Multiple myeloma prognosis data processing method

Abstract

The invention discloses a multiple myeloma prognostic data processing method, which relates to the field of medical information processing, and comprises the following steps: step 1, periodically acquiring a clonal plasma cell proportion, M protein concentration time sequence data, a bone metabolism marker and an imaging bone focus volume change coefficient, establishing a time alignment matrix, and establishing a time alignment matrix; irregular sampling intervals are unified through dynamic time warping; 2, performing wavelet decomposition on the M protein trajectory, extracting short-term fluctuation and long-term trend components, calculating a mutation intensity index of a cloned plasma cell index, and extracting bone focus spatial distribution characteristics through a 3D convolutional network; through a multi-dimensional analysis framework, data changes are monitored in real time, the causal relationship among different factors is quantified, the disease development of a patient is effectively predicted, a doctor is helped to visually understand the spatial distribution and risk degree of bone lesions, then the disease deterioration risk is recognized as early as possible, and personalized treatment suggestions are provided for patients of different risk groups.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information processing, and in particular to a method for processing prognosis data of multiple myeloma. Background Art

[0002] Multiple myeloma is a malignant tumor of the lymphatic system, mainly caused by abnormal plasma cell clone proliferation, which often leads to clinical manifestations such as bone lesions, anemia, renal impairment and immunosuppression. Therefore, its prognosis assessment is extremely important. Modern medical research pays more and more attention to understanding the disease mechanism in a data-driven way. The cooperation between clinical researchers and bioinformaticians is getting closer. Because the treatment effect and disease progression vary, clinicians have an increasing demand for dynamic monitoring of patients' conditions and personalized treatment. With the aging population and the increase in patients with related cancers, the treatment and management of multiple myeloma is facing economic pressure, which in turn requires doctors to be able to manage patients with greater efficiency and effectiveness, driving the need for technological advances and data-driven prognostic models; Early identification of bone lesion risk can help doctors adjust treatment plans and take appropriate intervention measures in advance, thereby improving patients' prognosis. However, traditional methods are often unable to effectively integrate data of different types and sources, resulting in information loss during the analysis process. They are unable to effectively capture the dynamic changes of clonal plasma cells and bone metabolism indicators in multiple myeloma, and lack sufficient accuracy in predicting changes in patients' conditions. Summary of the invention

[0003] 1. Technical issues to be resolved In view of the above-mentioned shortcomings of the prior art, the present invention provides a multiple myeloma prognosis data processing method, which can effectively solve the problems of the prior art.

[0004] (II) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: The present invention discloses a multiple myeloma prognosis data processing method, comprising the following steps: Step 1: Periodically obtain the time series data of clonal plasma cell ratio, M protein concentration, bone metabolism markers, and imaging bone lesion volume change coefficient, establish a time alignment matrix, and use dynamic time regularization to unify irregular sampling intervals; Step 2: Perform wavelet decomposition on the M protein trajectory to extract short-term fluctuation and long-term trend components, calculate the mutation intensity index of clonal plasma cell indicators, and extract the spatial distribution characteristics of bone lesions through a 3D convolutional network; Step 3: Based on the data obtained in step 2, a dual-stream analysis model is constructed to quantify the influence weight of the output clonal evolution signal on the bone microenvironment; Step 4: Analyze the correlation coefficient between the changes in clonal plasma cells and the occurrence of bone lesions through key causal verification links; Step 5: Based on the dual-stream analysis model obtained in step 3, the probability of regression, stabilization or progression of the patient's disease in the next several cycles is predicted; Step 6: Based on the correlation coefficient obtained in step 4 and the prediction result obtained in step 5, a joint risk index is defined, and patients are stratified into low, medium and high risk groups based on the joint risk index, triggering corresponding warning prompts; Step 7: Generate a bone lesion heat map based on the patient's joint risk index and the spatial distribution characteristics of bone lesions, and mark the key bone areas related to risk; Step 8: Based on the marked critical bone areas, index them in the preset opinion library and generate intervention suggestions in order based on the risk level.

[0005] Furthermore, in the dual-stream analysis model in step 3, stream one is used to process the M protein concentration and clonal plasma cell time series, and stream two is used to analyze bone metabolism markers and imaging features. The output feature vectors of the two streams are concatenated in dimension to form a comprehensive feature vector, and the fused features are input into a fully connected layer for further feature extraction and processing.

[0006] Furthermore, in the critical causal verification link of step 4, a longitudinal causal discovery algorithm is used. Under the condition of controlling treatment interference, a double robust estimation is used to quantify the independent effect of clonal evolution on bone lesions, and the increase in the risk of bone lesions corresponding to different growth rates of M protein is obtained and integrated as the correlation coefficient.

[0007] Furthermore, the dual-stream analysis model in step 5 outputs the transition probability of the clonal state through the Bayesian neural network, and combines the Monte Carlo Dropout technology to evaluate the prediction confidence. When the probability of progression exceeds the corresponding threshold, a graded warning is triggered, and a state transition matrix is ​​constructed to identify the high-risk transition window period. The calculation formula for the probability of regression, stability or progression of the patient's disease in the next few cycles is: ; In the formula, represents the predicted probability of the patient being in state s in the time window t months, s∈{regress, stable, progressive}, Represents the number of Monte Carlo Dropout sampling, represents the state-specific weight matrix of state s, represents the LSTM hidden state vector, represents the feature encoder, represents the longitudinal time series feature matrix, represents the bias term of state s, Represents the preset penalty intensity coefficient, represents the prediction standard deviation, Represents the correction of the predicted probability in the form of exponential decay.

[0008] Furthermore, the prediction standard deviation The expression is: ; In the formula, represents the original predicted probability of state s at the kth sampling time, reflecting the prediction result of a single forward propagation, Represents the probability mean of K samples, eliminates the baseline prediction probability after random disturbance, calculates the discreteness of the prediction result, quantifies the uncertainty level of the model's judgment on the current patient status, and restores the variance dimension to the same scale as the original probability.

[0009] Furthermore, the longitudinal time series feature matrix The expression is: ; In the formula, Represents the M protein kinetics index, including serum concentration, monthly change rate and area under the curve, Representative clonal plasma cell characteristics, including the proportion of abnormal clonal plasma cells with CD138, flow cytometry clonality index, and cytogenetic risk score, represents bone metabolism indicators, including CTX / NTX ratio, MRI lesion volume and PET-CT uptake value, and u represents the time slice dimension, which indicates the time mark of data collection and analysis.

[0010] Furthermore, the working logic of the dual-stream analysis model in step 5 for analyzing the probability of regression, stabilization or progression of the patient's disease in the next several cycles is: The patient's continuous monitoring data is input, spatial features are extracted through a three-layer residual network, and a multi-head temporal attention mechanism is superimposed to capture cross-period dependencies; Input the encoded features into the LSTM network and update the hidden state; Enable the Dropout layer and perform several random forward passes; Calculate prediction statistics and perform probability calibration; Feed the final prediction result back to the LSTM hidden state; Generate a visual trajectory map and mark key turning points.

[0011] Furthermore, in step 6, a combined risk index is defined, and the patients are stratified into low, medium and high risk based on the combined risk index, including: Step 61: quantifying the proportion of clonal plasma cells, M protein concentration, bone metabolism markers, and bone lesion volume change indexes into a unified scoring standard; Step 62: assign appropriate weights to each selected indicator based on the impact and effectiveness of the indicator on the bone microenvironment and disease progression; Step 63: multiply the quantified value of each indicator by the corresponding weight, and then add them up to calculate the joint risk index of each patient; Step 64: According to the calculated combined risk index, a critical value is established based on ROC curve analysis to stratify patients into low, medium and high risk groups.

[0012] Furthermore, the bone lesion heat map generated in step 7 includes: spatial distribution characteristics, lesion risk level, lesion characteristics, time dynamics and clinically relevant marker information.

[0013] Furthermore, the preset opinion library in step 8 actively adds intervention treatment strategies manually and supports cloud sharing.

[0014] (III) Beneficial effects Compared with the known prior art, the technical solution provided by the present invention has the following beneficial effects: By integrating the proportion of clonal plasma cells, M protein concentration and bone metabolism markers, a multi-dimensional analysis framework is formed. The complex pathological mechanism of multiple myeloma is identified with a comprehensive analysis framework. Through time series analysis and dynamic time warping technology, data changes are monitored in real time, and the patient's disease progression is effectively predicted, thereby identifying the risk of disease worsening early.

[0015] Wavelet transform and 3D convolutional network are used to obtain more accurate data features, effectively extract short-term fluctuations and long-term trends, and enhance the predictive ability of the model. Through the longitudinal causal discovery algorithm, the causal relationship between different factors can be quantified under the condition of controlling relevant interference factors, thus ensuring the reliability of the model prediction results.

[0016] Based on the combined risk index, patients can be divided into low, medium, and high risk groups, and personalized treatment recommendations can be provided for patients in different risk groups to enhance the targeted clinical treatment. The generated hotspot maps and visual trajectory maps can help doctors intuitively understand the spatial distribution and risk level of bone lesions, thereby providing important reference for clinical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 The figure is a flow chart of defining the combined risk index in step 6 of the present invention and stratifying patients into low, medium and high risk groups based on the combined risk index. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 making creative work are within the scope of protection of the present invention.

[0020] The present invention will be further described below in conjunction with the embodiments.

[0021] Embodiment 1: The multiple myeloma prognosis data processing method of this embodiment is as follows: Figure 1 As shown, the following steps are included: Step 1: Periodically obtain the time series data of clonal plasma cell ratio, M protein concentration, bone metabolism markers, and imaging bone lesion volume change coefficient, establish a time alignment matrix, and use dynamic time regularization to unify irregular sampling intervals; Step 2: Perform wavelet decomposition on the M protein trajectory to extract short-term fluctuation and long-term trend components, calculate the mutation intensity index of clonal plasma cell indicators, and extract the spatial distribution characteristics of bone lesions through a 3D convolutional network; Step 3: Based on the mutation intensity index of the clonal plasma cell index obtained in step 2 and the spatial distribution characteristics of bone lesions, a dual-stream analysis model is constructed to quantify the influence weight of the output clonal evolution signal on the bone microenvironment; In the dual-stream analysis model, stream one is used to process the M protein concentration and clonal plasma cell time series, and stream two is used to analyze bone metabolism markers and imaging features. The output feature vectors of the two streams are spliced ​​in dimension to form a comprehensive feature vector, and the fused features are input into a fully connected layer for further feature extraction and processing; Step 4: Analyze the correlation coefficient between the changes in clonal plasma cells and the occurrence of bone lesions through key causal verification links; In the key causal verification phase, a longitudinal causal discovery algorithm was used to quantify the independent effect of clonal evolution on bone lesions using double robust estimation under the condition of controlling treatment interference, and the magnitude of the increase in bone lesion risk corresponding to different magnitudes of M protein growth rate was obtained and integrated as the correlation coefficient; Step 5: Based on the dual-stream analysis model obtained in step 3, the probability of regression, stabilization or progression of the patient's disease in the next several cycles is predicted; The working logic of the dual-stream analysis model for analyzing the probability of regression, stabilization or progression of a patient's disease in the next few cycles is as follows: The patient's continuous monitoring data is input, spatial features are extracted through a three-layer residual network, and a multi-head temporal attention mechanism is superimposed to capture cross-period dependencies; Input the encoded features into the LSTM network and update the hidden state; Enable the Dropout layer and perform several random forward passes; Calculate prediction statistics and perform probability calibration; Feed the final prediction result back to the LSTM hidden state; Generate visual trajectory maps and mark key turning points; Step 6: Based on the correlation coefficient obtained in step 4 and the prediction result obtained in step 5, a joint risk index is defined, and patients are stratified into low, medium and high risk groups based on the joint risk index, triggering corresponding warning prompts; Step 7: Generate a bone lesion heat map based on the patient's joint risk index and the spatial distribution characteristics of bone lesions, and mark the key bone areas related to risk; The bone lesion heat map includes: Spatial distribution characteristics: coordinate information of each bone part, indicating the specific bone area associated with the lesion; Lesion risk level: The bone lesion area is divided into low, medium and high risk levels according to the joint risk index, and the statistical confidence level of the risk value is provided to indicate the reliability of the prediction results; Lesion characteristics: the number of lesions detected in a specific area, indicating the prevalence of bone lesions; the severity of the lesions, with larger lesions generally corresponding to higher risk; the type of lesions; Time dynamics: Combine the time dimension to show the changes in hot spots at different time points; Clinically relevant biomarker information: statistical information on biomarkers relevant to specific risk areas, as well as specific patient data; Step 8: Based on the marked key bone areas, index them in the preset opinion library and generate intervention suggestions in order based on the risk level; the preset opinion library actively adds intervention treatment strategies through manual work and supports cloud sharing.

[0022] Compared with existing technologies, by establishing a time alignment matrix and dynamic time warping, irregular sampling data of clonal plasma cell ratios, M protein concentrations, bone metabolism markers and imaging data from different sources can be uniformly processed, greatly improving the efficiency of data utilization. The use of wavelet decomposition technology to extract short-term fluctuations and long-term trends, combined with spatial feature extraction of 3D convolutional networks, can more comprehensively capture the development dynamics of M protein and bone lesions, which is often difficult to achieve in traditional methods. Through the dual-stream analysis mechanism, two different data streams can be integrated to form a comprehensive feature vector, which enhances the system's ability to capture complex pathological changes. The introduction of longitudinal causal discovery algorithms and dual robust estimation not only focuses on correlation, but also can quantitatively analyze the independent effects of clonal plasma cell changes on bone lesions, providing a deeper understanding of biology. The residual network and LSTM combined with the multi-head time attention mechanism can effectively extract the complex features of the time series and provide more accurate disease prediction, breaking through the limitations of traditional models in processing time series data. Generating a heat map of bone lesions based on spatial distribution characteristics can intuitively display high-risk areas, provide effective visualization support for clinical intervention, enhance attention to high-risk areas of bone lesions, and based on detailed risk stratification and early warning mechanisms, provide patients with personalized intervention strategies, thereby improving the pertinence and effectiveness of clinical treatment.

[0023] Embodiment 2: In other aspects, this embodiment provides another optimization strategy, specifically a formula for calculating the probability of regression, stabilization or progression of the patient's disease in the next several cycles. The dual-stream analysis model outputs the transition probability of the clone state through the Bayesian neural network, and the Monte Carlo Dropout technology is combined to evaluate the prediction confidence. When the probability of progression exceeds the corresponding threshold, a graded warning is triggered, and a state transition matrix is ​​constructed to identify the high-risk turning window period. The calculation formula for the probability of regression, stabilization or progression of the patient's disease in the next several cycles is specifically: ; In the formula, represents the predicted probability of the patient being in state s in the time window t months, s∈{regress, stable, progressive}, Represents the number of Monte Carlo Dropout sampling, represents the state-specific weight matrix of state s, represents the LSTM hidden state vector, represents the feature encoder, specifically a three-layer residual network, represents the longitudinal time series feature matrix, represents the bias term of state s, Represents the preset penalty intensity coefficient, represents the prediction standard deviation, represents the correction of the predicted probability through exponential decay; Prediction standard deviation The expression is: ; In the formula, represents the original predicted probability of state s at the kth sampling time, reflecting the prediction result of a single forward propagation, Represents the probability mean of K samplings, eliminates the baseline prediction probability after random disturbance, calculates the discreteness of the prediction results, quantifies the uncertainty level of the model's judgment on the current patient status, and restores the variance dimension to the same scale as the original probability; Longitudinal time series feature matrix The expression is: ; In the formula, Represents the M protein kinetics index, specifically serum concentration, monthly change rate and area under the curve, Represents the characteristics of clonal plasma cells, specifically the proportion of abnormal clonal plasma cells with CD138, flow cytometry clonality index, and cytogenetic risk score. represents bone metabolism indicators, specifically CTX / NTX ratio, MRI lesion volume and PET-CT uptake value, and u represents the time slice dimension, which indicates the time stamp of data collection and analysis.

[0024] Embodiment 3: In this embodiment, as Figure 2 As shown, a combined risk index is defined, and patients are stratified into low, intermediate, and high risk based on the combined risk index, including: Step 61: Quantify the clonal plasma cell ratio, M protein concentration, bone metabolism markers, and bone lesion volume change indicators into a unified scoring standard; standardize the clonal plasma cell ratio, M protein concentration, bone metabolism markers, and bone lesion volume change indicators, and convert data of different units and dimensions into a unified scoring standard. By unifying the scoring standard, the problem of inconsistent units and magnitudes among indicators is solved, so that data from different sources and types can be seamlessly compared and analyzed, and the impact of multiple biomarkers can be considered simultaneously instead of relying on a single indicator, thereby improving the comprehensiveness and accuracy of risk assessment; Step 62: assign appropriate weights to each selected indicator based on the impact and effectiveness evaluation of the indicator on the bone microenvironment and the disease progression; assign corresponding weights to each indicator based on the impact and clinical effectiveness evaluation of each selected indicator on the bone microenvironment and the disease progression. Compared with fixed weights, dynamic weight assignment based on clinical data and expert evaluation can better reflect the actual situation and enhance the flexibility and adaptability of the model. Through reasonable weight assignment, it can highlight indicators that have a greater impact on the patient's prognosis, thereby increasing the predictive ability of the model; Step 63: Multiply the quantitative value of each indicator by the corresponding weight, and then add them up to calculate the joint risk index for each patient; simplifying the complex biomarkers and their effects into a single quantitative indicator makes it easier to track and explain the changes in the patient's status. This calculation method is clear and easy to understand, which facilitates clinicians to quickly obtain comprehensive risk information for patients; Step 64: Based on the calculated combined risk index, the critical value is established based on the ROC curve analysis, and the patients are stratified into low-, medium-, and high-risk groups. The critical value is determined using the ROC curve, which provides an intuitive diagram of the true positive rate and the false positive rate to ensure the scientificity and reliability of the stratification. The critical value can be re-evaluated and adjusted according to the actual situation for different patient groups or disease progression stages, which helps to monitor and manage patient risks in real time.

[0025] Compared with existing technologies, establishing a combined risk index that integrates multiple prognostic indicators can help to achieve more refined risk stratification and improve the accuracy and personalization of clinical decision-making.

[0026] The above embodiment integrates the proportion of clonal plasma cells, M protein concentration and bone metabolism markers to form a multi-dimensional analysis framework, identifies the complex pathological mechanism of multiple myeloma with a comprehensive analysis framework, and monitors data changes in real time through time series analysis and dynamic time warping technology, and effectively predicts the progression of the patient's condition, thereby identifying the risk of disease deterioration early. Wavelet transform and 3D convolutional network are used to obtain more accurate data features, effectively extract short-term fluctuations and long-term trends, and enhance the prediction ability of the model. Through the longitudinal causal discovery algorithm, the causal relationship between different factors can be quantified under the condition of controlling related interference factors, ensuring the reliability of the model prediction results. Based on the joint risk index, patients can be divided into low, medium and high risk groups, thereby providing personalized treatment recommendations for patients in different risk groups, enhancing the targeted clinical treatment, and generating hot spot maps and visual trajectory maps to help doctors intuitively understand the spatial distribution and risk level of bone lesions, thereby providing important references for clinical decision-making; The specific working principle is: regularly collect data such as the proportion of clonal plasma cells, M protein concentration, and bone metabolism markers, and establish a time alignment matrix. Use dynamic time warping to solve the irregular sampling problem between data, perform wavelet decomposition on the M protein time series data to extract the fluctuation and trend components of different periods, and calculate the variation intensity index of clonal plasma cells at the same time. Use a 3D convolutional network to extract the spatial distribution characteristics of bone lesions. A two-stream analysis model was established. Stream 1 processed the time series of M protein and clonal plasma cells, and stream 2 processed bone metabolism indicators and imaging data. The feature vectors were merged and input into the fully connected layer for feature extraction. A longitudinal causal discovery algorithm was applied to control treatment interference, and a dual robust estimator was used to analyze the causal relationship between clonal plasma cell changes and bone disease. Based on a dual-stream analysis model, a Bayesian neural network was used to predict the future development of the patient's condition. Monte Carlo Dropout was introduced to evaluate the confidence of the prediction. If the probability of progression exceeds the threshold, an alarm will be triggered. The proportion of clonal plasma cells, M protein concentration and bone metabolism markers are quantified and weighted to calculate the joint risk index. Patients are divided into low, medium and high risk groups based on ROC curve analysis. A hotspot map is generated based on the joint risk index and the spatial distribution characteristics of bone lesions to mark high-risk key bone areas. Intervention suggestions are generated based on the patient's risk level in the preset opinion library, and cloud sharing is supported. In summary, the present invention has constructed an intelligent decision-making mechanism with clinical interpretability through the deep integration of dynamic time series modeling and causal inference. Compared with the traditional static model, the system innovatively integrates the dynamic evolution of clonal plasma cells, M protein concentration trajectory and multidimensional indicators of bone metabolism, and uses a dual-stream LSTM network to couple tumor clone dynamics and bone microenvironment spatiotemporal characteristics, which significantly improves the characterization ability of biological processes. By introducing the longitudinal mediation effect model and dual robust estimation; It also effectively verified the causal relationship between clonal expansion and bone destruction, providing a theoretical basis for targeted intervention. By combining the Bayesian neural network with the Monte Carlo Dropout technology, it not only outputs the probability distribution of future disease states, but also dynamically quantifies the prediction confidence. When the uncertainty reaches the threshold, the manual review mechanism is automatically triggered, greatly reducing the risk of misjudgment. The risk stratification system integrates the clonal mutation intensity index and the spatial clustering pattern of lesions to generate a heat map of bone lesions, achieving dynamic adaptation of prognosis prediction and treatment pathways. It continuously optimizes feature weights through reinforcement learning, and transforms dynamic monitoring of biomarkers into actionable clinical decisions, providing a technical blueprint for reference for the transformation of the clinical research paradigm of multiple myeloma from association analysis to causal inference.

[0027] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention 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. Such modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing prognostic data of multiple myeloma, characterized in that: The following steps are involved: Step 1: Periodically obtain the time series data of clonal plasma cell ratio, M protein concentration, bone metabolism markers, and imaging bone lesion volume change coefficient, establish a time alignment matrix, and use dynamic time regularization to unify irregular sampling intervals; Step 2: Perform wavelet decomposition on the M protein trajectory to extract short-term fluctuation and long-term trend components, calculate the mutation intensity index of clonal plasma cell indicators, and extract the spatial distribution characteristics of bone lesions through a 3D convolutional network; Step 3: Based on the data obtained in step 2, a dual-stream analysis model is constructed to quantify the influence weight of the output clonal evolution signal on the bone microenvironment; Step 4: Analyze the correlation coefficient between the changes in clonal plasma cells and the occurrence of bone lesions through key causal verification links; Step 5: Based on the dual-stream analysis model obtained in step 3, the probability of regression, stabilization or progression of the patient's disease in the next several cycles is predicted; Step 6: Based on the correlation coefficient obtained in step 4 and the prediction result obtained in step 5, a joint risk index is defined, and patients are stratified into low, medium and high risk groups based on the joint risk index, triggering corresponding warning prompts; Step 7: Generate a bone lesion heat map based on the patient's joint risk index and the spatial distribution characteristics of bone lesions, and mark the key bone areas related to risk; Step 8: Based on the marked critical bone areas, index them in the preset opinion library and generate intervention suggestions in order based on the risk level.

2. The multiple myeloma prognosis data processing method according to claim 1, characterized in that: In the dual-stream analysis model in step 3, stream one is used to process the M protein concentration and clonal plasma cell time series, and stream two is used to analyze bone metabolism markers and imaging features. The output feature vectors of the two streams are spliced ​​in dimension to form a comprehensive feature vector, and the fused features are input into a fully connected layer for further feature extraction and processing.

3. The multiple myeloma prognosis data processing method according to claim 1, characterized in that: In the critical causal verification link of step 4, a longitudinal causal discovery algorithm is used. Under the condition of controlling treatment interference, a double robust estimation is used to quantify the independent effect of clonal evolution on bone lesions, and the increase in the risk of bone lesions corresponding to different growth rates of M protein is obtained and integrated as the correlation coefficient.

4. The multiple myeloma prognosis data processing method according to claim 1, characterized in that: In step 5, the dual-stream analysis model outputs the transition probability of the clone state through the Bayesian neural network, and combines the Monte Carlo Dropout technology to evaluate the prediction confidence. When the probability of progression exceeds the corresponding threshold, a graded warning is triggered, and a state transition matrix is ​​constructed to identify the high-risk transition window period. The calculation formula for the probability of regression, stability or progression of the patient's disease in the next few cycles is: ； In the formula, represents the predicted probability of the patient being in state s in the time window t months, s∈{regress, stable, progressive}, Represents the number of Monte Carlo Dropout sampling, represents the state-specific weight matrix of state s, represents the LSTM hidden state vector, represents the feature encoder, represents the longitudinal time series feature matrix, represents the bias term of state s, Represents the preset penalty intensity coefficient, represents the prediction standard deviation, Represents the correction of the predicted probability in the form of exponential decay.

5. The multiple myeloma prognosis data processing method according to claim 4, characterized in that: The prediction standard deviation The expression is: ； In the formula, represents the original predicted probability of state s at the kth sampling time, reflecting the prediction result of a single forward propagation, Represents the probability mean of K samples.

6. The multiple myeloma prognosis data processing method according to claim 4, characterized in that: The longitudinal time series feature matrix The expression is: ； In the formula, Represents the M protein kinetics index, including serum concentration, monthly change rate and area under the curve, Representative clonal plasma cell characteristics, including the proportion of abnormal clonal plasma cells with CD138, flow cytometry clonality index, and cytogenetic risk score, represents bone metabolism indicators, including CTX / NTX ratio, MRI lesion volume and PET-CT uptake value, and u represents the time slice dimension, which indicates the time mark of data collection and analysis.

7. The multiple myeloma prognosis data processing method according to claim 1, characterized in that: The working logic of the dual-stream analysis model in step 5 for analyzing the probability of regression, stabilization or progression of the patient's disease in the next several cycles is: The patient's continuous monitoring data is input, spatial features are extracted through a three-layer residual network, and a multi-head temporal attention mechanism is superimposed to capture cross-period dependencies; Input the encoded features into the LSTM network and update the hidden state; Enable the Dropout layer and perform several random forward passes; Calculate prediction statistics and perform probability calibration; Feed the final prediction result back to the LSTM hidden state; Generate a visual trajectory map and mark key turning points.

8. The multiple myeloma prognosis data processing method according to claim 1, characterized in that: The combined risk index is defined in step 6, and the low, medium and high risk stratification of the patient based on the combined risk index includes: Step 61: quantifying the proportion of clonal plasma cells, M protein concentration, bone metabolism markers, and bone lesion volume change indexes into a unified scoring standard; Step 62: assign appropriate weights to each selected indicator based on the impact and effectiveness of the indicator on the bone microenvironment and disease progression; Step 63: multiply the quantified value of each indicator by the corresponding weight, and then add them up to calculate the joint risk index of each patient; Step 64: According to the calculated combined risk index, a critical value is established based on ROC curve analysis to stratify the patients into low, medium and high risk groups.

9. The multiple myeloma prognosis data processing method according to claim 1, characterized in that: The bone lesion heat map generated in step 7 includes: spatial distribution characteristics, lesion risk level, lesion characteristics, time dynamics and clinically relevant marker information.

10. The multiple myeloma prognosis data processing method according to claim 1, characterized in that: The preset opinion library in step 8 actively adds intervention treatment strategies manually and supports cloud sharing.

Images