A digital tumor prevention and treatment management platform and management method

Abstract

The present invention discloses a digital tumor prevention and treatment management platform and management method, which relates to the field of smart medical technology, including real-time collection of multimodal data of patients, and performing standardized cleaning and spatiotemporal alignment processing to generate a spatiotemporal labeled data set; inputting multimodal joint feature representation into a counterfactual reasoning engine, simulating the potential effects under different treatment interventions, calculating individual treatment effects, and generating a set of counterfactual rehabilitation recommendations; constructing a dynamic causal reinforcement decision model based on the set of counterfactual rehabilitation recommendations, continuously receiving real-time physiological data feedback from patients through a proximal strategy optimization algorithm, and dynamically adjusting the treatment plan. The present invention significantly improves the accuracy and clinical credibility of personalized tumor treatment plan recommendations through spatiotemporal alignment and feature fusion of multimodal data, combined with tumor biological mechanisms and machine learning algorithms.

Description

Technical Field

[0001] The present invention relates to the field of smart medical technology, and in particular to a digital tumor prevention and treatment management platform and management method. Background Art

[0002] Currently, the combined analysis of medical imaging omics, genomic sequencing, and wearable device monitoring data is gradually shifting from single-biomarker analysis to multi-dimensional dynamic assessment. Existing technologies use deep neural networks to construct correlation models between imaging features and gene expression, and combine them with clinical decision tree algorithms to generate treatment recommendations. These technologies have demonstrated promising applications in tumor staging prediction and drug sensitivity analysis.

[0003] However, existing methods still face technical bottlenecks in cross-modal data integration and causal mechanism mining: on the one hand, the high spatial resolution of medical images is significantly different from the time series characteristics of gene sequencing and wearable devices. Traditional registration algorithms find it difficult to eliminate the temporal and spatial scale deviations between multi-source data, resulting in a high error rate in modeling the dynamic association between image features and gene mutations; on the other hand, existing causal inference models lack a constraint mechanism for the topological relationship of cross-modal feature space. When analyzing complex causal chains such as gene mutations driving lesion evolution and patient behavior affecting treatment response, they are easily interfered by confounding factors and lead to misjudgment, which directly affects the clinical adaptability and safety threshold control of individualized treatment plans. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a digital tumor prevention and treatment management method to solve the problems of insufficient spatiotemporal alignment accuracy of multimodal tumor data and high causal association misjudgment rate.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides a digital tumor prevention and treatment management method, which includes real-time collection of multimodal data of patients, and standardized cleaning and spatiotemporal alignment processing to generate a spatiotemporal labeled data set; inputting the spatiotemporal labeled data set into a spatiotemporal causal alignment network, using a temporal convolutional network to extract time series features, and constructing a multimodal causal association graph through a causal discovery algorithm to identify the causal chain between image features, gene mutations and patient behaviors, and outputting a causal strength quantization matrix; based on the causal strength quantization matrix, screening associated nodes whose causal strength exceeds a preset causal strength threshold, forcing the alignment of image slice feature vectors and gene mutation embedding vectors through a cross-modal contrast loss function to generate a multimodal joint feature representation that integrates causal constraints; inputting the multimodal joint feature representation into a counterfactual reasoning engine to simulate the potential effects under different treatment interventions, calculate individual treatment effects, and generate a set of counterfactual rehabilitation recommendations; based on the counterfactual rehabilitation recommendation set, constructing a dynamic causal reinforcement decision model, continuously receiving real-time physiological data feedback from patients through a proximal strategy optimization algorithm, and dynamically adjusting the treatment plan.

[0008] As a preferred solution of the digital tumor prevention and treatment management method described in the present invention, the standardized cleaning and spatiotemporal alignment processing includes standardizing the image features of the lesion area and using a dynamic time warping algorithm to align the vital sign monitoring data with the behavior log timeline.

[0009] As an optimal solution of the digital tumor prevention and treatment management method described in the present invention, the causal discovery algorithm refers to establishing causal pointing rules based on the timestamp priority principle, and constructing a multimodal causal association graph through conditional independence test.

[0010] As a preferred solution of the digital tumor prevention and treatment management method of the present invention, the cross-modal contrast loss function includes the following steps:

[0011] Based on the causal strength quantization matrix, causal edges are screened and the image slice feature vectors and gene mutation embedding vectors are extracted as positive sample pairs.

[0012] Generate a set of negative sample pairs by uniformly sampling from non-causally related cross-modal nodes;

[0013] Based on the constructed set of positive and negative sample pairs, a causal strength weighted contrastive learning framework is adopted to construct a contrastive loss function. The contrastive loss is minimized through the Adam optimizer, and the distance decrease rate of the positive sample pairs is monitored.

[0014] As a preferred solution of the digitalized tumor prevention and treatment management method of the present invention, the counterfactual reasoning engine includes the following steps:

[0015] Based on the causal edges retained in the causal strength quantification matrix, the binary features of gene mutations and lesion volume were selected as intervenible variables;

[0016] The intervention variables were input into the doubly robust estimation framework and combined with the gradient boosting regression tree to predict the potential outcomes. The counterfactual prediction bias was calibrated by the real data residuals and the mean absolute error was verified.

[0017] As a preferred solution of the digital tumor prevention and treatment management method described in the present invention, the multimodal data includes medical images, gene sequencing results, continuous vital signs monitoring data and patient self-reported behavior logs.

[0018] As a preferred embodiment of the digital tumor prevention and treatment management method described in the present invention, the counterfactual rehabilitation suggestion set includes a drug dosage adjustment plan, a follow-up cycle optimization strategy and a behavioral intervention plan.

[0019] In the second aspect, the present invention provides a digital cancer prevention and treatment management platform, including a data preprocessing module, a causal modeling module, a feature fusion module, a counterfactual deduction module and a dynamic decision-making module. The data preprocessing module is used to collect multimodal data of patients in real time, and perform standardized cleaning and spatiotemporal alignment processing to generate a spatiotemporal labeled data set; the causal modeling module is used to input the spatiotemporal labeled data set into a spatiotemporal causal alignment network, use a temporal convolutional network to extract time series features, and construct a multimodal causal association graph through a causal discovery algorithm to identify the causal chain between image features, gene mutations and patient behaviors, and output a causal strength quantization matrix; the feature fusion module , which is used to screen the associated nodes whose causal strength exceeds the preset causal strength threshold based on the causal strength quantification matrix, and force the alignment of the image slice feature vector and the gene mutation embedding vector through the cross-modal contrast loss function to generate a multimodal joint feature representation that integrates causal constraints; the counterfactual inference module is used to input the multimodal joint feature representation into the counterfactual reasoning engine, simulate the potential effects under different treatment interventions, calculate the individual treatment effects, and generate a set of counterfactual rehabilitation recommendations; the dynamic decision-making module is used to construct a dynamic causal reinforcement decision-making model based on the counterfactual rehabilitation recommendation set, continuously receive real-time physiological data feedback from patients through the proximal strategy optimization algorithm, and dynamically adjust the treatment plan.

[0020] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the digital tumor prevention and treatment management method as described in the first aspect of the present invention is implemented.

[0021] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the digital tumor prevention and treatment management method as described in the first aspect of the present invention.

[0022] The beneficial effects of the present invention are: by processing multimodal medical images, genome sequencing data and clinical monitoring data through spatiotemporal alignment and feature fusion, the dynamic relationship of biomarkers with significant clinical correlation in tumor evolution is accurately captured, and the reliability of causal inference is improved. By calculating the covariance matrix of gene mutation frequency, lesion volume change rate and physiological index fluctuation, the spatiotemporal correlation characteristics of multimodal data are extracted, and multi-dimensional features with clear pathological characterization are extracted from them, thereby enhancing the adaptability and generalization ability of dynamic decision-making models to individualized treatment plans. By combining regularization constraints based on tumor biology with causal reasoning models, integrating pathological mechanisms with machine learning algorithms, the biological rationality and interpretability of causal reasoning models are enhanced, so that the treatment recommendation results output by the dynamic decision-making model have verifiable clinical credibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 This is a flow chart of the digital cancer prevention and treatment management method.

[0025] Figure 2 Flowchart of multimodal data processing.

[0026] Figure 3 Flowchart of causal feature fusion.

[0027] Figure 4 Flowchart optimized for dynamic decision making. DETAILED DESCRIPTION

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0030] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0031] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides a digital cancer prevention and treatment management method, including:

[0032] S1: Collect multimodal data of patients in real time, perform standardized cleaning and spatiotemporal alignment processing to generate a spatiotemporal labeled dataset.

[0033] Specifically, the following steps are included:

[0034] S1.1: Deploy medical imaging equipment, gene sequencers, wearable devices, and mobile applications to synchronize clocks using the Network Time Protocol (NTP) and automatically generate standard timestamps for date and time representation (ISO8601) when collecting multimodal data.

[0035] The multimodal data includes medical images, gene sequencing results, continuous vital signs monitoring data and patient self-reported behavior logs.

[0036] S1.2: Use the Elastix toolkit to perform multi-resolution B-spline registration on medical images, align different modality images to the patient's anatomical coordinate system, annotate the lesion area, and extract image features (e.g., CT value, volume) of the lesion area.

[0037] Based on the timestamp difference between medical images and gene sequencing results, gene mutations are associated with lesion areas.

[0038] For example, when the timestamp difference is ≤ 24 hours, the determination is valid.

[0039] S1.3: Dynamic time warping was used to align continuous vital sign monitoring data (sampled at 10 Hz) with the patient's self-reported behavioral log timeline. Vital sign data were resampled to 1 Hz, and missing values were filled using cubic spline interpolation.

[0040] S1.4: Perform Z-score standardization on the imaging features of the lesion area, convert the gene mutation data into binary features, remove abnormal vital signs (such as heart rate >220 bpm), and perform interpolation.

[0041] S1.5: Store the date and time representation method (ISO 8601) standard timestamp, annotated lesion area, processed image features and gene mutation data, aligned continuous vital sign monitoring data and behavioral log data into HDF5 files, grouped by patient ID.

[0042] Among them, HDF5 file is the specific storage form of spatiotemporal labeled dataset.

[0043] S2: The spatiotemporal labeled dataset is input into the spatiotemporal causal alignment network, temporal features are extracted using a temporal convolutional network, and a multimodal causal association graph is constructed through a causal discovery algorithm to identify the causal chain between imaging features, gene mutations, and patient behaviors, and output a causal strength quantification matrix.

[0044] Specifically, the following steps are included:

[0045] S2.1: Read the time series of continuous vital sign monitoring data and aligned behavioral log data from the spatiotemporally labeled dataset (HDF5 file).

[0046] For the time series of continuous vital sign monitoring data (such as heart rate, blood oxygen saturation), the dilated causal convolution layer of the temporal convolutional network (TCN) is used to set the dilation factor .

[0047] in, is the number of layers, ; A causal convolution layer with a convolution kernel size of 3 outputs a multi-scale temporal feature vector.

[0048] The aligned behavioral log data (such as medication and exercise) is one-hot encoded and aligned to a 1 Hz sampling rate along the time axis to generate behavioral event sequence features.

[0049] Example: Input heart rate sequence length 7200 (2 hours of data), after three layers of TCN output 256-dimensional time series feature vector. The behavioral event "taking medication" is encoded as [0, 1, 0, ..., 0], and after alignment, a 7200-dimensional sparse sequence is generated.

[0050] S2.2: Use the PC algorithm (Peter-Clark algorithm) to perform conditional independence tests on imaging features, gene mutations, and behavioral event sequence features.

[0051] Establish causal pointing rules based on the timestamp priority principle: based on the representation method of recorded date and time (ISO8601) standard timestamp, several pre-variables Timestamp is earlier than the result variable (If the gene sample collection time is earlier than the behavioral event time), add a directed edge → .

[0052] Specifically, the G-test statistic (the significance level was set at 0.05) was used to test the independence of variables.

[0053] For example, the gene mutation "EGFR L858R" (timestamp 14:25:30Z) → behavioral event "increased fatigue" (timestamp 14:30:00Z), G-test p=0.03, add a causal edge.

[0054] Based on all causal edges that pass the test ( → ), constructs a multimodal causal association graph in the form of a directed acyclic graph, where nodes represent variables such as image features, gene mutations, and behavioral events, and edges represent statistically verified causal relationships.

[0055] S2.3: Based on the constructed multimodal causal association graph, the extracted image features and the generated gene mutation binary features are forced to be aligned.

[0056] Define triplet loss (anchor, positive and negative).

[0057] Specifically, anchor: randomly selected image slice feature vector; positive sample: gene mutation embedding causally associated with the anchor; negative sample: randomly sampled non-related gene mutation embedding.

[0058] Loss Function , expressed as:

[0059] ;

[0060] Where, represents the anchor point feature vector, represents the positive sample feature vector, represents the negative sample feature vector, represents the L2 norm distance, is a preset constant threshold (such as γ=0.5), which is used to control the minimum difference between the positive and negative sample distances. Indicates taking the maximum value of the input value and 0;

[0061] For each edge in the multimodal causal graph → , calculating causal strength based on conditional mutual information :

[0062] ;

[0063] in, for The parent node set of represents conditional mutual information;

[0064] It should be noted that conditional mutual information is a measure of two variables (such as and ) given other variables The core logic is: when it is known that The parent node set of hour, right The amount of additional information contributed.

[0065] Setting a causal strength threshold ,reserve , filtering low-significance associations (such as The edges are removed), generating a causal strength quantization matrix.

[0066] It should be noted that setting the causal strength threshold First, we need to calculate each edge based on the constructed multimodal causal relationship graph → Causal strength , obtain the complete causal strength distribution; by analyzing the causal strength The distribution characteristics of the values are plotted, and a histogram is drawn to observe its central tendency and dispersion. Based on the inflection point characteristics of the distribution curve, a significant high-value interval (such as the top 30% high-value interval) is selected as the retention range. The corresponding numerical threshold of 0.7 is calculated from the distribution. The retained The causal edges of the causal graph are constructed to verify its medical rationality; finally, the causal strength threshold is determined. Set as .

[0067] Store the causal strength quantization matrix as a Numpy file (.npz).

[0068] S3: Based on the causal strength quantization matrix, the associated nodes whose causal strength exceeds the preset causal strength threshold are screened, and the image slice feature vector and the gene mutation embedding vector are forced to be aligned through the cross-modal contrast loss function to generate a multimodal joint feature representation that integrates causal constraints.

[0069] Specifically, the following steps are included:

[0070] S3.1: Read the generated causal strength quantization matrix (.npz file).

[0071] Image features (CT values, volumes, etc.) and converted gene mutation binary features are used as the original features to be aligned.

[0072] Based on a preset causal strength threshold , retain all satisfied causal edge.

[0073] Example: Preserving causal strength Edges with significant M=0.62 (such as gene mutation "EGFR L858R" → imaging feature "abnormal CT value area" extracted in S1.2) were removed, and low-significance edges with M=0.62 (such as gene mutation "ALK fusion" → behavioral event "intensified cough").

[0074] Only retained causal edges are included, where the node set covers image slice feature vectors, gene mutations, and behavioral event sequence features.

[0075] S3.2: Generate a set of negative sample pairs by uniformly sampling from non-causally related cross-modal nodes.

[0076] Among them, the non-causal cross-modal nodes refer to the nodes that meet the causal strength in the multimodal causal relationship graph constructed by the present invention. The cross-modal variable combination of conditions (such as gene mutation characteristics and imaging characteristics) has no significant causal relationship verified by the PC algorithm.

[0077] Specifically, positive sample pair generation: If there is an edge in the simplified causal graph → (e.g. gene mutation → image slice feature vector), then extract the gene mutation feature and image slice feature vector respectively. and The feature vectors of constitute the positive sample pair set.

[0078] Negative sample pair generation: traverse all cross-modal node combinations and filter those that meet The gene mutation-image slice feature vector pairs are selected from the screening results through uniform sampling, and the negative sample pair set that matches the number of positive sample pairs is selected to ensure data balance.

[0079] Based on the constructed positive and negative sample pairs, a contrastive learning framework with causal strength weighting is used to construct a contrastive loss function. , expressed as:

[0080] ;

[0081] Where, represents the set of positive sample pairs, represents the set of negative sample pairs, represents the gene mutation embedding vector (gene mutation binary feature), Represents the image slice feature vector (lesion area image feature extraction), represents the negative sample distance threshold, represents the cross-modal feature vector in the negative sample pair, Represents the anchor feature index in the loss function, Represents the contrast feature index in the negative sample pair, Represents the target feature index in the positive sample pair;

[0082] It should be noted that lesion area image feature extraction refers to the process of locating the lesion area from medical images (such as CT / MRI) and extracting its quantitative features; the image slice feature vector value refers to the final output form of "lesion area image feature extraction", that is, the high-dimensional vector after encoding by the neural network.

[0083] S3.3: Input the lesion region image features (CT value, volume) extracted by the medical image registration toolkit (Elastix) into a pre-trained 3D image classification neural network (3D ResNet-50) model (with ImageNet pre-trained weights), outputting a 512-dimensional feature vector.

[0084] Input positive sample pairs (gene embeddings, image features) and negative sample pairs for alignment training;

[0085] Optimization goal: minimize the defined contrast loss L, so that the distance between positive samples is reduced and the distance between negative samples is increased; training parameters: Adam optimizer (learning rate ), batch size 32, and 1000 iterations.

[0086] Monitor the rate of decrease in the distance between the positive pairs (calculated based on the loss function in S3.2). If the rate of decrease is < 1% within 10 consecutive epochs, convergence is considered.

[0087] S3.4: Concatenate the aligned image slice feature vectors with the gene embedding vectors, input the fully connected layer for dimensionality reduction, and output a 128-dimensional joint feature vector. , the expression is:

[0088] ;

[0089] Where, represents the weight matrix (initialization see He normal distribution), represents the bias term;

[0090] Randomly select samples and visualize the joint feature space using t-distributed stochastic neighbor embedding (t-SNE) to check whether strong causal association nodes (such as gene mutations and imaging features retained in S2.3) are clustered;

[0091] For example, the distance between the gene mutation data "EGFR L858R" (S1.4) and the imaging feature "abnormal CT value area" in the joint feature space should be significantly smaller than that of unrelated nodes.

[0092] The joint feature representation Save the HDF5 file generated by S1.5 to expand the original dataset.

[0093] S4: Input the multimodal joint feature representation into the counterfactual reasoning engine to simulate the potential effects of different treatment interventions, calculate the individual treatment effects, and generate a set of counterfactual rehabilitation recommendations.

[0094] Specifically, the following steps are included:

[0095] S4.1: Read multimodal joint feature representation (128-dimensional in HDF5 file) ) and the causal strength quantization matrix (.npz file).

[0096] Causal edges preserved in causal strength quantization matrix , screen the set of variables that can be intervened.

[0097] Define intervention and outcome variables.

[0098] Specifically, the definition of intervention variables is: selecting the root node in the causal graph that directly affects vital signs or behavioral logs (such as binary features of gene mutations and lesion volume); the definition of outcome variables is: vital sign monitoring data (such as blood glucose levels and blood oxygen saturation) and key events in behavioral logs (such as exercise duration and medication compliance).

[0099] Establish a mapping function between intervention variables and outcome variables.

[0100] Specifically, for each intervention variable For example, for a binary feature of the gene mutation "EGFR L858R", define its adjustable range (such as the presence / absence of the mutation); for a continuous intervention variable (such as the lesion volume in an imaging feature), define the adjustment range (such as ±10% of the current value).

[0101] Example: Intervention Variable =Gene mutation "EGFR L858R" (0 or 1), outcome variable = Mean continuous blood glucose monitoring value (mg / dL)

[0102] S4.2: Extract historical intervention-outcome pairs from spatiotemporally labeled datasets.

[0103] Specifically, the input multimodal joint feature representation (128 dimensions) and the intervention variable The current status of the device; output vital sign monitoring data (such as blood glucose time series) and behavior log events (such as the number of exercises).

[0104] A doubly robust estimation framework is used to train the potential outcome prediction model.

[0105] Specifically, we use the Gradient Boosting Regressor to predict the intervention variable Adjusted potential outcomes; calibrate the counterfactual forecast bias based on the residuals between the actual observations and the predicted outcomes.

[0106] To further validate the potential outcome prediction model, the training set and validation set were divided into two groups (8:2), and the mean absolute error (MAE) was calculated to evaluate the prediction accuracy. If the prediction error exceeded the clinically acceptable range (e.g., blood glucose prediction error >5 mg / dL), the sample size was increased or the depth of the regression tree was adjusted.

[0107] It should be noted that the "clinically acceptable range" refers to the acceptable error boundary of a certain indicator (such as blood sugar, tumor volume, etc.) specified in medical practice or diagnosis and treatment guidelines when making clinical decisions, ensuring that the prediction results have practical application value.

[0108] Generate a set of candidate treatment options based on the validated potential outcome prediction model.

[0109] The candidate treatment plan set includes drug intervention plan, behavioral intervention plan basis and follow-up strategy plan. Specifically: Drug intervention plan: If there is a strong causal edge between gene mutation and vital signs The system extracts the adjustable dosage range of the corresponding drug from the clinical knowledge base. For behavioral intervention plans, a safe adjustment interval of ±2 standard deviations is calculated for behavioral variables with significant causal relationships (e.g., exercise duration → blood sugar). For follow-up strategies, the system predicts the rate of change of lesion volume and stratifies the reexamination interval according to a gradient. This outputs a set of structured candidate plans for counterfactual simulation.

[0110] S4.3: For each candidate treatment, modify the intervention variable The value of .

[0111] For example, some predictions are related to gene mutations (such as the use of EGFR inhibitors), and the gene mutation characteristics are changed from 1 to 0 (assuming that the drug inhibits the expression of mutations); some predictions are related to imaging features (such as radiotherapy to shrink lesions), and the lesion volume characteristics are reduced by 10%.

[0112] Calculate the difference between the potential outcome after the intervention and the current state outcome.

[0113] Weighted summation of multiple outcome variables (e.g., blood glucose, blood oxygen, exercise duration) (e.g., blood glucose weight 0.6, blood oxygen weight 0.3). Screening for interventions that rank in the top 10% of individual treatment effects (ITE) (e.g., "a metformin dose of 5 mg can reduce blood glucose by 15 mg / dL");

[0114] Among them, "ITE" is the abbreviation of "Individual Treatment Effect", which refers to a quantitative indicator of the personalized effect difference produced by different treatment plans for a specific patient compared with the baseline status.

[0115] In combination with the constraints of medical guidelines, exclude suggestions that exceed the reasonable range (such as dosage adjustments exceeding the upper limit of the pharmacopoeia);

[0116] Among them, medical guideline constraints refer to the boundaries of treatment plan selection clearly stipulated in the diagnosis and treatment guidelines by authoritative medical institutions (such as the Chinese Medical Association, NCCN, etc.), including mandatory rules such as drug dosage range, contraindication list and treatment indications.

[0117] The data are classified by intervention type and stored as a JSON file as a set of counterfactual rehabilitation recommendations.

[0118] Among them, the set of counterfactual rehabilitation recommendations includes the following specific intervention plans.

[0119] Specifically, the drug dosage adjustment plan involves dynamically adjusting the dosage based on the individual patient's condition, taking into account drug metabolism characteristics and changes in clinical indicators. For example, for patients with hypercholesterolemia, the statin dosage can be appropriately increased to optimize lipid regulation. Liver function and creatine kinase levels are also monitored to ensure medication safety.

[0120] Follow-up cycle optimization strategy: Based on medical imaging data and lesion dynamics, individualized review frequency is determined. If accelerated lesion progression is detected, the imaging follow-up interval is shortened to more closely track disease progression and adjust treatment plans in a timely manner.

[0121] Behavioral intervention plans: Integrate patients' daily behavioral data with physiological indicators to design adaptive health management plans. For example, if a decrease in blood oxygen saturation is detected, we may recommend increasing the duration of low-intensity aerobic exercise and combining it with breathing exercises to gradually improve cardiopulmonary function.

[0122] S5: Based on the set of counterfactual rehabilitation recommendations, a dynamic causal reinforcement decision model is constructed, and the patient's real-time physiological data feedback is continuously received through the proximal strategy optimization algorithm to dynamically adjust the treatment plan.

[0123] Specifically, the following steps are included:

[0124] S5.1: Construct a state space based on real-time physiological data and a set of counterfactual rehabilitation recommendations, and define the executable action space.

[0125] Specifically, the patient's real-time physiological data (such as blood glucose and blood oxygen saturation), current treatment parameters (such as drug dosage and follow-up period), key intervention variables in the counterfactual rehabilitation recommendation set (such as gene mutation status and lesion volume adjustment recommendations), and the retained edges in the causal strength quantification matrix are extracted from the HDF5 file. .

[0126] The patient's real-time physiological data, current treatment plan parameters, key intervention variables, and the retained edges in the causal strength quantization matrix are aligned according to timestamps to form a state vector.

[0127] The state vector contains the following dimensions: vital sign values, drug dosages, binary states of gene mutations, lesion volumes, and causal edge strength values.

[0128] Furthermore, according to the JSON file of the counterfactual rehabilitation suggestion set, the executable actions are mapped into a discrete action space.

[0129] Among them, action types include: drug dosage adjustment, follow-up period adjustment, and behavioral intervention parameters.

[0130] Example: The state vector contains a blood glucose level of 126 mg / dL, a blood oxygen saturation of 98%, a metformin dose of 500 mg / day, the EGFR L858R gene mutation at state 1, a lesion volume of 3.2 cm³, and a causal edge strength of 0.85 for the causal edge "EGFR L858R → Abnormal CT Region." Possible action types include "Metformin dose +5 mg," "Follow-up period shortened by 3 days," and "Exercise duration increased by 15 minutes."

[0131] S5.2: Initialize the policy and value network parameters based on the state space and action space.

[0132] During the specific construction: the policy network adopts a two-layer fully connected structure. The first fully connected layer (FC1-256, ReLU activation) maps the state vector to a 256-dimensional feature space, and the second fully connected layer (FC2-128, Softmax output) generates the action probability distribution.

[0133] Furthermore, the value network uses the first fully connected layer (FC1-256, ReLU activation) and the second fully connected layer (FC2-1, linear output) with the same hierarchical configuration but independent parameters for state value estimation.

[0134] Among them, the network structure is consistent with the policy network, and the parameters are initialized independently.

[0135] Configure the Adam optimizer: policy network learning rate 1e-4, value network learning rate 3e-4, batch size 32, discount factor γ = 0.99.

[0136] S5.3: Perform multi-objective reward-driven policy training based on network parameters and real-time data.

[0137] The patient's latest physiological data is collected in real time (through wearable devices and HDF5 files), the state vector is updated, the policy network is input to generate the action probability distribution, and the action type is selected according to the softmax probability sampling.

[0138] Monitor the changes in vital signs after performing the action type and calculate the mean square error (MSE) with the target value.

[0139] Causal edge between gene mutation and risk based on causal strength quantification matrix , calculate the probability increment of the cumulative risk of gene mutation, and add it to the total reward after weighting.

[0140] The state transition records are stored in independent datasets of HDF5 files, partitioned by timestamp.

[0141] S5.4: Implement strategy optimization and dynamic security verification based on experience replay data.

[0142] Randomly sample batches of data (batch size 32) from the experience replay data and calculate the advantage function.

[0143] Calculate the policy gradient loss and limit the probability ratio of the new and old policies to the range of [0.8, 1.2] to prevent policy mutations.

[0144] Minimize the prediction error of the value function and calculate the mean squared error (MSE) loss.

[0145] If the action type causes physiological indicators to exceed the limit (such as heart rate > 180bpm), it will be terminated immediately and revert to the last effective strategy.

[0146] S5.5: Based on the optimization results, update the policy parameters and trigger periodic retraining.

[0147] Save the trained policy network weights and value network weights to separate groups of HDF5 files.

[0148] Record action execution logs (timestamp, action type, actual effect, reward value) to the / logs group of the HDF5 file, partitioned by patient ID and date.

[0149] When the amount of experience playback data reaches 1,000 or the patient's physiological state changes significantly (such as the detection of new gene mutations), retraining of the entire data is triggered.

[0150] This embodiment also provides a digital cancer prevention and treatment management platform, including: a data preprocessing module, a causal modeling module, a feature fusion module, a counterfactual deduction module and a dynamic decision module. The data preprocessing module is used to collect multimodal data of patients in real time, and perform standardized cleaning and spatiotemporal alignment processing to generate a spatiotemporal labeled data set; the causal modeling module is used to input the spatiotemporal labeled data set into the spatiotemporal causal alignment network, use the time convolution network to extract time series features, and construct a multimodal causal association graph through a causal discovery algorithm to identify the causal chain between image features, gene mutations and patient behavior, and output a causal strength quantization matrix; the feature fusion module is used to Based on the causal strength quantification matrix, the associated nodes whose causal strength exceeds the preset causal strength threshold are screened, and the image slice feature vector and the gene mutation embedding vector are forced to be aligned through the cross-modal contrast loss function to generate a multimodal joint feature representation that integrates causal constraints; the counterfactual inference module is used to input the multimodal joint feature representation into the counterfactual reasoning engine, simulate the potential effects under different treatment interventions, calculate the individual treatment effects, and generate a set of counterfactual rehabilitation recommendations; the dynamic decision-making module is used to construct a dynamic causal reinforcement decision-making model based on the counterfactual rehabilitation recommendation set, continuously receive real-time physiological data feedback from patients through the proximal strategy optimization algorithm, and dynamically adjust the treatment plan.

[0151] This embodiment also provides a computer device suitable for the digital tumor prevention and treatment management method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the digital tumor prevention and treatment management method proposed in the above embodiment.

[0152] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0153] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the digital tumor prevention and treatment management method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0154] In summary, the present invention processes multimodal medical images, genome sequencing data and clinical monitoring data through spatiotemporal alignment and feature fusion to accurately capture the dynamic relationship of biomarkers with significant clinical correlation in tumor evolution and improve the reliability of causal inference. By calculating the covariance matrix of gene mutation frequency, lesion volume change rate and physiological index fluctuation, the spatiotemporal correlation characteristics of multimodal data are extracted, and multi-dimensional features with clear pathological characterization are extracted from them, thereby enhancing the adaptability and generalization ability of dynamic decision-making models to individualized treatment plans. By combining regularization constraints based on tumor biology with causal reasoning models, integrating pathological mechanisms with machine learning algorithms, the biological rationality and interpretability of causal reasoning models are enhanced, so that the treatment recommendation results output by the dynamic decision-making model have verifiable clinical credibility.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A digital cancer prevention and treatment management method, characterized by: include, Collect multimodal data of patients in real time, perform standardized cleaning and spatiotemporal alignment processing to generate spatiotemporal labeled datasets; The spatiotemporal labeled dataset is input into the spatiotemporal causal alignment network, temporal convolutional networks are used to extract temporal features, and a multimodal causal association graph is constructed through a causal discovery algorithm to identify the causal chain between imaging features, gene mutations and patient behavior, and output a causal strength quantification matrix. Based on the causal strength quantification matrix, we screen out associated nodes whose causal strength exceeds a preset causal strength threshold. We then use a cross-modal contrast loss function to enforce alignment between the image slice feature vector and the gene mutation embedding vector, generating a multimodal joint feature representation that incorporates causal constraints. The multimodal joint feature representation is input into the counterfactual reasoning engine to simulate the potential effects of different treatment interventions, calculate individual treatment effects, and generate a set of counterfactual rehabilitation recommendations; Based on the set of counterfactual rehabilitation recommendations, a dynamic causal reinforcement decision model is constructed. Through the proximal strategy optimization algorithm, real-time physiological data feedback from patients is continuously received to dynamically adjust the treatment plan. The cross-modal contrast loss function includes the following steps: Based on the causal strength quantization matrix, causal edges are screened and the image slice feature vectors and gene mutation embedding vectors are extracted as positive sample pairs. Generate a set of negative sample pairs by uniformly sampling from non-causally related cross-modal nodes; Based on the constructed set of positive and negative sample pairs, a causal strength weighted contrastive learning framework is adopted to construct a contrastive loss function. The contrastive loss is minimized through the Adam optimizer, and the distance decrease rate of the positive sample pairs is monitored. The counterfactual reasoning engine comprises the following steps: Based on the causal edges retained in the causal strength quantification matrix, the binary features of gene mutations and lesion volume were selected as intervenible variables; The intervention variables were input into the doubly robust estimation framework and combined with the gradient boosting regression tree to predict the potential outcomes. The counterfactual prediction bias was calibrated by the real data residuals and the mean absolute error was verified.

2. The digitalized tumor prevention and management method according to claim 1, characterized in that: The standardized cleaning and spatiotemporal alignment processing includes standardizing the image features of the lesion area and aligning the vital sign monitoring data with the behavior log time axis using a dynamic time warping algorithm.

3. The digitalized tumor prevention and management method according to claim 1, wherein: The causal discovery algorithm refers to establishing causal pointing rules based on the timestamp priority principle and constructing a multimodal causal association graph through conditional independence testing.

4. The digitalized tumor prevention and management method according to claim 1, wherein: The multimodal data includes medical images, gene sequencing results, continuous vital sign monitoring data, and patient self-reported behavior logs.

5. The digital cancer prevention and management method according to claim 1, wherein: The counterfactual rehabilitation recommendation set includes drug dosage adjustment plan, follow-up period optimization strategy and behavioral intervention plan.

6. A digitalized tumor prevention and management platform, based on the digitalized tumor prevention and management method according to any one of claims 1 to 5, characterized in that: Including data preprocessing module, causal modeling module, feature fusion module, counterfactual deduction module and dynamic decision module, The data preprocessing module is used to collect multimodal data of patients in real time, perform standardized cleaning and spatiotemporal alignment processing, and generate a spatiotemporal labeled data set; The causal modeling module is used to input the spatiotemporal labeled dataset into the spatiotemporal causal alignment network, extract temporal features using a temporal convolutional network, and construct a multimodal causal association graph through a causal discovery algorithm to identify the causal chain between imaging features, gene mutations and patient behavior, and output a causal strength quantification matrix; The feature fusion module is used to screen associated nodes whose causal strength exceeds a preset causal strength threshold based on the causal strength quantification matrix, and to generate a multimodal joint feature representation that integrates causal constraints by forcing alignment of image slice feature vectors and gene mutation embedding vectors through a cross-modal contrast loss function; The counterfactual inference module is used to input the multimodal joint feature representation into the counterfactual inference engine, simulate the potential effects of different treatment interventions, calculate individual treatment effects, and generate a set of counterfactual rehabilitation recommendations; The dynamic decision-making module is used to build a dynamic causal reinforcement decision-making model based on the counterfactual rehabilitation suggestion set, continuously receive real-time physiological data feedback from patients through the proximal strategy optimization algorithm, and dynamically adjust the treatment plan.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the digital tumor prevention and treatment management method described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the digital tumor prevention and treatment management method according to any one of claims 1 to 5 are implemented.

Images