Modeling method of lupus nephritis prediction model

Abstract

The invention discloses a modeling method of a lupus nephritis prediction model, and belongs to the technical field of intelligent medical treatment. The modeling method comprises the following steps: S1, respectively collecting flow cytometry detection data of lupus nephritis patients and healthy control personnel; s2, performing data cleaning and conversion on the flow cytometry detection data, and converting non-numerical features into digits; carrying out implication on the missing value by adopting a k nearest neighbor algorithm from an implication packet; s3, random sampling is carried out on the cleaned and converted data set, and samples are divided into a training set and a verification set according to the proportion of 7: 3; s4, a training subset and a test set are divided from the training set, and a specified number of features are iteratively selected by adopting an RFE method; and S5, performing a classification task by adopting TabPFNClassifier, performing training from features selected from the training data set, then evaluating the performance of the model, performing model training by utilizing detection data, and performing evaluation to construct a lupus nephritis prediction model with high accuracy.

Description

A modeling method for predicting lupus nephritis Technical Field

[0001] The present invention belongs to the field of intelligent medical technology, and specifically relates to a modeling method for a lupus nephritis prediction model. Background Art

[0002] Systemic lupus erythematosus (SLE) is a chronic autoimmune disease that affects multiple organ systems, particularly women of childbearing age. Lupus nephritis (LN), one of the most common and serious complications of SLE, affects approximately 50% of SLE patients and is a leading cause of death in SLE patients. Early identification of kidney damage in SLE patients is crucial for improving prognosis.

[0003] Currently, the initial clinical diagnosis of lupus nephritis (LN) primarily relies on indicators such as urine protein, urine protein-to-creatinine ratio, and cellular casts. However, pathological biopsy after renal puncture remains the gold standard for diagnosis. However, this invasive procedure is highly invasive for patients, has low patient acceptance, and results are delayed. Before a definitive diagnosis of systemic lupus erythematosus (SLE) is established, urine protein and creatinine testing are essential tools for diagnosing LN. While accurate, 24-hour urine protein quantification is often complicated, subject to poor patient compliance, and sample contamination, which can compromise interpretation. Urine creatinine testing is significantly affected by diet and can fluctuate significantly during the day. Therefore, the search for less invasive and simpler diagnostic methods is urgent.

[0004] With the widespread adoption of electronic medical records in recent years, the data accumulated within them has continued to grow. This includes a wealth of descriptive information reflecting a patient's physical condition and disease progression, including diagnoses, medications, examinations, and surgeries from admission to discharge. For example, Chinese patent CN 111986814A mines potentially valuable information from this vast amount of descriptive case information to assist and guide clinical treatment. However, its accuracy and persuasiveness still need improvement.

[0005] Therefore, it is necessary to rebuild a model that can accurately predict lupus nephritis based on the shortcomings of the existing diagnostic models. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a modeling method for a lupus nephritis prediction model. The present invention constructs a lupus nephritis prediction model with high accuracy by performing flow cytometry detection on lupus nephritis patients and healthy people, and using the detection data for model training and evaluation.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] In a first aspect, the present invention provides a method for modeling a lupus nephritis prediction model, comprising the following steps:

[0009] S1. Collect flow cytometry data from lupus nephritis patients and healthy controls;

[0010] S2. Perform data cleaning and transformation on the flow cytometry data, converting non-numerical features into numbers; imputation of missing values ​​using the k-nearest neighbor algorithm from the impute package;

[0011] S3. Randomly sample from the cleaned and transformed dataset and split the samples into training and validation sets in a ratio of 7:3.

[0012] S4. Split the training set into a training subset and a test set, and iteratively select a specified number of features using the RFE method;

[0013] S5. Use TabPFNClassifier for classification tasks, train it on the features selected from the training dataset, and then evaluate the performance of the model.

[0014] Furthermore, the flow cytometry data in S1 is constructed as follows:

[0015] (1) Take 5 mL of EDTA-anticoagulated whole blood from patient 2 and dilute it to 5 mL with PBS. Add 5 mL of ficoll and 5 mL of diluted blood sample to a 15 mL centrifuge tube and centrifuge.

[0016] (2) After the centrifugation in step (1), aspirate the PBMCs and add PBS to the centrifuge volume of 15 mL;

[0017] (3) After the centrifugation in step (2), remove the supernatant, add 5 mL of red blood cell lysis buffer and gently blow the cells apart. After 2 minutes, add PBS to 15 mL and centrifuge;

[0018] (4) After centrifugation in step (3), remove the supernatant and resuspend the cells in 100 μL flow cytometry staining buffer. Then, filter the cell suspension using a 70-mesh sieve and transfer it to a 1.5 mL eppendorf tube. Add 4 μL HumanTruStainFcX to each tube. TM , incubate at room temperature for 5 min;

[0019] Then, three incubation antibody combinations were added: 1. APC CD14 and PerCP-Cy5.5 CD16; 2. APC CD8a, PE-Cy7 CD38, PE HLA-DR, and FITC PD-1; 3. PerCP-Cy5.5 CD3, APC CD4, FITC CD8a, APC-Cy7CD45RA, and PE-Cy7 CD27. The mixture was gently flicked to mix and incubated at 4°C for 25 minutes.

[0020] (5) After step (4), 1 mL of PBS was added to each tube to wash the unbound antibody, the supernatant was discarded by centrifugation, 300 μL of PBS was added to each tube, and the tubes were transferred to flow cytometry tubes. The fluorescence compensation adjustment was completed on the flow cytometer, and the test results were analyzed;

[0021] (6) The flow cytometric analysis method includes data acquisition, analysis and recording; wherein multiple data channels record the light scattering and fluorescence emitted by each particle when it passes through the sensing area from each detector, and the analysis system classifies and counts the particles, wherein each particle is presented as a series of digitized parameter values.

[0022] Further, the flow cytometer is configured to be triggered at selected parameters to distinguish target particles from background and noise;

[0023] The trigger is used to detect a preset threshold value of a parameter, and is used as a means of detecting a particle passing through the laser beam, and an event exceeding the preset threshold value of the selected parameter is detected;

[0024] For particles or other components in the assay medium whose response is below a threshold, no data is collected. The trigger parameter may be detecting forward scattered light caused by the particle passing through the light beam, and then detecting and collecting light scattering and fluorescence data of the particle.

[0025] Furthermore, the non-numerical feature conversion in S2 uses LabelEncoder of sklearn.preprocessing to encode the non-numerical target label into a numerical value from 0 to n_classes-1.

[0026] Furthermore, the k-nearest neighbor algorithm in S2 finds the nearest k samples for each missing value and fills the missing value with the mean or mode of these neighbors.

[0027] Furthermore, a random seed set.seed(123) is set during random sampling in S3 to ensure the repeatability of the results.

[0028] Furthermore, the test set size in S4 is 33%.

[0029] Furthermore, the cell types finally included in the model construction in S4 are CD14nCD16pcell, CD14pCD16ncell, Act CD8T, Ex CD8T, CD8T, CD8T

[0030] Furthermore, the evaluation indicators in S5 include: accuracy, confusion matrix, precision, recall, F1-Score and AUC.

[0031] Contains at least the following beneficial technical effects:

[0032] By performing flow cytometry on lupus nephritis patients and healthy individuals, and using the test data for model training and evaluation, a highly accurate lupus nephritis prediction model was constructed. This can assist in the early diagnosis of LN in a minimally invasive manner, not only reducing dependence on renal puncture, but also avoiding the potential risks and inconveniences associated with it. At the same time, it provides SLE patients with more accurate and timely diagnosis, which is more accurate than predictions based on case data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the flow gate;

[0034] Figure 2 shows the model construction and performance verification diagram;

[0035] FIG3 is a ROC curve diagram of cell-based diagnosis. DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0037] Example

[0038] S1. Data Collection

[0039] Flow cytometry data were obtained from 181 lupus nephritis patients and 57 healthy subjects; specifically:

[0040] (1) Take about 2.5 mL of EDTA anticoagulated whole blood from the patient and dilute it with PBS to 5 mL. Add 5 mL of ficoll to a 15 mL centrifuge tube, then slowly layer 5 mL of the diluted blood sample, and then centrifuge at 1440 RPM for 20 min.

[0041] (2) After centrifugation, four layers can be observed in the centrifuge tube: from top to bottom, the plasma and platelet layer, the PBMC layer, the ficol layer, the polymorphonuclear leukocyte layer, and the red blood cell layer. The PBMC (buffy coat layer) is aspirated, of which 70-90% are lymphocytes (70-90%), of which T cells account for the vast majority of lymphocytes (45-70%). PBS is then added to 15 mL, and the tube is centrifuged at 500 x g for 5 min.

[0042] (3) After centrifugation, remove the supernatant and add about 5 mL of red blood cell lysis buffer to gently blow the cells apart. After 2 minutes, add PBS to 15 mL and then centrifuge at 500 × g for 5 minutes.

[0043] (4) After centrifugation, remove the supernatant and resuspend the cells in about 100 μL of flow cytometry staining buffer. Then, filter the cell suspension using a 70-mesh sieve and transfer it to a 1.5 mL eppendorf tube and add Human TruStain FcX TM (Fc Receptor Blocking Solution), add 4 μL to each tube and incubate at room temperature for 5 minutes; then add three incubation antibody combinations: 1. APCCD14 and PerCP-Cy5.5 CD16; 2. APC CD8a, PE-Cy7 CD38, PE HLA-DR and FITC PD-1; 3. PerCP-Cy5.5 CD3, APC CD4, FITC CD8a, APC-Cy7 CD45RA and PE-Cy7 CD27, and incubate at 4°C for 25 minutes.

[0044] (5) After incubation, add 1 mL of PBS to each tube to wash away unbound antibodies and centrifuge at 0.5 x g for 5 min. Discard the supernatant, add 300 μL of PBS to each tube, transfer to a flow cytometer, and perform flow cytometric analysis. Complete fluorescence compensation adjustment and analyze the test results. All test indicators are strictly set up with blank, isotype, FMO, negative, and positive control groups.

[0045] (6) The flow cytometric analysis method includes data acquisition, analysis and recording; wherein multiple data channels record the light scattering and fluorescence emitted by each particle when it passes through the sensing area from each detector, and the analysis system classifies and counts the particles, wherein each particle is presented as a series of digitized parameter values.

[0046] The flow cytometer is configured to trigger at selected parameters to distinguish target particles from background and noise;

[0047] The trigger is used to detect a preset threshold value of a parameter, and is used as a means of detecting a particle passing through the laser beam, and an event exceeding the preset threshold value of the selected parameter is detected;

[0048] For particles or other components in the assay medium whose response is below a threshold, no data is collected. The trigger parameter may be detecting forward scattered light caused by the particle passing through the light beam, and then detecting and collecting light scattering and fluorescence data of the particle.

[0049] Figure 1 (A) shows the flow cytometry gating process of CD14nCD16pcell (CD14-CD16+cell) and CD14pCD16ncell (CD14+CD16-cell);

[0050] Figure 1(B) shows the flow cytometry gating process for Ex CD8T (CD38+PD-1+CD8+T) cells and Act CD8T (CD38+HLA-DR+CD8+T);

[0051] Figure 1(C) shows the flow cytometry gating process for Naive CD8+ T cells.

[0052] S2. Data cleaning and transformation

[0053] Perform data cleaning and transformation on the flow cytometry data, use LabelEncoder from sklearn.preprocessing to encode non-numeric target labels into values ​​from 0 to n_classes-1; imputation is performed on missing values ​​using the k-nearest neighbor algorithm from the impute package;

[0054] S3. Divide the training set and validation set

[0055] Randomly sample from the cleaned and transformed dataset and split the samples into training and validation sets in a ratio of 7:3;

[0056] S4. Feature selection

[0057] The training set is divided into training subsets and test sets, and the RFE method is used to iteratively select a specified number of features.

[0058] The training features include:

[0059] CD14pCD16pcell (CD14 + CD16 + )、CD14nCD16pcell(CD14 - CD16 + )、CD14nCD16ncell(CD14 - CD16 - )、CD14pCD16ncell(CD14 + CD16 - ), NK(HLA - DR - CD56 + )、Pdc(CD 14 - CD 16 - CD11c int CD 123 + ) and cDCs (CD14 - CD16 - CD11c + CD123 - HLA - DR +);CD4T(CD19 - CD3 + CD8 - CD4 + )、CD8T(CD19 - CD3 + CD4 - CD8 + )、 CD8T(CD3 + CD4 - CD8 + CD27 + CD45RA + )、E M R A C D 8T(C D 3 + CD 4 - C D 8 + C D 27 - C D 45R A + )、C MC D 8T(C D 3 + C D 4 - C D 8 + C D 27 + C D 45R A - )、E MCD 8T(C D 3 + C D 4 - C D 8 + C D 27 - C D 45R A - )、 C D 4T(C D 3 + C D 8 - C D 4 + C D27 + C D 45R A + )、E M R AC D 4T(C D 3 + C D 4 + C D 8 - C D 27 - C D 45R A + )、C MC D 4T(C D3 + C D 4 + C D 8 - C D 27 + C D 45R A - )、E MCD4T(CD3 + CD4 + CD8 - CD27 - CD45RA -)、CXCR5⁺CD4⁺T (CXCR5 + CD4 + )、Tfh (CXCR5 + CD4 + PD - 1 + )、Eff CD4T (CD4 + CD25low)、Treg (CD4 + CD25 + )、Th1 (CD4 + CD25 low CXCR3 + CD196 - )、Th2 (CD4 + CD25 low CXCR3 - CD196 - )、Th17 (CD4 + CD25 low CXCR3 - CD196 + )、CD3⁺T (CD3 + )、DNT (CD3 + CD4 - CD8 - )DPT (CD3 + CD4 + CD8 + )、Act CD8T (CD8 + CD38 + HLA - DR + [[ID=

[64] ])、Ex CD8T (CD8 + CD38 + HLA - DR + );CD19⁺B (CD19 + )、ASC (CD19 + CD38 + CD27 + )、non - ASC (CD19[[ID=A3]] + excluding CD38 + CD27 + )、TransB (CD1 + CD38 + CD27 - CD24 + !>)、SWM B (CD19 + IgD - CD27[[ID=A1]] + It should be noted that there may be some inaccuracies in the original text, especially in the part where the tags seem to be a bit jumbled. I have tried my best to translate according to the rules while maintaining the integrity of the tags. If possible, it is recommended to check and correct the original text for a more accurate translation.)、UN S WM B(CD 19 + IgD + CD 27 + )、DNB(CD 19 + IgD - CD 27 - ), B(CD 19 + IgD + CD 27 - ), DN1(CD19 + IgD - CD27 - CD21 + CXCR5 + ), DN2(CD19 + IgD - CD27 - CD21 - CXCR5 int ) and ABC(CD 19 + Ig D - CD 27 - CD 21 - CX CR 5 in t CD 11c + ).

[0060] Recursive feature elimination (RFE) was used to screen features for inclusion in the lupus nephritis diagnostic model. RFE selects features by recursively considering smaller and smaller feature sets. The goal of RFE is to select the most important features by recursively reducing the size of the feature set examined. It uses model accuracy to identify which features contribute most to the predictions.

[0061] The working process of RFE is as follows:

[0062] 1. Train a model and use the model's coef_ or feature_importances_ attribute to get the importance of each feature;

[0063] 2. Based on the obtained feature importance, remove the least important feature from the current feature set.

[0064] 3. Repeat steps 1 and 2 until the desired number of features is reached.

[0065] 4. At each step, RFE trains a new model and evaluates the importance of the remaining features. This process continues until the preset number of features is reached.

[0066] This approach can effectively reduce the number of features while retaining the features that contribute most to the model.

[0067] The features included in the above methods include CD14nCD16pcell, CD14pCD16ncell, Act CD8T (CD38+HLA-DR+CD8+T), Ex CD8T (CD38+PD-1+CD8+T), CD8T, CD8T.

[0068] S5. Training model and evaluation

[0069] TabPFNClassifier is used for classification tasks, trained on features selected from the training dataset, and then the performance of the model is evaluated.

[0070] 1. Input feature vector construction

[0071] Data source: Flow cytometry test results of patients, including immune cell subset proportion data.

[0072] Feature dimension: Input vector (x∈R d ), where (d) is the number of immune cell characteristics.

[0073] 2. Forward propagation calculation

[0074] Step 1: Input the preprocessed feature vector into the Transformer encoder of TabPFNClassifier and capture the nonlinear interaction relationship h between immune cells through the multi-head attention mechanism enc :

[0075] [h enc =TransformerEncoder(x norm )]

[0076] Step 2: Generate predicted probability distribution through the probabilistic inference layer:

[0077] [p(y|x)=softmaxx((W·h enc +b)]

[0078] Where (W) and (b) are the classification layer parameters, and (y∈{0,1}) represents lupus nephritis negative / positive.

[0079] 3. Real-time inference efficiency

[0080] The model only requires a single forward propagation in a single inference, with a response time of <10ms (based on GPU acceleration), making it suitable for clinical instant diagnosis scenarios.

[0081] 4. Hierarchical decision logic:

[0082] 1. If the model output is (p(y=1|x)>0.7), lupus nephritis is directly diagnosed;

[0083] 2. If the model output is (0.4 ≤ p(y = 1 | x) ≤ 0.7), additional renal biopsy verification is recommended;

[0084] 3. If the model output is (p(y=1|x)>0.4), exclude the diagnosis based on clinical signs.

[0085] Figure 2 (A) shows CD14pCD16n cell / CD14nCD16p ROC curves of the training and test data of the CD8T / Act CD8T / ExCD8T cell combined diagnosis model for lupus nephritis;

[0086] Figure 2(B) shows the confusion matrix of the training set and test set data;

[0087] Figure 2 (C) shows CD14pCD16n cell / CD14nCD16p ROC curve of the validation set data of the CD8T / Act CD8T / ExCD8T cell combined diagnosis model for lupus nephritis;

[0088] Figure 2(D) is the confusion matrix of the validation set data.

[0089] Figure 3 is a ROC curve diagram of the diagnosis of 6 types of cells alone; A, B, C, D, E, and F represent the ROC curves of CD14pCD16n cells, CD14nCD16p cells, Naive CD8T, CD8T, Act CD8T, and Ex CD8T, respectively. The diagnostic results show that the diagnostic effect of the 6 types of cells alone is not as good as the CD14pCD16n cell / CD14nCD16p combination in the present invention. Combined diagnostic effect of CD8T / Act CD8T / Ex CD8T cells.

[0090] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A modeling method for a lupus nephritis prediction model, characterized in that: The following steps are involved: S1. Collect flow cytometry data from lupus nephritis patients and healthy controls; S2. Perform data cleaning and transformation on the flow cytometry data, converting non-numerical features into numbers; imputation of missing values ​​using the k-nearest neighbor algorithm from the impute package; S3. Randomly sample from the cleaned and transformed dataset and split the samples into training and validation sets in a ratio of 7:

3. S4. Split the training set into a training subset and a test set, and iteratively select a specified number of features using the RFE method; S5. Use TabPFNClassifier for classification tasks, train it on the features selected from the training dataset, and then evaluate the performance of the model.

2. The modeling method according to claim 1, characterized in that The flow cytometry data in S1 is constructed as follows: (1) Take 5 mL of EDTA-anticoagulated whole blood from patient 2 and dilute it to 5 mL with PBS. Add 5 mL of ficoll and 5 mL of diluted blood sample to a 15 mL centrifuge tube and centrifuge. (2) After the centrifugation in step (1), aspirate the PBMCs and add PBS to the centrifuge volume of 15 mL; (3) After the centrifugation in step (2), remove the supernatant, add 5 mL of red blood cell lysis buffer and gently blow the cells apart. After 2 minutes, add PBS to 15 mL and centrifuge; (4) After the centrifugation in step (3), remove the supernatant and resuspend the cells in 100 μL of flow cytometry staining buffer. Then, filter the cell suspension using a 70-mesh sieve and transfer it to a 1.5 mL eppendorf tube. Add 4 μL of HumanTruStainFcX to each tube. TM , incubate at room temperature for 5 min; Then, three incubation antibody combinations were added:

1. APC CD14 and PerCP-Cy5.5 CD16; 2. APC CD8a, PE-Cy7 CD38, PE HLA-DR and FITC PD-1; 3. PerCP-Cy5.5 CD3, APC CD4, FITC CD8a, APC-Cy7CD45RA and PE-Cy7 CD27; (5) After step (4), 1 mL of PBS was added to each tube to wash the unbound antibody, the supernatant was discarded by centrifugation, 300 μL of PBS was added to each tube, and the tubes were transferred to flow cytometry tubes. The fluorescence compensation adjustment was completed on the flow cytometer, and the test results were analyzed; (6) The flow cytometric analysis method includes data acquisition, analysis and recording; wherein multiple data channels record the light scattering and fluorescence emitted by each particle when it passes through the sensing area from each detector, and the analysis system classifies and counts the particles, wherein each particle is presented as a series of digitized parameter values.

3. The modeling method according to claim 2, characterized in that The flow cytometer is configured to trigger at selected parameters to distinguish target particles from background and noise; The trigger is used to detect a preset threshold value of a parameter, and is used as a means of detecting a particle passing through the laser beam, and an event exceeding the preset threshold value of the selected parameter is detected; For particles or other components in the assay medium whose response is below a threshold, no data is collected. The trigger parameter may be detecting forward scattered light caused by the particle passing through the light beam, and then detecting and collecting light scattering and fluorescence data of the particle.

4. The modeling method according to claim 1, characterized in that The non-numeric feature conversion in S2 uses LabelEncoder of sklearn.preprocessing to encode the non-numeric target label into a numerical value from 0 to n_classes-1.

5. The modeling method according to claim 1, characterized in that The k-nearest neighbor algorithm in S2 finds the nearest k samples for each missing value and fills the missing value with the mean or mode of these neighbors.

6. The modeling method according to claim 1, characterized in that The random seed set.seed(123) is set during random sampling in S3 to ensure the repeatability of the results.

7. The modeling method according to claim 1, characterized in that The test set size in S4 is 33%.

8. The modeling method according to claim 1, characterized in that: The cell types finally included in the model construction in S4 are CD14nCD16pcell, CD14pCD16ncell, Act CD8T, Ex CD8T, CD8T, CD8T.

9. The modeling method according to claim 1, characterized in that: The evaluation indicators in S5 include: accuracy, confusion matrix, precision, recall, F1-Score and AUC.

Images