Construction method of recommendation model for individually adding IS during low-dose immune tolerance induction treatment, equipment, medium and program product

Abstract

The invention provides a construction method of a recommendation model for individually adding an IS during low-dose immune tolerance induction treatment, equipment, a medium and a program product, and relates to the field of intelligent medical treatment. The method comprises the following steps: acquiring clinical data X, a treatment scheme T and a treatment result Y of a training set sample; training a filling model capable of predicting treatment results of different samples by using the clinical data and the treatment results; training a weighting model capable of predicting the treatment receiving possibility of different samples by using the clinical data and the treatment scheme; calculating a DR estimator based on the padding model and the weighting model; calculating a causal effect value according to the DR estimator; and training a regression model based on the clinical data and the causal effect value to obtain a recommendation model. According to the method, the recommendation model capable of helping the clinician to determine whether to add the IS individually during the ITI treatment period is established by utilizing causal reasoning and machine learning, and the judgment of the model and the judgment of the clinician have relatively satisfactory consistency.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medicine, and more particularly, to a method, device, medium and program product for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy. Background Art

[0002] Hemophilia A (HA) is an X-linked coagulation disorder disease, usually treated by factor (F) VIII replacement therapy. As a major complication, 25 - 30% of severe HA (SHA) patients develop inhibitory FVIII-specific antibodies (FVIII inhibitors). Inhibitors render factor replacement therapy ineffective. Currently, immune tolerance induction (ITI) is the only way to eradicate inhibitors. Most HA patients with inhibitors will develop "immune tolerance" to FVIII after ITI, with a success rate between 60% and 90%. However, some patients have difficulty tolerating and / or are non-responsive to ITI.

[0003] As an alternative second-line treatment method, it is recommended to add further interventions during ITI, such as immunosuppressive agents (IS). Rituximab is an anti-CD20 antibody that is hypothesized to promote the induction of immune tolerance in drug-resistant inhibitor cases by rapidly depleting B lymphocytes. However, the appropriate criteria for the combined use of IS during ITI for "poor ITI responders" remain unclear.

[0004] Considering economic conditions, high-dose ITI regimens have not been widely applied. Therefore, we previously added IS to "poor ITI responders" in LD-ITI (becoming LD-ITI+IS), with a success rate as high as 62.5% and no complications. However, the LDITI+IS regimen does not guarantee success for all "poor ITI responders". On the contrary, some patients may achieve success with only the LDITI regimen without IS. Therefore, the management experience of adding IS during ITI is challenging and limited. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the prior art. To this end, the present invention provides a method, device, medium and program product for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy; the method of the present invention establishes a recommendation model that can help clinicians decide whether to add IS personalized during ITI treatment by using causal inference and machine learning, and the judgment of this model has relatively satisfactory consistency with the judgment of clinicians.

[0006] The first aspect of the present application discloses a method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy, the method comprising:

[0007] S101, Obtain the clinical data X, treatment plan T, and treatment outcome Y of the training set samples;

[0008] S102, Use the clinical data and treatment outcomes to train a filling model that can predict the treatment outcomes of different samples;

[0009] S103, Use the clinical data and treatment plan to train a weighted model that can predict the likelihood of different samples receiving treatment;

[0010] S104, Calculate the DR estimator based on the filling model and the weighted model; Calculate the causal effect value according to the DR estimator;

[0011] S105, Train a regression model based on the clinical data X and the causal effect value to obtain a recommendation model.

[0012] In some embodiments, the treatment plan includes whether IS is added during the LD-ITI treatment process, T = 1 indicates receiving treatment, and T = 0 indicates not receiving treatment;

[0013] Optionally, the treatment outcome includes whether the LD-ITI treatment is successful, Y = 1 indicates a successful outcome, and Y = 0 indicates an unsuccessful outcome;

[0014] Optionally, the clinical data includes: the inhibitor titer immediately before ITI, the historical peak inhibitor titer, the age at the start of ITI, and the interval time.

[0015] In some embodiments, the expression of the filling model is f(X,T);

[0016] Optionally, the expression of the weighted model is g(X)=P(T = 1|X); where T = 1 indicates receiving treatment;

[0017] Optionally, the calculation formula of the DR estimator is: ; where 1(T = t) is an indicator function, which is equal to 1 when T is equal to t, and 0 otherwise; T is a random variable representing "intervention", and t is the specific value of this random variable.

[0018] In some embodiments, the calculation formula of the causal effect value is: .

[0019] In some embodiments, the regression model includes any one of the following: linear regression, vector regression, random forest regression, gradient boosting regression tree.

[0020] The second aspect of the present application discloses a recommendation method for personalized addition of IS during low-dose immune tolerance induction treatment, and the method includes:

[0021] S201, Obtain the clinical data of the subject;

[0022] S202, Input the clinical data into the recommendation model constructed by the method described in the first aspect of the present application to calculate the causal effect value;

[0023] S203, Recommend a treatment plan for the subject according to the causal effect value; when the causal effect value is greater than 0, output an auxiliary prediction result that the subject accepts treatment; when the causal effect value is less than 0, output an auxiliary prediction result that the subject does not accept treatment.

[0024] In some embodiments, the clinical data includes any one or several of the following: the inhibitor titer immediately before ITI, the historical peak inhibitor titer, the age at the start of ITI, and the interval time.

[0025] The third aspect of the present application discloses a computer device, which includes: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the above method.

[0026] The fourth aspect of the present application discloses a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0027] The fifth aspect of the present application discloses a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0028] The present application has the following beneficial effects:

[0029] The present application innovatively discloses a method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy. This method uses causal inference methods to estimate the target causal relationship (to solve confounding bias and achieve unbiased estimation of causal relationships), as Figure 6 shown; machine learning methods are used to simulate the correlation between variables. Specifically, a machine learning model is used to capture the correlation between treatment and outcome, which includes causal relationships and spurious correlations caused by clinical characteristics; subsequently, causal inference methods are used to accurately identify causal relationships from the obtained overall statistical associations; finally, a regression model is used to learn the mapping relationship from clinical characteristics to causal relationships, so as to be able to predict the causal relationship and its corresponding optimal treatment strategy for any newly admitted patient. Different from traditional machine learning algorithms that focus on accurately predicting outcomes using many input features, this solution aims to recommend the most favorable treatment plan for each patient. Description of the Drawings

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those skilled in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0031] Figure 1 It is a schematic flowchart of the method provided in the first aspect of the embodiment of the present invention;

[0032] Figure 2 It is a schematic flowchart of the method provided in the second aspect of the embodiment of the present invention;

[0033] Figure 3 It is a schematic diagram of the computer device provided in the embodiment of the present invention;

[0034] Figure 4 It is a schematic diagram of the architecture of the exemplary computing device provided in the embodiment of the present invention;

[0035] Figure 5 It is a schematic diagram of the storage medium provided in the embodiment of the present invention;

[0036] Figure 6 It is a schematic diagram of the process of training the recommendation model provided in the embodiment of the present invention and its personalized results predicted under different IS treatment plans;

[0037] Figure 7 It is a schematic diagram of the relationship among the clinical data X, treatment plan T, and treatment result Y provided in the embodiment of the present invention;

[0038] Figure 8 It is a research flowchart provided in the embodiment of the present invention. Detailed implementation manners

[0039] To enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention.

[0040] In some of the processes described in the specification, claims, and above-mentioned drawings of the present invention, a plurality of operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent a sequence, and do not limit that "first" and "second" are of different types.

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0042] Ideally, we aim to compare the potential outcomes of all possible treatments (such as with or without the addition of IS) using observational (historical) data. If the two potential outcomes of all patients are obtained, the difference between the different potential outcomes, that is, the causal effect, can be calculated. By checking whether the causal effect of a specific patient is greater than 0, an individualized optimal treatment plan can be obtained. However, we can observe the outcomes with or without the addition of IS, but not both simultaneously. This is because after making a treatment choice, we cannot turn back time to revoke the treatment. In other words, the goal becomes estimating the causal effect (or the missing potential outcome) based on the observed potential outcomes. Unfortunately, the treatment choices and outcomes of patients usually depend on their clinical characteristics, resulting in the treatment effect being confounded by the clinical characteristics and posing a challenge to accurately estimating the causal effect. Traditional observational research methods (such as correlation analysis) are difficult to mitigate this confounding and cannot achieve an unbiased estimate of the causal effect.

[0043] Figure 1 It is a schematic flowchart of a method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction treatment provided by an embodiment of the present invention. Specifically, the method includes the following steps:

[0044] S101: Obtain the clinical data X, treatment plan T, and treatment outcome Y of the training set samples; the training set samples include a treatment group and a control group;

[0045] In some embodiments, the treatment plan includes whether to add IS during the LD-ITI treatment process, T = 1 indicates receiving the treatment, and T = 0 indicates not receiving the treatment;

[0046] In some embodiments, the treatment outcome includes whether the LD-ITI treatment is successful. Y = 1 indicates a successful outcome, and Y = 0 indicates an unsuccessful outcome.

[0047] In some embodiments, the clinical data includes: the immediate inhibitor titer before ITI (referring to the inhibitor level measured immediately before the start of immune tolerance induction therapy (ITI)), the historical peak inhibitor titer (or the previous highest inhibitor titer, referring to the highest inhibitor concentration ever reached in the patient's medical history), the age at the start of ITI, and the interval time (from inhibitor diagnosis to ITI initiation).

[0048] S102. Use the clinical data and treatment outcomes to train an imputation model that can predict the treatment outcomes of different samples.

[0049] In some embodiments, the expression of the imputation model is f(X, T).

[0050] S103. Use the clinical data and treatment regimens to train a weighted model that can predict the likelihood of different samples receiving treatment.

[0051] In some embodiments, the expression of the weighted model is g(X) = P(T = 1|X); where T = 1 indicates receiving treatment.

[0052] S104. Calculate the DR estimator based on the imputation model and the weighted model; calculate the causal effect value according to the DR estimator.

[0053] In some embodiments, the calculation formula of the DR estimator is: ; where 1(T = t) is an indicator function that equals 1 when T equals t and 0 otherwise; T is a random variable representing "intervention", and t is the specific value of this random variable.

[0054] In some embodiments, the calculation formula of the causal effect value is: .

[0055] S105. Train a regression model based on the clinical data X and the causal effect value to obtain a recommendation model.

[0056] In some embodiments, the regression model includes any one of the following: linear regression, support vector regression (SVR), random forest regression, gradient boosting regression tree (GBRT).

[0057] The second aspect of the present application discloses a recommendation method for personalized addition of IS during low-dose immune tolerance induction therapy, the method including:

[0058] S201. Obtain the clinical data of the subject.

[0059] In some embodiments, the clinical data includes any one or more of the following: inhibitor titer immediately before ITI, historical peak inhibitor titer, age at ITI start, and interval time.

[0060] In some embodiments, the term "subject" or "test subject" or "test sample" used herein refers to any animal (e.g., mammal), including but not limited to humans, non-human primates, rodents, etc., that will be the recipient of a specific treatment. Generally, the terms "subject" and "patient" are used interchangeably herein when referring to human subjects. Preferably, the subject is a human.

[0061] In some embodiments, the test sample is a patient clinically used for prognostic evaluation.

[0062] S202. Input the clinical data into the recommendation model constructed by the method described in the first aspect of the present application to calculate the causal effect value.

[0063] S203. Recommend a treatment plan for the subject according to the causal effect value; when the causal effect value is greater than 0, output an auxiliary prediction result that the subject accepts the treatment; when the causal effect value is less than 0, output an auxiliary prediction result that the subject does not accept the treatment.

[0064] In some embodiments, the auxiliary prediction result includes but is not limited to the form of a paper or electronic report. This result is only obtained by intelligent machines based on the relevant data of the subject and is only for reference by medical staff and not the final diagnosis result of the subject.

[0065] In some embodiments, the threshold 0 is obtained by training the training set samples, which can be a specific threshold or an interval range, and the specific form is not specifically limited in this embodiment.

[0066] Figure 3 is a schematic diagram of a computer device provided by an embodiment of the present invention, as Figure 3 shown, the device 2000 may include: one or more processors 2010 and one or more memories 2020; wherein, computer-readable code is stored in the memory, and when the computer-readable code is run by the one or more processors, the above-described method can be executed.

[0067] The processor in this embodiment may be an integrated circuit chip with signal processing capabilities. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, operations, and logic block diagrams disclosed in the embodiments of the present disclosure. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc., and may be of the X86 architecture or the ARM architecture.

[0068] Generally speaking, the various exemplary embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that can be executed by a controller, a microprocessor, or other computing devices. When aspects of the embodiments of the present disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented as non-limiting examples in hardware, software, firmware, dedicated circuits or logic, general hardware or a controller or other computing devices, or some combination thereof.

[0069] For example, the method or device according to the embodiments of the present disclosure may also be implemented by means of Figure 4 the architecture of the computing device 3000 shown. As Figure 4 shown, the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input / output component 3060, a hard disk 3070, etc. The storage device in the computing device 3000, such as the ROM 3030 or the hard disk 3070, may store various data or files used for the processing and / or communication of the methods provided by the present disclosure and the program instructions executed by the CPU. The computing device 3000 may also include a user interface 3080. Of course, Figure 4 the architecture shown is only exemplary, and when implementing different devices, one or more components shown in the Figure 4 computing device may be omitted according to actual needs.

[0070] The embodiments of the present invention also provide a computer-readable storage medium, such as Figure 5As shown, it is a schematic diagram of a storage medium 4000 provided by an embodiment of the present invention. Computer-readable instructions 4010 are stored on the computer storage medium 4020. When the computer-readable instructions 4010 are run by a processor, the method according to the embodiments of the present disclosure described with reference to the above figures can be executed. The computer-readable storage medium in the embodiments of the present disclosure may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM). It should be noted that the memory for the methods described herein is intended to include but not be limited to these and any other suitable types of memory. It should be noted that the memory for the methods described herein is intended to include but not be limited to these and any other suitable types of memory.

[0071] The embodiments of the present disclosure also provide a computer program product or system, including a computer program, which implements the steps of the above method when executed by a processor.

[0072] In some embodiments, this embodiment also discloses a construction system for a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy. The system includes:

[0073] A first data acquisition module, configured to obtain clinical data X, treatment plan T, and treatment outcome Y of training set samples.

[0074] A filling model training module, configured to train a filling model that can predict the treatment outcomes of different samples using clinical data and treatment outcomes.

[0075] A weighted model training module, configured to train a weighted model that can predict the likelihood of different samples receiving treatment using clinical data and treatment plans.

[0076] A causal effect value training module, configured to calculate a DR estimator based on the filling model and the weighted model; calculate a causal effect value according to the DR estimator.

[0077] A recommended model training module, which is used to or configured to train a regression model based on the clinical data X and the causal effect value to obtain a recommended model.

[0078] In some embodiments, the present embodiment also discloses a recommendation system for personalized addition of IS during low-dose immune tolerance induction therapy. The system includes:

[0079] A second data acquisition module, which is used to or configured to acquire the clinical data of the subject.

[0080] A causal effect value calculation module, which is used to or configured to input the clinical data into the recommendation model constructed by the method described in the first aspect of the present application to calculate the causal effect value.

[0081] A result prediction module, which is used to or configured to recommend a treatment plan for the subject according to the causal effect value; when the causal effect value is greater than 0, output an auxiliary prediction result that the subject accepts the treatment; when the causal effect value is less than 0, output an auxiliary prediction result that the subject does not accept the treatment. Specific embodiments

[0083] 1 Methods and materials

[0084] 1.1 Study design and population

[0085] This single-center retrospective cohort study was conducted at the BCH Hemophilia Comprehensive Care Center in China. We recruited children with SHA who received LD-ITI treatment with or without IS between September 2016 and December 2023, with a follow-up period of ≥24 months. We used four key pre-ITI clinical characteristics extracted from the patients' electronic medical records and trained and developed a recommendation model using causal inference to help clinicians decide whether to personalize the addition of IS during LD-ITI treatment.

[0086] Inclusion criteria included: (i) SHA (FVIII coagulation activity [FVIII:C] < 1% before inhibitor production); (ii) age ≤ 14 years at the start of ITI; (iii) high-titer inhibitor (≥5 BU / mL); (iv) receiving LD-ITI with or without IS, with a follow-up period of ≥24 months. Exclusion criteria included: (i) having congenital or acquired bleeding defects other than hemophilia A; (ii) having immune diseases.

[0087] 1.2 Data collection

[0088] Clinical, laboratory, and treatment data were collected from medical records and interviews. Baseline clinical and laboratory information was collected before ITI. During ITI, patients were seen every 1 to 2 weeks until the inhibitor titer trended downward after the initial peak caused by early repeated FVIII exposure, and then monthly until the end of the study. After successful ITI and during the prophylaxis phase, patients continued to be seen every 3 months.

[0089] 1.3 ITI Treatment

[0090] All patients received LD-ITI (FVIII 50 IU / kg, every other day) and plasma-derived factor VIII / von Willebrand factor concentrate (pd-FVIII / VWF). Once a patient achieved success (as defined below), the FVIII dose was slowly reduced to 25 - 30 IU / kg, three times a week, to continue prophylaxis. During ITI, immune suppressive agents (IS) (rituximab and prednisone) were added according to the following criteria (to become LDITI+IS): (1) if the inhibitor titer ≥ 40 BU / mL before or during ITI alone, (2) no downward trend (if the inhibitor decline during ITI was < 20% within the first 3 months of ITI).

[0091] IS included rituximab 375 mg / m2 / week (maximum dose 600 mg) for 4 weeks, combined with prednisone 2 mg / kg / day (maximum dose 60 mg) for 1 month, and then tapered over 6 weeks. After rituximab treatment for infection prophylaxis, intravenous immunoglobulin (IVIG, 200 mg / kg per month for 6 months) was given within 6 months to compensate for acquired IgG deficiency.

[0092] 1.4 Outcome Definitions

[0093] Success: The inhibitor titer was negative and the FVIII recovery rate ≥ 66% expected at the time of analysis. We were unable to conduct half-life studies on every patient who achieved an FVIII recovery rate ≥ 66%, so we did not include a half-life > 6 hours in the definition of success. Partial success: The inhibitor titer was negative, but the FVIII recovery rate < the expected 66%; Failure: The success criteria were not met. Relapse: After success, the inhibitor titer was > 0.6 BU / mL in two consecutive measurements, at least one month apart, regardless of whether the patient received subsequent FVIII prophylaxis.

[0094] 1.5 Laboratory Tests

[0095] During ITI, inhibitor titers were measured every two weeks using the Bethesda assay (Nijmegen modification) at BCH until a stable decline in the inhibitor titer was observed, and then monthly.

[0096] Once the inhibitor titer is negative for two consecutive times (with a one-month interval), in vivo FVIII recovery is performed. After the in vivo FVIII recovery rate ≥ 66%, the monitoring frequency is reduced to once every three months.

[0097] 1.6 Development process of causal inference

[0098] To assist clinical decision-making, we strive to establish a causal relationship model rather than a correlation model between treatment and outcome, so causal inference methods are adopted. Causal inference is used to evaluate the causal impact of treatment regimens on outcomes, where causal impact refers to the difference in outcomes under different interventions. In this article, the intervention is defined as whether to add IS during LD-ITI treatment, and the outcome is defined as whether the patient's ITI treatment is successful. The baseline clinical characteristics of the enrolled patients are collected, including the immediate inhibitor titer before ITI, the peak historical inhibitor titer, the age at the start of ITI, and the interval time (from inhibitor diagnosis to the start of ITI). Based on these characteristics, this article analyzes the impact of IS on the outcome and develops a model to personalized recommend a more favorable treatment plan for patients according to the causal effect estimation.

[0099] Estimating causal effects using observational data always faces two challenges. One challenge is the missing data problem mentioned above, that is, we can only observe the response of a patient to a specific treatment, but cannot observe both responses at the same time, which is also known as the fundamental problem of causal inference. Another challenge is the difference in characteristics between the treatment group and the control group in the observational data, resulting in differences in outcomes that are not entirely attributable to the treatment. For example, the condition of patients in the control group is often more optimistic than that of patients in the treatment group, and they are more likely to achieve successful outcomes without treatment. In this case, simply comparing the experimental results between the two groups may underestimate the causal effect of the treatment, resulting in bias.

[0100] To address the above two issues, we adopt the double-robust (DR) method to estimate the causal effect, which is one of the most effective causal inference methods and has achieved remarkable success in practical applications. Specifically, the DR estimator consists of two parts: an imputation model and a weighting model. The imputation model aims to fill in the missing outcomes of patients, while the weighting model balances the feature distributions of the treatment and control groups through sample weighting. Statistical theory shows that if the propensity scores or estimated weights are accurate, the DR estimator can serve as an unbiased estimator of the target causal effect. Based on the learned DR estimator, we can calculate the causal effect of the intervention with the addition of IS on the outcomes of all patients, which corresponds to the difference in potential outcomes with or without the application of IS. For the patients in the training set, we can directly compare whether their corresponding causal effect estimates exceed zero to determine which treatment is more effective. To accommodate any new patients, we adopt a regression model that estimates the causal effect using clinical features, enabling us to predict the outcome differences of any patient under different treatment regimens and thus recommend better treatment strategies.

[0101] Categorical variables are presented as frequencies and percentage values and are compared using the chi-square or Fisher's exact chi-square test. Continuous variables are presented as means and 95% confidence intervals (CIs) and are compared using Student's t-test (for normal distributions) or the Mann-Whitney U test (for non-normal distributions). Reported p-values are two-sided, and a p-value < 0.05 is considered statistically significant. All statistical analyses were performed using version 26.0 of SPSS (IBM Corp., Armonk, NY, USA).

[0102] 1.7 Double-robust design

[0103] First, we illustrate the relationship among the feature X, treatment T, and outcome Y in Figure 7 which explains the data generation process. As shown in Figure 7As shown, the result Y is jointly determined by the feature X and the treatment plan T. In addition, the treatment T received by patients is often affected by their respective characteristics. Without loss of generality, we assume that both the intervention variable T and the outcome variable Y are binary variables taking values 0 and 1. Among them, T = 1 indicates receiving treatment, otherwise 0; Y = 1 indicates a successful outcome, otherwise 0. As described in the previous section, the DR estimator is used to estimate the causal relationship. Without loss of generality, we represent the imputation model as f, whose inputs include the feature X and the treatment T. We aim to use the imputation model to predict the outcomes of different patients with different X and T, and the mathematical expression is f(X,T). By minimizing the difference between the output of the imputation model and the true outcome, the imputation model f(X,T) can be trained in a data-driven manner. For the weighted model represented as g(X), the input is the feature X, and the output is a probability value between 0 and 1, representing the likelihood of receiving treatment given the patient's characteristics, denoted as g(X)=P(T = 1|X);

[0104] Similarly, the weighted model g(x) is trained by fitting the treatment assignment in the training data. Based on the above two trained models, the estimated potential outcome with feature X and treatment t in the DR estimator is derived as ; where 1(T = t) is an indicator function that equals 1 when T equals t and 0 otherwise. Obviously, when a patient with feature Xi is observed to receive treatment t, the estimated potential outcome corresponding to treatment t is adjusted from the observed outcome through the imputation and weighted models. While the potential outcome corresponding to the unobserved treatment (1 - t) is directly generated by the imputation model.

[0105] Finally, the target causal effect CE(Xi) caused by the DR estimator can be expressed as the average difference between the estimated potential outcomes in the population with feature Xi. The calculation method is: ;

[0106] According to the definition of CE(Xi), CE(Xi) being greater than 0 means that patients with the corresponding feature Xi have a higher success probability when receiving treatment t = 1 than when receiving treatment t = 0. In this case, the best treatment for the patient is t = 1. In other words, the positive or negative value of CE(Xi) represents the advantages and disadvantages of different treatment methods. In addition, to predict the best treatment plan for newly admitted patients, we developed a regression model to establish the mapping relationship from patient characteristics to CE(Xi), which helps with the best treatment advice. Specifically, given the patient feature X as the input and the previously mentioned CE(X) as the output, we can use the training data to learn a regression model, such as linear regression. Therefore, for any given feature X, we can predict its corresponding causal effect (CE) to determine whether treatment should be performed.

[0107] 1.8 Experimental details and model specifications

[0108] The entire experimental dataset included 195 patients, and the four key clinical characteristics, designated IS treatments, and final outcomes of each patient were recorded in detail. Specifically, the four characteristics of the immediate inhibitor titer before ITI, historical peak inhibitor titer, age at the start of ITI, and interval time were regarded as the model input X, while the LD-ITI or LD-ITI+IS treatments and outcomes were regarded as T and Y, respectively. Five-fold cross-validation was used to verify the effectiveness of the proposed model. We established a decision tree as the basic model and used an ensemble learning method based on gradient boosting to implement the propensity model, filling model, and prediction model.

[0109] 2 Results

[0110] 2.1 Study population and baseline clinical characteristics

[0111] Among the 291 screened patients, 96 (33.0%) patients did not meet the inclusion criteria, as Figure 8 shown. A total of 195 SHA patients with high-titer inhibitors were included and received LD-ITI treatment, and the median follow-up period after the start of ITI was 4.4 years. Among the 195 patients, 145 (74.4%) patients achieved success, and 50 (25.6%) patients failed to eradicate the inhibitor. Among the 195 patients, 105 (53.8%) patients received LDITI, and 90 (46.2%) patients received LD-ITI+IS. A total of 13 (9.0%) patients relapsed at a median of 3.7 months after success. Among the 13 relapsed patients, 8 (61.5%) patients received LD-ITI+IS treatment.

[0112] In the results, in the median (range), the interval time in the successful group was shorter (2.7 months vs. 5.7 months, p = 0.030), and the age at the start of ITI was younger (3.1 years vs. 6.0 years, p < 0.000) compared with the unsuccessful group. In addition, compared with the unsuccessful group, the immediate inhibitor titer before ITI (13.5 BU / mL vs. 28.7 BU / mL, p = 0.015), historical peak inhibitor titer (21.3 BU / mL vs. 66.9 BU / mL, p = 0.043), and peak inhibitor titer during ITI (55.7 BU / mL vs. 146.4 BU / mL, p < 0.000) in the successful group were lower.

[0113] When focusing on the 145 patients who achieved success, we compared the baseline characteristics of patients with successful LD-ITI (n = 82) and patients with successful LD-ITI+IS (n = 63). Compared with patients with successful LD-ITI, patients with successful LD-ITI+IS had higher inhibitor titers immediately before ITI (42.2 BU / mL vs. 5.0 BU / mL, p<0.000), higher historical peak inhibitor titers (64.0 BU / mL vs. 10.4 BU / mL, p<0.000), and higher peak inhibitor titers during ITI (89.5 BU / mL vs. 39.2 BU / mL, p<0.000). No significant differences were found in the age characteristics (age at initial inhibitor, age at ITI start and interval) between the LD-ITI group and the LD-ITI+IS group.

[0114] Among the 50 patients in whom inhibitors could not be cleared, 23 patients (46%) received LD-ITI and 27 patients (54%) received LD-ITI+IS. Similarly, compared with patients who failed LD-ITI, patients who failed LD-ITI+IS had significantly higher inhibitor characteristics (immediate inhibitor titer before ITI, historical peak inhibitor titer, and peak inhibitor titer during ITI), and similar age characteristics.

[0115] 2.2 Recommendation model through causal inference

[0116] Since the experimental data was retrospective, all patients in the dataset received the treatment recommended by the clinician. When the treatment recommended by the model was different from that recommended by the clinician, the corresponding potential outcomes could not be observed, making it impossible to evaluate the accuracy of these recommendations. Based on this, to evaluate the accuracy of the model's recommendations, we considered the subset of patients for whom the model's recommendations were consistent with those of the doctor. Among 195 subjects, 119 patients (61.0%) received treatment recommendations that were consistent with those of the clinician. Among the 119 patients, 92 patients (77.3%) complied with the treatment plan recommended by the model and achieved success. In comparison, the clinician recommended treatment plans for 195 patients, and 145 patients achieved success, with an overall success rate of 74.5%. This indicates that our proposed recommendation model is comparable to clinicians in recommending treatment plans for patients.

[0117] 2.3 Rationality and potential mechanism of model recommendations

[0118] We conducted an in-depth analysis by comparing the characteristics of IS-treated patients recommended by the DR model with those recommended by clinicians. This comparison aimed to verify the rationality of the model's recommendations and explore its potential recommendation mechanism. Specifically, subgroup analyses were performed, focusing on four different subgroups. Among 195 patients, 90 (46.2%) received IS treatment in LD-ITI under the advice of clinicians (as Group A). 55.4% (108 / 195) of the patients were recommended by the model to add IS treatment in LD-ITI (as Group B). Among the 108 patients, 61 (56.5%) were recommended by both the model and clinicians (as Group B1), and 47 (43.5%) were recommended only by the model but not by clinicians (as Group B2).

[0119] No significant differences in inhibitor characteristics (median immediate titer before ITI: 45.0 BU / mL vs. 60.3 BU / mL, p = 0.869; historical peak titer: 69.7 BU / mL vs. 76.0 BU / mL, p = 0.356) and age characteristics (age at the start of ITI: 3.5 years vs. 3.1 years, p = 0.492; interval: 3.8 months vs. 4.5 months, p = 0.645) were found between Group A and Group B1. By combining Group A, which received treatment recommended by clinicians, with Group B2, which received treatment recommended by the model but not by clinicians, it was observed that Group B2 had significantly lower inhibitor characteristics (immediate titer before ITI: 15.7 BU / mL vs. 45.0 BU / mL, p = 0.001; historical peak titer: 10.7 BU / mL vs. 69.7 BU / mL, p = 0.001) and higher age characteristics (age at the start of ITI: 8.6 years vs. 3.5 years, p = 0.012; interval: 12.2 months vs. 3.8 months, p = 0.009).

[0120] Finally, ITI is the only clinically proven effective strategy for clearing inhibitors, especially for high-titer inhibitors. Although non-factor therapies (such as emicizumab) are increasingly used in patients who fail the first ITI treatment, in developing countries, emicizumab has not been included in the medical insurance. Reports show that only 2.8% of HA patients with inhibitors can use this expensive product for prophylaxis. For patients with "poor ITI response", appropriate use of IS during LD-ITI is an important attempt in developing countries with limited financial resources, as these countries cannot afford higher-dose ITI regimens or non-factor therapies. However, current guidelines on the use of IS are lacking and inaccurate. The criteria for adding IS during ITI treatment remain to be explored. To our knowledge, this is the first work to establish a recommendation model using causal inference methods, aiming to help clinicians make correct decisions on combining IS during LD-ITI treatment.

[0121] The model includes four key baseline clinical characteristics of patients, namely the immediate inhibitor titer before ITI, the historical peak inhibitor titer, the age at the start of ITI, and the interval. According to a large number of previous studies, these four clinical characteristics are important indicators of ITI outcomes. Multiple studies have shown that patients with a higher inhibitor profile (immediate inhibitor titer before ITI, historical peak inhibitor titer) are more likely to fail. When the model detects patients with a higher inhibitor profile, through a comprehensive algorithm, a more aggressive treatment approach will be recommended, such as combining with IS, to conform to current international treatment practices.

[0122] Notably, 61.0% of the patients received treatment recommendations that were consistent with those of the clinicians, indicating its potential reliability. Although this proportion seems low, the group with inconsistent treatment recommendations gives us the opportunity to identify predictive factors for adding IS treatment that may have been overlooked before. The high similarity of the characteristics between Group A and Group B1 indicates that the decisions of this model are not random, like flipping a coin, but extract useful clinical knowledge from the data, making its recommendations similar to those of the clinicians. This indicates the rationality of the model's decisions.

[0123] The comparison between Group A and Group B2 showed that when the inhibitor titer-related characteristics were relatively low, clinicians tended not to recommend IS treatment regardless of age-related characteristics such as the starting age or interval of ITI. On the contrary, when the patient's age was relatively older but the inhibitor titer was relatively low, the model still recommended adding IS. Based on the above information, it can be inferred that a relatively older starting age or longer interval of ITI may be new and previously unnoticed indicators for combining with IS treatment. Although previous guidelines generally did not recommend older age as a criterion for increasing treatment intensity, some studies have shown that a relatively older starting age and longer interval of ITI would lead to a decrease in the success rate of ITI. When the older age characteristic was detected, our model recommended receiving IS during LD-ITI. In the future, we plan to prospectively conduct a large-sample cohort study to determine the cut-off value of age characteristics.

[0124] In this study, the enrolled patients could be regarded as "low ITI risk" patients because the median pre-ITI titer immediately before was 16.8 BU / mL. A total of 74.4% of high-titer inhibitor SHA patients received LD-ITI+IS treatment under the advice of clinicians and achieved success. Assuming according to the treatment recommendations of the model, the success rate could be as high as 77.3%. Since the early 2000s, rituximab has been used to achieve tolerance in inhibitor patients with a "low risk profile". Rituximab binds to B cells, resulting in peripheral B cell depletion, but does not affect plasma cells or T cells. Rituximab enabled 40% (6 / 15) of SHA patients to obtain a negative inhibitor titer. The inhibitor was analyzed in the UK series, in which 80% (12 / 15) of the patients received FVIII simultaneously, and 27% (4 / 15) of the patients were used in combination with prednisolone or other IS. It may not be surprising that the success rate of the UK series seems lower than ours. Because all the patients included in the UK series had experienced ITI failure, and 33.3% of them had experienced more than one failure. However, in this study, when rituximab was used when the IS criteria were met, there was no evidence of waiting for ITI failure. In addition, our IS regimen included rituximab and prednisolone, not just rituximab. However, early identification of patients with a "poor risk profile" and the combined use of IS seems to improve the success rate. It is worth noting that no adverse reactions related to immunosuppressants such as infection were found in routine IVIG injections.

[0125] As pointed out by the comment of E. Carlos, the application of artificial intelligence (AI) in hemophilia is still in its early stages. Machine learning models in hemophilia have been used in the following aspects and achieved encouraging results: predicting disease severity, identifying factor V as an essential regulator of thrombin generation in mild to moderate HA, and determining predictors of long-term annualized bleeding rates. So far, we first developed this recommendation model based on baseline clinical information (immediate inhibitor titer before ITI, historical inhibitor titer, age and interval at the start of ITI) to add IS during LD-ITI. The application of AI currently mainly focuses on predicting ITI outcomes. But we hope that the AI model can provide useful treatment recommendations. This requires developing an optimal treatment recommendation model based on the characteristics of patients, rather than just focusing on outcome prediction. To establish a treatment plan recommendation model, traditional machine learning aims to adapt to clinicians' decisions based on past data. That is, under the judgment of clinicians, the mapping relationship between patient characteristics and treatment plans is learned without considering efficacy. However, the performance of the model largely depends on the level of clinicians, which requires clinicians to be very experienced in the disease for the model to perform perfectly. Limited by the characteristics of rare diseases, the model will inevitably make wrong decisions. To sum up, traditional machine learning models cannot effectively complete the task of recommending the optimal treatment plan for patients. Therefore, we adopt a method combining causal inference and machine learning to solve the above problems. Causal inference has developed rapidly in the past decade and has been widely applied in fields such as statistics, computer science, and medicine. It can accurately estimate the impact of different treatment plans on efficacy and recommend the optimal plan.

[0126] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code that contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0127] In general, the various example embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device. When aspects of the embodiments of the present disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented as non-limiting examples in hardware, software, firmware, dedicated circuits or logic, general hardware or a controller or other computing device, or some combination thereof.

[0128] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0129] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0130] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0131] The example embodiments of the present disclosure described in detail above are merely illustrative and not restrictive. Those skilled in the art should understand that various modifications and combinations can be made to these embodiments or their features without departing from the principles and spirit of the present disclosure, and such modifications should fall within the scope of the present disclosure.

Claims

1. A method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy, characterized in that: The method comprises: S101, obtaining clinical data X, treatment plan T, and treatment result Y of training set samples; S102, using clinical data and treatment outcomes to train a filling model that can predict treatment outcomes for different samples; S103, using clinical data and treatment plans to train a weighted model that can predict the likelihood of a sample receiving treatment; S104, calculating a DR estimate based on the filling model and the weighted model; calculating a causal effect value based on the DR estimate; S105, training a regression model based on the clinical data X and the causal effect value to obtain a recommendation model.

2. The method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy according to claim 1, characterized in that: The treatment plan includes whether IS is added during the LD-ITI treatment process, where T = 1 means receiving treatment and T = 0 means not receiving treatment; Optionally, the treatment result includes whether the LD-ITI treatment is successful, Y=1 indicates a successful result, and Y=0 indicates an unsuccessful result; Optionally, the clinical data include: inhibitor titer immediately before ITI, historical peak inhibitor titer, ITI start age, and interval time.

3. The method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy according to claim 1, characterized in that: The expression of the filling model is f(X,T); Optionally, the expression of the weighted model is g(X)=P(T=1|X); wherein T=1 indicates receiving treatment; Optionally, the calculation formula of the DR estimate is: ; Among them, 1(T=t) is the indicator function, which is equal to 1 when T is equal to t, otherwise it is 0; T is the random variable representing "intervention", and t is the specific value of the random variable.

4. The method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy according to claim 3, characterized in that: The calculation formula of the causal effect value is: .

5. The method for constructing a recommendation model for personalized addition of IS during low-dose immune tolerance induction therapy according to claim 1, characterized in that: The regression model includes any one of the following: linear regression, vector regression, random forest regression, gradient boosting regression tree.

6. A recommended method for personalized addition of IS during low-dose immune tolerance induction therapy, characterized in that: The method comprises: S201, obtaining clinical data of the subjects; S202, inputting the clinical data into a recommendation model constructed by the method according to any one of claims 1 to 5, and calculating a causal effect value; S203, recommending a treatment plan for the subject according to the causal effect value; when the causal effect value is greater than 0, outputting an auxiliary prediction result that the subject accepts the treatment; when the causal effect value is less than 0, outputting an auxiliary prediction result that the subject does not accept the treatment.

7. The method for recommending personalized addition of IS during low-dose immune tolerance induction therapy according to claim 6, characterized in that: The clinical data include any one or more of the following: inhibitor titer immediately before ITI, historical peak inhibitor titer, ITI start age, and interval time.

8. A computer device, characterized in that: The device comprises: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Images