Risk prediction method, device and equipment based on diagnostic data, and storage medium

Abstract

A risk prediction method and device based on diagnostic data, equipment and storage medium, comprising: obtaining standardized feature data by preprocessing the multi-dimensional feature data of the to-be-evaluated object at the diagnosis period; obtaining a basic risk score based on the standardized feature data and a preset basic survival model; obtaining a treatment response proxy score based on the standardized feature data and a treatment response prediction model; determining a case where any one of the RAS pathway mutation state or the TP53 mutation state is missing as unknown molecular state; based on the basic risk score, the treatment response proxy score, the RAS pathway mutation state, the TP53 mutation state and the molecular state unknown indicating variable, calculating the final relative risk score through the risk correction model to predict the risk stratification level of the patient, solving the problem that the risk stratification cannot identify the high heterogeneity within the subtype.

Description

Technical Field

[0001] This application relates to the field of data prediction technology, specifically to a risk prediction method, apparatus, device, and computer-readable storage medium based on diagnostic data. Background Technology

[0002] Acute myeloid leukemia (AML) is a group of hematologic malignancies with significant clinical and molecular heterogeneity. KMT2A gene rearrangements constitute a specific genetic abnormality in AML, with t(9;11)(p21.3;q23.3) forming a KMT2A::MLLT3 fusion, which is one of the more common fusion types in adult KMT2A rearranged AML. The World Health Organization's fifth edition classification includes related cases in AML with KMT2A rearrangements, and current NCCN and ELN guidelines uniformly classify them as intermediate-risk groups. However, significant differences remain in treatment response, relapse risk, and survival outcomes among patients with the same genetic subtype, making it difficult to fully identify individual heterogeneity within a subtype based solely on genetic abnormalities.

[0003] Current clinical prognostic and risk stratification systems for t(9;11) AML have significant shortcomings, primarily due to their inability to account for the high internal heterogeneity of this subtype. Existing research reports on its prognostic outcomes are highly controversial, generally agreeing that its prognosis is better than other KMT2A rearrangement subtypes such as t(6;11) and t(10;11), but worse than core-binding factor leukemia. Clinical patient prognoses exhibit marked polarization; some patients can achieve long-term disease-free survival with intensive chemotherapy combined with allogeneic hematopoietic stem cell transplantation, while others develop primary drug resistance, early relapse, and extremely poor overall survival. Relying solely on the t(9;11) cytogenetic marker cannot accurately predict individual disease progression.

[0004] Existing research has confirmed that factors such as age, initial white blood cell count, RAS pathway and FLT3-ITD co-mutation, and post-remission treatment regimen are associated with the prognosis of this subtype, but a unified and standardized risk assessment model has not yet been established. Furthermore, previous studies have suffered from issues such as mixed patient populations, iterative treatment regimens, and inconsistent standards for monitoring minimal residual disease, leading to poor consistency and limited reference value among research findings. In the current context of precision medicine, first-line clinical trials of novel menin inhibitors combined with chemotherapy are gradually being conducted, but the optimal treatment regimen for t(9;11)AML remains unclear. Clinical practice urgently needs to overcome the limitations of existing single risk stratification methods, explore potential prognostic targets within the subtype, and construct a refined risk stratification system. Summary of the Invention

[0005] This application provides a risk prediction method, apparatus, device, and computer-readable storage medium based on diagnostic data, which can solve the technical problem in the prior art that risk stratification cannot identify the high heterogeneity within the subtype.

[0006] In a first aspect, embodiments of this application provide a risk prediction method based on diagnostic data, including: The multidimensional feature data of the subject to be evaluated during the diagnostic period are preprocessed to obtain standardized feature data, which includes at least clinical feature data, laboratory test data and molecular genetic data. Based on the standardized feature data and the pre-set basic survival model, a basic risk score is obtained; Based on the standardized feature data and the treatment response prediction model, a treatment response proxy score is obtained. The treatment response prediction model is a multi-task model trained under joint supervision of survival outcome and treatment response. Obtain the RAS pathway mutation status, TP53 mutation status, and corresponding missing status indicator variables of the object to be evaluated, and determine the case where either the RAS pathway mutation status or the TP53 mutation status is missing as an unknown molecular state. Based on the baseline risk score, the treatment response proxy score, the RAS pathway mutation status, the TP53 mutation status, and the unknown molecular status indicator variable, the final relative risk score is calculated using a risk-corrected model. The patient's risk stratification level is predicted based on the final relative risk score.

[0007] In conjunction with the first aspect, in one implementation, obtaining the basic risk score based on the standardized feature data and the pre-set basic survival model includes: The pre-set basic survival model includes a first elastic network penalized Cox regression model and a second elastic network penalized Cox regression model. Extract core diagnostic variables from the standardized feature data, input them into the trained first elastic network penalized Cox regression model, and calculate the core basic risk score using the pre-fixed first regression coefficients. Extended diagnostic features are extracted from the standardized feature data and input into the trained second elastic net penalized Cox regression model. The survival risk score of the extended diagnostic features is calculated using the pre-fixed second regression coefficients. The core basic risk score and the extended diagnostic feature survival risk score are standardized by calling the pre-fixed mean and standard deviation respectively. The standardized core basic risk score and the extended diagnostic feature survival risk score are combined according to the preset component weights to obtain the basic risk score.

[0008] In conjunction with the first aspect, in one implementation, obtaining the treatment response proxy score based on the standardized feature data and the treatment response prediction model includes: The treatment response prediction model is a multi-task Cox neural network model, which includes a shared feature encoding layer, a survival risk residual prediction branch connected to the shared feature encoding layer, and a treatment response prediction branch connected to the shared feature encoding layer. The standardized feature data is input into the shared feature encoding layer to obtain the shared feature representation; The shared feature representation is input into the survival risk residual prediction branch and the treatment response prediction branch respectively to obtain the survival risk residual characterization and the predicted logit corresponding to complete remission in the first course of treatment or complete remission with incomplete hematological recovery. The predicted logit corresponding to complete remission (either complete or incomplete recovery) in the first treatment course is determined as the original surrogate score of the treatment response. The original surrogate score of the treatment response is then subjected to residualization and standardization to obtain the treatment response surrogate score.

[0009] In conjunction with the first aspect, in one implementation, the step of obtaining the final relative risk score through a risk-corrected model based on the baseline risk score, the treatment response proxy score, RAS pathway mutation status, TP53 mutation status, and unknown molecular status indicators includes: The risk correction model is a regularized Cox proportional hazards model. An interactive feature vector is constructed, which includes a standardized baseline risk score, a standardized treatment response proxy score, and product terms of the treatment response proxy score with RAS pathway mutation status, TP53 mutation status, and unknown molecular status indicator variables, respectively. The interaction feature vector is input into a regularized Cox proportional hazards model with fixed parameters. The model then calls the pre-stored fixed regression coefficients of each feature item to calculate a linear prediction value, and the linear prediction value is determined as the final relative risk score.

[0010] In conjunction with the first aspect, in one implementation, the preprocessing of the acquired multidimensional feature data of the object to be evaluated during the diagnostic period to obtain standardized feature data includes: The clinical characteristic data include at least one of the following: age, sex, physical performance status score, liver and spleen parameters, and secondary factors; The laboratory test data include at least one of the following: white blood cell count, hemoglobin, platelet count, bone marrow blast cell ratio, lactate dehydrogenase level, and coagulation function indicators. The molecular genetic data includes at least one of the following: cytogenetic karyotype analysis results, qualitative results of fusion genes, and gene mutation detection results; Missing values ​​in the multidimensional feature data are filled using multiple interpolation or median imputation. One-hot encoding or embedded encoding is used to digitize the categorical variables in the multidimensional feature data; The continuous variables in the multidimensional feature data are made dimensionless by using Z-score standardization or logarithmic transformation to obtain the standardized feature data.

[0011] In conjunction with the first aspect, in one implementation, the training process of the pre-set basic survival model includes: Obtain target multidimensional feature data for the diagnosis period of historical case cohorts; Preset core diagnostic variables are extracted from the target multidimensional feature data to construct a first training feature matrix, and a first elastic net penalized Cox regression model is trained using the elastic net penalized Cox regression algorithm. Extended diagnostic features are extracted from the target multidimensional feature data to construct a second training feature matrix, and a second elastic net penalized Cox regression model is trained using the elastic net penalized Cox regression algorithm. Cross-validation was used to determine the regularization strength and the mixing ratio of L1 and L2 regularization terms for the two elastic net penalized Cox regression models. Using the determined model parameters, the two models were refitted on the historical case cohort, and the feature preprocessing parameters, regression coefficients, and risk score standardization parameters were solidified to obtain the pre-set basic survival model.

[0012] In conjunction with the first aspect, in one implementation, the training process of the treatment response prediction model includes: A multi-task Cox neural network was constructed, comprising a shared feature encoding layer, a survival risk residual prediction branch, and a treatment response prediction branch. The target standardized feature data of the training queue during the diagnosis period is obtained as input data, the survival time and survival status corresponding to the input data are obtained as the main supervision label, and the treatment response data of the first and second courses of treatment are obtained as auxiliary supervision labels. The input data is fed into the shared feature encoding layer to obtain the shared feature representation, and the survival risk residual characterization and treatment response prediction logit are obtained through the survival risk residual prediction branch and the treatment response prediction branch, respectively. The standardized extended diagnostic feature survival risk score is summed with the survival risk residual representation at a preset ratio to obtain the neural network survival risk output; the Cox partial likelihood loss is calculated based on the neural network survival risk output and survival follow-up data, and the binary cross-entropy loss is calculated based on the treatment response prediction logit and auxiliary supervision label. The total loss is obtained by summing the product of the Cox partial likelihood loss, the preset auxiliary task weights, and the binary cross-entropy loss, along with the regularization term. The total loss is optimized using a gradient optimization algorithm until a preset number of iterations is reached or an early stopping condition is triggered, resulting in a trained treatment response prediction model. The network parameters of the model are then solidified.

[0013] Secondly, embodiments of this application provide a risk prediction device based on diagnostic data, the risk prediction device based on diagnostic data comprising: The first acquisition module is used to preprocess the multidimensional feature data of the subject to be evaluated during the diagnostic period to obtain standardized feature data. The multidimensional feature data includes at least clinical feature data, laboratory test data and molecular genetic data. The second acquisition module is used to acquire a basic risk score based on the standardized feature data and the pre-set basic survival model; The third acquisition module is used to acquire the treatment response proxy score based on the standardized feature data and the treatment response prediction model. The treatment response prediction model is a multi-task model trained under joint supervision based on survival outcome and treatment response. The fourth acquisition module is used to acquire the RAS pathway mutation status, TP53 mutation status, and corresponding missing status indicator variables of the object to be evaluated, and to determine the case where either the RAS pathway mutation status or the TP53 mutation status is missing as an unknown molecular status; based on the basic risk score, the treatment response proxy score, the RAS pathway mutation status, the TP53 mutation status, and the unknown molecular status indicator variables, the final relative risk score is calculated through a risk correction model; The prediction module is used to predict the patient's risk stratification level based on the final relative risk score.

[0014] Thirdly, embodiments of this application provide a risk prediction device based on diagnostic data. The risk prediction device based on diagnostic data includes a processor, a memory, and a risk prediction program based on diagnostic data stored in the memory and executable by the processor. When the risk prediction program based on diagnostic data is executed by the processor, it implements the steps of the risk prediction method based on diagnostic data as described above.

[0015] Fourthly, embodiments of this application provide a computer-readable storage medium storing a risk prediction program based on diagnostic data, wherein when the risk prediction program based on diagnostic data is executed by a processor, it implements the steps of the risk prediction method based on diagnostic data as described above.

[0016] The beneficial effects of the technical solutions provided in this application include: By preprocessing the multidimensional feature data of the subjects to be evaluated at the diagnostic stage, standardized feature data is obtained. This multidimensional feature data includes at least clinical feature data, laboratory test data, and molecular genetic data. Based on the standardized feature data and a pre-set basic survival model, a basic risk score is obtained. Based on the standardized feature data and a treatment response prediction model, a treatment response surrogate score is obtained. The treatment response prediction model is a multi-task model trained under joint supervision of survival outcome and treatment response. The RAS pathway mutation status, TP53 mutation status, and corresponding deletion status indicator variables of the subjects to be evaluated are obtained, and the absence of either the RAS pathway mutation status or the TP53 mutation status is defined as an unknown molecular status. Based on the basic risk score, the treatment response surrogate score, the RAS pathway mutation status, the TP53 mutation status, and the unknown molecular status indicator variables, a final relative risk score is calculated using a risk correction model. The patient's risk stratification level is predicted based on the final relative risk score, thus solving the technical problem in related technologies where risk stratification cannot identify the high heterogeneity within this subtype. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the first embodiment of the risk prediction method based on diagnostic data in this application; Figure 2 This is a schematic diagram of the functional modules of an embodiment of the risk prediction device based on diagnostic data according to this application; Figure 3 This is a schematic diagram of the hardware structure of the risk prediction device based on diagnostic data involved in the embodiments of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0020] In a first aspect, embodiments of this application provide a risk prediction method based on diagnostic data.

[0021] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the risk prediction method based on diagnostic data according to this application. Figure 1 As shown, risk prediction methods based on diagnostic data include: Step S10: Preprocess the multidimensional feature data of the subject to be evaluated during the diagnostic period to obtain standardized feature data. The multidimensional feature data includes at least clinical feature data, laboratory test data and molecular genetic data. Exemplary, multidimensional characteristic data should include at least clinical characteristic data, laboratory test data, and molecular genetic data. Clinical characteristic data should include at least one of the following: age, sex, performance status score, liver and spleen parameters, and secondary factors. Laboratory test data should include at least one of the following: white blood cell count, hemoglobin, platelet count, bone marrow blast cell percentage, lactate dehydrogenase level, and coagulation function indicators. Molecular genetic data should include at least one of the following: cytogenetic karyotype analysis results, qualitative results of fusion genes, and gene mutation detection results. Preprocessing should include one of the following: multiple imputation or median imputation, one of the following: one-hot encoding or embedding encoding, and one of the following: Z-score normalization or logarithmic transformation.

[0022] Specifically, the process of preprocessing the multidimensional feature data of the object to be evaluated during the diagnostic period to obtain standardized feature data includes: The clinical characteristic data include at least one of the following: age, sex, physical performance status score, liver and spleen parameters, and secondary factors; The laboratory test data include at least one of the following: white blood cell count, hemoglobin, platelet count, bone marrow blast cell ratio, lactate dehydrogenase level, and coagulation function indicators. The molecular genetic data includes at least one of the following: cytogenetic karyotype analysis results, fusion gene qualitative results, and gene mutation detection results; missing values ​​in the multidimensional feature data are filled using multiple imputation or median imputation; categorical variables in the multidimensional feature data are digitized using one-hot encoding or embedding encoding; and continuous variables in the multidimensional feature data are dimensionless using Z-score standardization or logarithmic transformation to obtain the standardized feature data.

[0023] As an example, multiple imputation or median imputation is used to fill in missing values ​​in multidimensional feature data. For instance, multiple imputation is used to fill in missing values ​​for some indicators, while median imputation is used for continuous variables with a missing percentage of less than 5%, effectively preserving sample size and avoiding information bias caused by missing data. For categorical variables in clinical feature data, one-hot encoding is used to convert them into binary vectors. Categorical variables include sex, gene mutation status, and FAB typing. For example, male sex is converted to [1,0] or female sex is converted to [0,1]. This eliminates spurious ordinal relationships between categorical variables, facilitating processing by neural networks and regression models. For continuous variables in laboratory test data, due to their significant dimensional differences (e.g., white blood cell count is at the 10^9 / L level while hemoglobin is at the g / L level), direct input into the model can lead to difficulties in gradient descent. Continuous variables include white blood cell count and lactate dehydrogenase. This embodiment utilizes the pre-determined and fixed mean and standard deviation for each continuous variable during the model training phase, and processes the continuous variables of the object to be evaluated according to the Z-score standardization formula. For the j-th continuous variable, z_j = (x_j - μ_j) / σ_j is calculated, where x_j is the variable value of the object to be evaluated, and μ_j and σ_j are the mean and standard deviation fixed during the training phase, respectively. The standardization parameters are not recalculated based on individual objects to be evaluated or newly added application data during the application process. For indicators with skewed distributions, logarithmic transformation is preferentially applied before standardization.

[0024] Step S20: Based on the standardized feature data and the pre-set basic survival model, obtain a basic risk score; The exemplary pre-defined baseline survival model comprises a first elastic net penalized Cox regression model and a second elastic net penalized Cox regression model. The first elastic net penalized Cox regression model uses pre-defined core diagnostic variables as input to obtain the core baseline risk score S_core; the second elastic net penalized Cox regression model uses extended diagnostic features as input to obtain the extended diagnostic feature survival risk score S_skip.

[0025] In the application phase, core diagnostic variables and extended diagnostic features are extracted from the standardized feature data of the objects to be evaluated. Let the fixed regression coefficients of the first elastic net penalized Cox regression model be β_i^core, and the fixed regression coefficients of the second elastic net penalized Cox regression model be β_k^skip, then calculate respectively:

[0026] The pre-stored mean and standard deviation of the scores are used to standardize S_core and S_skip, respectively. In this embodiment, the neural network survival risk residual representation is further residualized and standardized relative to S_core and S_skip, and the processed survival risk residual representation is combined with the standardized S_skip to obtain the extended diagnostic feature enhancement score S_enh; then, the standardized S_core and S_enh are combined according to the preset component weights and standardized again to obtain the base risk score S_base.

[0027] During the model training phase, multidimensional feature data from the diagnosis period of the historical case cohort were preprocessed to obtain the corresponding survival time and survival status. A first elastic network penalized Cox regression model was trained based on the core diagnostic variables, and a second elastic network penalized Cox regression model was trained based on the extended diagnosis time features. The regularization strength and L1 / L2 mixture ratio of the two models were determined through cross-validation. The models were then refitted on the complete training cohort using the determined parameters. The feature preprocessing parameters, the two sets of regression coefficients, the mean score, the standard deviation of the score, and the component weights were then solidified for use in the application phase.

[0028] Step S30: Based on the standardized feature data and the treatment response prediction model, obtain the treatment response proxy score. The treatment response prediction model is a multi-task model trained under joint supervision based on survival outcome and treatment response. As an example, the treatment response prediction model is a multi-task Cox neural network model, which inputs standardized feature data into the treatment response prediction model to obtain treatment response proxy scores.

[0029] Specifically, obtaining the treatment response surrogate score based on the standardized feature data and the treatment response prediction model includes: the treatment response prediction model is a multi-task Cox neural network model; inputting the standardized feature data into the shared bottom feature extraction layer of the multi-task Cox neural network model; obtaining the first course CR / CRi prediction logit output by the treatment response prediction branch, and determining the prediction logit as the original treatment response surrogate score.

[0030] Exemplary multi-task Cox neural network models are constructed for overall survival and relapse-free survival, respectively. For any subject to be evaluated, the multidimensional feature data from the diagnostic period is first converted into a standardized feature vector using the feature preprocessing parameters fixed during model training. The neural network input vector x is then extracted from this vector using the fixed feature selection index. This input vector x is input into a shared feature encoding layer, passed through a fully connected layer with 8 hidden units, and transformed using a Tanh activation function to obtain a shared feature representation h. A survival risk residual prediction branch linearly maps h to obtain the survival risk residual representation R_NN. A treatment response prediction branch maps h to obtain the predicted logit for the treatment response in the first and second treatment cycles, respectively. The predicted logit corresponding to the CR / CRi of the first treatment cycle is used as the original surrogate score for the treatment response. This original surrogate score is then residualized and standardized relative to the baseline risk score and the extended diagnostic feature enhancement score to obtain the treatment response surrogate score S_N. The above forward computation only inputs the multidimensional feature data from the diagnostic period, excluding the actual treatment plan, actual treatment response, or minimal residual disease outcome of the subject to be evaluated.

[0031] Specifically, a multi-task Cox neural network is constructed, comprising a shared feature encoding layer, a survival risk residual prediction branch, and a treatment response prediction branch. Standardized target feature data from the diagnostic period of the training cohort is obtained as input data. Survival time and survival status corresponding to the input data are obtained as primary supervision labels, and treatment response data from the first and second treatment courses are obtained as auxiliary supervision labels. The input data is fed into the shared feature encoding layer to obtain a shared feature representation. The survival risk residual representation and treatment response prediction logit are obtained through the survival risk residual prediction branch and the treatment response prediction branch, respectively. The standardized extended diagnostic feature survival risk... The score is summed with the survival risk residual representation of the preset ratio to obtain the neural network survival risk output; the Cox partial likelihood loss is calculated based on the neural network survival risk output and survival follow-up data, and the binary cross-entropy loss is calculated based on the treatment response prediction logit and auxiliary supervision labels; the product of the Cox partial likelihood loss, the preset auxiliary task weights and the binary cross-entropy loss, and the regularization term are summed to obtain the total loss; the total loss is optimized using a gradient optimization algorithm until a preset number of iterations is reached or an early stopping condition is triggered to obtain the trained treatment response prediction model, and the network parameters of the model are solidified.

[0032] The exemplary training process of the treatment response prediction model includes: acquiring multidimensional feature data of the diagnosis period of the training cohort; preprocessing the multidimensional feature data using missing value handling parameters, categorical variable encoding rules, mean, and standard deviation obtained from the fitting within the training data; filtering the preprocessed features within the training data based on survival associations to obtain a training feature matrix containing 25 selected features; acquiring the survival time and survival status corresponding to the training subjects; and constructing auxiliary task labels based on the treatment responses of the first and second treatment courses, respectively. The training feature matrix is ​​input into the shared feature encoding layer to obtain the shared feature representation. The survival risk residual representation is output by the survival risk residual prediction branch, and the treatment response prediction logit for two treatment cycles is output by the treatment response prediction branch. The standardized extended diagnostic feature survival risk score is summed with the survival risk residual representation of a preset ratio to obtain the neural network survival risk output used to calculate the Cox partial likelihood loss; the Cox partial likelihood loss is calculated based on the neural network survival risk output, survival time, and survival status; the binary cross-entropy loss is calculated based on the two treatment response prediction logit and the corresponding auxiliary task label; for training objects with missing auxiliary task labels, they are masked to prevent them from participating in the calculation of the corresponding auxiliary loss. The Cox partial likelihood loss, the 0.3x treatment response auxiliary loss, and the regularization term are summed to obtain the total loss. The Adam optimizer is used to optimize the total loss with a learning rate of 0.005, a maximum of 250 epochs, and an early stopping mechanism. After training, the network parameters of the shared feature encoding layer, the survival risk residual prediction branch, and the treatment response prediction branch are fixed.

[0033] In this embodiment, the shared feature encoding layer includes a fully connected layer with 8 hidden units and a Tanh activation function, with an auxiliary task weight of 0.3. Multi-task Cox neural network models are trained for both overall survival and relapse-free survival outcomes.

[0034] Step S40: Obtain the RAS pathway mutation status, TP53 mutation status and corresponding missing status indicator variables of the object to be evaluated, and determine the case where either the RAS pathway mutation status or the TP53 mutation status is missing as an unknown molecular state. Step S50: Based on the baseline risk score, the treatment response proxy score, the RAS pathway mutation status, the TP53 mutation status, and the unknown molecular status indicator variable, obtain the final relative risk score through a risk correction model; The exemplary risk-adjusted model is specifically the regularized Cox proportional hazards model. This model aims to comprehensively consider a patient's baseline risk, potential treatment sensitivity, and the interactive effects of specific gene mutations, thereby outputting a more accurate final relative risk score.

[0035] Specifically, the step of obtaining the final relative risk score based on the baseline risk score, the treatment response proxy score, the RAS pathway mutation status, the TP53 mutation status, and the molecular state unknown indicator variable through a risk correction model includes: obtaining the RAS pathway mutation status, the TP53 mutation status, and the corresponding missing status indicator variable of the subject to be evaluated, and determining the absence of either the RAS pathway mutation status or the TP53 mutation status as a molecular state unknown; the risk correction model is a regularized Cox proportional hazards model; constructing an interaction feature vector, which includes the standardized baseline risk score, the standardized treatment response proxy score, and product terms formed by the treatment response proxy score and the RAS pathway mutation status, the TP53 mutation status, and the molecular state unknown indicator variable; inputting the interaction feature vector into the regularized Cox proportional hazards model with fixed parameters, calling the pre-stored fixed regression coefficients of each feature item to calculate the linear predicted value, and determining the linear predicted value as the final relative risk score.

[0036] This example demonstrates the acquisition of a core baseline risk score, an extended diagnostic feature survival risk score, a neural network residual risk representation, and a treatment response surrogate score. The core baseline risk score S_core is output from the CoreCoxnetBaseline model; the extended diagnostic feature survival risk score S_skip is output from an elastic-net Coxmodel incorporating broader diagnostic features; and the neural network residual risk representation R_NN is output from the survival prediction branch of a multi-task neural network. All scores are calculated from diagnostic variables; actual treatment protocols, actual treatment responses, and post-treatment MRD are not used as inputs for model inference.

[0037] Variable Encoding and Standardization. To eliminate dimensional differences between different scores, each continuous score was Z-score standardized using the mean and standard deviation determined in the training queue. This standardization operation is denoted by Z(·). First, the neural network residual risk representation was residualized relative to the base risk score and the extended diagnostic feature survival risk score, resulting in R_NN_perp. Further, a diagnostic wide-feature enhanced score was constructed: S_enh = Z[Z(S_skip) + γ × Z(R_NN_perp)] Where γ is the preset weight for the residual risk representation of the neural network. Subsequently, the basic risk score is fused with the diagnostic wide feature enhancement score to obtain the basic risk score: S_base = Z[(1 λ) × Z(S_core) + λ × S_enh] Where λ is the fusion weight of the extended diagnostic feature survival risk score. In this embodiment, λ=0.5 and γ=0.3 are used for the OS model, and λ=0.8 and γ=0 for the RFS model.

[0038] Generation of Neural Network Treatment Response Surrogate Score. The multi-task neural network is trained primarily on survival outcomes, with treatment responses in the first and second courses of treatment serving as auxiliary supervision tasks. In this embodiment, the predicted logit corresponding to the CR / CRi auxiliary task in the first course is used to construct the neural network treatment response surrogate score. This predicted logit is then residualized and standardized relative to S_base and S_enh to obtain: S_N = Z{Res[L_N|S_base,S_enh]} Where L_N is the treatment response prediction logit output by the neural network, and Res(·) represents the residual processing performed using the training data. This treatment response surrogate score is inferred from the variables at diagnosis time, and the model does not require the input of the patient's actual post-treatment response during inference.

[0039] Molecular state variable encoding. RAS pathway mutations and TP53 mutations are binary encoded, denoted as G_RAS and G_TP53 respectively. The presence of the corresponding mutation is encoded as 1, and the absence of the mutation is encoded as 0. Missing state indicator variables U_RAS and U_TP53 are set, and an indicator variable for unknown molecular state is defined: G_UNK = max(U_RAS, U_TP53) Calculate the interaction terms for the first risk branch. Multiply the standardized neural network treatment response surrogate score by the RAS pathway mutation, TP53 mutation, and the unknown molecular state indicator variable to construct the following interaction terms: I_RAS = S_N × G_RAS I_TP53 = S_N × G_TP53 I_UNK = S_N × G_UNK The aforementioned interaction term is used to characterize the differences in survival risk caused by the treatment response-related risk information learned by the neural network under different RAS / TP53 molecular backgrounds.

[0040] Assemble the interaction feature vector of the first risk branch. Combine the basic risk score, the neural network treatment response proxy score, and the corresponding interaction terms in a predetermined order to obtain: X_L = [S_base, S_N, I_RAS, I_TP53, I_UNK] The interaction feature vector X_L is input into a regularized Cox proportional hazards model with pre-defined parameters. The pre-stored fixed regression coefficients are then used to calculate the final relative risk score using the following formula:

[0041] Wherein, S_L is the final relative risk score; S_base is the standardized baseline risk score; S_N is the treatment response surrogate score after residualization and standardization; I_RAS=S_NG_RAS, representing the product of the treatment response surrogate score and the RAS pathway mutation status; I_TP53=S_NG_TP53, representing the product of the treatment response surrogate score and the TP53 mutation status; I_UNK=S_NG_UNK, representing the product of the treatment response surrogate score and the molecular state unknown indicator variable; β_L1 to β_L5 are the regression coefficients of the above 5 features fitted during the model training phase and fixed before application.

[0042] Final risk score calculation. In this embodiment, the risk score \(S_L\) calculated by the aforementioned regularized Cox proportional hazards model is determined as the final relative risk score, that is: \[ S_{final}=S_L \] In this context, treatment response data serves only as auxiliary supervisory labels during the training phase of the multi-task Cox neural network; when calculating \(S_{final}\), only multidimensional feature data available during the diagnostic period of the subject to be evaluated is input, without inputting the actual treatment plan, actual treatment response, or minimal residual disease results after treatment.

[0043] In other alternative implementations, a regularized classification model trained based on treatment response labels can be set up, and additional surrogate scores generated by the model based on multidimensional feature data from the diagnosis period can be used for risk correction. The feature composition, fusion method, and weights of this alternative implementation are determined by the training data and do not constitute a necessary component of the current final risk score in this embodiment.

[0044] Vector Input and Model Calculation. The constructed interaction feature vector X_L is input into the regularized Cox proportional hazards model with fixed parameters. The model then uses pre-stored fixed regression coefficients to calculate a linear prediction value, which is used as the final relative risk score. After model training is complete, the data preprocessing parameters, sub-model parameters, score standardization parameters, and risk-corrected model regression coefficients are all fixed, and refitting is not performed during the inference phase.

[0045] Step S60: Predict the patient's risk stratification level based on the final relative risk score.

[0046] As an example, the final relative risk score is compared with multiple thresholds to obtain the corresponding risk stratification level. For example, the 33rd and 67th percentiles of the final relative risk score distribution of the training cohort are used as stratification thresholds; below the 33rd percentile is low risk, from the 33rd to the 67th percentile is medium risk, and above or equal to the 67th percentile is high risk.

[0047] In this embodiment, the multidimensional feature data of the subject to be evaluated at the diagnostic stage are preprocessed to obtain standardized feature data, which includes at least clinical feature data, laboratory test data, and molecular genetic data. Based on the standardized feature data and a pre-set basic survival model, a basic risk score is obtained. Based on the standardized feature data and a treatment response prediction model, a treatment response surrogate score is obtained, whereby the treatment response prediction model is a multi-task model trained under joint supervision of survival outcome and treatment response. The RAS pathway mutation status, TP53 mutation status, and corresponding deletion status indicator variables of the subject to be evaluated are obtained, and the absence of either the RAS pathway mutation status or the TP53 mutation status is defined as an unknown molecular status. Based on the basic risk score, the treatment response surrogate score, the RAS pathway mutation status, the TP53 mutation status, and the unknown molecular status indicator variables, a final relative risk score is calculated using a risk correction model. The patient's risk stratification level is predicted based on the final relative risk score, thus solving the technical problem in related technologies where risk stratification cannot identify the high heterogeneity within this subtype.

[0048] Secondly, embodiments of this application also provide a risk prediction device based on diagnostic data.

[0049] In one embodiment, reference is made to Figure 2 , Figure 2 This is a schematic diagram of the functional modules of an embodiment of the risk prediction device based on diagnostic data according to this application. Figure 2 As shown, the risk prediction device based on diagnostic data includes: The first acquisition module 10 is used to preprocess the multidimensional feature data of the subject to be evaluated during the diagnosis period to obtain standardized feature data. The multidimensional feature data includes at least clinical feature data, laboratory test data and molecular genetic data. The second acquisition module 20 is used to acquire a basic risk score based on the standardized feature data and the preset basic survival model; The third acquisition module 30 is used to acquire a treatment response proxy score based on the standardized feature data and the treatment response prediction model. The treatment response prediction model is a multi-task model trained under joint supervision based on survival outcome and treatment response. The fourth acquisition module 40 is used to acquire the RAS pathway mutation status, TP53 mutation status and corresponding missing status indicator variables of the object to be evaluated, and to determine the case where either the RAS pathway mutation status or the TP53 mutation status is missing as an unknown molecular state. Calculation module 50 is used to calculate the final relative risk score based on the basic risk score, the treatment response proxy score, the RAS pathway mutation status, the TP53 mutation status, and unknown molecular status indicator variables, using a risk correction model. The prediction module 60 is used to predict the patient's risk stratification level based on the final relative risk score.

[0050] Furthermore, in one embodiment, the second acquisition module 20 is used for: Core diagnostic variables and extended diagnostic features are extracted from standardized feature data. The core diagnostic variables are input into a first elastic network penalized Cox regression model to obtain a core basic risk score. The extended diagnostic features are input into a second elastic network penalized Cox regression model to obtain an extended diagnostic feature survival risk score. The two scores are standardized by calling pre-stored scoring standardization parameters and combined according to preset component weights to obtain a basic risk score.

[0051] Furthermore, in one embodiment, the third acquisition module 30 is used for: The treatment response prediction model is a multi-task Cox neural network model, which includes a shared feature encoding layer, a survival risk residual prediction branch, and a treatment response prediction branch. The standardized feature data is input into the shared feature encoding layer to obtain the shared feature representation; The shared feature representation is input into the survival risk residual prediction branch and the treatment response prediction branch respectively to obtain the survival risk residual representation and the predicted logit corresponding to the first course of treatment CR / CRi; the predicted logit is determined as the original surrogate score of treatment response, and the treatment response surrogate score is obtained after residualization and standardization.

[0052] Furthermore, in one embodiment, the calculation module 50 is used for: The risk correction model is a regularized Cox proportional hazards model. An interactive feature vector is constructed, which includes a standardized baseline risk score, a standardized treatment response proxy score, and product terms of the treatment response proxy score with RAS pathway mutation status, TP53 mutation status, and unknown molecular status indicator variables, respectively. The interaction feature vector is input into a regularized Cox proportional hazards model with fixed parameters. The model then calls the pre-stored fixed regression coefficients of each feature item to calculate a linear prediction value, and the linear prediction value is determined as the final relative risk score.

[0053] Furthermore, in one embodiment, the first acquisition module 10 is used to: The clinical characteristic data include at least one of the following: age, sex, physical performance status score, liver and spleen parameters, and secondary factors; The laboratory test data include at least one of the following: white blood cell count, hemoglobin, platelet count, bone marrow blast cell ratio, lactate dehydrogenase level, and coagulation function indicators. The molecular genetic data includes at least one of the following: cytogenetic karyotype analysis results, qualitative results of fusion genes, and gene mutation detection results; Missing values ​​in the multidimensional feature data are filled using multiple interpolation or median imputation. One-hot encoding or embedded encoding is used to digitize the categorical variables in the multidimensional feature data; The continuous variables in the multidimensional feature data are dimensionless by using Z-score standardization or logarithmic transformation to obtain standardized feature data.

[0054] Furthermore, in one embodiment, the risk prediction device based on diagnostic data further includes a basic survival model training module, used for: Obtain target multidimensional feature data for the diagnosis period of historical case cohorts; Preset core diagnostic variables are extracted from the target multidimensional feature data to construct a first training feature matrix, and a first elastic net penalized Cox regression model is trained using the elastic net penalized Cox regression algorithm. Extended diagnostic features are extracted from the target multidimensional feature data to construct a second training feature matrix, and a second elastic net penalized Cox regression model is trained using the elastic net penalized Cox regression algorithm. The regularization parameters of the two elastic net penalized Cox regression models were determined by cross-validation. Using defined regularization parameters, first elastic net penalized Cox regression models and second elastic net penalized Cox regression models were fitted to the historical case cohort, and the feature preprocessing parameters, the two sets of regression coefficients, and the risk score standardization parameters were solidified.

[0055] Furthermore, in one embodiment, the risk prediction device based on diagnostic data further includes a treatment response prediction model training module, used for: A multi-task Cox neural network was constructed, comprising a shared feature encoding layer, a survival risk residual prediction branch, and a treatment response prediction branch. Obtain standardized feature data of the target population, survival follow-up labels, and auxiliary labels of treatment response during the first and second courses of treatment from the training cohort at the diagnostic period; The survival risk residual representation and treatment response prediction logit are obtained by using a shared feature coding layer, a survival risk residual prediction branch, and a treatment response prediction branch, respectively. The Cox partial likelihood loss is calculated based on the neural network survival risk output and survival follow-up label, and the binary cross-entropy loss is calculated based on the treatment response prediction logit and auxiliary label. Gradient optimization is performed on the total loss consisting of Cox partial likelihood loss, weighted treatment response auxiliary loss, and regularization term, and the trained network parameters are then fixed.

[0056] The functions of each module in the aforementioned risk prediction device based on diagnostic data correspond to the steps in the aforementioned risk prediction method embodiment based on diagnostic data, and their functions and implementation processes will not be described in detail here.

[0057] Thirdly, embodiments of this application provide a risk prediction device based on diagnostic data. The risk prediction device based on diagnostic data can be a personal computer (PC), a laptop computer, a server, or other device with data processing capabilities.

[0058] Reference Figure 3 , Figure 3 This is a schematic diagram of the hardware structure of a risk prediction device based on diagnostic data involved in an embodiment of this application. In this embodiment, the risk prediction device based on diagnostic data may include a processor, a memory, a communication interface, and a communication bus.

[0059] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0060] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting internal components of the diagnostic data-based risk prediction device, as well as interfaces used for interconnecting the diagnostic data-based risk prediction device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0061] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0062] The processor can be a general-purpose processor, which can call a risk prediction program based on diagnostic data stored in memory and execute the risk prediction method based on diagnostic data provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the risk prediction program based on diagnostic data is called can be referred to in the various embodiments of the risk prediction method based on diagnostic data in this application, and will not be repeated here.

[0063] Those skilled in the art will understand that Figure 3 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0064] Fourthly, embodiments of this application also provide a computer-readable storage medium.

[0065] The present application provides a computer-readable storage medium storing a risk prediction program based on diagnostic data, wherein when the risk prediction program based on diagnostic data is executed by a processor, it implements the steps of the risk prediction method based on diagnostic data as described above.

[0066] The method implemented when the risk prediction procedure based on diagnostic data is executed can be referred to in various embodiments of the risk prediction method based on diagnostic data in this application, and will not be repeated here.

[0067] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0068] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0069] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0070] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0071] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0072] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0073] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A risk prediction method based on diagnostic data, characterized in that, include: The multidimensional feature data of the subject to be evaluated during the diagnostic period are preprocessed to obtain standardized feature data, which includes at least clinical feature data, laboratory test data and molecular genetic data. Based on the standardized feature data and the pre-set basic survival model, a basic risk score is obtained; Based on the standardized feature data and the treatment response prediction model, a treatment response proxy score is obtained. The treatment response prediction model is a multi-task model trained under joint supervision of survival outcome and treatment response. Obtain the RAS pathway mutation status, TP53 mutation status, and corresponding missing status indicator variables of the object to be evaluated, and determine the missing status of either the RAS pathway mutation status or the TP53 mutation status as an indicator variable of unknown molecular status. Based on the baseline risk score, the treatment response proxy score, the RAS pathway mutation status, the TP53 mutation status, and the indicator variable with unknown molecular status, the final relative risk score is calculated using a risk-corrected model. The patient's risk stratification level is predicted based on the final relative risk score.

2. The risk prediction method based on diagnostic data as described in claim 1, characterized in that, The process of obtaining a basic risk score based on the standardized feature data and the pre-set basic survival model includes: The pre-set basic survival model includes a first elastic network penalized Cox regression model and a second elastic network penalized Cox regression model. Extract core diagnostic variables from the standardized feature data, input them into the trained first elastic network penalized Cox regression model, and calculate the core basic risk score using the pre-fixed first regression coefficients; Extended diagnostic features are extracted from the standardized feature data and input into the trained second elastic network penalized Cox regression model. The survival risk score of the extended diagnostic features is calculated using the pre-fixed second regression coefficients. The core basic risk score and the extended diagnostic feature survival risk score are standardized by calling the pre-fixed mean and standard deviation respectively. The standardized core basic risk score and the extended diagnostic feature survival risk score are combined according to the preset component weights to obtain the basic risk score.

3. The risk prediction method based on diagnostic data as described in claim 1, characterized in that, The step of obtaining a treatment response proxy score based on the standardized feature data and the treatment response prediction model includes: The treatment response prediction model is a multi-task Cox neural network model, which includes a shared feature encoding layer, a survival risk residual prediction branch connected to the shared feature encoding layer, and a treatment response prediction branch connected to the shared feature encoding layer. The standardized feature data is input into the shared feature encoding layer to obtain the shared feature representation; The shared feature representation is input into the survival risk residual prediction branch and the treatment response prediction branch respectively to obtain the survival risk residual characterization and the predicted logit corresponding to complete remission in the first course of treatment or complete remission with incomplete hematological recovery. The predicted logit corresponding to complete remission or complete remission with incomplete hematological recovery in the first course of treatment is determined as the original surrogate score of the treatment response. The original surrogate score of the treatment response is then subjected to residualization and standardization to obtain the treatment response surrogate score.

4. The risk prediction method based on diagnostic data as described in claim 1, characterized in that, The final relative risk score is obtained through a risk-corrected model based on the baseline risk score, the treatment response proxy score, RAS pathway mutation status, TP53 mutation status, and unknown molecular status indicators, including: The risk correction model is a regularized Cox proportional hazards model. An interactive feature vector is constructed, which includes a standardized baseline risk score, a standardized treatment response proxy score, and product terms of the treatment response proxy score with RAS pathway mutation status, TP53 mutation status, and unknown molecular status indicator variables, respectively. The interaction feature vector is input into a regularized Cox proportional hazards model with fixed parameters. The model then calls the pre-stored fixed regression coefficients of each feature item to calculate a linear prediction value, and the linear prediction value is determined as the final relative risk score.

5. The risk prediction method based on diagnostic data as described in claim 1, characterized in that, The process of preprocessing the multidimensional feature data of the subject to be evaluated during the diagnostic period to obtain standardized feature data includes: The clinical characteristic data include at least one of the following: age, sex, physical performance status score, liver and spleen parameters, and secondary factors; The laboratory test data include at least one of the following: white blood cell count, hemoglobin, platelet count, bone marrow blast cell ratio, lactate dehydrogenase level, and coagulation function indicators. The molecular genetic data includes at least one of the following: cytogenetic karyotype analysis results, qualitative results of fusion genes, and gene mutation detection results; Missing values ​​in the multidimensional feature data are filled using multiple interpolation or median imputation. One-hot encoding or embedded encoding is used to digitize the categorical variables in the multidimensional feature data; The continuous variables in the multidimensional feature data are made dimensionless by using Z-score standardization or logarithmic transformation to obtain standardized feature data.

6. The risk prediction method based on diagnostic data as described in claim 1, characterized in that, The training process of the pre-built basic survival model includes: Obtain target multidimensional feature data for the diagnosis period of historical case cohorts; Preset core diagnostic variables are extracted from the target multidimensional feature data to construct a first training feature matrix, and a first elastic net penalized Cox regression model is trained using the elastic net penalized Cox regression algorithm. Extended diagnostic features are extracted from the target multidimensional feature data to construct a second training feature matrix, and a second elastic net penalized Cox regression model is trained using the elastic net penalized Cox regression algorithm. Cross-validation was used to determine the regularization strength and the mixing ratio of L1 and L2 regularization terms for the two elastic net penalized Cox regression models. Using the determined model parameters, the two models were refitted on the historical case cohort, and the feature preprocessing parameters, regression coefficients, and risk score standardization parameters were solidified to obtain the pre-set basic survival model.

7. The risk prediction method based on diagnostic data as described in claim 1, characterized in that, The training process of the treatment response prediction model includes: A multi-task Cox neural network was constructed, comprising a shared feature encoding layer, a survival risk residual prediction branch, and a treatment response prediction branch. The target standardized feature data of the training queue during the diagnosis period is obtained as input data, the survival time and survival status corresponding to the input data are obtained as the main supervision label, and the treatment response data of the first and second courses of treatment are obtained as auxiliary supervision labels. The input data is fed into the shared feature encoding layer to obtain the shared feature representation, and the survival risk residual characterization and treatment response prediction logit are obtained through the survival risk residual prediction branch and the treatment response prediction branch, respectively. The standardized extended diagnostic feature survival risk score is summed with the survival risk residual representation at a preset ratio to obtain the neural network survival risk output; the Cox partial likelihood loss is calculated based on the neural network survival risk output and survival follow-up data, and the binary cross-entropy loss is calculated based on the treatment response prediction logit and auxiliary supervision label. The total loss is obtained by summing the product of the Cox partial likelihood loss, the preset auxiliary task weights, and the binary cross-entropy loss, along with the regularization term. The total loss is optimized using a gradient optimization algorithm until a preset number of iterations is reached or an early stopping condition is triggered, resulting in a trained treatment response prediction model. The network parameters of the model are then solidified.

8. A risk prediction device based on diagnostic data, characterized in that, The risk prediction device based on diagnostic data includes: The first acquisition module is used to preprocess the multidimensional feature data of the subject to be evaluated during the diagnostic period to obtain standardized feature data. The multidimensional feature data includes at least clinical feature data, laboratory test data and molecular genetic data. The second acquisition module is used to acquire a basic risk score based on the standardized feature data and the pre-set basic survival model; The third acquisition module is used to acquire the treatment response proxy score based on the standardized feature data and the treatment response prediction model. The treatment response prediction model is a multi-task model trained under joint supervision based on survival outcome and treatment response. The fourth acquisition module is used to acquire the RAS pathway mutation status, TP53 mutation status and corresponding missing status indicator variables of the object to be evaluated, and to determine the case where either the RAS pathway mutation status or the TP53 mutation status is missing as an unknown molecular state. The calculation module is used to calculate the final relative risk score based on the basic risk score, the treatment response proxy score, the RAS pathway mutation status, the TP53 mutation status, and unknown molecular status indicator variables, using a risk correction model. The prediction module is used to predict the patient's risk stratification level based on the final relative risk score.

9. A risk prediction device based on diagnostic data, characterized in that, The diagnostic data-based risk prediction device includes a processor, a memory, and a diagnostic data-based risk prediction program stored in the memory and executable by the processor, wherein when the diagnostic data-based risk prediction program is executed by the processor, it implements the steps of the diagnostic data-based risk prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a risk prediction program based on diagnostic data, wherein when the risk prediction program based on diagnostic data is executed by a processor, it implements the steps of the risk prediction method based on diagnostic data as described in any one of claims 1 to 7.

Images