sampling frequency is set higher than the Nyquist frequency to avoid frequency aliasing and the appearance of incorrect low-frequency signals in the sampling results, which would cause spectral distortion. The Nyquist frequency is equal to half of the sampling rate.

[0042] In addition, the mapping relationship is specifically implemented through a data mapping mechanism. The time series of phase change direction is used as input features, and a stochastic Senli model is used to model the association between the phase change direction and the corresponding disease stage label. The correlation coefficient between the time series features of phase change direction and each disease stage is calculated by giving a feature importance assessment or node division contribution, which is used to quantify the degree of correlation.

[0043] In addition, during the specific implementation process, the patient's multimodal physiological signals include electrocardiogram (ECG), electroencephalogram (EEG), and blood oxygen saturation signals.

[0044] Step 12: Calculate the mutual information between each local frequency domain feature and the label: in, Indicates the first Entropy of local frequency domain features Represents the entropy of the label. Indicates the first The joint entropy of local frequency domain features and labels, This represents the correlation coefficient between phase change and disease stage.

[0045] The correlation coefficient ranges from 0 to 1. When the correlation coefficient is close to 1, it indicates that the time series feature is highly correlated with the current disease stage. When the correlation coefficient is close to 0, it indicates that the time series feature is weakly correlated with the current disease stage and makes no significant contribution.

[0046] Step 13: Dynamically adjust the initial weights of each local frequency domain feature based on the mutual information: in, This indicates the number of segments in step 11.

[0047] Step 14: Weighted and merged local frequency domain features to obtain global frequency domain features: in, Indicates differences The initial weights of the values ​​are proportional.

[0048] Step 2: Adaptive frequency band selection. Using a genetic algorithm, the optimal combination of frequency bands in the global frequency domain features is selected.

[0049] Step 21: Define the initial population and construct the fitness function: in, The AUC of the ROC curve is represented by... Indicates the F1 score. Indicates the first Stability coefficient of the disease class.

[0050] It should be noted that the initial frequency band is a global frequency domain feature.

[0051] Additionally, it should be noted that the stability coefficient was obtained using a correlation method based on frequency band consistency, focusing on the similarity of trends and patterns exhibited between the same frequency bands of different patients.

[0052] In practice, the Pearson correlation coefficient is used to compare the similarity of characteristic curves of different patients in the same frequency band.

[0053] Step 22: Define a frequency band combination using chromosomes, and then perform genetic operations to generate a new frequency band combination.

[0054] The new frequency band combination is achieved by automatically searching the frequency bands containing stronger predictive information through the algorithm, discovering features with optimal predictive value, and optimizing the predictive performance of the model.

[0055] Step 23: For each frequency band combination, calculate the classification performance based on the training set, and obtain... and ,Will and The fitness score is obtained by weighting: in, express The weight, express The weight, Indicates the first Stability coefficient of the disease class.

[0056] In the specific implementation process, and The value range is 0~1 for easy normalization, and the initial value is 0.5. In the initial state, the two weights are given equal importance, and during training... It is used to control the overall numerical range, which facilitates optimization.

[0057] Step 24: Iterate and optimize until the fitness converges to obtain the optimal frequency band combination.

[0058] Step 3: Extract the frequency domain template of the disease, construct a disease-specific frequency domain feature library (DSFP), and calculate the matching score: Step 31: Using large-scale clinical data, extract frequency domain templates for each disease based on patient signals from different diseases: in, Indicates the first Frequency domain templates for similar diseases The dominant frequency, representing the spectral characteristics This represents the amplitude spectrum of the spectral features in the frequency domain. The phase spectrum represents the spectral characteristics in the frequency domain.

[0059] Step 32: Construct a disease-specific frequency domain feature library (DSFP) using frequency domain templates from different diseases. Step 33: Normalize the individual frequency bands and frequency domain templates of the input optimal frequency band combination: in, Represents the frequency domain characteristics of a single frequency band. This represents the frequency domain template.

[0060] Step 34: Calculate the matching score: in, Represents the frequency domain characteristics of a single frequency band. This represents the frequency domain template.

[0061] The matching score is used to determine the likelihood of the current disease.

[0062] Step 4: Train the time series prediction model based on the optimal frequency band combination obtained in Step 2 and the matching score obtained in Step 3. Use the trained prediction model to predict the disease progression and learn the pattern of disease changes over time.

[0063] Specifically, learning the patterns of disease changes over time involves combining short-term, medium-term, and long-term frequency domain features to dynamically capture multi-dimensional changes in the disease process.

[0064] In the specific implementation process, the training time-series prediction model uses historical patient data and performs supervised learning based on disease process labels. At the same time, the loss function used in the training process adopts cross-entropy loss.

[0065] Step 5: Use the trained prediction model to predict disease progression. Specifically, collect multimodal physiological signals from a new patient, convert the physiological data into time-frequency domain features using Morlet wavelet transform, perform adaptive frequency band selection to obtain the frequency domain features of the optimal frequency band, divide the time domain features into time scales, calculate the matching scores of the frequency domain features at different time scales, and input the matching scores and the frequency domain features of the optimal frequency band into the trained time-series prediction model to obtain the prediction results of future disease progression.

[0066] The different time scales are short-term window, medium-term window and long-term window.

[0067] It should be noted that in the process of obtaining the prediction results of future disease progression, the mutual information calculated in the matching score is also used to calculate the weight value, and further dynamically weighted and fused multi-scale features to optimize the temporal correlation.

[0068] In the specific implementation process, the short-term window is 10 seconds, the medium-term window is 5 minutes, and the long-term window is 1 hour. Meanwhile, the time series prediction model uses LSTM, and the input features are the frequency domain features of the time series signal at multiple time scales. Furthermore, recursive calculations are performed in the hidden layer to capture time correlations, and the time series of disease progression is predicted in the output layer.

[0069] Additionally, it should be noted that when making predictions, the Morlet wavelet transform preserves both frequency domain features and time domain features, i.e., time dimension information, while obtaining frequency domain features. Specifically, the Morlet wavelet transform has good time-frequency localization features, and the combination of complex exponents and Gaussian window functions enables it to accurately locate the time position of feature occurrences while performing spectral analysis.

[0070] like Figure 2 As shown, a disease progression prediction system based on frequency domain features is provided, including: a time-series prediction model training module and a real-time prediction model application module. The time-series prediction module application includes multi-dimensional frequency domain feature modeling, adaptive frequency band selection, construction of a disease-specific frequency domain feature library (DSFP), and training the time-series prediction module. The real-time prediction model application module includes data acquisition, obtaining the optimal frequency band, calculating the matching score with the disease-specific frequency domain feature library (DSFP), and predicting the future disease progression.

[0071] Example 2: Predicting the progression of Alzheimer's disease based on frequency domain features, including: Collect patients' electrocardiogram (ECG) and electroencephalogram (EEG) signals, convert the physiological data into time-frequency domain features using Morlet wavelet transform, and extract... While performing phase synchronization characteristics of the band, adaptive frequency band selection is carried out to obtain the frequency domain characteristics of the optimal frequency band. The time domain characteristics are divided into time scales, and the matching scores of the frequency domain characteristics at different time scales are calculated in combination with disease-specific frequency domain templates. The trained time series prediction model is used to capture the spectral changes in the patient's disease progression. Finally, edge devices are deployed to monitor the signal in real time and notify doctors to intervene when abnormalities occur.

[0072] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0073] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A disease progression prediction method based on frequency domain features, characterized in that, include: Step 1: Multidimensional frequency domain feature modeling, converting the original physiological signal into global frequency domain features; Step 2: Adaptive frequency band selection, using a genetic algorithm to select the optimal combination of frequency bands from the global frequency domain features; Step 3: Extract the frequency domain template of the disease, construct a disease-specific frequency domain feature library (DSFP), and calculate the matching score; Step 4: Train the time series prediction model based on the optimal frequency band combination obtained in Step 2 and the matching score obtained in Step 3, and use the trained prediction model to predict the disease progression. Step 5: Use the trained prediction model to predict disease progression.

2. The disease progression prediction method based on frequency domain features according to claim 1, characterized in that, Step 1 includes: Step 11: Acquire the patient's multimodal physiological signals, perform data preprocessing, segment the preprocessed signals, smooth the segment boundaries using a Hanning window, and convert the segmented signals into local frequency domain features using a Fast Fourier Transform. Meanwhile, a mapping relationship between the phase change direction time series and disease staging was established using a random forest model; Step 12: Calculate the mutual information between each local frequency domain feature and the label: in, Indicates the first Entropy of local frequency domain features Represents the entropy of the label. Indicates the first The joint entropy of local frequency domain features and labels, This represents the correlation coefficient between phase change and disease stage; Step 13: Dynamically adjust the initial weights of each local frequency domain feature based on the mutual information: in, Indicates the number of segments; Step 14: Weighted and merged local frequency domain features to obtain global frequency domain features: in, Indicates different The initial weights of the values ​​are proportional.

3. The disease progression prediction method based on frequency domain features according to claim 1, characterized in that, Step 2 includes: Step 21: Define the initial population and construct the fitness function: in, The AUC of the ROC curve is represented by... Indicates the F1 score. Indicates the first Stability coefficient of similar diseases; Step 22: Define a frequency band combination using chromosomes, and then perform genetic operations to generate a new frequency band combination; Step 23: For each frequency band combination, calculate the classification performance based on the training set, and obtain... and ,Will and The fitness score is obtained by weighting: in, express The weight, express The weight, Indicates the first Stability coefficient of similar diseases; Step 24: Iterate and optimize until the fitness converges to obtain the optimal frequency band combination.

4. The disease progression prediction method based on frequency domain features according to claim 1, characterized in that, Step 3 includes: Step 31: Using large-scale clinical data, extract frequency domain templates for each disease based on patient signals from different diseases: in, Indicates the first Frequency domain templates for similar diseases The dominant frequency, representing the spectral characteristics This represents the amplitude spectrum of the spectral features in the frequency domain. The phase spectrum represents the spectral characteristics in the frequency domain. Step 32: Construct a disease-specific frequency domain feature library (DSFP) using frequency domain templates from different diseases. Step 33: Normalize the individual frequency bands and frequency domain templates of the input optimal frequency band combination: in, Represents the frequency domain characteristics of a single frequency band. Represents a frequency domain template; Step 34: Calculate the matching score: According to claim 1, a disease progression prediction method based on frequency domain features is characterized in that step 5 specifically involves: collecting multimodal physiological signals from a new patient, converting the physiological data into time-frequency domain features through Morlet wavelet transform, performing adaptive frequency band selection, obtaining the frequency domain features of the optimal frequency band, dividing the time domain features into time scales, calculating the matching scores of the frequency domain features at different time scales, and inputting the matching scores and the frequency domain features of the optimal frequency band into the trained time-series prediction model to obtain the prediction results of future disease progression.

5. A disease progression prediction system based on frequency domain features, characterized in that, include: The system includes a time-series prediction model training module and a real-time prediction model application module. The time-series prediction module application includes multi-dimensional frequency domain feature modeling, adaptive frequency band selection, construction of a disease-specific frequency domain feature library (DSFP), and training the time-series prediction module. The real-time prediction model application module includes data acquisition, obtaining the optimal frequency band, calculating the matching score with the disease-specific frequency domain feature library (DSFP), and predicting the future disease progression.

Images