An acute lymphoblastic leukemia scoring system

Abstract

This invention belongs to the field of acute lymphoblastic leukemia (ALL) diagnosis and treatment technology, and discloses an ALL scoring system. The feature library construction module establishes an associated feature system based on subtype molecular biological differences, uses LASSO regression and subtype stratified analysis to screen core features and determine weights, and excludes redundant cross-subtype information. The feature mining module targets the entire treatment cycle, dividing the collection nodes into induction, consolidation, and maintenance phases, capturing temporal correlation features such as the rate of MRD decline and gene expression change rate, breaking the limitation of relying solely on static data at the time of diagnosis. It not only fits the individual characteristics of patients with different subtypes, but also reflects the dynamic changes in the treatment process, improving the accuracy of prognostic assessment. A real-time corrected score is obtained by calculating the baseline score and the temporal feature adjustment score. Patient data is updated weekly through the electronic medical record interface, and appropriate solutions are output for different situations such as low-risk, intermediate-risk, and very high-risk.

Description

Technical Field

[0001] This invention belongs to the field of acute lymphoblastic leukemia diagnosis and treatment technology, specifically an acute lymphoblastic leukemia scoring system. Background Technology

[0002] Acute lymphoblastic leukemia (ALL) is a common hematologic malignancy with a high incidence in children and a relatively poor prognosis in adults. Clinical practice focuses on precise risk stratification to guide individualized treatment, aiming to balance efficacy and treatment toxicity. While existing scoring systems can achieve basic risk stratification, they still face the following technical challenges:

[0003] Existing systems only perform single-dimensional summation scoring of clinical indicators (age, white blood cell count) and molecular markers (fusion genes, etc.), without considering the interaction effect between genes and clinical indicators. For example, even if children and adult patients have the same high white blood cell count and specific fusion genes, their prognoses are still significantly different. Moreover, only the MRD value at the time of diagnosis is used, without taking into account the changes in MRD throughout the treatment process, which is crucial for relapse assessment.

[0004] Leukemia is also associated with osteoclasts, particularly in acute lymphoblastic leukemia (ALL). Leukemia cells secrete factors like RANKL, which overactivate osteoclasts, leading to osteolytic lesions (manifested as bone pain and osteoporosis, especially common in children). This represents a direct erosion of the bone by leukemia. There is also an indirect link between leukemia and gout, which is a manifestation of tumor lysis syndrome. During chemotherapy in ALL patients, a large number of leukemia cells are rapidly destroyed, releasing uric acid and causing secondary hyperuricemia, thus triggering gout. However, this differs from primary metabolic gout.

[0005] The existing system only outputs low / medium / high risk labels, without stratified treatment pathways. Patients with different gene mutations within the same high-risk group respond differently to chemotherapy, which can easily lead to overtreatment or undertreatment. It also does not combine gene characteristics to predict adverse reactions, nor does it have a prognosis-treatment effect feedback channel, resulting in a lag in dynamic decision-making. Summary of the Invention

[0006] The purpose of this invention is to provide an acute lymphoblastic leukemia scoring system to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an acute lymphoblastic leukemia scoring system, comprising a feature library construction module, a feature mining module, a feature screening module, an adjustment module, an early warning and intervention module, a training module, and a desensitization and iteration module;

[0008] Preferably, the feature library construction module constructs a classification association feature system based on the molecular biological differences of ALL subtypes, and clarifies the core association features and weight ratios of each subtype; it connects to a multi-center ALL subtype database, and the data dimensions include clinical indicators, molecular markers, and immunophenotypes; it uses LASSO regression combined with subtype stratified analysis, firstly screening features that are significantly associated with prognosis within a single subtype through LASSO regression, then comparing the differences in features between different subtypes through stratified analysis, excluding redundant features across subtypes, and finally determining the weight of each feature through 5-fold cross-validation;

[0009] The feature library is stored in a five-level structure: subtype, feature category, weight value, data source, and validation results.

[0010] Output a subtype-specific associated feature library and a feature weight table. The feature weight table includes feature name, subtype affiliation, weight value, data dimension, and number of validation centers.

[0011] Preferably, the feature mining module is based on the subtype-specific associated feature library and feature weight table of the feature library construction module. For the core associated features of each subtype that have been determined, it performs time-series dynamic mining throughout the entire treatment cycle to capture the dynamic correlation changes between gene expression data, clinical response indicators and MRD, and generate time-series associated features. The collection nodes are divided according to the treatment stage. During the induction period, three key nodes are set: d15, d33 and d45. During the consolidation period, data is collected every 3 weeks, and during the maintenance period, data is collected every 2 months. Each time, MRD values, gene expression data and clinical response indicators are collected simultaneously.

[0012] A sliding window time-series analysis and linear fitting were employed, with the sliding window size set to two consecutive treatment cycles. The slope of MRD decline was calculated, and the association features between genes and MRD were generated by combining the gene expression trends during the same period. The time-series association features were updated weekly, with the update trigger condition being that the amount of newly collected data reached 5% of the number of cases in that subtype.

[0013] The rate of change in gene expression was calculated using the logarithm of the post-treatment node gene expression level / baseline gene expression level, as shown in the following formula:

[0014]

[0015] in, The expression level of the target gene at a certain collection point during the treatment phase (detected by real-time quantitative PCR, with GAPDH as the internal reference gene, and the relative expression level calculated using the 2^-ΔΔCt method).

[0016] The baseline gene expression level at diagnosis; the time point is defined as: induction period with day 15 as the baseline. d33 is d45 is The consolidation period is divided into three-week intervals. The maintenance period is in two-month intervals. -0.5 indicates a significant downregulation of gene expression. 0.5 indicates that gene expression is significantly upregulated.

[0017] Output subtype time-series associated feature set, including MRD decrease slope, gene expression change rate, and clinical response time-series curve.

[0018] The slope of the MRD descent was calculated using linear regression fitting, as shown in the following formula:

[0019]

[0020] in, The number of data collection nodes within the sliding window is fixed at 2, corresponding to the data collection data of 2 consecutive treatment cycles;

[0021] For the first The time of each data collection node is based on the patient's date of diagnosis. The unit is days;

[0022] For the first The MRD values ​​of each acquisition node need to be logarithmically transformed first. The transformation formula is: value This is to avoid the inability to calculate when the MRD value is 0;

[0023] If the calculation yields A negative value indicates that the MRD value decreases over time. The larger the absolute value, the faster the MRD decreases; if A positive value indicates that the MRD value increases with treatment time, suggesting a poor treatment response.

[0024] Preferably, the feature screening module uses the subtype basic association features of the feature library construction module and the subtype temporal association feature set of the feature mining module as input to construct a candidate feature pool, screen causal features that affect the prognosis, and eliminate false associations; a two-step method of univariate causal verification and multivariate correlation filtering is adopted. In the first step, univariate causal verification uses Kaplan-Meier survival analysis to calculate the log-rank P value of each feature and the event-free survival rate. If P < 0.05, it is determined to be directly associated with the prognosis.

[0025] The second step, multivariate correlation filtering, uses the Pearson correlation coefficient to calculate the correlation coefficient between any two directly related features. A correlation coefficient > 0.8 is considered highly co-occurring, and features with weak clinical significance are removed. Through the above process, 8 to 10 core causal association features are retained, with the retention criterion being a causal effect value > 0.3, and consistent on three independent center datasets.

[0026] Output a list of core causal association features of the subtype and the causal effect determination results. The list includes feature name, causal effect value, number of validation centers, and reason for exclusion.

[0027] The causal effect value was calculated using a Cox proportional hazards regression model, with the model parameters set as follows:

[0028] Event time was defined as event-free survival (EFS, the time from diagnosis to relapse, death, or last follow-up). Outcome variables were binary, 1 = event occurred, 0 = no event occurred. Covariates were univariate features that passed causal validation. The model used maximum likelihood estimation to solve for regression coefficients. The hazard ratio (HR) was used as a quantitative indicator of the causal effect value. HR > 1 indicated that the feature was a poor prognostic factor, and HR < 1 indicated that the feature was a good prognostic factor. After model fitting, the proportional hazards hypothesis was tested using the Schönfeld residual test (P > 0.05 was considered met). If the hypothesis was not met, time-dependent covariates were introduced to adjust the model.

[0029] Preferably, the adjustment module is based on the subtype core causal association feature list of the feature screening module, constructs a feature, score and plan matching rule library, establishes a real-time linkage mechanism of associated features → score correction → plan adjustment, and dynamically optimizes the treatment plan; it has a built-in matching rule library of more than 20 ALL treatment scenarios, and the rule formulation needs to combine the core feature weights and the clinical guideline recommended plan;

[0030] The system uses rule-based matching and clinical threshold judgment to calculate and correct the score. The base score is the sum of the weights of each core feature, and the corrected score is the base score plus the time-series feature adjustment score. The corrected score is updated weekly based on the real-time time-series features of the feature mining module. The system integrates the real-time data interface of electronic medical records to obtain patients' blood routine, liver and kidney function, treatment and medication records, and adverse reaction occurrences in real time. At the same time, the system sets physician intervention permissions. When the adjustment of the plan involves changes in the treatment intensity level, it must be reviewed and confirmed by the attending physician before it is implemented.

[0031] Outputs real-time corrected patient scores and appropriate treatment plans, including score composition, plan name, medication dosage, dosing cycle, and monitoring indicators.

[0032] Preferably, the early warning and intervention module takes the patient's real-time correction score and appropriate treatment plan as the core input, and combines clinical sign data and multi-omics data to construct an adverse reaction early warning model; a weighted allocation method is used for multimodal fusion, first standardizing the data of each dimension, and then calculating the comprehensive risk value according to the weight. The weight of high-risk features is set to 40%, the weight of clinical indicators is set to 35%, and the weight of multi-omics data is set to 25%. An early warning is triggered when the comprehensive risk value is >0.7.

[0033] The adverse reaction early warning model adopts the Gradient Boosting Decision Tree (GBDT) model, with the following specific structure: 100 decision trees, a maximum depth of 5 for each decision tree, a learning rate of 0.05, a minimum number of sample splits of 20, and a minimum number of leaf nodes of 10. The model input features are standardized multi-dimensional data. The way each dimension of data is mapped to the model input features is as follows: clinical sign data (12 features), multi-omics data (8 features), real-time corrected scores (1 feature), and adaptive treatment plan coding features (numerical features coded as 0-4 according to treatment intensity level, 1 feature) are concatenated to form a 22-dimensional input feature vector.

[0034] The specific steps of multimodal fusion are as follows:

[0035] ① Data standardization: Z-score standardization is adopted, and the formula is: z=(x-μ) / σ, where x is the original data, μ is the mean of the feature in the training set, and σ is the standard deviation of the feature in the training set;

[0036] ② Weight allocation basis: Based on the feature importance analysis results of multicenter clinical data, the average importance of each dimension feature is calculated by the feature gain value of the GBDT model. The average gain value of high-risk features accounts for 38%-42%, clinical indicators account for 33%-37%, and multi-omics data accounts for 23%-27%. Finally, the median value is taken as the fixed weight.

[0037] ③ Calculation of comprehensive risk value: First, the standardized data of each dimension are multiplied by their corresponding weights, and then the sum is obtained to obtain the original risk value. Finally, the data is normalized to the interval [0, 1] by the sigmoid function. The formula is: Comprehensive risk value = 1 / (1 + e^(-original risk value)), where the original risk value = weighted sum of high-risk features + weighted sum of clinical indicators + weighted sum of multi-omics data.

[0038] The intervention protocol library is stored according to the type of adverse reaction, risk level and intervention measures. The protocol library is updated quarterly with the latest clinical research evidence. The early warning results are output in the form of a pop-up window in the EMR system and pushed to the attending physician's mobile terminal at the same time, showing the risk level, correlation characteristics, recommended intervention measures and evidence level.

[0039] Output adverse reaction warning results and intervention plans to the adjustment module and training module.

[0040] Preferably, the training module divides federated nodes based on the ALL subtype database of the feature library construction module. The node types are divided into three categories: children's B-ALL center, adults' T-ALL center, and mixed subtype center, with no less than 3 nodes in each category. At the same time, the local adverse reaction warning and intervention data of the warning and intervention module are input to realize the collaborative training of subtype-specific sub-models.

[0041] Each node trains a sub-model of the corresponding subtype's associated features based on local subtype data and feedback data from the early warning and intervention module using a logistic regression algorithm. The training parameters are set to 100 iterations, a learning rate of 0.01, and a regularization coefficient of 0.001. Five-fold cross-validation is used to evaluate the model performance. Before aggregation, the feature consistency of each sub-model is verified (core feature overlap rate > 90%). During aggregation, weights are allocated according to the proportion of data volume of each node (data volume proportion = number of cases in that node / total number of cases in all nodes). The global model parameters are calculated by weighted average.

[0042] Output the sub-models and feature weight update results of global collaborative training, including sub-model parameters, feature weight adjustment values, and model validation metrics.

[0043] The specific structure of the logistic regression model is as follows: the input layer is an 8-10 dimensional core causal association feature vector, the model includes a bias term (initial value set to 0), the activation function is the sigmoid function, and the output is the prognostic risk probability (0-1 interval).

[0044] Feature preprocessing steps before model training:

[0045] ① Missing value imputation: Multiple imputation was used, with 5 imputations performed. An imputation model was constructed based on the distribution characteristics of clinical indicators and molecular markers of this subtype.

[0046] ② Feature standardization: min-max standardization is adopted, and the formula is: x'=(x-x_min) / (x_max-x_min), where x_min is the minimum value of the feature in the local training set, and x_max is the maximum value of the feature in the local training set. After standardization, the feature value range is [0, 1].

[0047] ③ Feature encoding: For categorical features, one-hot encoding is used to convert them into numerical features;

[0048] The specific steps of model training are as follows:

[0049] Data partitioning: The local subtype data is divided into training and validation sets in a 7:3 ratio. Stratified sampling is used to ensure that the incidence of prognostic events is consistent between the training and validation sets.

[0050] Initialization parameters: Set the regression coefficients to an all-zero vector and the bias term to 0.

[0051] Iterative training: The stochastic gradient descent optimization algorithm is used to minimize the cross-entropy loss function. In each iteration, 32 samples are randomly selected as batch data. The update formula is: θ=θ-η×∇L(θ), where θ is the model parameter, η is the learning rate (0.01), and ∇L(θ) is the gradient of the loss function.

[0052] Early stopping strategy: When the AUC of the validation set shows no improvement for 5 consecutive iterations, stop training and retain the optimal model parameters;

[0053] Model evaluation: In addition to AUC, accuracy (ACC), sensitivity (SEN), and specificity (SPE) are also calculated. The calculation method for the evaluation metrics is as follows:

[0054] ACC=(TP+TN) / (TP+TN+FP+FN)

[0055] SEN = TP / (TP + FN)

[0056] SPE = TN / (TN + FP)

[0057] TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative.

[0058] Preferably, the desensitization and iteration module takes the global subtype submodel and feature weights of the training module as its core, and synchronously inputs the highly sensitive gene features involved in the feature library construction module and feature screening module, as well as the clinical feedback data from the early warning and intervention module, to achieve privacy protection and cross-center iterative optimization of sensitive gene features, and ensure the security of multi-center data collaboration and the continuous effectiveness of the model.

[0059] Feature transformation and fixed noise addition are used to desensitize highly sensitive gene features. First, the mutant or wild type of the gene is converted into a mutation risk index. Then, fixed noise is added during feature aggregation. After desensitization, the consistency of feature prediction needs to be verified by confusion matrix.

[0060] The anonymized features and model parameters are transmitted using SSL / TLS 1.3 encryption, and the node identity is verified through digital certificates during transmission. Cross-center iteration is triggered quarterly based on feedback data from the early warning and intervention module. The iteration process is as follows: the leading center initiates an iteration request → each node verifies the validity of the current features based on the latest local data → feedback of verification results → the leading center calculates the feature weight adjustment value (adjustment range ≤ 20% of the original weight) → synchronizes and updates to all nodes → verifies the global model performance.

[0061] Output the de-identified cross-center shared feature set and the updated global model to the feature library construction module and adjustment module.

[0062] The beneficial effects of this invention are as follows:

[0063] 1. The feature library construction module of this invention establishes an associated feature system based on the molecular biological differences of subtypes. It uses LASSO regression and subtype hierarchical analysis to screen core features and determine weights, and excludes redundant information across subtypes. The feature mining module targets the entire treatment cycle, dividing the collection nodes into induction, consolidation and maintenance phases, and capturing time-series associated features such as the slope of MRD decline and gene expression change rate, breaking the limitation of relying solely on static data at the time of diagnosis. It not only fits the individual characteristics of patients with different subtypes, but also reflects the dynamic changes in the treatment process, effectively improving the accuracy of prognostic assessment.

[0064] 2. The adjustment module of this invention is based on core causal features and constructs a rule library for matching more than 20 treatment scenarios, establishing a linkage mechanism of real-time correlation features → score correction → treatment plan adjustment. It calculates the real-time corrected score by calculating the basic score and the time-series feature adjustment score, and updates patient data weekly through the electronic medical record interface. It outputs appropriate plans for different situations such as low risk, intermediate risk, and very high risk, such as VDP standard chemotherapy for low risk and CD19 CAR-T therapy for very high risk. At the same time, it sets doctor intervention permissions to ensure that the adjustment of treatment intensity must be reviewed, avoiding over-treatment or under-treatment caused by the homogenization of treatment within the same risk group, and realizing the dynamic optimization of individualized plans.

[0065] 3. The early warning module of this invention integrates clinical sign data and multi-omics data to construct a model, and uses a weighted allocation method to calculate the comprehensive risk value. High-risk features account for 40%, clinical indicators account for 35%, and multi-omics data account for 25%. When the risk value exceeds 0.7, an early warning is triggered, and the intervention plan is simultaneously pushed to the EMR system and the doctor's mobile terminal to reduce the incidence of grade III-IV adverse reactions. The training module divides federated nodes according to subtypes and trains sub-models collaboratively based on local data. It combines SSL / TLS 1.3 encryption and feature conversion technology to protect the privacy of highly sensitive genes. The desensitization and iteration module updates the global model every quarter based on clinical feedback, forming a complete workflow of assessment, treatment, early warning and iteration, which not only ensures the security of multi-center data collaboration, but also continuously improves the model's generalization ability. Attached Figure Description

[0066] Figure 1 This is a flowchart of the acute lymphoblastic leukemia scoring system of the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] like Figure 1 As shown, this embodiment of the invention provides an acute lymphoblastic leukemia scoring system, including a feature library construction module, a feature mining module, a feature screening module, an adjustment module, an early warning and intervention module, a training module, and a desensitization and iteration module. The specific implementation of each module is as follows:

[0069] The feature library construction module constructs a classification association feature system based on the molecular biological differences of ALL subtypes, which solves the prognostic prediction bias caused by the "one-size-fits-all" approach of general modeling, and clarifies the core association features and weight ratios of each subtype; ALL subtypes include childhood B-ALL, adult T-ALL, and mixed subtypes;

[0070] The database is connected to a multi-center ALL subtype database. The database must contain no fewer than 1,000 cases for each subtype, and the data dimensions include clinical indicators, molecular markers, and immunophenotypes. LASSO regression combined with subtype stratified analysis is used. First, LASSO regression is used to screen features that are significantly associated with prognosis within a single subtype. Then, stratified analysis is used to compare the differences in features between different subtypes, exclude redundant features across subtypes, and finally determine the weight of each feature through 5-fold cross-validation.

[0071] Clinical indicators include age, white blood cell count, prednisone response, and bone marrow remission status; molecular markers include fusion genes such as BCR-ABL1 and ETV6-RUNX1, and mutant genes such as TP53 and IKZF1; immunophenotype refers to the expression levels of CD20, CD4, CD8, and CD19.

[0072] The feature library is stored in a five-level structure: subtype, feature category, weight value, data source and verification results. For example, in children with B-ALL, the weight of "ETV6-RUNX1 fusion and MRD negative conversion speed (d15→d33)" is 40%, and in adults with T-ALL, the weight of "PI3K mutation and CD4 / CD8 double positivity" is set to 1.2. Both are marked as having data verified by more than 3 tertiary hospitals.

[0073] Output a subtype-specific associated feature library and feature weight table. The feature weight table includes fields such as feature name, subtype affiliation, weight value, data dimension, and number of validation centers.

[0074] The feature mining module is based on the subtype-specific associated feature library and feature weight table of the feature library construction module. It performs time-series dynamic mining of the entire treatment cycle for the core associated features that have been determined for each subtype, captures the dynamic association changes of gene expression data, clinical response indicators and MRD, and generates time-series associated features.

[0075] Data collection points are divided according to the treatment stage. During the induction period, three key points are set: day 15, day 33, and day 45. During the consolidation period, data are collected every 3 weeks, and during the maintenance period, data are collected every 2 months. Each time, MRD values, gene expression data, and clinical response indicators are collected simultaneously.

[0076] A sliding window time-series analysis and linear fitting were employed, with the sliding window size set to include time-series data from two consecutive acquisition nodes. The slope of MRD decline was calculated, for example, the slope from d15 to d33 and from d33 to d45. Gene-MRD association features were generated by combining concurrent gene expression trends (e.g., the magnitude of increase or decrease in IKZF1 expression). The time-series association features were updated weekly, triggered when the amount of newly acquired data reached 5% of the number of cases in that subtype. MRD values ​​were detected using flow cytometry or NGS, with a sensitivity of at least 10. -4 Gene expression data were obtained by detecting the expression level of target genes using real-time quantitative PCR; clinical response indicators included white blood cell count and platelet recovery time.

[0077] Output subtype time-series associated feature set, including fields such as MRD decline slope, gene expression change rate, and clinical response time-series curve.

[0078] The feature screening module takes the subtype basic association features from the feature library construction module and the subtype temporal association feature set from the feature mining module as input to construct a candidate feature pool, which includes 15 clinical indicators, 20 molecular markers, and 15 temporal features, for a total of 50 association features. It screens causal features that affect prognosis and eliminates false associations, such as high white blood cell count, which is often co-occurring with TP53 mutation but has no direct causal relationship, thereby improving the prediction accuracy of the scoring model.

[0079] A two-step approach of univariate causal validation and multivariate correlation filtering was adopted. In the first step, univariate causal validation used Kaplan-Meier survival analysis to calculate the log-rank P-value of each feature with event-free survival (EFS). P < 0.05 was considered to be directly associated with the prognosis. For example, the log-rank P < 0.001 for TP53 mutation was considered to be a directly associated feature.

[0080] The second step, multivariate correlation filtering, uses the Pearson correlation coefficient to calculate the correlation coefficient between any two directly related features. A correlation coefficient > 0.8 indicates high co-occurrence, and features with weak clinical significance are removed. For example, high white blood cell count, after univariate causal verification (Kaplan-Meier survival analysis), has a log-rank P ≥ 0.05, indicating no direct association with prognosis and is directly removed. TP53 mutation, with a log-rank P < 0.001, enters multivariate correlation filtering, and since its correlation coefficient with other features is all < 0.8, it is retained as a core causal feature.

[0081] The above process retains 8 to 10 core causal association features. The retention criterion is a causal effect value > 0.3. The causal effect value is calculated using a Cox proportional hazards regression model and is consistent on three independent central datasets. For example, “TP53 mutation”, “MRD clearance delay of more than 45 days”, and “IKZF1 deletion” all meet the validation requirements.

[0082] Output a list of core causal association features of the subtype and the causal effect determination results. The list includes feature name, causal effect value, number of validation centers, and reason for exclusion.

[0083] The adjustment module is based on the subtype core causal association feature list of the feature screening module, combined with the NCCNALL guidelines (2024 edition) and the Chinese ALL diagnosis and treatment guidelines (2024 edition), to build a feature, score and plan matching rule base, establish a real-time linkage mechanism of associated features → score correction → plan adjustment, dynamically optimize the treatment plan, and avoid overtreatment or undertreatment caused by fixed plans.

[0084] It has a built-in matching rule library for more than 20 ALL treatment scenarios. The rule formulation needs to combine the weight of core features and the recommended clinical guidelines. For example, "when the low-risk modified score is ≤80 points (modified score = base score + time feature adjustment score, the base score is the sum of the weights of each core feature) and the slope of MRD decrease is ≥0.8, the corresponding VDP standard chemotherapy regimen (vincristine + daunorubicin + prednisone)"; "when the intermediate-risk score is 101-120 points and there is IKZF1 deletion, the corresponding VDP combined with rituximab regimen"; "when the very high-risk score is ≥141 points and MRD is persistently positive, the corresponding CD19CAR-T therapy followed by allogeneic hematopoietic stem cell transplantation regimen";

[0085] The weight ranges for the core features of each subtype are as follows:

[0086] In the pediatric B-ALL subtype, ETV6-RUNX1 fusion (weight 15-20), MRD clearance time ≤33 days (weight -10 to -5), and IKZF1 deletion (weight 10-15) were observed.

[0087] In the adult T-ALL subtype, PI3K mutation (weight 12-18), CD4 / CD8 double positivity (weight 8-12), and MRD decrease slope ≥0.8 (weight -8 to -3) are significant factors.

[0088] In mixed subtypes, TP53 mutation (weight 20-25), anthracycline tolerance (weight -5 to -2), and white blood cell count >50×10⁶ are also associated with these subtypes. 9 / L (weight 10-15);

[0089] The scoring stratification threshold standards are as follows: low risk corrected score ≤ 80 points, medium risk corrected score 81-120 points, high risk corrected score 121-140 points, and very high risk corrected score ≥ 141 points.

[0090] The calculation rules for the time-series characteristic adjustment score are as follows: for every 0.1 increase in the slope of MRD decrease, the adjustment score is -2; for every 0.2 decrease in gene expression change rate (after treatment / before treatment), the adjustment score is +3; for each clinical response indicator that meets the target (white blood cell decrease rate ≥50%, platelet recovery time ≤7 days), the adjustment score is -1; for each indicator that does not meet the target, the adjustment score is +2.

[0091] The system employs rule-based matching and clinical threshold judgment to calculate a corrected score. The base score equals the sum of the weights of each core feature, and the corrected score equals the base score plus the time-series feature adjustment score. The adjustment score is determined based on MRD changes and gene expression trends, ranging from -20 to +20. The corrected score is updated weekly based on real-time time-series features from the feature mining module. The system integrates a real-time electronic medical record data interface that supports the HL7FHIR standard to obtain real-time patient blood routine, liver and kidney function, treatment and medication records, and adverse reaction occurrences. The response time from feature update to protocol output is less than 1 hour. Simultaneously, physician intervention permissions are set. When protocol adjustments involve changes in treatment intensity (such as switching from standard chemotherapy to targeted therapy), the attending physician must review and confirm before execution.

[0092] Outputs real-time corrected patient scores and appropriate treatment plans, including score composition, plan name, medication dosage, dosing cycle, and monitoring indicators.

[0093] The specific technical parameters of the electronic medical record interface are as follows:

[0094] ① The data transmission protocol uses HTTPS, port number 443;

[0095] ② The data format is JSON, the data size of a single request is ≤10MB, and the response timeout is ≤30 seconds;

[0096] ③ The data update frequency is as follows: clinical indicators are automatically synchronized once every 6 hours, treatment medication records are synchronized in real time, and adverse reaction data are synchronized within ≤10 minutes after manual entry;

[0097] Physician intervention authority grading standards:

[0098] ① Attending physician: Can approve adjustments to standard chemotherapy regimen dosage and increases in the frequency of monitoring indicators;

[0099] ② Associate chief physicians and above: can approve changes to treatment plan types, adjustments to extremely high-risk plans, and temporary corrections to scoring thresholds; changes in authority must be submitted through the hospital information system and will take effect after being reviewed by the department head, and operation logs must be retained for ≥5 years.

[0100] The early warning and intervention module uses the patient's real-time correction score and appropriate treatment plan as the core input, and combines clinical signs data and multi-omics data to construct an adverse reaction early warning model, identify treatment toxicity risks in advance, and reduce the incidence of grade III-IV adverse reactions. Clinical signs data include blood routine, liver and kidney function, troponin I, and electrocardiogram indicators. Multi-omics data include metabolomics such as the concentration of mercaptopurine metabolites and proteomics such as the level of inflammatory factors.

[0101] Multimodal fusion was performed using a weighted allocation method. First, the data of each dimension were standardized, such as converting neutrophil counts into Z-scores. Then, the comprehensive risk value was calculated according to the weights. The weight of high-risk features was set at 40%, the weight of clinical indicators was set at 35%, and the weight of multi-omics data was set at 25%. An alert was triggered when the comprehensive risk value was >0.7.

[0102] The intervention protocol library is stored according to the type of adverse reaction, risk level and intervention measures. The protocol library is updated quarterly with the latest clinical research evidence. The early warning results are output in the form of a pop-up window in the EMR system and pushed to the attending physician's mobile terminal at the same time, showing the risk level, correlation characteristics, recommended intervention measures and evidence level.

[0103] Adverse reactions include bone marrow suppression, cardiotoxicity, and liver damage; risk levels are categorized as low, medium, and high; interventions include drug replacement, dosage adjustment, and symptomatic treatment.

[0104] The adverse reaction warning results and intervention plans are fed back to the adjustment module to optimize subsequent treatment plans, and serve as local clinical feedback data for the training module to conduct federated training.

[0105] The standard for early warning response time is:

[0106] ① The time from the completion of the comprehensive risk value calculation to the display of the EMR pop-up window is ≤1 minute;

[0107] ②Push notifications to doctors via mobile devices take ≤3 minutes to process;

[0108] ③ If the doctor confirms the intervention plan within 2 hours, it will be automatically escalated to the department head;

[0109] Follow-up frequency settings after early warning:

[0110] ① High-risk warning (overall risk value > 0.85): Monitor related indicators once every 4 hours;

[0111] ② Medium-risk warning (0.7 < comprehensive risk value ≤ 0.85): Monitor once every 12 hours;

[0112] ③ Low-risk warning (0.6≤Comprehensive risk value≤0.7): Monitor once a day; follow-up data needs to be transmitted back to the warning model in real time for dynamic correction of the risk value until the risk value ≤0.6 and the warning is lifted.

[0113] Examples of specific intervention plans for adverse reactions are as follows:

[0114] Bone marrow suppression (neutrophil count <0.5×10⁻⁶) 9 / L): Chemotherapy drugs are suspended, and granulocyte colony-stimulating factor (G-CSF) is administered at a dose of 5 μg / kg, subcutaneously once daily, until the neutrophil count is ≥1.5 × 10⁻⁶. 9 / L;

[0115] Cardiotoxicity (troponin I > 0.1 ng / mL): Discontinue anthracyclines and replace with vincristine (1.4 mg / m², once a week, maximum dose 2 mg), while administering dexrazoxane at a dose 10 times the cumulative dose of anthracyclines, via intravenous infusion.

[0116] Liver injury (ALT > 100 U / L): Reduce the methotrexate dose to 70% of the original dose, add hepatoprotective drugs (diammonium glycyrrhizinate, dose 150 mg, orally 3 times a day), monitor liver function every 3 days until ALT ≤ 50 U / L.

[0117] All intervention protocols must be labeled with their level of evidence; for example, the level of evidence for G-CSF in bone marrow suppression is IA.

[0118] The training module divides federated nodes based on the ALL subtype database of the feature library construction module. The node types are divided into three categories: children's B-ALL center, adults' T-ALL center, and mixed subtype center, with no less than 3 nodes in each category. At the same time, the local adverse reaction warning and intervention data of the warning and intervention module are input, including the warning accuracy rate and the effectiveness of intervention measures, so as to realize the collaborative training of subtype-specific sub-models and improve the generalization of the global model.

[0119] Each node trains a sub-model of the corresponding subtype's associated features based on local subtype data and feedback data from the early warning and intervention module using a logistic regression algorithm. The training parameters are set to 100 iterations, a learning rate of 0.01, and a regularization coefficient of 0.001. Five-fold cross-validation is used to evaluate model performance, with an AUC ≥ 0.85 considered acceptable. Before aggregation, the feature consistency of each sub-model is verified (core feature overlap rate > 90%). During aggregation, weights are assigned according to the proportion of data volume of each node. The proportion of data volume = number of cases in that node / total number of cases in all nodes. For example, if the number of cases in three pediatric B-ALL nodes are 1200, 1000, and 900 respectively, and the total number of cases is 3100, the weights are 0.39, 0.32, and 0.29 respectively. The global model parameters are calculated by weighted average.

[0120] The statistical period for the number of cases at a node is the effective cases diagnosed in the past 5 years and followed up for at least 1 year, excluding cases with lost follow-up (follow-up rate <80%) and data tampering (inconsistent after hash verification). If the number of cases at a certain type of node is less than 300, a node merging strategy is adopted: two adjacent centers in the same region are merged into one joint node. The number of cases after merging must be ≥500. The weight of the joint node is calculated as the sum of the weights of the two nodes before merging. If the number of cases after merging is still less than 300, the training of this type of node is suspended and restarted after the number of new cases reaches the target. During the suspension period, the historical best sub-model is used as a temporary replacement.

[0121] Output the subtype sub-models and feature weight update results of global collaborative training, including sub-model parameters, feature weight adjustment values, and model validation metrics (AUC, sensitivity, specificity).

[0122] The local data screening criteria for federal training are as follows: Included cases must meet the following criteria:

[0123] ① Confirmed as the corresponding ALL subtype (confirmed by morphological, immunological, cytogenetic, and molecular biological typing);

[0124] ② The treatment period is ≥3 months, and the data for key milestones (induction period d15, d33, d45) are complete;

[0125] ③ No serious underlying diseases (such as severe heart failure, liver failure); exclude cases with more than 2 missing core features (such as MRD value, high-sensitivity gene mutation status);

[0126] The acceptable threshold standards for model validation metrics are: AUC ≥ 0.85, sensitivity ≥ 0.80, specificity ≥ 0.75, and accuracy ≥ 0.78. If a single-node sub-model fails to meet the standards, local data needs to be re-screened (missing data needs to be collected or outliers need to be removed) and training needs to be repeated until the thresholds are met. After global model aggregation, additional independent multi-center test sets (≥ 500 cases, covering all subtypes) are required for validation to ensure that the difference between the test set metrics and the training set metrics is < 5%.

[0127] The desensitization and iteration module takes the global subtype submodel and feature weights of the training module as its core, and synchronously inputs the highly sensitive gene features involved in the feature library construction module and feature screening module, as well as the clinical feedback data from the early warning and intervention module, to achieve privacy protection and cross-center iterative optimization of sensitive gene features, and ensure the security of multi-center data collaboration and the continuous effectiveness of the model.

[0128] Based on the HIPAA privacy protection list and clinical value in the ALL field, TP53, IKZF1, BCR-ABL1, and NOTCH1 were identified as highly sensitive genes. Clinical feedback data included early warning accuracy, intervention effectiveness, and patient prognosis follow-up results.

[0129] Feature transformation and fixed noise addition were used to desensitize highly sensitive gene features. First, the mutant / wildtype gene was converted into a mutation risk index, with mutants assigned a value of 8-10 and wildtypes assigned a value of 0-2. The index calculation rules were stored locally by each center and not transmitted across centers. Then, fixed noise was added during feature aggregation, with a noise intensity of 0.3 for sensitive features and 0.1 for non-sensitive features. The noise type was Gaussian noise. After desensitization, the consistency of prediction with the original features was verified by calculating indicators such as prediction accuracy and recall, ensuring that the difference in prediction accuracy between the two was <5%.

[0130] The anonymized features and model parameters are transmitted using SSL / TLS 1.3 encryption, and the node identity is verified through digital certificates during transmission. Cross-center iteration is triggered quarterly based on feedback data from the early warning and intervention module. The iteration process is as follows: the leading center initiates an iteration request → each node verifies the validity of the current features based on the latest local data → feedback of verification results → the leading center calculates the feature weight adjustment value (adjustment range ≤ 20% of the original weight) → synchronizes and updates to all nodes → verifies the global model performance.

[0131] The selection criteria for the lead center in cross-center iteration are: ① ≥1500 cases of this subtype in the past 3 years; ② AUC of model validation is consistently ≥0.88; ③ Possess HL7FHIR level 5 data exchange capability; the lead center rotates every 2 years, and before the rotation, it must pass the consensus vote of all nodes (agreement rate ≥70%).

[0132] The evaluation criteria for iteration effectiveness include: ① The improvement in AUC of the global model at each node is ≥0.02; ② After adjusting the weights of the core features, the difference in the prognostic accuracy of the three independent validation sets is ≤3%; ③ The consistency of prediction between the desensitized features and the original features (Kappa coefficient) is ≥0.85. If the above criteria are not met, the iteration needs to be restarted, with a maximum of 3 iterations. If the criteria are still not met, the feature library reconstruction process will be initiated.

[0133] The de-identified cross-center shared feature set and the updated global model are output. On the one hand, they are fed back to the feature library construction module to optimize the weight values ​​of the subtype feature library. On the other hand, they are synchronized to the adjustment module to update the matching rule library of features, scores and schemes.

[0134] The specific calculation rules for the mutation risk index are determined based on the prognostic impact of the gene in the ALL subtype. The index assignment standards for each highly sensitive gene are as follows: TP53 mutant is assigned 10, wild type is assigned 0; IKZF1 deletion is assigned 8, wild type is assigned 2; BCR-ABL1 fusion is assigned 9, wild type is assigned 1; NOTCH1 mutant is assigned 7, wild type is assigned 3.

[0135] After the index is calculated, it needs to be used through the formula. Normalize, The original index, This is the minimum value of the gene index. The maximum value of the gene index is given by the normalized index range of [0, 1].

[0136] The fixed noise is Gaussian noise, and the noise parameters are set as follows: noise mean μ=0 and variance σ for high-sensitivity gene characteristics. 2 =0.09, corresponding to a noise intensity of 0.3; the noise mean μ=0 and variance σ of non-sensitive gene features. 2 =0.01, corresponding to a noise intensity of 0.1;

[0137] The formula for adding noise is: , These are the normalized eigenvalues. It is a Gaussian random variable that meets the above parameters.

[0138] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An acute lymphoblastic leukemia scoring system, characterized in that, It includes a feature library construction module, a feature mining module, a feature selection module, an adjustment module, an early warning and intervention module, a training module, and a desensitization and iteration module; Feature library construction module: Establish a classification association feature system based on the molecular biological differences of ALL subtypes, clarify the core association features and weight ratios of each subtype, and output the association feature library and feature weight table; Feature mining module: Performs time-series mining of the entire treatment cycle based on the associated feature library, collects data by treatment stage, and generates time-series associated features; Feature selection module: Establishes a candidate feature pool, filters direct causal features, eliminates spurious associations, calculates effect values, and outputs a list of core causal features and the judgment results; Adjustment module: Establish a rule base for features, scores and treatment plans, and establish a linkage mechanism to calculate and correct scores. Integrate EMR real-time interface to obtain data and output real-time corrected scores and appropriate treatment plans for patients. Early warning and intervention module: integrates multi-source data to calculate a comprehensive risk value, triggers early warnings and outputs intervention plans, and feeds back data to the treatment adjustment and federated training modules; Training module: Divide federated nodes according to subtype, train subtype sub-models, use equal weight aggregation to generate a global model, and output the global sub-model and feature weight update results; Desensitization and Iteration Module: Highly sensitive genes are desensitized using feature transformation and fixed noise. Data is transmitted with encryption. Based on clinical feedback, iterative optimization is performed to output a desensitized shared feature set and an updated global model to optimize the subtype feature library and matching rule library.

2. The acute lymphoblastic leukemia scoring system according to claim 1, characterized in that, The feature library construction module constructs a classification association feature system based on the molecular biological differences of ALL subtypes, and clarifies the core association features and weight ratios of each subtype; it connects to a multi-center ALL subtype database, and the data dimensions include clinical indicators, molecular markers, and immunophenotypes; it adopts LASSO regression combined with subtype stratified analysis, firstly using LASSO regression to screen features that are significantly associated with prognosis within a single subtype, and then using subtype stratified analysis to compare the differences in features between different subtypes, excluding redundant features across subtypes, and determining the weight of each feature through 5-fold cross-validation; Output a subtype-specific associated feature library and feature weight table.

3. The acute lymphoblastic leukemia scoring system according to claim 2, characterized in that, The feature mining module is based on a subtype-specific association feature library and feature weight table. It performs time-series dynamic mining of the core association features of each subtype throughout the entire treatment cycle, captures the dynamic association changes of gene expression data, clinical response indicators and MRD, and generates time-series association features. Data collection nodes are divided according to the treatment stage, and MRD values, gene expression data, and clinical response indicators are collected simultaneously. Sliding window time series analysis and linear fitting were used to calculate the slope of MRD decline, and the association features between genes and MRD were generated by combining the gene expression trends during the same period. Periodically update the temporal correlation features and output the subtype temporal correlation feature set.

4. The acute lymphoblastic leukemia scoring system according to claim 3, characterized in that, The feature selection module takes the subtype basic association features and the subtype temporal association feature set as input to construct a candidate feature pool, select causal features that affect the prognosis, and eliminate false associations; it adopts a two-step method of univariate causal verification and multivariate correlation filtering to select features, retain the core causal association features that are consistent after multi-center verification, and outputs a list of core causal association features of the subtype and the causal effect determination result.

5. The acute lymphoblastic leukemia scoring system according to claim 4, characterized in that, The adjustment module is based on the subtype core causal relationship feature list, constructs a feature, score and solution matching rule base, and establishes a real-time linkage mechanism of relationship feature → score correction → solution adjustment; The system calculates and corrects scores using rule-based matching and clinical threshold judgment, integrates real-time data interfaces of electronic medical records to obtain relevant patient data, sets doctor intervention permissions, and outputs real-time corrected scores and appropriate treatment plans for patients.

6. The acute lymphoblastic leukemia scoring system according to claim 5, characterized in that, The early warning and intervention module takes the patient's real-time corrected score and appropriate treatment plan as the core input, and combines clinical sign data and multi-omics data to construct an adverse reaction early warning model; it uses a weighted allocation method to perform multimodal fusion, and then calculates the comprehensive risk value according to the weights and triggers the early warning. The intervention plan library is stored by category and updated regularly; the early warning results are simultaneously output to the EMR system and the mobile terminal of the attending physician, and the adverse reaction early warning results and intervention plans are output to the adjustment module and training module.

7. The acute lymphoblastic leukemia scoring system according to claim 6, characterized in that, The training module divides federated nodes based on the ALL subtype database and inputs local adverse reaction warning and intervention data to achieve collaborative training of subtype-specific sub-models; Each node trains a corresponding sub-model of associated features based on local subtype data and feedback data using logistic regression algorithm, and evaluates model performance using 5-fold cross-validation. The sub-models of each node are integrated by using an aggregation method that allocates weights according to the proportion of data volume of each node, and the global model parameters are calculated by weighted averaging. The results of global collaborative training of subtype sub-models and feature weight updates are then output.

8. The acute lymphoblastic leukemia scoring system according to claim 7, characterized in that, The desensitization and iteration module takes the global subtype model and feature weights as its core, and simultaneously inputs highly sensitive gene features and early warning and clinical feedback data. Feature transformation and fixed noise addition were used to desensitize highly sensitive gene features; An encrypted protocol was used to transmit desensitized features and model parameters, and cross-center iterative optimization was triggered based on clinical feedback. Output the de-identified cross-center shared feature set and the updated global model to the feature library construction module and adjustment module.

9. The acute lymphoblastic leukemia scoring system according to claim 8, characterized in that, The feature weight table includes feature name, subtype affiliation, weight value, data dimension, and number of validation centers.

10. The acute lymphoblastic leukemia scoring system according to claim 9, characterized in that, The subtype time-related feature set includes the MRD decrease slope, gene expression change rate, and clinical response time curve.

Images