Method and system for establishing artificial intelligence screening model for leukemia based on blood routine data

Abstract

The application provides a leukemia artificial intelligence preliminary screening model based on blood routine data and a system for establishing the same. The method for establishing the leukemia artificial intelligence preliminary screening model comprises the following steps: S1: collecting electronic medical record data of leukemia patients and non-leukemia patients and processing the same to obtain a plurality of training databases and a verification database; S2: training the plurality of training databases by using an analysis module to obtain a plurality of corresponding rule sets in sequence, and combining and summarizing the rule sets with a false negative rate of less than 5% to obtain a final rule set, i.e., a leukemia risk screening model; S3: verifying the verification database by using the leukemia risk screening model; and S4: performing multiple iterations on the leukemia risk screening model to obtain an optimal model of the leukemia risk screening model. The leukemia artificial intelligence preliminary screening model established by the application can quickly and accurately screen out leukemia.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medicine, in particular to a leukemia screening technology based on AI technology. BACKGROUND

[0002] Leukemia is a malignant proliferative disease of hematopoietic stem cells, mainly characterized by uncontrolled malignant proliferation of immature blood cells in bone marrow, and is one of the top ten malignant tumors in China. The occurrence of the disease is related to genetic and environmental pollution and differences in living habits of people in different regions. According to the statistics in 2019, the male mortality rate is 4.5 / 10 million, which is higher than the female patients of 3.8 / 10 million. Children and adolescents are prone to occur in children under 5 years old and 5-14 years old, and acute lymphoblastic leukemia is the most serious, accounting for the first place of incidence rate and mortality rate of children's tumors. As can be seen, leukemia has caused serious damage to people's health, and also brought serious burden to society and family.

[0003] Acute leukemia patients often do not know that they have leukemia themselves, and usually go to the hospital for treatment according to the initial symptoms, such as cold and fever, gum bleeding, subcutaneous hemorrhage, anemia, etc. The preliminary diagnosis of leukemia currently mainly relies on the examination of white blood cell morphology. Due to the high work intensity, fatigue caused by continuous work and other reasons, the efficiency of manual inspection is low. Therefore, rapid and accurate screening of leukemia has great clinical significance for saving the lives of patients. SUMMARY

[0004] In order to overcome the above technical defects, the first aspect of the present application provides a method for establishing a leukemia artificial intelligence preliminary screening model based on blood routine data, which comprises:

[0005] Step S1: using an analysis module to train a plurality of training databases respectively, obtaining a plurality of corresponding rule sets in turn, and combining and summarizing a plurality of rule sets with false negative rate <5% to obtain a final rule set, i.e. a leukemia risk screening model;

[0006] Step S2: the leukemia risk screening model is used to statistically obtain the false negative rate of the verification database. If the false negative rate of the rule set is <5%, it is judged that the rule set meets the verification standard. If the false negative rate of the rule set is ≥5%, return to step S2, adjust the setting before starting the rule training, start the rule training again to obtain a new rule set, until the false negative rate <5% is met;

[0007] Step S3: the leukemia risk screening model is iterated for multiple times, and the optimal model of the leukemia risk screening model under the condition of a specific iteration number and a calculation time limit is found out through multiple rounds of model calculation.

[0008] Further, before step S1, further comprising:

[0009] Step S11: preparing medical record data; collecting patient electronic medical records from a hospital electronic medical record platform, and collecting electronic medical records of patients diagnosed with leukemia and non-leukemia patients;

[0010] Step S12: medical feature extraction; extracting leukemia medical features and medical feature values from qualified electronic medical record data obtained in step S11; the leukemia medical features include basic information, blood routine test items, blood routine test parameters, and XN instrument Q-Flag information;

[0011] Step S13: feature data standardization and data cleaning; performing feature data standardization on the leukemia clinical and test big data obtained in step S12, removing data with missing values, and obtaining standard data samples, the standard data samples including a standard medical feature data set and a standard diagnosis result set, the data in the standard medical feature data set and the standard diagnosis result set being in one-to-one correspondence;

[0012] Step S14: feature screening; using Laboman AI software to perform statistics and rule mining on the standard data, and screening out leukemia-related feature parameters, the leukemia-related feature parameters including: total number of samples, true positive sample number, true negative sample number, predicted positive sample number, predicted negative sample number, false negative rate, false positive rate, true negative rate, true positive rate, release rate, total efficiency, positive predictive value, negative predictive value, specificity, sensitivity, Youden's index, false negative rate-confidence interval, T-test, and key case coverage rate, 19 statistical indicators;

[0013] Step S15: data set splitting; pre-splitting the data, and dividing the data into several training databases and one verification database.

[0014] Further, the blood routine test items include 35 specific feature fields, as shown in the following table:

[0015]

[0016]

[0017] The blood routine test parameters include 52 specific feature fields, as shown in the following table:

[0018]

[0019]

[0020]

[0021] The XN instrument Q-Flag information includes 14 specific feature fields, as shown in the following table:

[0022] Item Name Explanation Blasts / Abn Lympho? Possibility of blasts / possibility of abnormal lymphocytes Blasts? Possibility of blasts Abn Lympho? Possibility of abnormal lymphocytes Left Shift? Possibility of left shift Atypical Lympho? Possibility of atypical lymphocytes NRBC Present High NRBC count IG Present Increased immature granulocytes RBC Agglutination? Possibility of RBC agglutination Turbidity / HGB Interf? Possibility of celiac disease affecting HGB Iron Deficiency? Possibility of iron deficiency HGB Defect? Possibility of HGB defect Fragments? Possibility of RBC fragmentation Anemia Anemia PLT Clumps? Possibility of PLT clumps .

[0023] Further, the step S13 comprises:

[0024] S131 feature data standardization: a standard library and a leukemia professional database are established, a blood routine item, a research parameter, Q-Flag information and other medical entity name standard library are artificially constructed based on the leukemia professional database, the standard library contains a standard name and similar names of each medical name, and is coded for unique identification, forming a feature standard library, medical features and medical feature values extracted in step S12 are replaced by the feature standard library for different expressions of the same feature, unified feature description is obtained, and standardized medical feature data is obtained. Specifically, the replacement of the medical feature value is mainly the replacement of symbols, letters, characters, units and medical codes, and the symbols, letters, characters, units and medical codes are unified.

[0025] S132 data cleaning: the data is preprocessed as follows: (1) label feature data and label information; (2) discard the specimen with more missing key feature data information; (3) discard the specimen with missing label information; (4) fill the average value of the field with the average value of the field; for the field with a very large missing proportion, such as >=80%, discard the field.

[0026] Further, the plurality of training databases comprises a training database A, a training database B and a training database C, and the step S1 comprises: (1) before starting rule training, 7 settings are pre-set in the analysis module: model training intensity, combination condition number, maximum training round, false negative rate threshold, running time limit, automatic merging of redundant rules is opened or closed, and key hits are opened or closed; the training database A is imported into the analysis module; then the rule training is started, and the initial rule set initial A is obtained after the running is completed; for the obtained rule set initial A, three indexes of single rule overfitting, single rule false negative rate and single rule false positive rate are used to optimize the rule; (2) according to the mined rule set, the statistical result is viewed, if the rule set with false negative rate lower than 5% is judged as a rule set meeting the training standard; if the false negative rate is not lower than 5%, return to step (1), adjust the settings before starting rule training, modify the combination condition number, the false negative rate threshold and the running time limit, start the rule training again, until the false negative rate is lower than 5%, and the final rule set A is obtained; (3) repeat step S1, and obtain rule set B and rule set C by using database B and database C respectively; (4) finally, the rule set A, the rule set B and the rule set C are summarized to obtain the final rule set, which is called a leukemia risk screening model.

[0027] The second aspect of the present application provides a system for establishing a leukemia artificial intelligence preliminary screening model based on blood routine data, which comprises:

[0028] an analysis module for training the several training databases, sequentially obtaining several corresponding rule sets, and combining and summarizing several rule sets with a false negative rate of less than 5% to obtain a final rule set, i.e., a leukemia risk screening model; and for using the leukemia risk screening model to statistically obtain a false negative rate of a verification database, and judging a rule set with a false negative rate of less than 5% as a verification standard rule set, and if the false negative rate of the rule set is greater than or equal to 5%, adjusting the settings before starting the rule training, starting the rule training again to obtain a new rule set, and repeating until the false negative rate is less than 5%.

[0029] an iteration module for iterating the leukemia risk screening model multiple times to find the optimal model of the leukemia risk screening model under a specific iteration number and a calculation time limit through multiple rounds of model calculation.

[0030] Further, the analysis module comprises n groups of sub-models, n groups of statistical modules and a group of selection modules; each of the sub-models comprises a decision tree algorithm module, the sub-models divide the feature data, continuously construct suitable rules according to the decision tree algorithm module, and repeatedly perform the steps of the decision tree algorithm module several times to select qualified models in the current round, and finally statistically select a group of optimal models through the cooperative calculation of the n groups of sub-models; the statistical module is used to take the rule groups mined by a group of models as input, calculate various statistical indexes according to the actual labeled label data in the sample set; and the selection module is used to take the statistical indexes of a series of statistical modules as input, and select a group of sub-models with optimal false negative rate and false positive rate indexes from the sub-models corresponding to the series of statistical data.

[0031] Further, the system for establishing a leukemia artificial intelligence preliminary screening model based on blood routine data further comprises a data acquisition and processing module, which comprises:

[0032] a medical record data acquisition module for acquiring patient electronic medical records from a hospital electronic medical record platform, and collecting electronic medical records of patients diagnosed with leukemia and patients not diagnosed with leukemia;

[0033] a medical feature extraction module for extracting leukemia medical features from the qualified electronic medical record data obtained by the medical record data acquisition module, and extracting medical features and medical feature values; the leukemia medical features comprise basic information, blood routine items, blood routine research parameters and XN instrument Q-Flag information; the blood routine research parameters comprise 52 specific feature fields, as shown in the following table:

[0034]

[0035]

[0036]

[0037] The XN instrument Q-Flag information includes 14 specific feature fields, as shown in the following table:

[0038] Item Name Explanation Blasts / Abn Lympho? Possibility of blasts / possibility of abnormal lymphocytes Blasts? Possibility of blasts Abn Lympho? Possibility of abnormal lymphocytes Left Shift? Possibility of left shift Atypical Lympho? Possibility of atypical lymphocytes NRBC Present High NRBC count IG Present Increased immature granulocytes RBC Agglutination? Possibility of RBC agglutination Turbidity / HGB Interf? Possibility of celiac disease affecting HGB Iron Deficiency? Possibility of iron deficiency HGB Defect? Possibility of HGB defect Fragments? Possibility of RBC fragmentation Anemia Anemia PLT Clumps? Possibility of PLT clumps

[0039] The feature data standardization and data cleaning module is used for feature data standardization of leukemia clinical and test big data, removes data with missing values, and obtains standard data samples, which include a standard medical feature data set and a standard diagnosis result set, and the data in the standard medical feature data set and the standard diagnosis result set are in one-to-one correspondence.

[0040] The feature screening module is used for statistics and rule mining of the standard data by the Laboman AI software, and screens out feature parameters related to leukemia, including total specimen number, true positive specimen number, true negative specimen number, predicted positive specimen number, predicted negative specimen number, false negative rate, false positive rate, true negative rate, true positive rate, release rate, total efficiency, positive predictive value, negative predictive value, specificity, sensitivity, Youden index, false negative rate-confidence interval, T test, key case coverage, and 19 statistical indexes.

[0041] The data set splitting module is used for pre-splitting of data, and the data is equally divided into a plurality of training databases and a verification database.

[0042] Further, the feature data standardization and data cleaning include:

[0043] The feature data standardization module is used for establishing a standard library and a leukemia professional database, manually constructing a blood routine project, research parameters, Q-Flag information and other medical entity term standard library based on the leukemia professional database, the standard library containing standard names and similar names of each medical term and being coded for unique identification, forming a feature standard library, replacing different expressions of the same feature in the extracted medical features and medical feature values against the feature standard library, unifying feature description, obtaining standardized medical feature data, and specifically, replacing symbols, letters, words, units and medical codes of the medical feature values, and unifying symbols, letters, words, units and medical codes.

[0044] A data cleaning module is configured to preprocess the data as follows: (1) labeling feature data and label information; (2) discarding a specimen if a large amount of key feature data information of the specimen is missing; (3) discarding a specimen if label information of the specimen is missing; (4) filling a null field with an average value of the field; and (5) discarding a field if a missing proportion of the field is greater than or equal to 80%.

[0045] Further, the training databases include a training database A, a training database B and a training database C, and the analysis module is configured to: (1) before starting rule training, pre-setting in the analysis module: model training intensity, combination condition number, maximum training round, false negative rate threshold, running time limit, automatic merging of redundant rules, and 7 settings; importing the training database A into the analysis module; then starting rule training, and obtaining a preliminary rule set A after the running ends; using three indexes of single rule overfitting, single rule false negative rate and single rule false positive rate to optimize the rule set A; (2) according to the mined rule set, if the false negative rate is less than 5%, the rule set is determined as a qualified rule set; if the false negative rate is greater than or equal to 5%, returning to step (1) to adjust the settings before starting rule training, modifying the combination condition number, the false negative rate threshold and the running time limit, and starting rule training again until the false negative rate is less than 5%, and finally obtaining a final rule set A; (3) repeating steps (1) and (2) to obtain rule sets B and C by using the database B and the database C respectively; and (4) finally, the rule sets A, B and C are combined to obtain a final rule set, which is referred to as a leukemia risk screening model.

[0046] After the above technical scheme is adopted, compared with the prior art, the following beneficial effects are achieved:

[0047] The application utilizes the learning ability of the decision tree algorithm to divide the data into a plurality of training data groups and a verification data group, and the plurality of training data groups are used for establishing rules, so that a plurality of sets of rules can be obtained, the rules are finally combined to reduce the false negative rate, and the false negative rate is less than 1.5%, which is much better than the statistical requirement that the false negative rate is less than 5%. The leukemia artificial intelligence preliminary screening model established by the application can quickly and accurately screen leukemia, and has great clinical significance for saving lives of patients. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A flowchart of a leukemia artificial intelligence preliminary screening method based on blood routine data according to the application;

[0049] Figure 2 A flowchart of establishing a leukemia artificial intelligence preliminary screening model;

[0050] Figure 3A structure diagram of a system for establishing a leukemia artificial intelligence preliminary screening model based on blood routine data. DETAILED DESCRIPTION

[0051] The advantages of the present application are further illustrated in conjunction with the accompanying drawings and specific embodiments. Those skilled in the art should understand that the specific description below is illustrative rather than limiting, and should not limit the scope of protection of the present application.

[0052] As shown in Figure 1-2 , the present embodiment provides a leukemia artificial intelligence preliminary screening method based on blood routine data, which includes steps 1-3.

[0053] Step 1: Establish a leukemia risk screening model, including steps 11-14.

[0054] As shown in Figure 3 , the present embodiment establishes a leukemia risk screening model through a system for establishing a leukemia artificial intelligence preliminary screening model based on blood routine data, which includes a data acquisition and processing module, an analysis module, and an iteration module.

[0055] The method for establishing a leukemia risk screening model using the above-mentioned system includes steps 11-14:

[0056] Step 11: Data acquisition and processing module: collect electronic medical record data of leukemia patients and non-leukemia patients, extract medical features and medical feature values, standardize and clean the feature data, select feature parameters related to leukemia, and finally divide the data into several training databases and a validation database according to time period, thereby serving as the specimen to be analyzed.

[0057] Step 11 includes steps 111-115:

[0058] Step 111: Prepare medical record data; collect patient electronic medical records from hospital electronic medical record platform, and collect electronic medical records of patients diagnosed with leukemia and non-leukemia.

[0059] The method for collecting electronic medical records of patients diagnosed with leukemia is to compare the diagnosis results of electronic medical records with the disease names in the leukemia disease name database to obtain the electronic medical records of patients diagnosed with leukemia.

[0060] The method for collecting electronic medical records of non-leukemia patients is to exclude patients with unclear medical history and incomplete test data from the data of contemporaneous patients and medical personnel; the medical record data includes medical history records, test results, medical orders, surgical records, nursing records, and true diagnosis results, and the test results include medical features and threshold values.

[0061] Obtain qualified electronic medical record data;

[0062] The leukemia disease name database contains various medical disease names that can be determined as leukemia.

[0063] Step 112: medical feature extraction: performing leukemia medical feature extraction on the qualified electronic medical record data obtained in step 111, extracting medical features and medical feature values; the leukemia medical features include basic information, blood routine project, blood routine research parameter, XN instrument Q-Flag information.

[0064] The basic information table includes: gender, age, specimen number, date, medical record number, and 5 specific feature fields.

[0065] The blood routine project includes 35 specific feature fields, as shown in Table 1; the blood routine research parameter includes 52 specific feature fields, as shown in Table 2; and the XN instrument Q-Flag information includes 14 specific feature fields, as shown in Table 3.

[0066] Table 1: Blood routine project details

[0067]

[0068]

[0069] Table 2: Blood routine research parameter details

[0070]

[0071]

[0072]

[0073] Table 3: XN instrument Q-Flag information details

[0074]

[0075]

[0076] The medical feature values are the specific numerical values of each medical feature in the basic information, blood routine project, research parameter, and Q-Flag information features.

[0077] Obtain large data of leukemia clinical and examination.

[0078] Step 113: feature data standardization and data cleaning: the large data of leukemia clinical and test obtained in step 112 is subjected to feature data standardization, and data with missing values is removed to obtain standard data samples, the standard data samples include a standard medical feature data set and a standard diagnosis result set, and the data in the standard medical feature data set and the standard diagnosis result set are in one-to-one correspondence;

[0079] Step 113 includes 1131-1132:

[0080] Step 1131: feature data standardization: a standard library and a leukemia professional database are established by using conventional technical methods in the art, a blood routine item, a research parameter, a Q-Flag information and other medical entity name standard library are artificially constructed based on the leukemia professional database, the standard library contains standard names and similar names of each medical term, and is coded for unique identification, forming a feature standard library, the medical features and medical feature values extracted in step 112 are replaced by referring to the feature standard library, the feature description is unified, and the standardized medical feature data is obtained. Specifically, the replacement of the medical feature values mainly includes the replacement of symbols, letters, characters, units and medical codes, and the unification of symbols, letters, characters, units and medical codes;

[0081] Step 1132: data cleaning: the data is subjected to the following pretreatment: (1) labeling feature data and label information; (2) discarding the specimen with more missing key feature data information; (3) discarding the specimen with missing label information; (4) filling the average value of the field for the empty field; for the field with a very large missing proportion, such as ≥80%, the field is discarded.

[0082] Step 114: feature screening: the standard data is subjected to statistics and rule mining by using Laboman AI software to screen out feature parameters related to leukemia, the feature parameters related to leukemia include: total specimen number, true positive specimen number, true negative specimen number, predicted positive specimen number, predicted negative specimen number, false negative rate, false positive rate, true negative rate, true positive rate, release rate, total effective rate, positive predictive value, negative predictive value, specificity, sensitivity, Youden index, false negative rate-confidence interval, T test, key case coverage rate, 19 statistical indexes;

[0083] The data is summarized by using Excel, and the label set contains: specimen number, date, medical record number, positive label, leukemia positive case is marked as 1, and non-leukemia negative case is marked as 0.

[0084] The standard data is subjected to statistics and rule mining by using Laboman AI software to screen out parameters related to leukemia, and a set of judgment rules are mined.

[0085] The Laboman AI software is a Laboman AI rule system (blood version) software suitable for medical examination rules mining and analysis software used in hospitals, which can statistically analyze specimen result data and microscopic data collected by medical examination equipment and mine effective determination rules, and the main functions include: new experiment, import experiment, rule training, label determination rule setting, project group setting, etc.

[0086] The standard data is statistically analyzed and rule mining is performed using the Laboman AI software, and the rule mining includes 19 statistical indexes, including total specimen number, positive specimen number (real), negative specimen number (real), positive specimen number (predicted), negative specimen number (predicted), false negative rate, false positive rate, true negative rate, true positive rate, release rate, total efficiency, positive predictive value, negative predictive value, specificity, sensitivity, Youden index, false negative rate-confidence interval, T test, and key case coverage. These statistical items are commonly used statistical items in statistics, and the Laboman AI software belongs to the prior art.

[0087] The statistical analysis and rule mining are performed using the above statistical method and the Laboman AI software, and the false negative rate is obtained, and it is considered that the obtained rule has an extremely significant correlation with leukemia diagnosis when the false negative rate is less than 0.05%, and the model established by the rule set is reasonable.

[0088] The false negative rate is the ratio of the number of missed diagnoses of leukemia in the test data to the total number of test data;

[0089] The false positive rate is the ratio of the number of misdiagnosed leukemia in the test data to the total number of test data;

[0090] The true negative rate is the ratio of the number of true non-leukemia diagnosed by the leukemia risk screening model in the test data to the total number of test data;

[0091] The true positive rate is the ratio of the number of true leukemia diagnosed by the leukemia risk screening model in the test data to the total number of test data;

[0092] The release rate is the ratio of the sum of the number of true non-leukemia and the number of missed diagnosis of leukemia in the test data to the total number of test data;

[0093] The total efficiency is the ratio of the sum of the number of true non-leukemia and the number of true leukemia in the test data to the total number of test data;

[0094] The total effective rate is the ratio of the sum of the number of true non-leukemia and the number of true leukemia judged by the leukemia risk screening model in the test data to the total number of test data.

[0095] The positive predictive value = true positive number / (true positive number + false positive number) * 100%;

[0096] The negative predictive value = true negative number / (true negative number + false negative number) * 100%;

[0097] The sensitivity = true positive number / (true positive number + false negative number) * 100%;

[0098] The specificity = true negative number / (true negative number + false positive number) * 100%;

[0099] The Youden index = (sensitivity + specificity) - 1;

[0100] The false negative rate - confidence interval = the statistical result of 1000 times 50% random sampling;

[0101] The T test, Wherein is the sample mean, μ is the population mean, σ X is the sample standard deviation, and n is the sample size.

[0102] The key case coverage is the ratio of the cases marked by the rules to the total number of cases marked in the database.

[0103] Step 115: data set splitting: pre-cutting the data, and dividing the data into several training databases and a verification database.

[0104] The data is pre-cut, and the data is divided into four databases A, B, C and D according to time period. The first three databases A, B and C are used for rule establishment, and the last database D is used for rule verification.

[0105] Step 12: analysis module: using the analysis module to train the several training databases respectively, and sequentially obtaining several corresponding rule sets, and combining and summarizing several rules with false negative rate <5% to obtain the final rule set, i.e. leukemia risk screening model.

[0106] The analysis module is composed of n sets of sub-models, a statistical module, and a set of selection modules. The sub-models divide the feature data, continuously construct suitable rules according to the decision tree algorithm module, and repeat the algorithm module step n times to select the qualified model of the current round. Through the collaborative calculation of the n sets of sub-models, a set of optimal models is finally selected and counted. The statistical module takes the rule set mined by a set of models as input, calculates various statistical indicators such as false negative rate, false positive rate, specificity, sensitivity, etc. according to the actual labeled label data in the sample set. The selection module takes the statistical indicators of a series of statistical modules as input and selects a set of sub-models with the optimal false negative rate, false positive rate, etc. from the sub-models corresponding to a series of statistical data.

[0107] The several training databases include training database A, training database B, and training database C. The step 12 includes: (121) Excel data can be imported into Laboman AI software. In the import step, key cases can be marked, and key case coverage can be calculated during statistics; training database A is imported into Laboman AI, referred to as training data A, and this part of data is used for rule mining and establishment; before starting rule training, the following 7 settings are set in advance in the analysis module: model training intensity, combination condition number, maximum training round, false negative rate threshold, runtime limit, automatic merging of redundant rules on or off, and key hits on or off; training database A is imported into the analysis module; then rule training is started, and a preliminary rule set initial A is obtained after the running ends; for the obtained rule set initial A, single rule overfitting, single rule false negative rate, and single rule false positive rate are used to optimize the rules; (122) according to the mined rule set, the statistical results are viewed, and whether the false negative rate result is lower than 5% is focused on. If the false negative rate is lower than 5%, the rule set is judged as a training qualified rule set; if the false negative rate is not lower than 5%, step (121) is returned, and the settings before starting rule training are adjusted, and the combination condition number, the false negative rate threshold, and the runtime limit are focused on. Rule training is started again until the false negative rate is lower than 5%, and the final rule set A is obtained; (123) step 12 is repeated, and rule set B and rule set C are obtained by using database B and database C respectively; (124) finally, rule set A, rule set B, and rule set C are summarized to obtain the final rule set, referred to as the leukemia risk screening model. The rule set A is specifically shown in Table 4, the rule set B is specifically shown in Table 5, the rule set C is specifically shown in Table 6, and the final rule set is specifically shown in Table 7. The combination condition number is that each rule mined is combined by at most several parameters. The false negative rate threshold is the ratio of the number of white blood cells correctly predicted by the leukemia risk screening model to the total number of white blood cells predicted by the leukemia risk screening model in the test data.

[0108] Table 4 Rule set A

[0109] Serial Number Rule Details 1 [PLT-I] <= 130.0 AND [RBC] <= 3.015 AND [age] <= 49.5 2 [NEUT#&] <= 2.165 AND [RBC] <= 3.015 AND [age] > 49.5 3 [MO-WY]< = 53.5 AND [RBC] > 3.015 AND [age]< = 56.0 4 [LYMP#&] <= 0.105 AND [MO-WY] > 53.5 AND [RBC] > 3.015 5 [EO %] <= 30.85 AND [NE-WY] <= 1047.5 AND [NE-WZ] > 1045.5 6 [EO %] > 30.85 AND [NE-WY] <= 1047.5 7 [LY-WY] <= 1176.5 AND [MONO %] > 9.05 AND [NE-WY] > 1047.5 8 [MCV] <= 99.4 AND [NE-WY] > 382.0 AND [RDW-SD] <= 29.5 9 [MCV] <= 99.4 AND [NE-WZ] <= 53.0 AND [RDW-SD] > 29.5 10 [EO#] <= 0.005 AND [MCV] > 99.4 AND [NRBC%] <= 0.6 11 [MCV] > 99.4 AND [NRBC %] > 0.6 AND [age] <= 67.0

[0110] Table 5 Rule Set B

[0111] Serial Number Rule Details 1 [WBC] <= 1.935 AND [age] > 22.0 AND [age] <= 52.5 2 [WBC] <= 1.935 AND [age] > 52.5 3 [LY-Y] > 87.35 AND [NE-WX] <= 346.5 AND [WBC] > 1.935 4 [LY-X] > 84.05 AND [NE-WX] > 346.5 AND [WBC] > 1.935 5 [NE-WY] <= 664.0 AND [NE-WZ] > 845.5 AND [age] <= 50.5 6 [MacroR] > 4.0 AND [NE-WY] > 664.0 AND [age] <= 50.5 7 [MO-Z] <= 42.5 AND [NEUT#&] > 1.84 AND [age] > 50.5 8 [IG %] > 16.635 AND [MO-Z] > 42.5 AND [age] > 50.5 9 [EO %] <= 32.05 AND [LY-Z] > 85.8 AND [LYMPH #] <= 2.945 10 [EO %] <= 1.2 AND [LYMPH #] > 2.945 11 [EO %] > 32.05 AND [NE-SFL] > 140.5 12 [LY-WX] > 496.5 AND [MO-WY] <= 157.0 AND [age] <= 62.5 13 [MO-WY] > 157.0 AND [MO-Y] <= 142.5 AND [RDW-SD] <= 6.0 14 [MO-WY] > 157.0 AND [MO-Y] > 142.5

[0112] Table 6 Rule Set C

[0113]

[0114]

[0115] Table 7 Final Rule Set

[0116] Serial Number Rule 1 [PLT-I] <= 130.0 AND [RBC] <= 3.015 AND [age] <= 49.5 2 [NEUT#&] <= 2.165 AND [RBC] <= 3.015 AND [age] > 49.5 3 [MO-WY]< = 53.5 AND [RBC] > 3.015 AND [age]< = 56.0 4 [LYMP#&] <= 0.105 AND [MO-WY] > 53.5 AND [RBC] > 3.015 5 [EO %] <= 30.85 AND [NE-WY] <= 1047.5 AND [NE-WZ] > 1045.5 6 [EO %] > 30.85 AND [NE-WY] <= 1047.5 7 [LY-WY] <= 1176.5 AND [MONO %] > 9.05 AND [NE-WY] > 1047.5 8 [MCV] <= 99.4 AND [NE-WY] > 382.0 AND [RDW-SD] <= 29.5 9 [MCV] <= 99.4 AND [NE-WZ] <= 53.0 AND [RDW-SD] > 29.5 10 [EO#] <= 0.005 AND [MCV] > 99.4 AND [NRBC%] <= 0.6 11 [MCV] > 99.4 AND [NRBC %] > 0.6 AND [age] <= 67.0 12 [WBC] <= 1.935 AND [age] > 22.0 AND [age] <= 52.5 13 [WBC] <= 1.935 AND [age] > 52.5 14 [LY-Y] > 87.35 AND [NE-WX] <= 346.5 AND [WBC] > 1.935 15 [LY-X] > 84.05 AND [NE-WX] > 346.5 AND [WBC] > 1.935 16 [NE-WY] <= 664.0 AND [NE-WZ] > 845.5 AND [age] <= 50.5 17 [MacroR] > 4.0 AND [NE-WY] > 664.0 AND [age] <= 50.5 18 [MO-Z] <= 42.5 AND [NEUT#&] > 1.84 AND [age] > 50.5 19 [IG %] > 16.635 AND [MO-Z] > 42.5 AND [age] > 50.5 20 [EO %] <= 32.05 AND [LY-Z] > 85.8 AND [LYMPH #] <= 2.945 21 [EO %] <= 1.2 AND [LYMPH #] > 2.945 22 [EO %] > 32.05 AND [NE-SFL] > 140.5 23 [LY-WX] > 496.5 AND [MO-WY] <= 157.0 AND [age] <= 62.5 24 [MO-WY] > 157.0 AND [MO-Y] <= 142.5 AND [RDW-SD] <= 6.0 25 [MO-WY] > 157.0 AND [MO-Y] > 142.5 26 [NE-SSC] <= 165.7 AND [NEUT#&] <= 1.845 AND [RBC] <= 3.69 27 [NEUT#&] <= 1.845 AND [RBC] > 3.69 AND [RDW-CV] <= 11.8 28 [NEUT#&]>1.845 AND [RBC]<=3.575 AND [age]<=49.5 29 [MO-WX] <= 107.0 AND [NEUT#&] > 1.845 AND [age] > 49.5 30 [MONO %] <= 5.4 AND [NE-WX] <= 382.5 AND [NEUT #&] <= 2.98 31 [BA-D%] > 1.285 AND [NE-WX] <= 382.5 AND [NEUT#&] > 2.98 32 [LY-X] > 87.05 AND [NE-WX] > 382.5 AND [age] > 18.5

[0117] The method of leukemia judgment using the final rule set (i.e. leukemia risk screening model) in Table 7 is as follows: 1. Each rule has 2-3 item combination judgment, and they are in AND (and) relationship, that is, it needs to meet the judgment requirement at the same time to be a positive patient, which refers to a leukemia patient here; 2. The rule set is a combination of several rules, and each rule in the rule set is in OR (or) relationship, that is, it is a positive patient if any one rule is met, which refers to a leukemia patient here. For example, the first rule in the final rule set is: [PLT-I] <= 130.0 AND [RBC] <= 3.015 AND [age] <= 49.5, when the patient's test results meet ([PLT-I] <= 130.0 and [RBC] <= 3.015 and [age] <= 49.5) at the same time, it is suggested that the patient may be a leukemia patient, otherwise the second rule is verified, and when none of the 32 rules judges the patient to be a leukemia patient, the patient is suggested to be a non-leukemia patient.

[0118] Step 13: The leukemia risk screening model is used to statistically obtain the false negative rate of the verification database, and the rule set with a false negative rate < 5% is judged to be a verification qualified rule set, if the false negative rate of the rule set is ≥ 5%, then return to step 12, adjust the setting before starting the rule training, start the rule training again to obtain a new rule set, until the false negative rate < 5% is met.

[0119] The leukemia risk screening model performs statistics on the training database D obtained in step 115, and as the training data, the total number of samples, the number of positive samples (true), the number of negative samples (true), the number of positive samples (predicted), the number of negative samples (predicted), the false negative rate, the false positive rate, the true negative rate, the true positive rate, the release rate, the total efficiency, the positive predictive value, the negative predictive value, the specificity, the sensitivity, the Youden index, the false negative rate-confidence interval, the T test, and the key case coverage, 19 statistical indicators, can be obtained. The false negative rate result is mainly observed whether it is lower than 5%. If the false negative rate is lower than 5%, the rule set is judged as a verified rule set. If the false negative rate is not lower than 5%, return to step 12, adjust the settings before starting the rule training, and mainly modify the number of combined conditions, the false negative rate threshold, and the running time limit, start the rule training again to obtain a new rule set, until the false negative rate is lower than 5%.

[0120] Using the verification database D, the final rule set is verified, and the statistics show that the false negative rate is 1.30%.

[0121] Step 14: Iteration module: multiple iterations are performed on the leukemia risk screening model, and the optimal model of the leukemia risk screening model under the condition of a specific number of iterations and a calculation time limit is found out through multiple rounds of model calculation. The iteration module is multiple iterations of the analysis module, and aims to find out the optimal model under the condition of a specific number of iterations and a calculation time limit through multiple rounds of model calculation.

[0122] Step 2: Organize the data of patients to be screened.

[0123] The data of patients to be screened is organized according to the method of steps 112-115. The user data includes medical characteristics and medical characteristic values.

[0124] The leukemia medical characteristics include basic information, blood routine project, blood routine research parameter, and XN instrument Q-Flag information. The blood routine project has 35 specific characteristic fields, as shown in Table 1; the blood routine research parameter has 52 specific characteristic fields, as shown in Table 2; and the XN instrument Q-Flag information has 14 specific characteristic fields, as shown in Table 3.

[0125] The medical characteristic values are the specific numerical values of each medical characteristic in the basic information, blood routine project, blood routine research parameter, and XN instrument Q-Flag information characteristics.

[0126] Obtain the big data of leukemia clinical and examination.

[0127] Step 3: Substitute the data of patients to be screened into the leukemia risk screening model for calculation, and finally obtain the leukemia risk prediction result.

[0128] For example, the partial test results of a patient to be screened are: male, 41 years old, PLT-I = 120.0, WBC = 19.73, RBC = 2.97, HGB = 94, HCT = 0.296, MCV = 29.6, MCHC = 31.6, etc. According to the final rule set in Table 7, the patient is determined to be positive. The patient is clinically diagnosed to be positive.

[0129] The laboratory can quickly and accurately screen leukemia patients based on the blood routine results and using the leukemia risk screening model.

[0130] It should be noted that the embodiments of the present application have better implementation, and do not limit the present application in any form. Any skilled person in the art can change or modify the above disclosed technical content into equivalent effective embodiments without departing from the technical scheme of the present application. Any modification or equivalent change and modification of the above embodiments according to the technical essence of the present application still belongs to the scope of the technical scheme of the present application.

Claims

1. A method for establishing an artificial intelligence-based initial screening model for leukemia based on routine blood data, characterized in that, include: Step S1: Use the analysis module to train several training databases respectively, obtain several corresponding rule sets in sequence, and summarize several rule sets with false negative rates <5% to obtain the final rule set, i.e., the leukemia risk screening model. Step S2: The leukemia risk screening model statistically analyzes the validation database to obtain the false negative rate. The rule set with a false negative rate of <5% is judged as the validation standard met. If the false negative rate of the rule set is ≥5%, the settings before starting rule training are adjusted, and rule training is started again to obtain a new rule set until the false negative rate is <5%. Step S3: Iterate the leukemia risk screening model multiple times, and find the optimal model within a specific number of iterations and computation time constraints through multiple rounds of model calculation. Use a rule set as the leukemia risk screening model. The method for leukemia diagnosis using this rule set is as follows: each rule in the rule set is a combination of multiple items, which are related by AND, meaning that multiple criteria must be met simultaneously to determine if the patient is a leukemia positive patient; each rule in the rule set is related by OR, meaning that meeting any one rule indicates a leukemia positive patient. The rule set contains 32 rules as follows: Prior to step S1, the method further includes: Step S11: Prepare medical record data; collect patients' electronic medical records from the hospital's electronic medical record platform, and collect electronic medical records of patients diagnosed with leukemia and those without leukemia; Step S12: Medical Feature Extraction: Extract medical features for leukemia from the qualified electronic medical record data obtained in Step S11, extracting medical features and medical feature values; the leukemia medical features include basic information, blood routine test items, blood routine test parameters, and XN instrument Q-Flag information; Step S13: Feature data standardization and data cleaning: The big data of clinical and laboratory leukemia obtained in step S12 is standardized to remove data with missing values ​​and obtain standard data samples. The standard data samples include standard medical feature datasets and standard diagnostic result sets. The data in the standard medical feature datasets and standard diagnostic result sets have a one-to-one correspondence. Step S14: Feature Screening: Statistical analysis and rule mining are performed on the standard data using Laboman AI software to screen out feature parameters related to leukemia. These feature parameters include: total number of samples, number of true positive samples, number of true negative samples, number of predicted positive samples, number of predicted negative samples, false negative rate, false positive rate, true negative rate, true positive rate, release rate, overall effectiveness, positive predictive value, negative predictive value, specificity, sensitivity, Youden index, false negative rate confidence interval, T-test, and coverage rate of key cases, totaling 19 statistical indicators. Step S15: Dataset splitting: Pre-split the data, dividing it into several training databases and one validation database; The blood routine test parameters include 52 specific feature fields, as detailed in the table below: The XN instrument Q-Flag information includes 14 specific feature fields, as detailed in the table below: Step S13 includes: S131 Feature Data Standardization: Establish a standard library and a leukemia-specific database. Based on the leukemia-specific database, artificially construct a standard library of blood routine test items, blood routine test parameters, XN instrument Q-Flag information, and other medical entity terms. The standard library contains the standard name of each medical term and similar names that have appeared, and encodes them for unique identification, forming a feature standard library. For the medical features and medical feature values ​​extracted in step S12, replace the different expressions of the same feature with the feature standard library to unify the feature description and obtain standardized medical feature data. Specifically, the replacement of medical feature values ​​mainly involves the replacement of symbols, letters, characters, units, and medical codes to unify symbols, letters, characters, units, and medical codes. S132 Data Cleaning: The data is preprocessed as follows: (1) Label feature data and label information; (2) If a lot of key feature data information of the specimen is missing, the specimen is discarded; (3) If the label information of the specimen is missing, the specimen is discarded; (4) Fill null fields with the average value of the field; For fields with a very large missing ratio, such as ≥80%, the field is discarded.

2. The method for establishing an artificial intelligence-based initial screening model for leukemia based on routine blood data as described in claim 1, characterized in that, The training databases include training database A, training database B, and training database C. Step S1 includes: (1) Before starting rule training, pre-set the following in the analysis module: model training intensity, number of combination conditions, maximum number of training rounds, false negative rate threshold, runtime limit, automatic merging of redundant rules on or off, and key points must be hit on or off, 7 settings; import training database A into the analysis module; then start rule training, and after the run is completed, obtain the initial rule set A; for the obtained initial rule set A, use three indicators to optimize the rules: single rule overfitting, single rule false negative rate, and single rule false positive rate; (2) Based on the mined rule set, check the statistical results. If the rule set with a false negative rate of less than 5% is judged as a rule set that has met the training target, if the false negative rate is not lower than 5%, return to step (1), adjust the settings before starting rule training, modify the number of combination conditions, the false negative rate threshold, and the runtime limit, and start rule training again until the false negative rate is lower than 5%, and obtain the final rule set A; (3) Repeat step S1, and use database B and database C to obtain rule set B and rule set C respectively; (4) Finally, summarize rule set A, rule set B, and rule set C to obtain the final rule set, which is called the leukemia risk screening model.

3. A system for establishing an artificial intelligence-based initial screening model for leukemia based on routine blood data, characterized in that, include: The analysis module is used to train several training databases, sequentially obtain several corresponding rule sets, and summarize several rule sets with false negative rates <5% to obtain the final rule set, i.e., the leukemia risk screening model; it is also used to use the leukemia risk screening model to statistically analyze the validation database to obtain the false negative rate. The rule set with a false negative rate <5% is judged as the validation qualified rule set. If the false negative rate of the rule set is ≥5%, the settings before starting rule training are adjusted, and rule training is started again to obtain a new rule set until the false negative rate <5% is met; The iteration module is used to iterate the leukemia risk screening model multiple times, and find the optimal model of the leukemia risk screening model under specific iteration number and computation time constraints through multiple rounds of model calculation. The analysis module comprises n sub-models, n statistical modules, and a selection module. Each sub-model includes a decision tree algorithm module. The sub-model divides the feature data, continuously constructs suitable rules based on the decision tree algorithm module, and repeats the steps of the decision tree algorithm module several times to select the qualified model for the current round. Through the collaborative calculation of the n sub-models, a set of optimal models is finally statistically selected. The statistical module is used to take the rule set mined from a set of models as input and calculate various statistical indicators based on the actual labeled data in the sample set. The selection module is used to take the statistical indicators of a series of statistical modules as input and select a set of sub-models with the best false negative rate and false positive rate among the sub-models corresponding to a series of statistical data. The system further includes a data acquisition and processing module, which comprises: The medical record data acquisition module is used to collect patients' electronic medical records from the hospital's electronic medical record platform, and to collect electronic medical records of patients diagnosed with leukemia and those not diagnosed with leukemia. The medical feature extraction module is used to extract leukemia medical features from qualified electronic medical record data obtained by the medical record data acquisition module, and to extract medical features and medical feature values; the leukemia medical features include basic information, blood routine test items, blood routine test parameters, and XN instrument Q-Flag information; The blood routine test parameters include 52 specific feature fields, as detailed in the table below: The XN instrument Q-Flag information includes 14 specific feature fields, as detailed in the table below: The feature data standardization and data cleaning module is used to standardize the feature data of big data in clinical and laboratory studies of leukemia, remove data with missing values, and obtain standard data samples. The standard data samples include standard medical feature datasets and standard diagnostic result sets. The data in the standard medical feature datasets and standard diagnostic result sets have a one-to-one correspondence. The feature filtering module is used to perform statistical analysis and rule mining on standard data using Laboman AI software to filter out feature parameters related to leukemia. These leukemia-related feature parameters include: total number of samples, number of true positive samples, number of true negative samples, number of predicted positive samples, number of predicted negative samples, false negative rate, false positive rate, true negative rate, true positive rate, release rate, overall effectiveness, positive predictive value, negative predictive value, specificity, sensitivity, Youden index, false negative rate confidence interval, T-test, and coverage rate of key cases, totaling 19 statistical indicators. The dataset splitting module is used to pre-split the data, dividing it into several training databases and a validation database. Feature data standardization and data cleaning include: The feature data standardization module is used to establish a standard library and a leukemia-specific database. Based on the leukemia-specific database, a standard library of blood routine test items, blood routine test parameters, XN instrument Q-Flag information, and other medical entity terms is manually constructed. The standard library contains the standard name of each medical term and similar names that have appeared, and is coded for unique identification, forming a feature standard library. For the extracted medical features and medical feature values, different expressions of the same feature are replaced with the feature standard library to unify the feature description and obtain standardized medical feature data. Specifically, the replacement of medical feature values ​​mainly involves the replacement of symbols, letters, characters, units, and medical codes to unify symbols, letters, characters, units, and medical codes. The data cleaning module is used to perform the following preprocessing on the data: (1) labeling feature data and label information; (2) discarding specimens with a large amount of missing key feature data information; (3) discarding specimens with missing label information; (4) filling null fields with the average value of the field; and discarding fields with a very high missing rate, such as ≥80%.

4. The system for establishing an artificial intelligence-based initial screening model for leukemia based on routine blood data as described in claim 3, characterized in that, The training databases include training database A, training database B, and training database C. The analysis module is used for: (1) setting the following settings in the analysis module before starting rule training: model training intensity, number of combination conditions, maximum number of training rounds, false negative rate threshold, runtime limit, automatic merging of redundant rules on or off, and key rule must-hit on or off; importing training database A into the analysis module; then starting rule training, and obtaining the initial rule set A after the operation is completed; optimizing the obtained rule set A using three indicators: single rule overfitting, single rule false negative rate, and single rule false positive rate. Then; (2) Based on the mined rule set, check the statistical results. If the rule set with a false negative rate < 5%, it is judged as a rule set that has met the training target; if the false negative rate ≥ 5%, return to step (1), adjust the settings before starting rule training, modify the number of combination conditions, the false negative rate threshold, and the running limit, and start rule training again until the false negative rate < 5% is met, and the final rule set A is obtained; (3) Repeat (1) and (2), and use database B and database C to obtain rule set B and rule set C respectively; (4) Finally, summarize rule set A, rule set B, and rule set C to obtain the final rule set, which is called the leukemia risk screening model.

Images