Longitudinal analysis-based lung cancer chemotherapy and immunotherapy combined patient symptom prediction method

Abstract

The invention discloses a longitudinal analysis-based lung cancer chemotherapy combined immunotherapy patient symptom prediction method, which comprises the following steps: S1, collecting a multi-modal medical data set of a patient, and carrying out data preprocessing; S2, carrying out time sequence integration on the preprocessed multi-modal medical data set to construct a longitudinal analysis data set; s3, constructing a symptom weight adaptive modeling method based on an attention mechanism; s4, forming a symptom-treatment causal relationship evaluation model; s5, inputting real-time multi-modal medical data of the patient into the symptom-treatment causal relationship evaluation model, and classifying prediction results; s6, generating an individual symptom risk assessment report of the patient according to a prediction result; s7, the clinician dynamically optimizes the chemotherapy dose, the immunotherapy dose and the drug use time of the patient based on the individualized symptom risk assessment report. The invention provides an effective technical means for accurately optimizing a lung cancer chemotherapy and immunotherapy combined scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of lung cancer chemotherapy, and in particular to a method for predicting symptoms of patients undergoing lung cancer chemotherapy combined with immunotherapy based on longitudinal analysis. Background Art

[0002] With the rapid development of chemotherapy and immunotherapy for lung cancer, combined treatment has gradually become an important strategy to improve patient prognosis and enhance treatment efficacy. However, the combined treatment of lung cancer with chemotherapy and immunotherapy also brings complex symptom management issues in clinical applications, including nausea and vomiting, fatigue, immune-related adverse reactions, and specific side effects of targeted drugs. Accurately predicting possible symptoms and their probability of occurrence is of great significance for optimizing treatment plans, reducing side effects, and improving patients' quality of life.

[0003] In recent years, deep learning technology has been gradually applied to the field of medical data analysis, providing new technical means for symptom prediction. Some methods attempt to use time series models to capture the dynamic changes in patients' treatment process, but time series models often only focus on single-modal data and lack the ability to deeply integrate multimodal data. In addition, existing symptom prediction technologies have obvious deficiencies in modeling the correlation of multimodal data, fail to effectively deal with the heterogeneous associations between medical images and biomarkers, and it is difficult to use patients' self-reported symptom information to personalize the calibration of the model.

[0004] On the other hand, traditional methods are less applicable in combined treatment models. The combination of chemotherapy and immunotherapy will produce complex combined effects, and there are significant individual differences in their symptom manifestations. Existing technologies are difficult to fully capture the dynamic characteristics of patients' individualized treatment responses. In addition, existing models generally lack the ability to analyze causal relationships of treatment adjustment plans, and it is difficult to evaluate the impact of different treatment plans on the probability of symptom occurrence, thereby failing to provide clinicians with sufficient data support and optimized decision-making recommendations.

[0005] In summary, the existing technology has obvious deficiencies in the dynamic analysis of symptom prediction, the fusion processing of multimodal data, and the causal relationship modeling under combined treatment. The defects of the existing technology not only affect the accuracy and personalization of the prediction, but also limit the ability to optimize the combined treatment plan. A new method is urgently needed to solve the above problems. Summary of the invention

[0006] One purpose of the present invention is to propose a method for predicting symptoms of patients receiving chemotherapy and immunotherapy for lung cancer based on longitudinal analysis. The present invention provides an effective technical means for accurately optimizing the combined chemotherapy and immunotherapy regimen for lung cancer.

[0007] According to an embodiment of the present invention, a method for predicting symptoms of patients undergoing chemotherapy combined with immunotherapy for lung cancer based on longitudinal analysis comprises the following steps:

[0008] S1. Collect multimodal medical data sets of patients and perform data preprocessing;

[0009] S2, integrating the preprocessed multimodal medical dataset into time series to construct a longitudinal analysis dataset;

[0010] S3. Construct a symptom weight adaptive modeling method based on the attention mechanism. The symptom weight adaptive modeling method dynamically assigns weights to different modal data and different time node data in the longitudinal analysis data set;

[0011] S4. Construct a generative adversarial model of symptom-treatment causal network based on symptom weight adaptive modeling method, and form a symptom-treatment causal relationship evaluation model;

[0012] S5. Inputting the patient's real-time multimodal medical data into the symptom-treatment causal relationship assessment model to predict the symptoms that may occur during chemotherapy and immunotherapy in the future and the probability of symptom occurrence, and classifying the prediction results;

[0013] S6. Generate an individualized symptom risk assessment report for the patient based on the prediction results, including the probability of symptom occurrence, symptom severity and time point, and provide the individualized symptom risk assessment report to the clinician;

[0014] S7. Clinicians dynamically optimize chemotherapy doses, immunotherapy doses, and drug use duration for patients based on individualized symptom risk assessment reports.

[0015] Optionally, the S1 includes the following steps:

[0016] S11. Collect multimodal medical dataset D of patients image , the multimodal medical dataset includes the patient’s medical imaging data D image , Laboratory test data D lab and patient self-reported symptom data symptom :

[0017] The patient’s medical imaging data included CT images reflecting tumor changes, lung inflammation, and immune response;

[0018] Laboratory test data include CT images reflecting tumor changes, lung inflammation, and immune response;

[0019] Patient self-reported symptom data included patients’ subjective ratings of chemotherapy-induced nausea and vomiting, fatigue, and immune-related adverse reactions induced by combined immunotherapy;

[0020] S12, using Gaussian filtering method to process medical image data D image Perform image denoising to generate denoised medical image datasets

[0021] S13. Laboratory test data D lab Perform standardization to obtain a standardized laboratory test data set

[0022] S14. Patient self-reported symptom data symptom Outlier processing was performed, and the outlier range was defined as [Q1-1.5·IQR, Q3+1.5·IQR]. The values ​​outside the outlier range were replaced with the nearest boundary values ​​to obtain the processed patient self-reported symptom data set.

[0023] S15, fill the missing values ​​of all data sets using linear interpolation method;

[0024] S16. Medical imaging dataset Standardized laboratory test data set and cleaned patient self-reported symptom dataset Integrate to obtain the final preprocessed multimodal medical dataset D pre .

[0025] Optionally, S2 includes the following steps:

[0026] S21. Preprocessed multimodal medical dataset D pre Perform timestamp marking, the timestamp t i , i=1,2,…,n is the actual time when the patient data is collected, generating a timestamp data set T={t1,t2,…,t n};

[0027] S22. Align the preprocessed medical imaging dataset, the standardized laboratory test dataset, and the cleaned patient self-reported symptom dataset in the order of timestamps to construct a time series alignment matrix M. aligned :

[0028]

[0029] in, Respectively represent the timestamp t i Corresponding medical imaging data, laboratory test data, and patient self-reported symptom data;

[0030] S23, align the time series matrix M alignedPerform sparsity detection. If missing time point data is detected, fill it in using the linear interpolation method.

[0031] S24, based on the time series alignment matrix M aligned Reorder multimodal medical data by time dimension to generate longitudinal analysis dataset D long :

[0032] D long ={(t1,F1),(t2,F2),…,(t n ,F n )};

[0033] in, Indicates time point t i The multimodal feature vector under .

[0034] Optionally, S3 includes the following steps:

[0035] S31. From the longitudinal analysis of data set D long Extract each time point t i The multimodal feature vector F i , the multimodal feature vector F i Indicates time point t i The following medical imaging characteristics, laboratory test characteristics and patient self-reported symptom characteristics;

[0036] S32. Construct a causal attention mechanism based on causal reasoning method. The causal attention mechanism only focuses on the current time point t i and the features of the previous time points, excluding the influence of future time points on the feature weights of the current time point;

[0037] S33. Define the causal attention mask matrix M causal :

[0038]

[0039] Among them, M causal [i,j] represents time point t j For time point t i The causal relevance of is assigned a non-zero weight only when j≤i;

[0040] S34, based on the causal attention mask matrix M causal Calculate each time point t i The attention weight distribution W(F i ):

[0041]

[0042] Where Q = F i ·Wq represents the query matrix, K = F i ·W k represents the bond matrix, W q and W k are the query weight matrix and the key weight matrix respectively, d k represents the feature dimension, ⊙ represents the element-by-element multiplication, which is used to convert the causal mask matrix M causal Applied to attention calculation;

[0043] S35, according to the calculated attention weight distribution W(F i ) for time point t i The multimodal feature vector F i Perform weighted processing to generate weighted feature vector

[0044]

[0045] S36. Arrange the weighted feature vectors of all time points in time series order to generate a weighted longitudinal analysis data set

[0046]

[0047] Optionally, S4 includes the following steps:

[0048] S41. Define the symptom-treatment causal network generative adversarial model M GAN The symptom-treatment causal network generative adversarial model consists of a generator G and a discriminator D, where the input of the generator G is a weighted longitudinal analysis dataset. The output is the predicted symptom sequence that the patient may experience in the future time period. The input of the discriminator D is the real symptom data and the generated symptom data, and the output is the evaluation probability of the authenticity of the input data;

[0049] S42. Weighted longitudinal analysis of the dataset Perform time window segmentation and construct the generator input data sequence X G ;

[0050] S43. Define the output of generator G as the predicted symptom sequence Y G ;

[0051] S44. For the real symptom data sequence Y real Generate symptom data series Y G For comparison, input the discriminator D, and the discriminator outputs the probability P of the authenticity of the input data D ;

[0052] S45. Generate a joint loss function L for adversarial models using symptom-treatment causal networksGAN Optimize the generator and discriminator;

[0053] S46. After the generator optimization is completed, the generator is used to predict the patient's symptom data sequence Y in the future time period. G ,The generated results include symptom types and their occurrence probabilities;

[0054] S47, the symptom prediction result Y output by the generator G A symptom-treatment causal relationship evaluation model was formed by linking it with the patient's current treatment plan to analyze the impact of different treatment adjustments on the probability of symptom occurrence.

[0055] Optionally, the S47 includes the following steps:

[0056] S471. Define the symptom-treatment causal relationship assessment model M causal , the symptom-treatment causal relationship evaluation model is based on the symptom prediction result Y output by the generator G and the patient's current treatment regimen T as input to evaluate the impact of different treatment adjustments on the probability of symptom occurrence;

[0057] S472. Construct symptom prediction result Y G and the joint distribution P(Y G ,T):

[0058] P(Y G ,T)=P(Y G |T)·P(T);

[0059] Among them, P(Y G |T) represents the conditional probability distribution of symptom prediction results under treatment plan T, and P(T) represents the prior distribution of treatment plans;

[0060] S473. Perform intervention analysis on the probability of symptom occurrence under different treatment adjustment schemes T' and define the symptom distribution after intervention as P(Y G |do(T ′ )):

[0061] P(Y G |do(T ′ ))=Σ T P(Y G |T ′ )·P(T ′ );

[0062] Among them, do(T ′ ) represents the intervention operation on the treatment plan T';

[0063] S474, based on the symptom prediction result Y Gand the distribution of symptoms after intervention P(Y G |do(T ′ ))Calculate the causal effect ΔP of the treatment adjustment on the probability of symptom occurrence:

[0064] ΔP=P(Y G |do(T ′ ))-P(Y G |T);

[0065] Among them, ΔP represents the increase or decrease in the probability of symptom occurrence caused by different treatment adjustments;

[0066] S475. Use the causal impact analysis results ΔP to construct the contribution matrix C of treatment adjustment to symptom occurrence:

[0067] C[i,j]=ΔP j |T i ;

[0068] Among them, C[i,j] represents the i Next, the adjusted contribution to symptom j;

[0069] S476. Optimize the symptom-treatment causal relationship assessment model M based on the treatment adjustment contribution matrix C and the causal relationship analysis results causal , generate a causal relationship assessment report, including the impact of treatment adjustments on the probability of occurrence of specific symptoms, the comprehensive effects of different treatment plans and the optimal treatment adjustment recommendations.

[0070] Optionally, the S476 includes the following steps:

[0071] S4761, based on the treatment adjustment contribution matrix C and symptom prediction results Y G Extract each treatment plan T i The comprehensive symptoms under the impact index S i :

[0072]

[0073] Among them, C[i,j] represents the treatment plan T i The adjusted contribution to symptom j, m represents the total number of symptoms, S i For treatment plan T i The combined effects of the following symptoms;

[0074] S4762. Define the comprehensive effect index of the treatment plan E i , the effect of reducing the probability of comprehensive symptoms and the potential side effects of the treatment plan on patients:

[0075] E i =w1·S i -w2·Ri ;

[0076] Among them, w1 and w2 are the comprehensive effect weight coefficients, S i is the comprehensive symptom impact index, R i Indicates treatment plan TT i Side effect scores;

[0077] S4763, based on comprehensive effect index E i Sort all treatment options and select the best treatment option T opt :

[0078]

[0079] Among them, T opt It is the treatment plan with the highest comprehensive effect index;

[0080] S4764. Generate a causal relationship assessment report using the ranking results and the causal relationship analysis of the treatment plan. The assessment report includes the following contents:

[0081] the extent to which each treatment option affects the probability of specific symptoms occurring;

[0082] Ranking of comprehensive effect indicators of different treatment options;

[0083] Recommended optimal treatment plan T opt and its scope of application.

[0084] 8. The method for predicting symptoms of patients undergoing chemotherapy combined with immunotherapy for lung cancer based on longitudinal analysis according to claim 1, wherein S5 comprises the following steps:

[0085] S51, inputting the patient's real-time multimodal medical data into the symptom-treatment causal relationship evaluation model, combining the symptom prediction results output by the generator and the patient's current treatment plan, calculating the symptoms that the patient may experience in the future time period and their probability of occurrence, and generating the patient's symptom prediction results;

[0086] S53. Classify the patient's symptom prediction results, and the classification content includes:

[0087] Chemotherapy-induced nausea and vomiting, including acute and delayed reactions;

[0088] Fatigue caused by the combined effects of chemotherapy and immunotherapy;

[0089] Immune-related adverse reactions that may occur with combined immunotherapy include pneumonia, rash, and endocrine dysfunction;

[0090] Symptoms related to chemotherapy-induced bone marrow suppression, including infection risk and general fatigue caused by leukopenia and anemia;

[0091] Specific side effects caused by targeted drugs, including skin toxicity and liver damage;

[0092] S54. Analyze the patient's comprehensive symptom risk based on the classification results, dynamically adjust the risk level of each symptom, and generate a comprehensive symptom risk score based on the patient's individual characteristics;

[0093] S55. Generate a patient symptom prediction classification report based on the classification results and comprehensive symptom risk score. The report includes the predicted probability of occurrence and risk score of each symptom, time trend analysis of each symptom, distinction between acute and chronic reactions, and personalized symptom management recommendations for current treatment options, including intervention strategies for high-risk symptoms.

[0094] The beneficial effects of the present invention are:

[0095] (1) The present invention introduces a symptom-treatment causal relationship evaluation model, and establishes a causal chain between the treatment plan and the probability of symptom occurrence based on the patient's real-time multimodal medical data and the symptom prediction results output by the generator. Compared with traditional methods based on statistical regression or static deep learning models, the present invention can dynamically analyze the direct impact of treatment adjustments on symptom occurrence, and calculate the symptom distribution after intervention through causal reasoning technology, providing quantitative causal impact indicators. The causal-driven prediction method solves the problem of the lack of evaluation of the effect of treatment adjustments in the prior art, so that the model not only has higher prediction accuracy, but also has stronger interpretability, providing reliable data support for clinicians to optimize treatment strategies.

[0096] (2) The present invention constructs a symptom weight adaptive modeling method that can dynamically weight and process multimodal medical data by combining longitudinal analysis and attention mechanism. Compared with the existing unimodal time series model, the present invention can capture the heterogeneous associations of different modal features in multimodal data. In particular, the attention mechanism allocates weights to different time points and modal features, so that the model can focus on feature areas that are strongly correlated with the occurrence of symptoms. The dynamic fusion strategy overcomes the limitation of the simple stacking of modal fusion in the prior art that leads to information redundancy or omission, and significantly improves the model's ability to capture complex features. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0098] Figure 1 This is a flow chart of a method for predicting symptoms in patients undergoing chemotherapy combined with immunotherapy for lung cancer based on longitudinal analysis proposed by the present invention. DETAILED DESCRIPTION

[0099] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0100] refer to Figure 1 , a method for predicting symptoms of patients with lung cancer undergoing chemotherapy combined with immunotherapy based on longitudinal analysis, comprising the following steps:

[0101] S1. Collect multimodal medical data sets of patients and perform data preprocessing;

[0102] S2, integrating the preprocessed multimodal medical dataset into time series to construct a longitudinal analysis dataset;

[0103] S3. Construct a symptom weight adaptive modeling method based on the attention mechanism. The symptom weight adaptive modeling method dynamically assigns weights to different modal data and different time node data in the longitudinal analysis data set;

[0104] S4. Construct a generative adversarial model of symptom-treatment causal network based on symptom weight adaptive modeling method, and form a symptom-treatment causal relationship evaluation model;

[0105] S5. Input the patient's real-time multimodal medical data into the symptom-treatment causal relationship assessment model to predict the symptoms and the probability of symptom occurrence that the patient may experience during chemotherapy and immunotherapy in the future time period, and classify the prediction results;

[0106] S6. Generate an individualized symptom risk assessment report for the patient based on the prediction results, including the probability of symptom occurrence, symptom severity and time point, and provide the individualized symptom risk assessment report to the clinician;

[0107] S7. Clinicians dynamically optimize chemotherapy doses, immunotherapy doses, and drug use duration for patients based on individualized symptom risk assessment reports.

[0108] In this implementation, S1 includes the following steps:

[0109] S11. Collect multimodal medical dataset D of patients image , the multimodal medical dataset includes the patient’s medical imaging data D image , Laboratory test data D lab and patient self-reported symptom data symptom :

[0110] The patient’s medical imaging data included CT images reflecting tumor changes, lung inflammation, and immune response;

[0111] Laboratory test data include CT images reflecting tumor changes, lung inflammation, and immune response;

[0112] Patient self-reported symptom data included patients’ subjective ratings of chemotherapy-induced nausea and vomiting, fatigue, and immune-related adverse reactions induced by combined immunotherapy;

[0113] S12, using Gaussian filtering method to process medical image data D image Perform image denoising to generate denoised medical image datasets

[0114] S13. Laboratory test data D lab Perform standardization to obtain a standardized laboratory test data set

[0115] S14. Patient self-reported symptom data symptom Outlier processing was performed, and the outlier range was defined as [Q1-1.5·IQR, Q3+1.5·IQR]. The values ​​outside the outlier range were replaced with the nearest boundary values ​​to obtain the processed patient self-reported symptom data set.

[0116] S15, fill the missing values ​​of all data sets using linear interpolation method;

[0117] S16. Medical imaging dataset Standardized laboratory test data set and cleaned patient self-reported symptom dataset Integrate to obtain the final preprocessed multimodal medical dataset D pre .

[0118] In this implementation, S2 includes the following steps:

[0119] S21. Preprocessed multimodal medical dataset D pre Perform timestamp marking, timestamp t i , i=1,2,…,n is the actual time when the patient data is collected, generating a timestamp data set T={t1,t2,…,t n};

[0120] S22. Align the preprocessed medical imaging dataset, the standardized laboratory test dataset, and the cleaned patient self-reported symptom dataset in the order of timestamps to construct a time series alignment matrix M. aligned :

[0121]

[0122] in, Respectively represent the timestamp t i Corresponding medical imaging data, laboratory test data, and patient self-reported symptom data;

[0123] S23, align the time series matrix M aligned Perform sparsity detection. If missing time point data is detected, fill it in using the linear interpolation method.

[0124] S24, based on the time series alignment matrix M aligned Reorder multimodal medical data by time dimension to generate longitudinal analysis dataset D long :

[0125] D long ={(t1,F1),(t2,F2),…,(t n ,F n )};

[0126] in, Indicates time point t i The multimodal feature vector under .

[0127] In this implementation, S3 includes the following steps:

[0128] S31. From the longitudinal analysis of data set D long Extract each time point t i The multimodal feature vector F i , the multimodal feature vector F i Indicates time point t i The following medical imaging characteristics, laboratory test characteristics and patient self-reported symptom characteristics;

[0129] S32. Construct a causal attention mechanism based on causal reasoning method. The causal attention mechanism only focuses on the current time point t i and the features of the previous time points, excluding the influence of future time points on the feature weights of the current time point;

[0130] S33. Define the causal attention mask matrix M causal :

[0131]

[0132] Among them, M causal [i,j] represents time point t j For time point t i The causal relevance of is assigned a non-zero weight only when j≤i;

[0133] S34, based on the causal attention mask matrix M causal Calculate each time point t iThe attention weight distribution W(F i ):

[0134]

[0135] Where Q = F i ·W q represents the query matrix, K = F i ·W k represents the bond matrix, W q and W k are the query weight matrix and the key weight matrix respectively, d k represents the feature dimension, ⊙ represents element-by-element multiplication, which is used to convert the causal mask matrix M causal Applied to attention calculation;

[0136] S35, according to the calculated attention weight distribution W(F i ) for time point t i The multimodal feature vector F i Perform weighted processing to generate weighted feature vector

[0137]

[0138] S36. Arrange the weighted feature vectors of all time points in time series order to generate a weighted longitudinal analysis data set

[0139]

[0140] In this implementation, S4 includes the following steps:

[0141] S41. Define the symptom-treatment causal network generative adversarial model M GAN The symptom-treatment causal network generative adversarial model consists of a generator G and a discriminator D, where the input of the generator G is a weighted longitudinal analysis dataset. The output is the predicted symptom sequence that the patient may experience in the future time period. The input of the discriminator D is the real symptom data and the generated symptom data, and the output is the evaluation probability of the authenticity of the input data;

[0142] S42. Weighted longitudinal analysis of the dataset Perform time window segmentation and construct the generator input data sequence X G ;

[0143] S43. Define the output of generator G as the predicted symptom sequence Y G ;

[0144] S44. For the real symptom data sequence Y real Generate symptom data series YG For comparison, input the discriminator D, and the discriminator outputs the probability P of the authenticity of the input data D ;

[0145] S45. Generate a joint loss function L for adversarial models using symptom-treatment causal networks GAN Optimize the generator and discriminator;

[0146] S46. After the generator optimization is completed, the generator is used to predict the patient's symptom data sequence Y in the future time period. G ,The generated results include symptom types and their occurrence probabilities;

[0147] S47, the symptom prediction result Y output by the generator G A symptom-treatment causal relationship evaluation model was formed by linking it with the patient's current treatment plan to analyze the impact of different treatment adjustments on the probability of symptom occurrence.

[0148] In this implementation, S47 includes the following steps:

[0149] S471. Define the symptom-treatment causal relationship assessment model M causal , the symptom-treatment causal relationship evaluation model is based on the symptom prediction result Y output by the generator G and the patient's current treatment regimen T as input to evaluate the impact of different treatment adjustments on the probability of symptom occurrence;

[0150] S472. Construct symptom prediction result Y G and the joint distribution P(Y G ,T):

[0151] P(Y G ,T)=P(Y G |T)·P(T);

[0152] Among them, P(Y G |T) represents the conditional probability distribution of symptom prediction results under treatment plan T, and P(T) represents the prior distribution of treatment plans;

[0153] S473. Perform intervention analysis on the probability of symptom occurrence under different treatment adjustment schemes T' and define the symptom distribution after intervention as P(Y G |do(T ′ )):

[0154] P(Y G |do(T ′ ))=Σ T P(Y G |T ′ )·P(T ′ );

[0155] Among them, do(T ′ ) represents the intervention operation on the treatment plan T';

[0156] S474, based on the symptom prediction result Y G and the distribution of symptoms after intervention P(Y G |do(T ′ ))Calculate the causal effect ΔP of the treatment adjustment on the probability of symptom occurrence:

[0157] ΔP=P(Y G |do(T ′ ))-P(Y G |T);

[0158] Among them, ΔP represents the increase or decrease in the probability of symptom occurrence caused by different treatment adjustments;

[0159] S475. Use the causal impact analysis results ΔP to construct the contribution matrix C of treatment adjustment to symptom occurrence:

[0160] C[i,j]=ΔP j |T i ;

[0161] Among them, C[i,j] represents the i Next, the adjusted contribution to symptom j;

[0162] S476. Optimize the symptom-treatment causal relationship assessment model M based on the treatment adjustment contribution matrix C and the causal relationship analysis results causal , generate a causal relationship assessment report, including the impact of treatment adjustments on the probability of occurrence of specific symptoms, the comprehensive effects of different treatment plans and the optimal treatment adjustment recommendations.

[0163] In this implementation, S476 includes the following steps:

[0164] S4761, based on the treatment adjustment contribution matrix C and symptom prediction results Y G Extract each treatment plan T i The comprehensive symptoms under the impact index S i :

[0165]

[0166] Among them, C[i,j] represents the treatment plan T i The adjusted contribution to symptom j, m represents the total number of symptoms, S i For treatment plan T i The combined effects of the following symptoms;

[0167] S4762. Define the comprehensive effect index of the treatment plan E i , the effect of reducing the probability of comprehensive symptoms and the potential side effects of the treatment plan on patients:

[0168] E i =w1·S i -w2·R i ;

[0169] Among them, w1 and w2 are the comprehensive effect weight coefficients, S i is the comprehensive symptom impact index, R i Indicates treatment plan TT i Side effect scores;

[0170] S4763, based on comprehensive effect index E i Sort all treatment options and select the best treatment option T opt :

[0171]

[0172] Among them, T opt It is the treatment plan with the highest comprehensive effect index;

[0173] S4764. Generate a causal relationship assessment report using the ranking results and the causal relationship analysis of the treatment plan. The assessment report includes the following contents:

[0174] the extent to which each treatment option affects the probability of specific symptoms occurring;

[0175] Ranking of comprehensive effect indicators of different treatment options;

[0176] Recommended optimal treatment plan T opt and its scope of application.

[0177] 8. A method for predicting symptoms of patients undergoing chemotherapy combined with immunotherapy for lung cancer based on longitudinal analysis according to claim 1, characterized in that S5 comprises the following steps:

[0178] S51, inputting the patient's real-time multimodal medical data into the symptom-treatment causal relationship evaluation model, combining the symptom prediction results output by the generator and the patient's current treatment plan, calculating the symptoms that the patient may experience in the future time period and their probability of occurrence, and generating the patient's symptom prediction results;

[0179] S53. Classify the patient's symptom prediction results, and the classification content includes:

[0180] Chemotherapy-induced nausea and vomiting, including acute and delayed reactions;

[0181] Fatigue caused by the combined effects of chemotherapy and immunotherapy;

[0182] Immune-related adverse reactions that may occur with combined immunotherapy include pneumonia, rash, and endocrine dysfunction;

[0183] Symptoms related to chemotherapy-induced bone marrow suppression, including infection risk and general fatigue caused by leukopenia and anemia;

[0184] Specific side effects caused by targeted drugs, including skin toxicity and liver damage;

[0185] S54. Analyze the patient's comprehensive symptom risk based on the classification results, dynamically adjust the risk level of each symptom, and generate a comprehensive symptom risk score based on the patient's individual characteristics;

[0186] S55. Generate a patient symptom prediction classification report based on the classification results and comprehensive symptom risk score. The report includes the predicted probability of occurrence and risk score of each symptom, time trend analysis of each symptom, distinction between acute and chronic reactions, and personalized symptom management recommendations for current treatment options, including intervention strategies for high-risk symptoms.

[0187] Embodiment 1:

[0188] Example In the lung cancer ward of a large comprehensive cancer treatment center, a 62-year-old male patient developed mild nausea and fatigue symptoms after receiving a combination of chemotherapy and immunotherapy. The patient is receiving a combination of chemotherapy (docetaxel combined with cisplatin) and a PD-1 inhibitor (pembrolizumab). The doctor plans to continue to increase the chemotherapy dose in the next two weeks to control tumor progression. However, since the patient has a strong risk of adverse reactions to chemotherapy in the past, especially nausea, vomiting and fatigue symptoms are more prominent, the doctor decided to use the method of the present invention to comprehensively predict and dynamically analyze the patient's future symptom risks in order to optimize the treatment plan.

[0189] First, the hospital data center extracted relevant multimodal medical data from the patient's electronic medical record system, including:

[0190] The patient's chest CT scan before treatment recorded the specific location and volume of the tumor (the maximum diameter of the lesion was 3.2 cm and the tumor volume was 24.5 cm 3 );

[0191] Laboratory test data included hemoglobin concentration (10.8 g / dL), white blood cell count (3.2 × 10 9 / L), and inflammatory factor CRP (26.7mg / L);

[0192] Patients' self-reported symptom scores showed a 4 for nausea (on a 0-10 scale) and a 5 for fatigue.

[0193] After the above data are input into the system of the present invention, the model automatically completes data cleaning, standardization and feature extraction. In CT data processing, the system eliminates noise through Gaussian filtering, and focuses on the lesion area to extract the tumor morphology complexity and edge clarity characteristics. In laboratory data, the system finds that the CRP value is higher than the normal range (the standard range is <10 mg / L) through outlier detection, indicating that the patient may be at risk of inflammation.

[0194] Next, the system used a longitudinal analysis model to dynamically integrate the patient's data from the past six months, and combined the attention mechanism to weight different modal data and time nodes. The system determined that the patient's self-reported symptom scores had significant fluctuations in the past chemotherapy cycle (the nausea score peaked at 7 points on the 3rd day and dropped to 3 points on the 10th day). Therefore, the features of these nodes were weighted amplified. The model weight showed that the feature weight of nausea symptoms increased from the initial 0.3 to 0.7.

[0195] Subsequently, a generative adversarial network (GAN) was used to predict the symptoms that the patient might experience in the next two weeks. The generator simulated the evolution of the patient's symptoms under various treatment options. Under the condition of maintaining the current treatment option, the prediction results showed:

[0196] The probability of nausea symptoms was 85%, of which the probability of severe nausea (score>7 points) was 45%;

[0197] The probability of fatigue symptoms was 78%;

[0198] Immune-related rash occurred in 12%.

[0199] The discriminator verifies the authenticity of the generated results and finally outputs a prediction report.

[0200] Based on the symptom prediction results, the system further uses causal reasoning models to analyze the impact of different treatment adjustment plans on symptom occurrence. For example:

[0201] If the chemotherapy dose is reduced by 20%, the incidence of nausea will drop from 85% to 55%, and the probability of severe nausea will drop to 20%;

[0202] If the immunotherapy interval is extended to three weeks, the incidence of immune-related rash can be reduced to 5%, but the tumor control rate may decrease slightly (estimated to decrease by 2.5%).

[0203] Based on the analysis results, the system generated a symptom classification report, which clearly recorded the main symptoms that the patient might face in the next two weeks, including their severity and probability of occurrence. The doctor adjusted the treatment plan based on the report content, reducing the chemotherapy dose by 15%, and carried out drug intervention in advance to prevent and treat nausea and vomiting in the patient (using ondansetron and dexamethasone). In the following treatment cycle, the patient's actual symptoms were highly consistent with the predicted results: the nausea score did not exceed 6 points, the fatigue score was stable below 4 points, and no immune-related rash appeared.

[0204] Comparative data show the significant advantages of the method of the present invention over traditional methods. Compared with doctors who previously adjusted treatment plans based solely on experience, the method of the present invention accurately quantifies symptom risk and provides an optimization plan through causal analysis of different treatment adjustments. The actual symptom severity scores of patients were reduced by more than 30% on average compared with those who did not use this method. The effect of treatment plan adjustment was significantly improved, and the overall patient satisfaction reached 92%.

[0205] The examples fully demonstrate the feasibility and effectiveness of the method of the present invention in complex clinical scenarios, and provide reliable technical support for precision medicine.

[0206] The present invention introduces a symptom-treatment causal relationship evaluation model, and establishes a causal chain between the treatment plan and the probability of symptom occurrence based on the patient's real-time multimodal medical data and the symptom prediction results output by the generator. Compared with traditional methods based on statistical regression or static deep learning models, the present invention can dynamically analyze the direct impact of treatment adjustments on symptom occurrence, and calculate the symptom distribution after intervention through causal reasoning technology, providing quantitative causal impact indicators. The causal-driven prediction method solves the deficiency of the existing technology in the lack of evaluation of the effect of treatment adjustments, so that the model not only has higher prediction accuracy, but also has stronger interpretability, providing reliable data support for clinicians to optimize treatment strategies.

[0207] The present invention constructs a symptom weight adaptive modeling method that can dynamically weight and process multimodal medical data by combining longitudinal analysis and attention mechanism. Compared with the existing unimodal time series model, the present invention can capture the heterogeneous associations of different modal features in multimodal data, especially through the attention mechanism to allocate weights to different time points and modal features, so that the model can focus on feature areas that are strongly correlated with the occurrence of symptoms. The dynamic fusion strategy overcomes the limitation of simple stacking of modal fusion in the prior art that leads to information redundancy or omission, and significantly improves the model's ability to capture complex features.

[0208] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting symptoms of patients with lung cancer undergoing chemotherapy combined with immunotherapy based on longitudinal analysis, characterized in that: The steps include: S1. Collect multimodal medical data sets of patients and perform data preprocessing; S2, integrating the preprocessed multimodal medical dataset into time series to construct a longitudinal analysis dataset; S3. Construct a symptom weight adaptive modeling method based on the attention mechanism. The symptom weight adaptive modeling method dynamically assigns weights to different modal data and different time node data in the longitudinal analysis data set; S4. Construct a generative adversarial model of symptom-treatment causal network based on symptom weight adaptive modeling method, and form a symptom-treatment causal relationship evaluation model; S5. Inputting the patient's real-time multimodal medical data into the symptom-treatment causal relationship assessment model to predict the symptoms that may occur during chemotherapy and immunotherapy in the future and the probability of symptom occurrence, and classifying the prediction results; S6. Generate an individualized symptom risk assessment report for the patient based on the prediction results, including the probability of symptom occurrence, symptom severity and time point, and provide the individualized symptom risk assessment report to the clinician; S7. Clinicians dynamically optimize chemotherapy doses, immunotherapy doses, and drug use duration for patients based on individualized symptom risk assessment reports.

2. A method for predicting symptoms of patients with lung cancer undergoing chemotherapy combined with immunotherapy based on longitudinal analysis according to claim 1, characterized in that: The S1 comprises the following steps: S11. Collect multimodal medical dataset D of patients image , the multimodal medical dataset includes the patient’s medical imaging data D image , Laboratory test data D lab and patient self-reported symptom data symptom : The patient’s medical imaging data included CT images reflecting tumor changes, lung inflammation, and immune response; Laboratory test data include CT images reflecting tumor changes, lung inflammation, and immune response; Patient self-reported symptom data included patients’ subjective ratings of chemotherapy-induced nausea and vomiting, fatigue, and immune-related adverse reactions induced by combined immunotherapy; S12, using Gaussian filtering method to process medical image data D image Perform image denoising to generate denoised medical image datasets S13. Laboratory test data D lab Perform standardization to obtain a standardized laboratory test data set S14. Patient self-reported symptom data symptom Outlier processing was performed, and the outlier range was defined as [Q1-1.5·IQR, Q3+1.5·IQR]. The values ​​outside the outlier range were replaced with the nearest boundary values ​​to obtain the processed patient self-reported symptom data set. S15, fill the missing values ​​of all data sets using linear interpolation method; S16. Medical imaging dataset Standardized laboratory test data set and cleaned patient self-reported symptom dataset Integrate to obtain the final preprocessed multimodal medical dataset D pre .

3. The method for predicting symptoms of patients with lung cancer undergoing chemotherapy combined with immunotherapy based on longitudinal analysis according to claim 1, characterized in that: The S2 comprises the following steps: S21. Preprocessed multimodal medical dataset D pre Perform timestamp marking, the timestamp t i , i=1,2,…,n is the actual time when the patient data is collected, generating a timestamp data set T={t1,t2,…,t n }; S22. Align the preprocessed medical imaging dataset, the standardized laboratory test dataset, and the cleaned patient self-reported symptom dataset in the order of timestamps to construct a time series alignment matrix M. aligned : in, Respectively represent the timestamp t i Corresponding medical imaging data, laboratory test data, and patient self-reported symptom data; S23, align the time series matrix M aligned Perform sparsity detection. If missing time point data is detected, fill it in using the linear interpolation method. S24, based on the time series alignment matrix M aligned Reorder multimodal medical data by time dimension to generate longitudinal analysis dataset D long : D long ={(t1,F1),(t2,F2),…,(t n ,F n )}; in, Indicates time point t i The multimodal feature vector under .

4. The method for predicting symptoms of patients undergoing chemotherapy combined with immunotherapy for lung cancer based on longitudinal analysis according to claim 1, characterized in that: The S3 comprises the following steps: S31. From the longitudinal analysis of data set D long Extract each time point t i The multimodal feature vector F i , the multimodal feature vector F i Indicates time point t i The following medical imaging characteristics, laboratory test characteristics and patient self-reported symptom characteristics; S32. Construct a causal attention mechanism based on causal reasoning method. The causal attention mechanism only focuses on the current time point t i and the features of the previous time points, excluding the influence of future time points on the feature weights of the current time point; S33. Define the causal attention mask matrix M causal : Among them, M causal [i,j] represents time point t j For time point t i The causal relevance of is assigned a non-zero weight only when j≤i; S34, based on the causal attention mask matrix M causal Calculate each time point t i The attention weight distribution W(F i ): Where Q = F i ·W q represents the query matrix, K = F i ·W k represents the bond matrix, W q and W k are the query weight matrix and the key weight matrix respectively, d k represents the feature dimension, ⊙ represents element-by-element multiplication, which is used to convert the causal mask matrix M causal Applied to attention calculation; S35, according to the calculated attention weight distribution W(F i ) for time point t i The multimodal feature vector F i Perform weighted processing to generate weighted feature vector S36. Arrange the weighted feature vectors of all time points in time series order to generate a weighted longitudinal analysis data set 5. The method for predicting symptoms of patients with lung cancer undergoing chemotherapy combined with immunotherapy based on longitudinal analysis according to claim 1, characterized in that: The S4 comprises the following steps: S41. Define the symptom-treatment causal network generative adversarial model M GAN The symptom-treatment causal network generative adversarial model consists of a generator G and a discriminator D, where the input of the generator G is a weighted longitudinal analysis dataset. The output is the predicted symptom sequence that the patient may experience in the future time period. The input of the discriminator D is the real symptom data and the generated symptom data, and the output is the evaluation probability of the authenticity of the input data; S42. Weighted longitudinal analysis of the dataset Perform time window segmentation and construct the generator input data sequence X G ; S43. Define the output of generator G as the predicted symptom sequence Y G ; S44. For the real symptom data sequence Y real Generate symptom data series Y G For comparison, input the discriminator D, and the discriminator outputs the probability P of the authenticity of the input data D ; S45. Generate a joint loss function L for adversarial models using symptom-treatment causal networks GAN Optimize the generator and discriminator; S46. After the generator optimization is completed, the generator is used to predict the patient's symptom data sequence Y in the future time period. G ,The generated results include symptom types and their occurrence probabilities; S47, the symptom prediction result Y output by the generator G A symptom-treatment causal relationship evaluation model was formed by linking it with the patient's current treatment plan to analyze the impact of different treatment adjustments on the probability of symptom occurrence.

6. A method for predicting symptoms of patients undergoing chemotherapy combined with immunotherapy for lung cancer based on longitudinal analysis according to claim 5, characterized in that: The S47 comprises the following steps: S471. Define the symptom-treatment causal relationship assessment model M causal , the symptom-treatment causal relationship evaluation model is based on the symptom prediction result Y output by the generator G and the patient's current treatment regimen T as input to evaluate the impact of different treatment adjustments on the probability of symptom occurrence; S472. Construct symptom prediction result Y G and the joint distribution P(Y G ,T): P(Y G ,T)=P(Y G |T)·P(T); Among them, P(Y G |T) represents the conditional probability distribution of symptom prediction results under treatment plan T, and P(T) represents the prior distribution of treatment plans; S473. Perform intervention analysis on the probability of symptom occurrence under different treatment adjustment schemes T' and define the symptom distribution after intervention as P(Y G |do(T ′ )): P(Y G |do(T ′ ))=Σ T P(Y G |T ′ )·P(T ′ ); Among them, do(T ′ ) represents the intervention operation on the treatment plan T'; S474, based on the symptom prediction result Y G and the distribution of symptoms after intervention P(Y G |do(T ′ ))Calculate the causal effect ΔP of the treatment adjustment on the probability of symptom occurrence: ΔP=P(Y G |do(T ′ ))-P(Y G |T); Among them, ΔP represents the increase or decrease in the probability of symptom occurrence caused by different treatment adjustments; S475. Use the causal impact analysis results ΔP to construct the contribution matrix C of treatment adjustment to symptom occurrence: C[i,j]=ΔP j |T i ; Among them, C[i,j] represents the i Next, the adjusted contribution to symptom j; S476. Optimize the symptom-treatment causal relationship assessment model M based on the treatment adjustment contribution matrix C and the causal relationship analysis results causal , generate a causal relationship assessment report, including the impact of treatment adjustments on the probability of occurrence of specific symptoms, the comprehensive effects of different treatment plans and the optimal treatment adjustment recommendations.

7. A method for predicting symptoms of patients undergoing chemotherapy combined with immunotherapy for lung cancer based on longitudinal analysis according to claim 6, characterized in that: The S476 comprises the following steps: S4761, based on the treatment adjustment contribution matrix C and symptom prediction results Y G Extract each treatment plan T i The comprehensive symptoms under the impact index S i : Among them, C[i,j] represents the treatment plan T i The adjusted contribution to symptom j, m represents the total number of symptoms, S i For treatment plan T i The combined effects of the following symptoms; S4762. Define the comprehensive effect index of the treatment plan E i , the effect of reducing the probability of comprehensive symptoms and the potential side effects of the treatment plan on patients: From i =w1·S i -w2·R i ; Among them, w1 and w2 are the comprehensive effect weight coefficients, S i is the comprehensive symptom impact index, R i Indicates treatment plan TT i Side effect scores; S4763, based on comprehensive effect index E i Sort all treatment options and select the best treatment option T opt : Among them, T opt It is the treatment plan with the highest comprehensive effect index; S4764. Generate a causal relationship assessment report using the ranking results and the causal relationship analysis of the treatment plan. The assessment report includes the following contents: the extent to which each treatment option affects the probability of specific symptoms occurring; Ranking of comprehensive effect indicators of different treatment options; Recommended optimal treatment plan T opt and its scope of application.

8. The method for predicting symptoms of patients with lung cancer undergoing chemotherapy combined with immunotherapy based on longitudinal analysis according to claim 1, characterized in that: The S5 comprises the following steps: S51, inputting the patient's real-time multimodal medical data into the symptom-treatment causal relationship evaluation model, combining the symptom prediction results output by the generator and the patient's current treatment plan, calculating the symptoms that may occur in the patient in the future time period and their probability of occurrence, and generating the patient's symptom prediction results; S53. Classify the patient's symptom prediction results, and the classification content includes: Chemotherapy-induced nausea and vomiting, including acute and delayed reactions; Fatigue caused by the combined effects of chemotherapy and immunotherapy; Immune-related adverse reactions that may occur with combined immunotherapy include pneumonia, rash, and endocrine dysfunction; Symptoms related to chemotherapy-induced bone marrow suppression, including infection risk and general fatigue caused by leukopenia and anemia; Specific side effects caused by targeted drugs, including skin toxicity and liver damage; S54. Analyze the patient's comprehensive symptom risk based on the classification results, dynamically adjust the risk level of each symptom, and generate a comprehensive symptom risk score based on the patient's individual characteristics; S55. Generate a patient symptom prediction classification report based on the classification results and comprehensive symptom risk score. The report includes the predicted probability of occurrence and risk score of each symptom, time trend analysis of each symptom, distinction between acute and chronic reactions, and personalized symptom management recommendations for current treatment options, including intervention strategies for high-risk symptoms.

Images