Data processing device and equipment based on blood routine examination and storage medium

Abstract

The invention discloses a data processing device and equipment based on blood routine examination and a storage medium, and relates to the field of electrical digital data processing.The device performs unsupervised classification on a three-dimensional light scattering data set through a Gaussian mixture model based on three-dimensional light scattering signals to obtain lymphocyte subgroups, mononuclear cell subgroups and neutrophil subgroups; and extracting distribution center bits, percentile widths and variable coefficients of the three subgroups, integrating to obtain a feature quantification database, inputting the feature quantification database into a self-organizing mapping neural network, mapping high-dimensional features into discretized risk indexes, and outputting a risk level corresponding to each risk index. And finally, extracting all key parameters of the external blood routine sample and inputting the key parameters to the self-organizing mapping neural network, the risk index and the corresponding risk level. According to the method, the SOM is applied to data processing for the first time, a TRI supervision scoring mechanism is introduced on the basis of an unsupervised SOM layer, and disease mode exploration and clinical decision support are both considered.

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, and in particular to a data processing device, equipment and storage medium based on blood routine data. Background Technology

[0002] Currently, the main methods for processing blood routine samples or data related to cancer include: imaging, endoscopic imaging, and tumor marker data processing. However, all three methods have significant limitations. Imaging suffers from a noticeable time lag, and frequent testing carries a certain risk of cancer. Endoscopic imaging has a limited detection range, is expensive, can cause discomfort, and may lead to cross-infection. Tumor marker data processing suffers from low accuracy, insufficient specificity, and high false negative and false positive rates, resulting in low precision in processing blood routine samples. Summary of the Invention

[0003] The main objective of this application is to provide a data processing device, equipment, and storage medium based on routine blood tests to solve the problem of low data processing accuracy of routine blood test samples in the prior art.

[0004] To achieve the above objectives, this application provides the following technical solution:

[0005] A data processing device based on routine blood tests, the data processing device being applied to routine blood test samples, the data processing device comprising:

[0006] The three-dimensional light scattering dataset acquisition module is used to acquire several three-dimensional light scattering signals of the blood routine sample based on the FSC / SSC / SFL dimensions using the three-laser flow cytometry method, and to perform logarithmic transformation on all three-dimensional light scattering signals to obtain a three-dimensional light scattering dataset.

[0007] A three-dimensional light scattering dataset clustering module is used to perform unsupervised classification of the three-dimensional light scattering dataset using a Gaussian mixture model to obtain lymphocyte subsets, monocyte subsets, and neutrophil subsets.

[0008] The feature quantization database acquisition module is used to extract the distribution center position, percentile width, and coefficient of variation of each subgroup based on FSC / SSC / SFL dimensions, and integrate them to obtain the feature quantization database.

[0009] The risk index mapping module is used to input the feature quantization database into the self-organizing map neural network, map the high-dimensional features into discrete risk indices through a dynamic neighborhood radius decay mechanism, and output the risk level corresponding to each risk index.

[0010] The key parameter screening module is used to iterate the AUC value of the self-organizing map neural network based on a preset sample set and a preset validation set, and when the AUC value exceeds a preset percentage threshold, obtain the risk index that contributes the most to the AUC value as a key parameter based on the feature quantization database.

[0011] The routine blood sample detection module extracts all key parameters of the external routine blood sample and inputs them into the self-organizing map neural network. Based on the output of the self-organizing map neural network, the risk index and corresponding risk level of the external routine blood sample are obtained.

[0012] As a further improvement to this application, the three-dimensional light scattering dataset acquisition module includes:

[0013] The standard microsphere calibration unit is used to calibrate the detection optical path of the three-laser flow cytometry method using standard microspheres.

[0014] The three-dimensional light scattering signal acquisition unit is used to acquire several three-dimensional light scattering signals of a blood routine sample based on the FSC / SSC / SFL dimensions through the calibrated detection optical path;

[0015] The extreme value removal unit is used to remove extreme values ​​exceeding ±5SD in any dimension of FSC / SSC / SFL from all three-dimensional light scattering signals.

[0016] The missing value interpolation unit is used to interpolate the missing values ​​of all three-dimensional light scattering signals to obtain the corrected three-dimensional light scattering signals.

[0017] The scattering signal logarithmic transformation unit is used to convert the signal value of each corrected three-dimensional light scattering signal. conduct The conversion yields a three-dimensional light scattering data based on a corrected three-dimensional light scattering signal;

[0018] The scattering dataset integration unit is used to integrate all three-dimensional light scattering data to obtain a three-dimensional light scattering dataset.

[0019] As a further improvement to this application, the three-dimensional light scattering dataset clustering module includes:

[0020] The N×3 matrix transformation unit is used to convert a three-dimensional light scattering dataset into an N×3 matrix, where N is the number of cell events, each row represents one cell event, and the three columns represent the three-dimensional light scattering data of FSC / SSC / SFL, respectively.

[0021] The Gaussian mixture model definition unit is used to define the number of Gaussian components, covariance type, maximum number of iterations, and initialization strategy of the Gaussian mixture model based on a first preset strategy.

[0022] The Gaussian Mixture Model Training Unit is used to input an N×3 matrix into the Gaussian Mixture Model and call the GaussianMixture module of scikit-learn to train the Gaussian Mixture Model several times.

[0023] The contour coefficient acquisition unit is used to acquire contour coefficients once per training iteration via the GaussianMixture module.

[0024] The contour coefficient judgment unit is used to determine whether the contour coefficient of the current training iteration is greater than a preset threshold based on each training iteration.

[0025] The unsupervised classification result labeling unit is used to determine that the Gaussian mixture model has completed unsupervised classification if the result is positive. Based on the unsupervised classification result, the lymphocyte subset, monocyte subset, and neutrophil subset are labeled.

[0026] As a further improvement to this application, the feature quantization database acquisition module includes:

[0027] The scattering dataset splitting unit is used to split the three-dimensional light scattering dataset into independent sets of lymphocytes, monocytes, and neutrophils based on the classification results of lymphocyte subsets, monocyte subsets, and neutrophil subsets.

[0028] The distribution center position definition unit is used to obtain the center position of each subgroup in each dimension by means of the median, and the center position of a subgroup in one dimension is defined as a distribution center position.

[0029] The subgroup sorting unit is used to sort the values ​​of each dimension of each subgroup from smallest to largest, and select the 10th percentile to 90th percentile of the sorted dimension to form a span interval as the percentile width of the current dimension.

[0030] The coefficient of variation calculation unit is used to obtain the standard deviation and mean of the current dimension, and to obtain the coefficient of variation of the current dimension.

[0031] The subgroup data integration unit is used to integrate the distribution center position, percentile width, and coefficient of variation of all dimensions of all subgroups to obtain a feature quantification database.

[0032] As a further improvement to this application, the risk index mapping module includes:

[0033] The self-organizing map neural network definition unit is used to define the neuron weights, grid size, topology type, weight initialization method, initial neighborhood radius, decay strategy, and learning rate of the self-organizing map neural network based on a second preset strategy.

[0034] The optimal neuron definition unit is used to input the feature quantization database into the self-organizing map neural network to calculate the Euclidean distance between each input data and the weights of all neurons, and the neuron with the smallest Euclidean distance to the current input data is defined as the optimal matching unit.

[0035] The neighborhood radius dynamic decay unit is used to dynamically decay the neighborhood radius of each best-matching unit using a Gaussian function.

[0036] The best matching position marking unit is used to mark the best matching unit position of all healthy samples after the feature quantization database has been input.

[0037] The healthy cluster center marker cell is used to calculate the centroid of the best matching cell position for all healthy samples and mark it as the healthy cluster center;

[0038] The risk index definition unit is used to obtain the maximum Euclidean distance between all input data and the center of the health cluster, as well as the ratio of the Euclidean distance of each input data to the maximum Euclidean distance, and to define the ratio of each input data as a risk index.

[0039] Risk interval division unit, used as the half-open interval boundary with 0, 30, 60, 100 as low risk interval, medium risk interval, and high risk interval respectively;

[0040] The risk level output unit is used to obtain the risk range matched by each risk index, so as to output the risk level corresponding to each risk index.

[0041] As a further improvement to this application, the key parameter screening module includes:

[0042] Subsample set division unit, used to divide a preset sample set into several subsample sets;

[0043] The current fold AUC value calculation unit is used to input each fold subsample set into the self-organizing map neural network and calculate the current AUC value once based on training for one fold.

[0044] The network weight reset unit is used to reset the network weights once for each training fold.

[0045] The self-organizing map neural network termination unit is used to complete the training of the self-organizing map neural network when all subsample sets have been input or the AUC value exceeds a preset percentage threshold.

[0046] The initial validation AUC value acquisition unit is used to input the preset validation set into the trained self-organizing map neural network to obtain the initial validation AUC value;

[0047] The AUC value decrease rate acquisition unit is used to randomly shuffle one of the values ​​in the preset validation set and obtain the decrease rate of the initial validation AUC value.

[0048] The Spearman rank correlation coefficient acquisition unit is used to obtain the Spearman rank correlation coefficients between the feature quantification database and each risk index.

[0049] The key parameter definition unit is used to define key parameters when the decrease in the risk index is greater than a preset percentage threshold and the Spearman rank correlation coefficient is greater than a preset coefficient threshold.

[0050] As a further improvement to this application, the blood routine sample detection module includes:

[0051] The feature quantization database acquisition unit is used to process external blood routine samples through the three-dimensional light scattering dataset acquisition module to obtain the feature quantization database;

[0052] The feature quantization database filtering unit is used to retain data in the feature quantization database that are the same as all key parameter categories and input them into the self-organizing map neural network.

[0053] The health cluster center comparison unit is used to obtain the Euclidean distance between each data point and the health cluster center, as well as the ratio of the Euclidean distance of each data point to the maximum Euclidean distance, and to define the ratio of each data point as a risk index of the data.

[0054] The blood routine sample grading matching unit is used to obtain the risk range matched by the risk index of each data point, so as to output the risk level corresponding to the risk index of each data point.

[0055] To achieve the above objectives, this application also provides the following technical solutions:

[0056] An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the data processing apparatus as described above.

[0057] To achieve the above objectives, this application also provides the following technical solutions:

[0058] A storage medium storing program instructions that, when executed by a processor, enable the data processing apparatus as described above.

[0059] Beneficial effects:

[0060] This application uses a three-laser flow cytometry method to acquire several three-dimensional light scattering signals from routine blood samples based on the FSC / SSC / SFL dimensions, and performs logarithmic transformation on all three-dimensional light scattering signals to obtain a three-dimensional light scattering dataset. An unsupervised classification of the three-dimensional light scattering dataset is performed using a Gaussian mixture model to obtain lymphocyte, monocyte, and neutrophil subpopulations. The distribution center position, percentile width, and coefficient of variation of each subpopulation based on the FSC / SSC / SFL dimensions are extracted and integrated to obtain a feature quantization database. The feature quantization database is input into a self-organizing map neural network (SOM), which uses a dynamic neighborhood radius decay mechanism to map high-dimensional features into discrete risk indices, and outputs the risk level corresponding to each risk index. The AUC value of the SOM is iterated based on a preset sample set and a preset validation set. When the AUC value exceeds a preset percentage threshold, the risk index that contributes the most to the AUC value is obtained from the feature quantization database as a key parameter. All key parameters of external routine blood samples are extracted and input into the SOM, and the risk index and corresponding risk level of the external routine blood samples are obtained based on the output of the SOM. This application is the first to apply the topology-preserving property of SOM (Systematic Origin of Metrics) to data processing. Through low-dimensional discretization mapping of high-dimensional blood routine data, it directly correlates spatial location with cancer risk. Furthermore, it proposes a neighborhood update strategy based on Gaussian decay to improve the model's tolerance to instrument fluctuations and individual variability. Finally, based on the unsupervised SOM layer, it introduces a TRI (Treatment of Trials) supervised scoring mechanism to balance disease pattern exploration and clinical decision support. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the steps of an embodiment of the blood routine data processing device of this application.

[0062] Figure 2 This is a Mon_SSC_CenterPos single-index AUC curve after model training is completed in one embodiment of the blood routine data processing device of this application.

[0063] Figure 3 This is a box plot distribution of Mon_SSC_CenterPos from one embodiment of the blood routine data processing device of this application.

[0064] Figure 4 This is a box plot distribution of Lym_SSC_Width from one embodiment of the blood routine data processing device of this application.

[0065] Figure 5 This is a value map of all indicators for one embodiment of the blood routine data processing device of this application.

[0066] Figure 6This is a diagram illustrating the model training process of an embodiment of the blood routine data processing device of this application.

[0067] Figure 7 This is a diagram showing the model of an embodiment of the blood routine data processing device of this application in a 10x10 low-dimensional space.

[0068] Figure 8 This is a capability evaluation diagram of a model of a blood routine data processing device according to an embodiment of this application on a preset validation set.

[0069] Figure 9 This is a schematic diagram of the functional modules of an embodiment of the blood routine data processing device of this application;

[0070] Figure 10 This is a schematic diagram of the functional modules of the three-dimensional light scattering dataset acquisition module 1 in an embodiment of the blood routine data processing device of this application;

[0071] Figure 11 This is a schematic diagram of the functional modules of the three-dimensional light scattering dataset clustering module 2 in one embodiment of the blood routine data processing device of this application;

[0072] Figure 12 This is a schematic diagram of the function of the feature quantization database acquisition module 3 in one embodiment of the blood routine data processing device of this application;

[0073] Figure 13 This is a schematic diagram of the function of the risk index mapping module 4 in one embodiment of the blood routine data processing device of this application;

[0074] Figure 14 This is a functional block diagram of the key parameter screening module 5 in one embodiment of the blood routine data processing device of this application;

[0075] Figure 15 This is a schematic diagram of the functional modules of the blood routine sample detection module 6 in one embodiment of the blood routine data processing device of this application;

[0076] Figure 16 This is a schematic diagram of the structure of an embodiment of the electronic device of this application;

[0077] Figure 17 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0078] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0079] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0080] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same instance, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0081] like Figure 1 As shown, this embodiment provides an example of a data processing device based on routine blood tests. In this embodiment, the data processing device is applied to routine blood test samples.

[0082] Preferably, 2 mL of peripheral venous blood is collected for routine blood tests and treated with EDTA anticoagulation. The samples are stored at room temperature and the tests are completed within 4 hours after collection to ensure the accuracy of the results.

[0083] The data processing device includes the following steps:

[0084] Step S1: Several three-dimensional light scattering signals of blood routine samples are obtained based on the FSC / SSC / SFL dimensions using the three-laser flow cytometry method, and all three-dimensional light scattering signals are logarithmically transformed to obtain a three-dimensional light scattering dataset.

[0085] Preferably, a complete blood count analyzer equipped with three-laser flow cytometry can be used. The instrument must be capable of simultaneously detecting three dimensions: FSC (forward scattered light), SSC (side scattered light), and SFL (side fluorescence). Specifically, FSC reflects cell size, with larger cells producing a stronger FSC signal; SSC reflects intracellular complexity (e.g., granules, nuclear lobes); and SFL reflects the content of intracellular nucleic acid / granular material (e.g., peroxidase).

[0086] Preferably, the blood routine sample is automatically mixed, diluted, and injected into the flow chamber by the instrument to begin the testing process.

[0087] Preferably, the purpose of logarithmic transformation is to eliminate the dynamic range deviation of instrument detection and to linearize signals of different intensities.

[0088] Preferably, the pseudocode for step S1 is as follows:

[0089] #Instrument Raw Signal Acquisition

[0090] raw_signals=instrument.capture sample(

[0091] sample_id=PATIENT 001

[0092] cell count=20000,

[0093] Parameters=["FSC","SSC","SFL"] )

[0095] #Logarithmic transformation

[0096] processed signals=np.log10(raw_signals+1)

[0097] #Optical Path Calibration and Verification

[0098] calibration_status=verify_calibration(

[0099] reference microspheres="Flow Check Pro",

[0100] max_cv=0.03 )

[0102] #Dataset Construction

[0103] dataset={

[0104] metadata":{

[0105] "instrument":"Instrument model not displayed",

[0106] "sample_id":"PATIENT_001",

[0107] "acquisition time":"aaaa-bb-ccTxx:yy:zzZ"

[0108] },

[0109] "data":processed signals,

[0110] "quality metrics":{

[0111] "FSC_CV":calculate_cv(processed_signals[:,0])

[0112] "SSC CV":calculate_cv(processed_signals[:,1])

[0113] "SFL_CV":calculate_cv(processed signals[:,2])

[0114] }

[0115] }

[0116] It is worth noting that the letters in "aaaa-bb-ccTxx:yy:zzZ" only represent time and their meanings are not interchangeable with those of the symbols in other formulas in this embodiment.

[0117] Step S2: Unsupervised classification of the three-dimensional light scattering dataset is performed using a Gaussian mixture model to obtain lymphocyte subsets, monocyte subsets, and neutrophil subsets.

[0118] Preferably, the three-dimensional light scattering dataset (N×3 matrix, where N is the number of cell events) can be preprocessed, with each row representing a cell event and the three columns representing the logarithmic transformation values ​​of FSC / SSC / SFL. Z-score standardization is performed on the data in each dimension to eliminate the dimensional differences between different dimensions and ensure that the clustering effect is not affected by the scale of the variables.

[0119] Preferably, the Gaussian Mixture Model (GMM) configuration details are as follows:

[0120] ① The number of Gaussian components (GMM_n_components) = 3 (corresponding to the three major subsets of lymphocytes, monocytes, and neutrophils).

[0121] ② Covariance type (covariance_type) = "full" (allows each component to have an independent covariance matrix).

[0122] ③Maximum number of iterations (max_iter) = 200 (to ensure that the model converges fully).

[0123] ④ Initialization strategy (init_params) = "kmeans" (use K-means for initial parameter estimation).

[0124] Preferably, the output format can be set to a labeled dataset (N×4 matrix, the first three columns are FSC / SSC / SFL values, and the fourth column is the subgroup label), with the following label encoding example: 0=lymphocyte, 1=monocyte, 2=neutrophil.

[0125] Step S3: Extract the distribution center position, percentile width, and coefficient of variation of each subgroup based on FSC / SSC / SFL dimensions, and integrate them to obtain a feature quantification database.

[0126] Preferably, based on the data format of the labeled dataset described above, the dataset can be split into three independent subsets according to the labels: lymphocyte set, monocyte set, and neutrophil set. The pseudocode is as follows:

[0127] lym_data = data[data[:,3] == 0, 0:3] # Lymphocyte data

[0128] mon_data = data[data[:,3] == 1, 0:3] # Mononuclear cell data

[0129] neu_data = data[data[:,3] == 2, 0:3] # Neutrophil data

[0130] Preferably, a structured DataFrame can be constructed using Python pandas.

[0131] The resulting database includes 27 parameters:

[0132] ① Nine lymphocyte parameters:

[0133] Lym_FSC_CenterPos - Lymphocyte forward scattering center.

[0134] Lym_FSC_Width - the width of the forward scattering distribution of lymphocytes.

[0135] Lym_FSC_CV - Lymphocyte forward scattering variation coefficient.

[0136] Lym_SSC_CenterPos - Lymphocyte lateral scattering center position.

[0137] Lym_SSC_Width - the width of the lateral scattering distribution of lymphocytes.

[0138] Lym_SSC_CV - Lymphocyte lateral scattering variation coefficient.

[0139] Lym_SFL_CenterPos - Lymphocyte lateral fluorescence center position.

[0140] Lym_SFL_Width - the width of the lateral fluorescence distribution of lymphocytes.

[0141] Lym_SFL_CV - Lymphocyte lateral fluorescence variation coefficient.

[0142] ② Nine monocyte parameters:

[0143] Mon_FSC_CenterPos - Monocyte forward scattering center.

[0144] Mon_FSC_Width - The width of the forward scattering distribution of monocytes.

[0145] Mon_FSC_CV - Monocyte forward scattering variation coefficient.

[0146] Mon_SSC_CenterPos - Monocyte lateral scattering center.

[0147] Mon_SSC_Width - The width of the lateral scattering distribution of mononuclear cells.

[0148] Mon_SSC_CV - Mononuclear cell lateral scattering variation coefficient.

[0149] Mon_SFL_CenterPos - Monocyte lateral fluorescence center.

[0150] Mon_SFL_Width - The width of the lateral fluorescence distribution in monocytes.

[0151] Mon_SFL_CV - Mononuclear cell lateral fluorescence variation coefficient.

[0152] ③ Nine neutrophil parameters:

[0153] Neutrophil forward scattering center position.

[0154] Neutrophil forward scattering width.

[0155] Neutrophil forward scattering variation coefficient.

[0156] Neu_SSC_CenterPos - Neutrophil lateral scattering center.

[0157] Neutrophil lateral scattering distribution width.

[0158] Neutrophil lateral scattering variation coefficient.

[0159] Neutrophil lateral fluorescence center position.

[0160] Neutrophil lateral fluorescence distribution width.

[0161] Neutrophil lateral fluorescence variation coefficient.

[0162] Preferably, the characterization and clinical significance of the above 27 parameters are shown in Table 1 below (a standardized table of the 27 parameters):

[0163]

[0164]

[0165]

[0166]

[0167] Table 1: Standardized table of 27 parameters.

[0168] Step S4: Input the feature quantization database into the self-organizing map neural network, and map the high-dimensional features into discrete risk indices through the dynamic neighborhood radius decay mechanism, and output the risk level corresponding to each risk index.

[0169] Preferably, the 27 feature parameters summarized above are Z-score standardized, retaining the mean and standard deviation of the preset sample set for subsequent validation set standardization. The pseudocode is as follows:

[0170] from sklearn.preprocessing import StandardScaler

[0171] scaler = StandardScaler()

[0172] train_features = scaler.fit_transform(train_data)

[0173] test_features = scaler.transform(test_data)

[0174] Preferably, a Dropout mechanism (probability = 0.2) can be introduced during the training process to randomly mask some input parameters and simulate the noise environment of the instrument.

[0175] In summary, the self-organizing map neural network chosen in this embodiment has the following advantages:

[0176] ① Advantages of dynamic neighborhood:

[0177] Initial large radius: quickly establish the global topology.

[0178] Small radius in the later stage: fine-tuning the representation of local features.

[0179] Avoid underfitting or overfitting caused by traditional fixed radius.

[0180] ② Hexagonal topology selection:

[0181] It has 12% more neighboring neurons than a rectangular grid.

[0182] Smoother boundary transitions.

[0183] Risk gradients are easier to identify during clinical visualization.

[0184] ③ Dropout noise simulation:

[0185] During training, 20% of the input parameters are randomly masked.

[0186] Enhance the model's robustness to instrument fluctuations and individual differences.

[0187] In practical applications, AUC is improved by approximately 3.2%.

[0188] ④ Real-time processing optimization:

[0189] It uses a C++ core computing module.

[0190] The inference time for a single sample is less than 200ms.

[0191] Supports multi-threaded parallel processing.

[0192] Step S5: Iterate the AUC value of the self-organizing map neural network based on the preset sample set and preset validation set, and when the AUC value exceeds the preset percentage threshold, obtain the risk index that contributes the most to the AUC value as the key parameter based on the feature quantization database.

[0193] Step S6: Extract all key parameters of the external blood routine sample and input them into the self-organizing map neural network. Based on the output of the self-organizing map neural network, obtain the risk index and corresponding risk level of the external blood routine sample.

[0194] Preferably, extracting all key parameters from external blood routine samples requires re-executing steps S1 to S4 using the external blood routine samples as the execution subject. However, standardized parameters of a preset sample set, such as the mean and standard deviation, can be used to perform Z-score standardization on the external blood routine samples. The pseudocode is as follows:

[0195] # Load the scaler object saved during the training phase

[0196] import joblib

[0197] scaler = joblib.load('trained_scaler.pkl')

[0198] # Standardize the new sample

[0199] new_sample_scaled = scaler.transform([new_sample_values])

[0200] Next, load the trained SOM model file (containing the weight matrix, healthy cluster center coordinates, etc.) and use ONNX Runtime to accelerate the inference process, so that the processing time for a single sample is kept within 200ms. The pseudocode is as follows:

[0201] import onnxruntime as rt

[0202] sess = rt.InferenceSession("cancer_som_model.onnx")

[0203] input_name = sess.get_inputs()[0].name

[0204] output_name = sess.get_outputs()[0].name

[0205] pred = sess.run([output_name], {input_name: new_sample_scaled. astype(np.float32)})

[0206] In summary, this embodiment completed clinical validation using a pre-set sample set of 5,000 cases (1,593 healthy individuals and 3,407 cases of pan-cancer cancer), covering 14 types of solid tumors and hematological malignancies (lung cancer, breast cancer, colorectal cancer, etc.) and an independent pre-set validation set of 1,583 cases (1,220 healthy individuals and 361 cases of pan-cancer cancer), ensuring that the geographical, age, and gender distribution did not overlap with the pre-set validation set.

[0207] See Figure 2 In this embodiment, after the model training is completed, the single index AUC of Mon_SSC_CenterPos is 0.84 (95% CI: 0.78-0.89), which is significantly higher than CEA (AUC=0.66).

[0208] See Figure 3 , Figure 3 The box plot shows the distribution of Mon_SSC_CenterPos under different diseases or health conditions (Control=control, N=benign, L=lung cancer, BC=breast cancer, O=esophageal cancer, C=colon cancer, H=liver cancer, G=stomach cancer).

[0209] See Figure 4 , Figure 4 The box plot shows the distribution of Lym_SSC_Width under different diseases or health conditions (Control=control, N=benign, L=lung cancer, BC=breast cancer, O=esophageal cancer, C=colon cancer, H=liver cancer, G=stomach cancer).

[0210] See Figure 5 , Figure 5 This is a value map for all indicators, reflecting the key parameters obtained in this embodiment.

[0211] See Figure 6 and Figure 7 , Figure 6 and Figure 7 The images show the model training process and the model's display in a 10x10 low-dimensional space, respectively.

[0212] See Figure 8 , Figure 8 This is used to evaluate the model's capabilities on an independent validation set (preset validation set).

[0213] Further, in step S1, several three-dimensional light scattering signals of the blood routine sample are obtained based on the FSC / SSC / SFL dimensions using a three-laser flow cytometry method, and all three-dimensional light scattering signals are logarithmically transformed to obtain a three-dimensional light scattering dataset. Specifically, this includes the following steps:

[0214] Step S11: The detection optical path of the three-laser flow cytometry method is calibrated using standard microspheres.

[0215] Preferably, optical path calibration can be performed before the three-laser flow cytometry detection method is started. The optical path is calibrated using standard microspheres FlowCheck Pro to ensure that the light intensity stability error is less than 3%.

[0216] Step S12: Obtain several three-dimensional light scattering signals of the blood routine sample based on the FSC / SSC / SFL dimensions through the calibrated detection optical path.

[0217] Preferably, the quality control measures for the three-laser flow cytometry method are as follows:

[0218] ① Each batch of tests includes a quality control sample (mixed plasma from healthy individuals).

[0219] ② Monitor the CV value (coefficient of variation) of the FSC / SSC / SFL signals.

[0220] ③ Set the trigger threshold for signal acquisition to eliminate fragmentation and noise interference.

[0221] Step S13: Remove extreme values ​​exceeding ±5SD in any dimension of FSC / SSC / SFL from all three-dimensional light scattering signals.

[0222] Step S14: Interpolate the missing values ​​of all three-dimensional light scattering signals to obtain the corrected three-dimensional light scattering signals.

[0223] Step S15: For each corrected three-dimensional light scattering signal, the signal value is... conduct The conversion yields a three-dimensional light scattering data based on a corrected three-dimensional light scattering signal.

[0224] Step S16: Integrate all three-dimensional light scattering data to obtain a three-dimensional light scattering dataset.

[0225] Preferably, the specific construction method of the three-dimensional light scattering dataset is as follows:

[0226] ① Data acquisition parameters:

[0227] Each sample yields 10,000–50,000 valid cellular events.

[0228] Sampling rate: 1,000-5,000 events / second.

[0229] Signal resolution: 16-bit ADC (0-65,535 channels).

[0230] ② Data format specifications:

[0231] Each cell event is recorded as a (FSC, SSC, SFL) triplet.

[0232] Dataset structure: N×3 matrix (N is the number of cell events)

[0233] Data is standardized and saved in CSV or HDF5 format.

[0234] ③ Outlier handling:

[0235] Remove extreme values ​​exceeding ±5SD for any of the FSC / SSC / SFL dimensions.

[0236] Use the KNN algorithm to fill in missing values ​​(k=5 neighborhood interpolation).

[0237] Further, in step S2, unsupervised classification of the three-dimensional light scattering dataset is performed using a Gaussian mixture model to obtain lymphocyte subsets, monocyte subsets, and neutrophil subsets, including:

[0238] Step S21: Convert the three-dimensional light scattering dataset into an N×3 matrix, where N is the number of cell events, each row represents one cell event, and the three columns represent the three-dimensional light scattering data of FSC / SSC / SFL, respectively.

[0239] Step S22: Define the Gaussian component number, covariance type, maximum number of iterations, and initialization strategy of the Gaussian mixture model based on the first preset strategy.

[0240] Step S23: Input the N×3 matrix into the Gaussian mixture model and call the GaussianMixture module of scikit-learn to train the Gaussian mixture model several times.

[0241] Preferably, the pseudocode for model training is as follows:

[0242] from sklearn.mixture import GaussianMixture

[0243] gmm=GaussianMixture(n_components=3, covariance_type='full', max_iter=200)

[0244] gmm.fit(standardized_data)

[0245] Wherein, standardized_data is the standardized 3D data.

[0246] Step S24: Obtain contour coefficients once through the GaussianMixture module based on each training iteration.

[0247] Preferably, the clustering effect is evaluated by calculating the silhouette score.

[0248] Preferably, the pseudocode for obtaining the contour coefficients is as follows:

[0249] from sklearn.metrics import silhouette_score

[0250] labels = gmm.predict(standardized_data)

[0251] score = silhouette_score(standardized_data, labels)

[0252] Step S25: Based on each training iteration, determine whether the contour coefficient of the previous iteration is greater than a preset threshold. If so, proceed to step S26.

[0253] Preferably, a contour coefficient greater than 0.7 indicates good cell subpopulation separation, i.e., the preset threshold is 0.7.

[0254] Step S26: Determine that the Gaussian mixture model has completed unsupervised classification, and label lymphocyte subsets, monocyte subsets, and neutrophil subsets based on the unsupervised classification results.

[0255] Preferably, subpopulation identification can label three cell populations based on clustering results:

[0256] Lymphocytes (Lym): FSC moderate, SSC lowest, SFL moderate.

[0257] Monocytes (Mon): FSC is relatively high, SSC is moderate, and SFL is relatively high.

[0258] Neutrophils (Neu): FSC is the highest, SSC is the highest, and SFL is the lowest.

[0259] Preferably, subgroup characteristic verification can be performed by calculating the three-dimensional center coordinates of each cluster and comparing them with known light scattering characteristics of blood cells:

[0260] Lymphocytes: It was confirmed that the FSC (volume) was significantly smaller than that of neutrophils.

[0261] Monocytes: SFL (fluorescence) was confirmed to be significantly higher than the other two types.

[0262] Further, in step S3, the distribution center position, percentile width, and coefficient of variation of each subgroup based on the FSC / SSC / SFL dimensions are extracted and integrated to obtain a feature quantification database, which specifically includes the following steps:

[0263] Step S31: Based on the classification results of lymphocyte subsets, monocyte subsets, and neutrophil subsets, the three-dimensional light scattering dataset is split into mutually independent sets of lymphocytes, monocytes, and neutrophils.

[0264] Step S32: Obtain the center position of each dimension of each subgroup using the median, and define the center position of one dimension of a subgroup as a distribution center.

[0265] Preferably, the median is used instead of the mean to calculate the center position to avoid interference from outliers. The median is calculated separately for each dimension (FSC / SSC / SFL) of each subpopulation. This embodiment provides a pseudocode example for lymphocytes:

[0266] lym_fsc_center = np.median(lym_data[:,0]) # Center position of FSC dimension

[0267] lym_ssc_center = np.median(lym_data[:,1]) # Center position of SSC dimension

[0268] lym_sfl_center = np.median(lym_data[:,2]) # Center of SFL dimension

[0269] Step S33: Sort the values ​​of each dimension of each subgroup from smallest to largest, and select the 10th percentile to 90th percentile of the sorted dimension to form the span interval as the percentile width of the current dimension.

[0270] Preferably, the pseudocode for the percentile width is as follows:

[0271] def calc_width(values):

[0272] p10 = np.percentile(values, 10)

[0273] p90 = np.percentile(values, 90)

[0274] Return to page 90 - page 10

[0275] lym_fsc_width = calc_width(lym_data[:,0])

[0276] Step S34: Obtain the standard deviation and mean of the current dimension, and obtain the coefficient of variation of the current dimension.

[0277] Preferably, the pseudocode for the coefficient of variation is as follows:

[0278] def calc_cv(values):

[0279] std = np.std(values)

[0280] mean = np.mean(values) + 1e-6

[0281] return (std / mean) * 100

[0282] lym_fsc_cv = calc_cv(lym_data[:,0])

[0283] The addition of 1e-6 is to prevent division by zero errors.

[0284] Step S35: Integrate the distribution center position, percentile width, and coefficient of variation of all dimensions of all subgroups to obtain the feature quantification database.

[0285] Further, in step S4, the feature quantization database is input into the self-organizing map neural network, and the high-dimensional features are mapped into discretized risk indices through a dynamic neighborhood radius decay mechanism. The risk level corresponding to each risk index is then output, specifically including the following steps:

[0286] Step S41: Define the neuron weights, grid size, topology type, weight initialization method, initial neighborhood radius, decay strategy, and learning rate of the self-organizing map neural network based on the second preset strategy.

[0287] Preferably, the self-organizing map neural network structure is configured as follows:

[0288] Grid size: 10×10 two-dimensional competition layer (corresponding to 100 neurons).

[0289] Topology type: Hexagonal topology, which preserves topological relationships better than rectangular mesh.

[0290] Weight initialization: PCA initialization method is used to accelerate convergence.

[0291] Initial neighborhood radius: 7 (covering approximately 50% of neurons).

[0292] Decay strategy: Gaussian function decays exponentially, and the radius is reduced to 1 at the end of training.

[0293] Learning rate: Initially 0.5, then exponentially decays to 0.01 based on the number of epochs.

[0294] Step S42: Input the feature quantization database into the self-organizing map neural network to calculate the Euclidean distance between each input data and the weights of all neurons, and define the neuron with the smallest Euclidean distance to the current input data as the best matching unit.

[0295] Preferably, for each input sample, the Euclidean distance to the weights of all neurons is calculated, and the neuron with the smallest distance is selected as the best matching unit (BMU). The pseudocode is as follows:

[0296] def neighborhood_radius(epoch, max_epochs):

[0297] initial_radius = 7

[0298] final_radius = 1

[0299] return initial_radius *

[0300] (final_radius / initial_radius)**(epoch / max_epochs)

[0301] Preferably, a U-Matrix (unified distance matrix) can be calculated to visualize cluster boundaries, monitor quantization error (QE) and topology error (TE) to ensure training stability, and the termination condition is defined as the cluster structure similarity being greater than 0.98 for 20 consecutive iterations.

[0302] Step S43: Dynamically decay the neighborhood radius of each best-matching unit using a Gaussian function.

[0303] Step S44: After the feature quantization database is input, mark the best matching unit position for all healthy samples.

[0304] Step S45: Calculate the centroid of the best matching unit position for all healthy samples and mark it as the center of the healthy cluster.

[0305] Step S46: Obtain the maximum Euclidean distance between all input data and the center of the health cluster, and the ratio of the Euclidean distance of each input data to the maximum Euclidean distance, and define the ratio of each input data as a risk index.

[0306] Preferably, the risk index TRI = 100 × (D_current / D_max), where D_current is the Euclidean distance from the current sample BMU to the center of the healthy cluster, and D_max is the maximum Euclidean distance (normalized baseline) in the preset sample set.

[0307] Step S47: Use 0, 30, 60, and 100 as the semi-open interval boundaries of the low-risk interval, medium-risk interval, and high-risk interval, respectively.

[0308] Preferably, the risk levels are classified as follows:

[0309] Low risk: TRI < 30 (healthy reference range).

[0310] Medium risk: 30≤TRI<60 (follow-up observation required).

[0311] High risk: TRI≥60 (clinical intervention recommended).

[0312] Step S48: Obtain the risk range matched for each risk index, and output the risk level corresponding to each risk index.

[0313] Further, in step S5, the AUC value of the self-organizing map neural network is iterated based on a preset sample set and a preset validation set. When the AUC value exceeds a preset percentage threshold, the risk index that contributes the most to the AUC value is obtained as a key parameter based on the feature quantization database, including:

[0314] Step S51: Divide the preset sample set into several sub-sample sets.

[0315] Preferably, using the 5000-example preset sample set of this embodiment, 5-epoch cross-validation can be used, with each epoch containing 1000 examples. After each round of validation, the network weights are reinitialized to avoid local optima.

[0316] Step S52: Input each fold sample set into the self-organizing map neural network, and calculate the current AUC value once based on training one fold.

[0317] Preferably, this embodiment uses a pre-set sample set of 5,000 cases (1,593 healthy cases and 3,407 cases of pan-cancer), covering 14 types of solid tumors and hematological malignancies (lung cancer, breast cancer, colorectal cancer, etc.), and an independent pre-set validation set of 1,583 cases (1,220 healthy cases and 361 cases of pan-cancer). The model index is the area under the ROC curve (AUC), and the pre-set threshold is 85%, i.e., 0.85, as the minimum clinically acceptable standard.

[0318] Preferably, the auxiliary indicators can be set as a high-risk cutoff value TRI≥60, specificity (TRI<30), and Youden index (sensitivity + specificity -1).

[0319] Step S53: Reset the network weights once for each training fold.

[0320] Step S54: When all subsample sets have been input, or the AUC value exceeds the preset percentage threshold, the training of the self-organizing map neural network is complete.

[0321] Preferably, the current AUC value is calculated after each fold of training is completed, and the quantization error (QE) is recorded to reflect the model fit, the topology error (TE) is recorded to assess the topology preservation ability, and the AUC change trend of the validation set is recorded.

[0322] Step S55: Input the preset validation set into the trained self-organizing map neural network to obtain the initial validation AUC value.

[0323] Step S56: Randomly shuffle one of the values ​​in the preset validation set and obtain the decrease in the initial validation AUC value.

[0324] Preferably, one of the values ​​in the preset validation set is randomly shuffled, the decrease in the initial validation AUC value is obtained, and the decrease in AUC is observed. The greater the decrease, the higher the importance of the parameter.

[0325] Preferably, this can be achieved using the following pseudocode:

[0326] base_auc <- calculate_auc(model, validation_set)

[0327] for (param in all_params) {

[0328] shuffled_data <- shuffle_column(validation_set, param)

[0329] new_auc <- calculate_auc(model, shuffled_data)

[0330] importance[param] <- base_auc - new_auc

[0331] }

[0332] Step S57: Obtain the Spearman rank correlation coefficients between the feature quantification database and each risk index.

[0333] Preferably, the correlation coefficients between 27 parameters and TRI can be calculated, and parameters with |ρ|>0.3 are retained.

[0334] Step S58: When the decrease in the risk index is greater than a preset percentage threshold and the Spearman rank correlation coefficient is greater than a preset coefficient threshold, it is defined as a key parameter.

[0335] That is, simultaneously satisfying:

[0336] ① The importance of substitution leads to a decrease in AUC of ≥0.02.

[0337] ② The rank correlation coefficient |ρ| > 0.3.

[0338] Preferably, if the initial AUC < 0.8: expand the range of key parameters (|ρ| > 0.2); if the AUC > 0.9: tighten the criteria (|ρ| > 0.35), ultimately locking in 5 to 8 key parameters, such as Mon_SSC_CenterPos (monocyte lateral scattering center position, reflecting cell complexity) and Lym_SSC_Width (lymphocyte lateral scattering distribution width, indicating heterogeneity). See the following pseudocode for details:

[0339] # Model Training and Evaluation Cycle

[0340] best_auc <- 0

[0341] for (i in 1:50) {

[0342] model <- train_som(train_data, epochs=10)

[0343] current_auc <- evaluate_auc(model, valid_data)

[0344] if (current_auc > 0.85 & current_auc > best_auc) {

[0345] best_auc <- current_auc

[0346] important_params <- get_important_params(model)

[0347] save_model(model, "best_model.rds")

[0348] }

[0349] }

[0350] # Application of key parameters

[0351] final_model <- load_model("best_model.rds")

[0352] key_params <- c("Mon_SSC_CenterPos", "Lym_SSC_Width",

[0353] "Neu_SFL_CV", "Lym_FSC_CV",

[0354] "Mon_SFL_Width", "Neu_SSC_CenterPos",

[0355] "Lym_SFL_CenterPos")

[0356] Further, in step S6, all key parameters of the external blood routine sample are extracted and input into a self-organizing map neural network. Based on the output of the self-organizing map neural network, the risk index and corresponding risk level of the external blood routine sample are obtained. This specifically includes the following steps:

[0357] Step S61: The external blood routine sample is processed through steps S1 to S3 to obtain a feature quantification database.

[0358] Step S62: Retain the data in the feature quantization database that are the same as all key parameter categories and input them into the self-organizing map neural network.

[0359] For example, the 5 to 8 key parameters mentioned above.

[0360] Step S63: Obtain the Euclidean distance between each data point and the center of the healthy cluster, as well as the ratio of the Euclidean distance of each data point to the maximum Euclidean distance, and define the ratio of each data point as a risk index of the data.

[0361] Step S64: Obtain the risk range matched by the risk index of each data point, and output the risk level corresponding to the risk index of each data point.

[0362] Preferably, since the model has already been trained, it is only necessary to obtain the three-dimensional scattered light data of the blood routine sample and then select the key parameters to substitute into the model. The pseudocode is as follows:

[0363] def evaluate new_sample(sample data):

[0364] #1. Data Preprocessing

[0365] scaled_data=trained_scaler.transform([sample_data])

[0366] #2. SOM Model Inference

[0367] distances=np.linalg.norm(som_weights-scaled_data, axis=1)

[0368] bmu_index=np.argmin(distances)

[0369] bmu_coords=(bmu_index / / 10,bmu_index%10)

[0370] #3. TRI Calculation

[0371] D_current=np.linalg.norm(bmu_coords-health_center)

[0372] TRI=min(100,100*D_current / D_max)

[0373] #4. Risk Classification

[0374] if TRI<30:

[0375] risk_level="Low"

[0376] elif TRI<60:

[0377] risk_level="Medium"

[0378] else:

[0379] risk_level="High"

[0380] return {

[0381] "TRI":round(TRI,1),

[0382] "RiskLevel":risk_level,

[0383] "BMU Coord":bmu_coords,

[0384] "AbnormalParams":get_abnormal_params(scaled data)

[0385] }

[0386] # Helper function: Get abnormal parameters

[0387] Def get_abnormal_params(sample):

[0388] abnormal= {}

[0389] for i,param in enumerate(KEY PARAMS):

[0390] z_score=sample[0][i]

[0391] if abs(z_score)>2: # Exceeds 2 standard deviations

[0392] Abnormal[param] = {

[0393] "Value":sample[0][i],

[0394] "RefRange": f"{MEANS[i]-2*STDS[i]:.1f}-{MEANS[i]+2*STDS[i]:.1f}"

[0395] }

[0396] return abnormal

[0397] As can be seen from the above, new samples no longer need to go through steps S1 to S6 completely once, which makes the speed of processing new samples significantly increased after the model training is completed.

[0398] This embodiment uses a three-laser flow cytometry method to acquire several three-dimensional light scattering signals from a routine blood sample based on the FSC / SSC / SFL dimensions. All three-dimensional light scattering signals are logarithmically transformed to obtain a three-dimensional light scattering dataset. An unsupervised classification of the three-dimensional light scattering dataset is performed using a Gaussian mixture model to obtain lymphocyte, monocyte, and neutrophil subpopulations. The distribution center position, percentile width, and coefficient of variation for each subpopulation based on the FSC / SSC / SFL dimensions are extracted and integrated to obtain a feature quantization database. This feature quantization database is input into a self-organizing map neural network (SOM), which uses a dynamic neighborhood radius decay mechanism to map high-dimensional features into discrete risk indices and outputs the risk level corresponding to each risk index. The AUC value of the SOM is iterated based on a preset sample set and a preset validation set. When the AUC value exceeds a preset percentage threshold, the risk index that contributes the most to the AUC value is obtained from the feature quantization database as a key parameter. All key parameters of an external routine blood sample are extracted and input into the SOM, and the risk index and corresponding risk level of the external routine blood sample are obtained based on the output of the SOM. This embodiment is the first to apply the topology-preserving property of SOM to data processing. By mapping the low-dimensional discretization of high-dimensional blood routine data, it directly correlates spatial location with cancer risk. At the same time, it proposes a neighborhood update strategy based on Gaussian decay to improve the model's tolerance to instrument fluctuations and individual variations. Furthermore, based on the unsupervised SOM layer, a TRI supervised scoring mechanism is introduced to take into account both disease pattern exploration and clinical decision support.

[0399] Based on the above embodiments, see Figure 9 This embodiment provides an example of a data processing device based on blood routine tests. In this embodiment, the data processing device is applied to the data processing device as described in the above embodiment.

[0400] Specifically, the data processing device includes a three-dimensional light scattering dataset acquisition module 1, a three-dimensional light scattering dataset clustering module 2, a feature quantization database acquisition module 3, a risk index mapping module 4, a key parameter screening module 5, and a blood routine sample detection module 6, which are electrically or signalally connected in sequence.

[0401] The three-dimensional light scattering dataset acquisition module 1 is used to acquire several three-dimensional light scattering signals of blood routine samples based on FSC / SSC / SFL dimensions using a three-laser flow cytometry method, and performs logarithmic transformation on all three-dimensional light scattering signals to obtain a three-dimensional light scattering dataset; the three-dimensional light scattering dataset clustering module 2 is used to perform unsupervised classification of the three-dimensional light scattering dataset using a Gaussian mixture model to obtain lymphocyte subsets, monocyte subsets, and neutrophil subsets; the feature quantification database acquisition module 3 is used to extract the distribution center position, percentile width, and coefficient of variation of each subset based on FSC / SSC / SFL dimensions, and integrate them to obtain a feature quantification database; the risk index mapping module 4 uses... The feature quantization database is input into the self-organizing map neural network, and the high-dimensional features are mapped into discrete risk indices through a dynamic neighborhood radius decay mechanism. The risk level corresponding to each risk index is output. The key parameter screening module 5 is used to iterate the AUC value of the self-organizing map neural network based on a preset sample set and a preset validation set. When the AUC value exceeds a preset percentage threshold, the risk index that contributes the most to the AUC value is obtained as the key parameter based on the feature quantization database. The blood routine sample detection module 6 extracts all key parameters of the external blood routine sample and inputs them into the self-organizing map neural network. The risk index and corresponding risk level of the external blood routine sample are obtained based on the output of the self-organizing map neural network.

[0402] Further, see Figure 10 The three-dimensional light scattering dataset acquisition module 1 specifically includes a standard microsphere calibration unit, a three-dimensional light scattering signal acquisition unit, an extreme value elimination unit, a missing value interpolation unit, a scattering signal logarithmic transformation unit, and a scattering dataset integration unit, which are electrically or signal-connected in sequence. The scattering dataset integration unit is electrically or signal-connected to the three-dimensional light scattering dataset clustering module 2.

[0403] The system includes: a standard microsphere calibration unit for calibrating the detection optical path of the three-laser flow cytometry method using standard microspheres; a three-dimensional light scattering signal acquisition unit for acquiring several three-dimensional light scattering signals of a blood routine sample based on the FSC / SSC / SFL dimensions using the calibrated detection optical path; an extreme value removal unit for removing extreme values ​​exceeding ±5SD in any dimension of FSC / SSC / SFL from all three-dimensional light scattering signals; a missing value interpolation unit for interpolating missing values ​​in all three-dimensional light scattering signals to obtain corrected three-dimensional light scattering signals; and a scattering signal logarithmic transformation unit for converting the signal value of each corrected three-dimensional light scattering signal. conduct The conversion is based on a corrected three-dimensional light scattering signal to obtain three-dimensional light scattering data; the scattering dataset integration unit is used to integrate all three-dimensional light scattering data to obtain a three-dimensional light scattering dataset.

[0404] Further, see Figure 11 The three-dimensional light scattering dataset clustering module 2 specifically includes an N×3 matrix transformation unit, a Gaussian mixture model definition unit, a Gaussian mixture model training unit, a contour coefficient acquisition unit, a contour coefficient judgment unit, and an unsupervised classification result labeling unit, which are electrically or signal-connected in sequence. The N×3 matrix transformation unit is electrically or signal-connected to the scattering dataset integration unit, and the unsupervised classification result labeling unit is electrically or signal-connected to the feature quantization database acquisition module 3.

[0405] The system comprises the following components: an N×3 matrix transformation unit, which converts a 3D light scattering dataset into an N×3 matrix, where N represents the number of cell events, each row represents a cell event, and the three columns represent the 3D light scattering data for FSC / SSC / SFL, respectively; a Gaussian mixture model definition unit, which defines the number of Gaussian components, covariance type, maximum number of iterations, and initialization strategy of the Gaussian mixture model based on a first preset strategy; a Gaussian mixture model training unit, which inputs the N×3 matrix into the Gaussian mixture model and calls the GaussianMixture module of scikit-learn to train the Gaussian mixture model several times; a silhouette coefficient acquisition unit, which acquires silhouette coefficients once per training iteration using the GaussianMixture module; a silhouette coefficient judgment unit, which judges whether the silhouette coefficient of the current iteration is greater than a preset threshold per training iteration; and an unsupervised classification result labeling unit, which determines that the Gaussian mixture model has completed unsupervised classification if the result is positive, and labels lymphocyte subsets, monocyte subsets, and neutrophil subsets based on the unsupervised classification result.

[0406] Further, see Figure 12 The feature quantification database acquisition module 3 specifically includes a scattering dataset splitting unit, a distribution center location definition unit, a subgroup sorting unit, a coefficient of variation calculation unit, and a subgroup data integration unit that are electrically or signal-connected in sequence; the scattering dataset splitting unit is electrically or signal-connected to the unsupervised classification result labeling unit, and the subgroup data integration unit is electrically or signal-connected to the risk index mapping module 4.

[0407] The system comprises the following components: a scattering dataset splitting unit, which splits the three-dimensional light scattering dataset into independent sets of lymphocytes, monocytes, and neutrophils based on the classification results of lymphocyte, monocyte, and neutrophil subpopulations; a distribution center position definition unit, which obtains the center position of each dimension for each subpopulation using the median and defines the center position of one dimension for each subpopulation as a distribution center position; a subpopulation sorting unit, which sorts each dimension of each subpopulation from smallest to largest and selects the 10th to 90th percentiles of the sorted dimension as the percentile width of the current dimension; a coefficient of variation calculation unit, which obtains the standard deviation and mean of the current dimension and calculates the coefficient of variation of the current dimension; and a subpopulation data integration unit, which integrates the distribution center positions, percentile widths, and coefficients of variation of all dimensions of all subpopulations to obtain a feature quantification database.

[0408] Further, see Figure 13 The risk index mapping module 4 specifically includes a self-organizing mapping neural network definition unit, an optimal neuron definition unit, a neighborhood radius dynamic decay unit, an optimal matching position marking unit, a healthy cluster center marking unit, a risk index definition unit, a risk interval division unit, and a risk level output unit, which are electrically or signal-connected in sequence. The self-organizing mapping neural network definition unit is electrically or signal-connected to the subgroup data integration unit, and the risk level output unit is electrically or signal-connected to the key parameter screening module 5.

[0409] The self-organizing map neural network definition unit is used to define the neuron weights, grid size, topology type, weight initialization method, initial neighborhood radius, decay strategy, and learning rate of the self-organizing map neural network based on a second preset strategy. The optimal neuron definition unit is used to input the feature quantization database into the self-organizing map neural network to calculate the Euclidean distance between each input data and all neuron weights, and to define the neuron with the smallest Euclidean distance to the current input data as the optimal matching unit. The neighborhood radius dynamic decay unit is used to dynamically decay the neighborhood radius of each optimal matching unit using a Gaussian function. The optimal matching position marking unit is used to mark all healthy samples after the feature quantization database has been input. The optimal matching unit location is used; the healthy cluster center marking unit is used to calculate the centroid of the optimal matching unit location for all healthy samples and mark it as the healthy cluster center; the risk index definition unit is used to obtain the maximum Euclidean distance between all input data and the healthy cluster center, as well as the ratio of the Euclidean distance of each input data to the maximum Euclidean distance, and defines each input data ratio as a risk index; the risk interval division unit is used to define the half-open interval boundaries of the low-risk interval, medium-risk interval, and high-risk interval with 0, 30, 60, and 100 respectively; the risk level output unit is used to obtain the risk interval matched by each risk index and output the risk level corresponding to each risk index.

[0410] Further, see Figure 14 The key parameter screening module 5 specifically includes a subsample set division unit, a current AUC value calculation unit, a network weight reset unit, a self-organizing map neural network termination unit, an initial validation AUC value acquisition unit, an AUC value decrease rate acquisition unit, a Spearman rank correlation coefficient acquisition unit, and a key parameter definition unit, which are electrically or signal-connected in sequence. The subsample set division unit is electrically or signal-connected to the risk level output unit, and the key parameter definition unit is electrically or signal-connected to the blood routine sample detection module 6.

[0411] The system includes the following components: a subsample set division unit, which divides a preset sample set into several folded subsample sets; a current fold AUC value calculation unit, which inputs each folded subsample set into the self-organizing map neural network and calculates the current AUC value based on each fold; a network weight reset unit, which resets the network weights based on each fold; a self-organizing map neural network termination unit, which terminates the training of the self-organizing map neural network when all subsample sets have been input or the AUC value exceeds a preset percentage threshold; an initial validation AUC value acquisition unit, which inputs a preset validation set into the trained self-organizing map neural network to obtain the initial validation AUC value; an AUC value decrease rate acquisition unit, which randomly shuffles one of the values ​​in the preset validation set to obtain the decrease rate of the initial validation AUC value; a Spearman rank correlation coefficient acquisition unit, which obtains the Spearman rank correlation coefficients between the feature quantization database and each risk index; and a key parameter definition unit, which defines key parameters when the decrease rate of a risk index is greater than a preset percentage threshold and the Spearman rank correlation coefficient is greater than a preset coefficient threshold.

[0412] Further, see Figure 15 The blood routine sample detection module 6 specifically includes a feature quantification database acquisition unit, a feature quantification database screening unit, a healthy cluster center comparison unit, and a blood routine sample grade matching unit, which are electrically or signal-connected in sequence; the feature quantification database acquisition unit and the key parameter definition unit are electrically or signal-connected.

[0413] The feature quantization database acquisition unit processes external blood routine samples through the three-dimensional light scattering dataset acquisition module 1 to the feature quantization database acquisition module 3 to obtain a feature quantization database; the feature quantization database filtering unit retains data in the feature quantization database that are the same as all key parameter categories and inputs them into the self-organizing map neural network; the healthy cluster center comparison unit obtains the Euclidean distance between each data point and the healthy cluster center, as well as the ratio of the Euclidean distance of each data point to the maximum Euclidean distance, and defines the ratio of each data point as a risk index; the blood routine sample level matching unit obtains the risk interval matched by the risk index of each data point, and outputs the risk level corresponding to the risk index of each data point.

[0414] It should be noted that this embodiment is a functional module embodiment based on the above embodiments. For the preferred, extended, limited, exemplified and principle explanation parts of this embodiment, please refer to the above embodiments. This embodiment will not repeat them.

[0415] This embodiment uses a three-laser flow cytometry method to acquire several three-dimensional light scattering signals from a routine blood sample based on the FSC / SSC / SFL dimensions. All three-dimensional light scattering signals are logarithmically transformed to obtain a three-dimensional light scattering dataset. An unsupervised classification of the three-dimensional light scattering dataset is performed using a Gaussian mixture model to obtain lymphocyte, monocyte, and neutrophil subpopulations. The distribution center position, percentile width, and coefficient of variation for each subpopulation based on the FSC / SSC / SFL dimensions are extracted and integrated to obtain a feature quantization database. This feature quantization database is input into a self-organizing map neural network (SOM), which uses a dynamic neighborhood radius decay mechanism to map high-dimensional features into discrete risk indices and outputs the risk level corresponding to each risk index. The AUC value of the SOM is iterated based on a preset sample set and a preset validation set. When the AUC value exceeds a preset percentage threshold, the risk index that contributes the most to the AUC value is obtained from the feature quantization database as a key parameter. All key parameters of an external routine blood sample are extracted and input into the SOM, and the risk index and corresponding risk level of the external routine blood sample are obtained based on the output of the SOM. This embodiment is the first to apply the topology-preserving property of SOM to data processing. By mapping the low-dimensional discretization of high-dimensional blood routine data, it directly correlates spatial location with cancer risk. At the same time, it proposes a neighborhood update strategy based on Gaussian decay to improve the model's tolerance to instrument fluctuations and individual variations. Furthermore, based on the unsupervised SOM layer, a TRI supervised scoring mechanism is introduced to take into account both disease pattern exploration and clinical decision support.

[0416] Figure 16 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Figure 16 As shown, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71.

[0417] The memory 72 stores program instructions for implementing the blood routine-based data processing device of any of the above embodiments.

[0418] The processor 71 is used to execute program instructions stored in the memory 72 to perform data processing based on blood routine.

[0419] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also 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, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0420] Furthermore, Figure 17 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 17 The storage medium 8 in this embodiment stores program instructions 81 capable of implementing all the above-described methods or functional modules. These program instructions 81 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods or functional modules described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0421] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0422] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A data processing device based on routine blood tests, wherein the data processing device is applied to routine blood test samples, characterized in that, The data processing device includes: The three-dimensional light scattering dataset acquisition module is used to acquire several three-dimensional light scattering signals of the blood routine sample based on the FSC / SSC / SFL dimensions using the three-laser flow cytometry method, and to perform logarithmic transformation on all three-dimensional light scattering signals to obtain a three-dimensional light scattering dataset. A three-dimensional light scattering dataset clustering module is used to perform unsupervised classification of the three-dimensional light scattering dataset using a Gaussian mixture model to obtain lymphocyte subsets, monocyte subsets, and neutrophil subsets. The feature quantization database acquisition module is used to extract the distribution center position, percentile width, and coefficient of variation of each subgroup based on FSC / SSC / SFL dimensions, and integrate them to obtain the feature quantization database. The risk index mapping module is used to input the feature quantization database into the self-organizing map neural network, map the high-dimensional features into discrete risk indices through a dynamic neighborhood radius decay mechanism, and output the risk level corresponding to each risk index. The key parameter screening module is used to iterate the AUC value of the self-organizing map neural network based on a preset sample set and a preset validation set, and when the AUC value exceeds a preset percentage threshold, obtain the risk index that contributes the most to the AUC value as a key parameter based on the feature quantization database. The routine blood sample detection module extracts all key parameters of the external routine blood sample and inputs them into the self-organizing map neural network. Based on the output of the self-organizing map neural network, the risk index and corresponding risk level of the external routine blood sample are obtained.

2. The data processing apparatus according to claim 1, characterized in that, The 3D light scattering dataset acquisition module includes: The standard microsphere calibration unit is used to calibrate the detection optical path of the three-laser flow cytometry method using standard microspheres. The three-dimensional light scattering signal acquisition unit is used to acquire several three-dimensional light scattering signals of a blood routine sample based on the FSC / SSC / SFL dimensions through the calibrated detection optical path; The extreme value removal unit is used to remove extreme values ​​exceeding ±5SD in any dimension of FSC / SSC / SFL from all three-dimensional light scattering signals. The missing value interpolation unit is used to interpolate the missing values ​​of all three-dimensional light scattering signals to obtain the corrected three-dimensional light scattering signals. The scattering signal logarithmic transformation unit is used to convert the signal value of each corrected three-dimensional light scattering signal. conduct The conversion yields a three-dimensional light scattering data based on a corrected three-dimensional light scattering signal; The scattering dataset integration unit is used to integrate all three-dimensional light scattering data to obtain a three-dimensional light scattering dataset.

3. The data processing apparatus according to claim 1, characterized in that, The 3D light scattering dataset clustering module includes: The N×3 matrix transformation unit is used to convert a three-dimensional light scattering dataset into an N×3 matrix, where N is the number of cell events, each row represents one cell event, and the three columns represent the three-dimensional light scattering data of FSC / SSC / SFL, respectively. The Gaussian mixture model definition unit is used to define the number of Gaussian components, covariance type, maximum number of iterations, and initialization strategy of the Gaussian mixture model based on a first preset strategy. The Gaussian Mixture Model Training Unit is used to input an N×3 matrix into the Gaussian Mixture Model and call the GaussianMixture module of scikit-learn to train the Gaussian Mixture Model several times. The contour coefficient acquisition unit is used to acquire contour coefficients once per training iteration via the GaussianMixture module. The contour coefficient judgment unit is used to determine whether the contour coefficient of the current training iteration is greater than a preset threshold based on each training iteration. The unsupervised classification result labeling unit is used to determine that the Gaussian mixture model has completed unsupervised classification if the result is positive. Based on the unsupervised classification result, the lymphocyte subset, monocyte subset, and neutrophil subset are labeled.

4. The data processing apparatus according to claim 3, characterized in that, The feature quantization database acquisition module includes: The scattering dataset splitting unit is used to split the three-dimensional light scattering dataset into independent sets of lymphocytes, monocytes, and neutrophils based on the classification results of lymphocyte subsets, monocyte subsets, and neutrophil subsets. The distribution center position definition unit is used to obtain the center position of each subgroup in each dimension by means of the median, and the center position of a subgroup in one dimension is defined as a distribution center position. The subgroup sorting unit is used to sort the values ​​of each dimension of each subgroup from smallest to largest, and select the 10th percentile to 90th percentile of the sorted dimension to form a span interval as the percentile width of the current dimension. The coefficient of variation calculation unit is used to obtain the standard deviation and mean of the current dimension, and to obtain the coefficient of variation of the current dimension. The subgroup data integration unit is used to integrate the distribution center position, percentile width, and coefficient of variation of all dimensions of all subgroups to obtain a feature quantification database.

5. The data processing apparatus according to claim 1, characterized in that, The risk index mapping module includes: The self-organizing map neural network definition unit is used to define the neuron weights, grid size, topology type, weight initialization method, initial neighborhood radius, decay strategy, and learning rate of the self-organizing map neural network based on a second preset strategy. The optimal neuron definition unit is used to input the feature quantization database into the self-organizing map neural network to calculate the Euclidean distance between each input data and the weights of all neurons, and the neuron with the smallest Euclidean distance to the current input data is defined as the optimal matching unit. The neighborhood radius dynamic decay unit is used to dynamically decay the neighborhood radius of each best-matching unit using a Gaussian function. The best matching position marking unit is used to mark the best matching unit position of all healthy samples after the feature quantization database has been input. The healthy cluster center marker cell is used to calculate the centroid of the best matching cell position for all healthy samples and mark it as the healthy cluster center; The risk index definition unit is used to obtain the maximum Euclidean distance between all input data and the center of the health cluster, as well as the ratio of the Euclidean distance of each input data to the maximum Euclidean distance, and to define the ratio of each input data as a risk index. Risk interval division unit, used as the half-open interval boundary with 0, 30, 60, 100 as low risk interval, medium risk interval, and high risk interval respectively; The risk level output unit is used to obtain the risk range matched by each risk index, so as to output the risk level corresponding to each risk index.

6. The data processing apparatus according to claim 5, characterized in that, The key parameter filtering module includes: Subsample set division unit, used to divide a preset sample set into several subsample sets; The current fold AUC value calculation unit is used to input each fold subsample set into the self-organizing map neural network and calculate the current AUC value once based on training for one fold. The network weight reset unit is used to reset the network weights once for each training fold. The self-organizing map neural network termination unit is used to complete the training of the self-organizing map neural network when all subsample sets have been input or the AUC value exceeds a preset percentage threshold. The initial validation AUC value acquisition unit is used to input the preset validation set into the trained self-organizing map neural network to obtain the initial validation AUC value; The AUC value decrease rate acquisition unit is used to randomly shuffle one of the values ​​in the preset validation set and obtain the decrease rate of the initial validation AUC value. The Spearman rank correlation coefficient acquisition unit is used to obtain the Spearman rank correlation coefficients between the feature quantification database and each risk index. The key parameter definition unit is used to define key parameters when the decrease in the risk index is greater than a preset percentage threshold and the Spearman rank correlation coefficient is greater than a preset coefficient threshold.

7. The data processing apparatus according to claim 5, characterized in that, The routine blood sample testing module includes: The feature quantization database acquisition unit is used to process external blood routine samples through the three-dimensional light scattering dataset acquisition module to obtain the feature quantization database; The feature quantization database filtering unit is used to retain data in the feature quantization database that are the same as all key parameter categories and input them into the self-organizing map neural network. The health cluster center comparison unit is used to obtain the Euclidean distance between each data point and the health cluster center, as well as the ratio of the Euclidean distance of each data point to the maximum Euclidean distance, and to define the ratio of each data point as a risk index of the data. The blood routine sample grading matching unit is used to obtain the risk range matched by the risk index of each data point, so as to output the risk level corresponding to the risk index of each data point.

8. An electronic device, characterized in that, The device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements a data processing apparatus as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium stores program instructions, which, when executed by a processor, enable the data processing apparatus as described in any one of claims 1 to 7.

Images