Lung Disease Identification Tool and System Based on Salivary Glucose Chip Data

Abstract

This invention provides a lung disease identification tool and system based on salivary lectin chip data, comprising: a processor and a storage medium; the storage medium is used to execute the following steps when a stored program is loaded by the processor: using a lung lesion identification model to identify the lectin chip data of the saliva of a patient to be diagnosed, and determining whether the patient to be diagnosed is: a healthy volunteer, a patient with benign lung lesions, or a lung cancer patient. The lung disease identification tool of this invention, through a lung cancer diagnosis module, identifies the saliva of the patient to be diagnosed to determine whether the sample is a lung disease sample and confirms the type of lung cancer. The lung disease identification tool of this invention has the characteristics of convenient sampling and high sensitivity, is suitable for early and mid-stage screening of patients, and can quickly identify the lung lesions of subjects.

Description

Technical Field

[0001] This invention relates to the field of biocomputers, and in particular to a lung disease identification tool and system based on salivary lectin chip data. Background Technology

[0002] Lung cancer is a leading cause of death among cancer patients, accounting for nearly one-fifth of all cancer deaths. Existing saliva-based cancer screening primarily utilizes lectin chips. Lectins are non-immune, non-enzymatic sugar-binding proteins that specifically recognize and bind to specific glycan sequences in monosaccharides or glycans with particular structures. A lectin chip is a biochip made by immobilizing lectins from various sources onto an epoxidized chip substrate.

[0003] Current methods for identifying cancer using lectin chips primarily involve selecting a control group and comparing the lectin chip results of the clinical samples to be tested with those of the healthy control group. The Fold-change value of each lectin is obtained, using Fold-change > 1.5 and Fold-change < 0.67 as selection criteria. Fold-change > 1.5 indicates upregulated glycans, and Fold-change < 0.67 indicates downregulated glycans. Combinations of upregulated and downregulated glycans are then selected, and these combinations are used to diagnose specific types of cancer. However, this method utilizes incomplete lectin data, only applying data showing upregulated or downregulated glycan structures, and lacks practical application value in complex scenarios. Therefore, a highly accurate lung disease identification system is urgently needed. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention provides a lung cancer identification tool and system based on salivary lectin chip data, which is used to solve at least one of the aforementioned technical problems.

[0005] Specifically, the technical solution is as follows:

[0006] A lung disease identification tool based on salivary lectin chip data includes:

[0007] A storage medium for performing the following steps when the stored program is loaded by the processor:

[0008] The lung lesion identification model was used to identify the lectin chip data of the saliva of the patients to be diagnosed, and to determine whether the patients to be diagnosed were: healthy volunteers, patients with benign lung lesions, or patients with lung cancer.

[0009] The lung cancer patients mentioned include: small cell lung cancer patients, lung adenocarcinoma patients, or lung squamous cell carcinoma patients.

[0010] The aforementioned lung disease identification tool based on salivary lectin chip data also includes: a data acquisition module;

[0011] The acquisition module acquires external samples and, after performing the following steps, sends the external samples to the storage medium:

[0012] After centrifuging the collected whole saliva, the supernatant was collected and the precipitate was discarded.

[0013] The supernatant was then filtered, labeled with a fluorescent dye, and the free fluorescence was removed to obtain a fluorescently labeled pretreated saliva sample.

[0014] The aforementioned lung disease identification tool based on salivary lectin chip data also includes a preprocessing module:

[0015] The preprocessing module interacts with the acquisition unit and the storage medium to send the lectin chip data to the storage medium after performing the following steps:

[0016] The prepared lectin chip is cleaned to obtain a cleaned lectin chip.

[0017] The cleaned lectin chip was incubated with blocking buffer in a chip hybridization box and a rotation reaction was performed to obtain a blocked lectin chip.

[0018] The fluorescently labeled pretreated saliva sample was mixed with incubation buffer to prepare a loading system, which was then uniformly loaded onto a coverslip.

[0019] The sealed lectin chip was placed on the cover glass slide and incubated by rotation in the dark. After incubation, the slide was cleaned, centrifuged and dried to obtain the chip to be scanned.

[0020] The chip to be scanned is scanned, and the chip scanning results are normalized after image analysis to obtain the lectin chip data.

[0021] A lung disease identification system based on salivary lectin chip data includes:

[0022] The data acquisition module is used to acquire lectin chip data based on saliva samples;

[0023] The model building module interacts with the data acquisition module to form classification labels and features using the lectin chip data, thereby building and training a lung lesion identification model.

[0024] The lung disease identification tool described above interacts with the data acquisition module to identify the saliva of the patient to be diagnosed using a trained lung lesion identification model, thereby determining whether the patient to be diagnosed is a healthy volunteer, a patient with benign lung lesions, or a lung cancer patient.

[0025] The model building module includes:

[0026] The data preprocessing unit interacts with the data acquisition module to normalize the lectin chip data and obtain normalized lectin chip data.

[0027] The data classification unit interacts with the data preprocessing unit to randomly sort the normalized lectin chip data, obtain the features and labels of the normalized lectin chip data, and randomly select a portion of the normalized lectin chip data as the training set of the lung lesion identification model according to the labels, and another portion of the normalized lectin chip data as the test set of the lung lesion identification model.

[0028] The construction unit is connected to the lung disease identification tool to obtain the trained lung lesion identification model and store it within the lung disease identification tool.

[0029] The construction unit is connected to the data classification unit and is used to construct the lung lesion identification model based on the lectin chip data using the features and labels through any one of the following algorithms: K-nearest neighbor algorithm, support vector machine, multilayer perceptron, logistic regression, and random forest.

[0030] The construction unit receives the tags and features and uses them to construct the lung lesion identification model based on the lectin chip data through the following steps:

[0031] Using the Bootstraping method, m samples are randomly selected, and a total of n_tree sampling is performed to generate n_tree training sets;

[0032] For each of the n_tree training sets, train n_tree decision tree models respectively;

[0033] For a single decision tree model, the number of training sample features is set to n. When splitting at each node, features are selected for splitting based on information gain, information gain ratio, or Gini index until all training samples of that node belong to the same class.

[0034] The generated decision trees are combined into a random forest, and the final classification result is determined by voting among the multiple tree classifiers to obtain the lung lesion identification model.

[0035] The model building module further includes: a parameter selection and optimization unit;

[0036] The parameter selection and optimization unit interacts with the construction unit to adjust the lung lesion identification model through the following steps:

[0037] Adjust the parameter n_estimators that has the greatest impact on the lung lesion identification model, and use Gridsearch to explore n_estimators to find the optimal value;

[0038] The maximum depth of the decision tree is adjusted, and without changing the effect of the lung lesion identification model, a smaller max_depth is searched by performing a grid search to obtain the optimal max_depth parameter.

[0039] The minimum number of samples required to further split the internal nodes of the decision tree (min_samples_split) and the minimum number of samples required to split the leaf nodes (min_samples_leaf) are simultaneously tuned to obtain the optimal values ​​of min_samples_split and min_samples_leaf.

[0040] Tune the parameter max_features to obtain the optimal value of max_features.

[0041] The data acquisition module includes:

[0042] Collection unit, used to collect saliva samples;

[0043] A labeling unit, connected to the acquisition unit, is used to fluorescently label the saliva sample.

[0044] A lectin chip unit, connected to the labeling unit, is used to hold fluorescently labeled saliva samples;

[0045] An incubation unit, connected to the lectin chip unit, is used to incubate the lectin chip unit containing a fluorescently labeled saliva sample.

[0046] A data acquisition unit, connected to the incubation unit, is used to scan the incubated lectin chip unit and then perform image analysis to acquire the lectin chip data.

[0047] The data acquisition unit is connected to the data preprocessing unit and is used to acquire the normalized lectin chip data.

[0048] The present invention has at least the following beneficial effects:

[0049] The lung disease identification tool of this invention uses a lung lesion identification model to identify the saliva of the patient to be diagnosed, thereby determining whether the patient has lung disease; at the same time, it identifies the type of cancer in lung cancer patients; the lung disease identification tool of this invention has the characteristics of convenient sampling and high sensitivity, making it suitable for early and mid-stage screening of patients, and can quickly identify the lung lesion status of the subject.

[0050] The system described in this invention preprocesses and fluorescently labels the saliva sample using a data acquisition module, then performs lectin chip detection using a pre-made lectin chip to obtain lectin chip data. A model building module normalizes the lectin chip data to form classification labels and features, thereby constructing and training a lung lesion identification model. Finally, a lung disease identification tool uses the trained lung lesion identification model to identify the saliva of the patient to be diagnosed, determining whether the patient is a healthy volunteer, a patient with benign lung lesions, or a lung cancer patient. Furthermore, it simultaneously identifies the type of lung cancer patient: small cell lung cancer, adenocarcinoma, or squamous cell carcinoma. The system of this invention comprehensively utilizes lectin data and combines it with machine algorithms, offering the advantage of accurate detection. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a system block diagram of the system described in this invention.

[0053] Figure 2 This is a schematic diagram of a lectin chip.

[0054] Figure 3 This is a confusion matrix diagram of the KNN model.

[0055] Figure 4 This is the ROC curve of the KNN model.

[0056] Figure 5 This is a confusion matrix diagram of the SVM model.

[0057] Figure 6 This is the ROC curve for the SVM model.

[0058] Figure 7 This is a confusion matrix diagram of the MLP model.

[0059] Figure 8The ROC curve for the MLP model.

[0060] Figure 9 This is the confusion matrix diagram for the LR model.

[0061] Figure 10 This is the ROC curve for the LR model.

[0062] Figure 11 This is the confusion matrix diagram for the RF model.

[0063] Figure 12 This is the ROC curve for the RF model.

[0064] Among them, 100. Data acquisition module; 200. Model building module; 300. Lung disease identification tool;

[0065] 101. Acquisition unit; 102. Labeling unit; 103. Lectin chip unit; 104. Incubation unit; 105. Data acquisition unit;

[0066] 201. Data preprocessing unit; 202. Data classification unit; 203. Construction unit; 204. Parameter selection and optimization unit;

[0067] It needs to be clarified that: in Figure 3 , 5 In 7, 9, and 11:

[0068] HV: Healthy volunteers (HV)

[0069] BPD: Benign pulmonary disease (BPD)

[0070] SCLC: Small-cell lung cancer;

[0071] ADC: Lung cancers with adenocarcinoma (ADC);

[0072] SCC: Squamous cell carcinoma of the lung (SCC). Detailed Implementation

[0073] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0074] To address the shortcomings of existing technologies, this embodiment introduces machine learning for lung disease type identification, which can perfectly solve the practical application problem of lectin chip data and identify the type of lung lesion in the test sample, especially showing significant advantages in the identification of lung adenocarcinoma, lung squamous cell carcinoma, and small cell lung cancer. Specific implementation examples:

[0076] The present invention also provides an embodiment: such as Figure 1 A lung disease identification system based on saliva lectin chip data includes: a data acquisition module 100 for acquiring lectin chip data from saliva samples, a model building module 200, and a lung disease identification tool 300 as described above; wherein, the model building module 200 interacts with the data acquisition module 100 to form classification labels and features using the lectin chip data, thereby constructing and training a lung lesion identification model; the lung disease identification tool 300 interacts with the data acquisition module 100 to identify the type of lung lesion in the saliva of a patient to be diagnosed using the trained lung lesion identification model; specifically, the data acquisition module 100 is used to collect samples, and the model building module 200 is used to process the data collected by the data acquisition module 100 to obtain a lung lesion identification model.

[0077] Specifically, the model construction module 200 includes: a data preprocessing unit 201, a data classification unit 202, a construction unit 203, and a parameter selection and optimization unit 204; wherein, the data preprocessing unit 201 interacts with the data acquisition module to normalize the lectin chip data to obtain normalized lectin chip data; the data classification unit 202 interacts with the data preprocessing unit 201 to randomly sort the normalized lectin chip data, obtain the features and labels of the normalized lectin chip data, and randomly select according to the labels. A portion of the normalized lectin chip data is used as a training set, and another portion of the normalized lectin chip data is used as a test set. The construction unit 203 is connected to the data classification unit 202 and is used to construct the lung lesion identification model based on the features and the labels, and to train the lung lesion identification model using the training set to obtain the trained lung lesion identification model. The parameter selection and optimization unit 204 interacts with the construction unit 203 to adjust the parameters and the maximum number of features that have the greatest impact on the lung lesion identification model, thereby adjusting the lung lesion identification model.

[0078] The data acquisition module 100 includes: an acquisition unit 101, a labeling unit 102, a lectin chip unit 103, an incubation unit 104, and a data acquisition unit 105; wherein, the acquisition unit 101 is used to collect saliva samples; the labeling unit 102 is connected to the acquisition unit 101 and is used to fluorescently label the saliva samples; the lectin chip unit 103 is connected to the labeling unit 102 and is used to place the fluorescently labeled saliva samples; the incubation unit 104 is connected to the lectin chip unit 103 and is used to incubate the lectin chip unit 103 containing the fluorescently labeled saliva samples; the data acquisition unit 105 is connected to the incubation unit 104 and is used to scan the incubated lectin chip unit 103 and perform image analysis to acquire the lectin chip data.

[0079] like Figure 2 The fabrication process of the lectin chip unit 103 shown is as follows:

[0080] The following reagents were used for the lectin chip: GPTS (glucose-glucose-sulfurized silane reagent), Bradford reagent, protease inhibitors, DMSO, Tween-20, and hydroxylamine hydrochloride were purchased from Sigma-Aldrich (USA); Cy3 and Cy5 fluorescent dyes were purchased from Amherst (USA); Sephadex G-25 columns were purchased from GE Healthcare (USA); 384-well plates were purchased from Genetix (UK); bovine serum albumin (BSA) was purchased from Calbiochem (Germany); glass substrates were purchased from Gold Seal; 0.2 μm and 0.45 μm filter membranes were purchased from Sartorius (Japan); and other commonly used reagents were purchased from domestically produced analytical grade reagents. Thirty-seven lectins, listed in Table 1, were purchased from Vector and Sigma-Aldrich, respectively.

[0081] Table 1. Lectin Comparison Table

[0082]

[0083] The equipment used included: electric heating drying oven (Tianjin Taist Company); autoclave (TOMY Company, Japan); 5804R ultra-fast refrigerated centrifuge (Eppendorf Company, Germany); micro-volume nucleic acid and protein analyzer (Implen Company, Germany); 4000B biochip scanner (Axon Company, USA); chip spotting instrument (SmartArrayer48 spotting instrument from Biochip); and HL-2000 chip hybridization chamber (UVP Company, USA).

[0084] The fabrication process of lectin chips includes:

[0085] Untreated glass slides were washed three times with anhydrous ethanol for 10 min each time. After centrifugation, the slides were immersed in 250 mL of 10% NaOH solution and gently shaken on a shaker overnight in the dark. After the reaction, the slides were sonicated for 15 min, then washed four times with ultrapure water for 2 min each time, and twice with anhydrous ethanol for 2 min each time. After centrifugation, the slides were immersed in 200 mL of 10% GPTS solution and gently shaken on a shaker for 3 h in the dark. After the reaction, the slides were sonicated for 15 min and washed three times with anhydrous ethanol for 10 min each time. After centrifugation, the epoxidation modification of the chip was completed, and the modified glass slides were stored in a desiccator at 4 °C for later use. The spotting design of the prepared lectin chip is as follows. Figure 1 Each chip is divided into 4 matrices, each with a size of 12*10, and each sample point is repeated three times.

[0086] When collecting saliva samples in the collection unit 101, the following is included:

[0087] Healthy volunteers with no underlying medical conditions and who have not taken any medications within the past week are eligible. Individuals with confirmed colorectal, lung, liver, stomach, breast, or thyroid diseases are also eligible. Two hours after a meal, between 9 and 10 AM, gargle with saline solution three times and then quickly collect naturally secreted saliva. Collect at least 1 mL of saliva and immediately place it on ice. Add 1 μL of protease inhibitor per mL of saliva to prevent protein degradation.

[0088] Table 2 Clinical Sample Information Table

[0089]

[0090]

[0091] The working process of labeling unit 102 is as follows: Collected whole saliva is centrifuged at 12,000 rpm at 4℃ for 10 min, and the supernatant is collected and the precipitate discarded. The supernatant is then filtered through a 0.22 μm pore size filter membrane to remove bacteria and other microorganisms. The sample is labeled with Cy3 fluorescent dye and then desalted using a Sephadex G-25 desalting column to remove free fluorescence. The labeled proteins are then prepared for incubation on a lectin chip.

[0092] The operation of incubation unit 104 is as follows: The lectin chip prepared as described above is removed from the 4°C desiccator and allowed to warm to room temperature. First, the slide is washed once with PBST and once with PBS, each time for 3 minutes, and then centrifuged to dry. The lectin chip is incubated with 600 μL of blocking buffer in a chip hybridization cassette at 25°C for 1 hour by rotation. After blocking, the slide is washed twice with PBST and twice with PBS, each time for 3 minutes, and then centrifuged to dry. The blocked chip is scanned using a Genepix 4000B chip scanner to check the blocking effect.

[0093] 3 μg of fluorescently labeled salivary protein was mixed with incubation buffer to prepare a 600 μL loading system. This system was then evenly loaded onto a coverslip, and the sealed lectin chip was placed on top. The system was incubated in a chip hybridization instrument at 25°C in the dark for 3 hours. After incubation, the coverslip was washed twice with PBST and PBS, 5 min each time, and then centrifuged to dry. During the operation of data acquisition unit 105, the chip was scanned using a Genepix 4000B chip scanner. GenePix 3.0 software was used to analyze the chip scan results (images, circles, and dots), and then exported as a GPR file for further analysis.

[0094] The data preprocessing unit 201 preprocesses the data information, including signal and background value processing, effective data screening and processing, and data normalization, to obtain lectin chip data for subsequent analysis.

[0095] All cases are randomly sorted by data classification unit 202, and 70% of the data is randomly selected from each of the seven categories to be used as the training set for machine learning, while the remaining 30% is used as the test set for machine learning.

[0096] The present invention provides a preferred solution: the construction unit 203 uses random forest to construct a lung lesion identification model; the algorithm idea of ​​random forest is: a classifier that uses multiple decision trees to train and predict samples.

[0097] The implementation steps of this preferred solution are as follows:

[0098] We randomly sampled m samples with replacement from the original training set using the Bootstraping method, performing n_tree sampling operations in total. This generated n_tree training sets. For each of the n_tree training sets, we trained n_tree decision tree models. We used 37 lectins from the lectin chip data as features; the label refers to the disease type of each sample, such as lung cancer, stomach disease, etc.; each sample corresponds to one label and 37 feature variables.

[0099] For a single decision tree model, assuming the number of training sample features is n, the best feature is selected for splitting at each split based on information gain / information gain ratio / Gini index. Each tree knows that all training samples at that node belong to the same class. The generated decision trees are combined into a random forest, and the final classification result is determined by voting among the multiple tree classifiers. The output prediction result is compared with the label of the test set based on the features of the cases in the test set to measure the algorithm's performance. Then, the parameters are adjusted through the parameter selection and optimization unit 204, including: first adjusting the parameter n_estimators that has the greatest impact on the model, and using Gridsearch to explore n_estimators. The optimal value was found to be 47. The maximum tree depth was adjusted to search for a smaller `max_depth` without changing the model's performance, which simplifies the model. After a grid search, the optimal parameter for `max_depth` was found to be 12. Then, the minimum number of samples required for internal node splitting (`min_samples_split`) and the minimum number of samples required for leaf nodes (`min_samples_leaf`) were tuned together, resulting in the optimal values ​​for `min_samples_split` (2) and `min_samples_leaf` (1). Finally, the maximum number of features (`max_features`) was tuned, resulting in the optimal value for `max_features` (22).

[0100] This invention discloses an embodiment: a lung disease identification tool based on salivary lectin chip data, comprising: a storage medium and a processor; the storage medium is used to execute the following steps when a stored program is loaded by the processor:

[0101] The lung lesion identification model was used to identify the lectin chip data of the saliva of the patients to be diagnosed, and to determine whether the patients to be diagnosed were: healthy volunteers, patients with benign lung lesions, or patients with lung cancer.

[0102] Preferably, the lung cancer patient includes: a small cell lung cancer patient, a lung adenocarcinoma patient, or a lung squamous cell carcinoma patient.

[0103] For ease of use, the aforementioned lung disease identification tool based on saliva agglutinin chip data further includes: a collection module; the collection module collects external samples to obtain fluorescently labeled pre-treated saliva samples, specifically referring to the "working process of labeling unit 102" and the "preparation process of agglutinin chip unit 103" described in the embodiment of the "Lung Disease Identification System Based on Saliva Agglutinin Chip Data".

[0104] The collected saliva needs to be processed before it can be tested. Therefore, the aforementioned lung disease identification tool based on salivary lectin chip data also includes a preprocessing module. This preprocessing module interacts with the acquisition unit and the storage medium to acquire lectin chip data. Specific steps refer to the working process of the "incubation unit 104" and "data preprocessing unit 201" described in the embodiment of the "Lung Disease Identification System Based on Salivary Lung Chloride Chip Data". Using this lectin chip data, a pre-trained lung lesion identification model can be used to identify patients to be diagnosed, determining whether the patient is a healthy patient or a lung disease patient.

[0105] Preferably, the lung disease types include: benign lung lesions and lung cancer; more specifically, lung cancer includes: small cell lung cancer, lung adenocarcinoma, and lung squamous cell carcinoma.

[0106] The random forest modeling, testing, and application process is as follows:

[0107] Modeling: (1) Select n samples with replacement from the modeling set; (2) Randomly select k features from 37 lectin features, and build a decision tree for the selected samples using the selected k features. Each time the tree splits a node, randomly select a portion of these k features to find the optimal solution, apply it to the node, and split it. The splitting process for each node is the same until all nodes reach the leaf nodes. Each leaf node corresponds to a category label: healthy, benign lung disease, small cell lung cancer, lung adenocarcinoma, or lung squamous cell carcinoma.

[0108] (3) Repeat the above two steps m times to generate m decision trees, forming a random forest and completing the construction of the random forest model.

[0109] Test: (4) For the test set sample, after each tree makes a decision, the final vote confirms which category it belongs to. The label with the most votes is considered as the predicted label of the test sample, and it is compared with the real label of the test set sample to measure the performance of the model algorithm.

[0110] When using the above random forest model in practice: (5) For the sample to be identified, each tree makes a decision and finally votes to confirm which category it belongs to. The label with the most votes is considered to be the predicted label of the sample to be identified, and finally the identification result of the sample to be identified is obtained.

[0111] Accuracy verification:

[0112] Machine learning includes various methods. The accuracy of this embodiment is verified by comparing the modeling results of the logical forest in the construction unit 203 with those of the K-nearest neighbor algorithm, support vector machine, multilayer perceptron, and logistic regression.

[0113] The K-Nearest Neighbor algorithm is a simple, intuitive, and practical algorithm widely used in classification problems. Its main idea is to: calculate the distance between the current point and points in a known class dataset; sort the points by distance in ascending order; select the k points closest to the current point; determine the frequency of the classes of the k points; and return the class with the highest frequency among the k points as the predicted classification for the current point.

[0114] The implementation steps of the K-nearest neighbor algorithm are as follows:

[0115] A KNN classifier is constructed by inputting all features and labels of cases in the training set; the output predictions are compared with the labels of the test set using features of cases in the test set to measure the algorithm's performance; the most important parameters in KNN include the K value, weights, and the method of distance calculation.

[0116] K: Select the k points closest to the current point. If k is too small, it will not be able to resist interference. In this case, the bias of the model will decrease and the variance will increase, and the model will be prone to overfitting. If k is too large, it is equivalent to using training instances in a larger neighborhood for prediction. The approximate error of learning will increase, so it is not representative.

[0117] weights: Two forms, one without considering distance weights and one with considering distance weights;

[0118] Distance metrics P include Minkowski distance, Euclidean distance, Manhattan distance, etc.

[0119] By iterating through the values ​​of weights, p, and k using GridSearch, the results show that the optimal parameters are weights = "distance", p = 1, and k = 4.

[0120] like Figures 3-4 According to Table 3, the performance of the KNN model is as follows: accuracy: 73.79%; precision: 63.10%; recall: 64.30%; area under the ROC curve: 0.92; sensitivity: 73.78%; specificity: 93.44%.

[0121] Table 3 Summary of KNN Model Data

[0122] actual predict False negative False positive Sensitivity Specificity HV 48 42 6 3 87.5％ 94.47％ BPD 15 13 2 3 86.67％ 96.59％ ADC 17 15 2 13 88.24％ 84.88％ SCC 13 3 10 6 23.08％ 93.33％ SCLC 10 3 7 2 30％ 92.47％ Sum 103 76 27 27 Accuracy 73.79

[0123] Support Vector Machines (SVMs) are essentially a binary classification model that seeks a linear classifier in the feature space that maximizes the margin between two classes. This characteristic distinguishes it from the Perceptron Model (MLP). Furthermore, kernel tricks can be used to extend it to non-linear cases.

[0124] Implementation steps: Input case features from the training set. Since SVM essentially performs binary classification, an SVM needs to be designed between any two classes of samples. To achieve classification of 5 classes, 5*(5-1) / 2 = 10 SVMs need to be constructed. Input case features from the test set into the trained SVMs. The class with the most votes is the class of the unknown sample. Label prediction and comparison of the test set data.

[0125] Parameter selection and optimization:

[0126] C: Penalty coefficient, i.e., tolerance for error. If C is too large or too small, the generalization ability will be poor.

[0127] Kernel Functions: Commonly used kernel functions include linear kernels, polynomial kernels, radial basis function (RBF) kernels, and Fourier kernels. Kernel selection: The Cross-Validation method is used, where different kernel functions are tested, and the kernel function with the smallest induction error is the best. In this study, the best kernel function is RBF. After selecting RBF as the kernel, its built-in parameter Gamma determines the distribution of data after mapping to the new feature space. A larger Gamma results in fewer support vectors, while a smaller Gamma value results in more support vectors and a wider RBF width. The number of support vectors affects the training and prediction speed. The optimal parameters for RBF are selected using GridSearch: the optimal parameters are kernel = rbf, gamma = 3, and C = 15.

[0128] like Figures 5-6 According to Table 4, the performance of the SVM model is as follows: accuracy: 75.73%; precision: 59.89%; recall: 63.94%; area under the ROC curve: 0.85; sensitivity: 75.73%; specificity: 93.93%.

[0129] Table 4 Summary of SVM Model Data

[0130] actual predict False negative False positive Sensitivity Specificity HV 48 46 2 3 95.83％ 94.55％ BPD 15 15 0 4 100％ 98.86％ ADC 17 15 2 17 88.24％ 80.23％ SCC 13 2 11 0 15.38％ 100％ SCLC 10 0 10 1 0 98.92％ Sum 103 78 25 25 Accuracy 75.73％

[0131] The Multilayer Perceptron (MPP) algorithm is a shallow neural network algorithm consisting of an input layer, hidden layers, and an output layer. Each layer is composed of units. The input layer is fed in by the instance feature vectors of the training set, which are then passed to the next layer through the weights of the connection nodes. The output of the previous layer is the input of the next layer. The number of hidden layers is arbitrary, and there is only one output layer and one input layer.

[0132] Implementation steps: Construct an MLP model by inputting all features and labels of cases in the training set; compare the output prediction results with the labels of the test set by inputting features of cases in the test set to measure the algorithm performance.

[0133] Parameter Selection and Optimization: MLPs have over 20 parameters, but the most important are the number of hidden layers and the number of neurons in each hidden layer. Therefore, parameter selection primarily considers the weight optimization solver (slover) and `hidden_layer_sizes` (the number of hidden layers and neurons). Slover options include 'lbfgs', 'sgd', and 'adam'. 'lbfgs' is a quasi-Newton family of optimizers, which converges faster and performs better on small datasets. 'sgd' refers to stochastic gradient descent. 'adam' is a stochastic gradient-based optimizer that works well on relatively large datasets in terms of training time and validation scores.

[0134] Using GridSearch to select optimal parameters includes the choice of optimizer, the number of hidden layers, and the number of neurons: the optimal parameters obtained are slover = 'adam', hidden_layer_sizes(14,8), two hidden layers, the first layer has 14 neurons, and the second layer has 8 neurons.

[0135] like Figures 7-8 According to Table 5, the accuracy was 72.82%; the precision was 57.16%; the recall was 43.51%; the area under the ROC curve was 0.81; the sensitivity was 72.82%; and the specificity was 93.20%.

[0136] Table 5 Summary of MLP Model Data

[0137] actual predict False negative False positive Sensitivity Specificity HV 48 44 4 1 91.67％ 98.18％ BPD 15 15 0 7 100％ 92.05％ ADC 17 16 1 15 94.12％ 82.56％ SCC 13 0 13 1 0 98.89％ SCLC 10 0 10 4 0 95.70％ Sum 103 75 28 28 Accuracy 72.82％

[0138] Logistic Regression (LR) algorithm: LR belongs to supervised learning and is a "classification" algorithm. The regression of LR is essentially the probability of occurrence divided by the probability of non-occurrence and then the logarithm.

[0139] Implementation steps: Construct an LR model by inputting all features and labels of cases in the training set; compare the output prediction results with the labels of the test set by inputting features of cases in the test set to measure the algorithm performance.

[0140] Parameter selection and optimization:

[0141] Regularization parameter `penalty`: The `penalty` parameter can be selected as "l1" or "l2", corresponding to L1 and L2 regularization respectively. Algorithm selection parameter `solver`: When `penalty` is L1, `solver` can only be 'liblinear'; when `penalty` is L2, `solver` can be liblinear, lbfgs, newton-cg, or sag. Classification method selection parameter `multi_class`: It has two values: `ovr` and `multinomial`. In Python's sklearn, `LRlogisticRegressionCV` uses cross-validation to select the regularization coefficient `C`, therefore the coefficient of `C` does not need further optimization; the optimal parameters are `penalty = "l2", solver = "lbfgs", multi_class = "multinomial"`.

[0142] like Figures 9-10 According to Table 6, the performance of the LR model is as follows: accuracy: 69.90%; precision: 60.62%; recall: 56.96%; ROC area below: 0.81; sensitivity: 69.90%; specificity: 92.48%.

[0143] Table 6 Summary of LR Model Data

[0144] actual predict False negative False positive Sensitivity Specificity HV 48 39 9 0 81.25％ 100％ BPD 15 15 0 8 100％ 90.91％ ADC 17 13 4 10 76.47％ 88.37％ SCC 13 2 11 3 15.38％ 96.67％ SCLC 10 3 7 10 30％ 89.25％ Sum 103 72 31 31 Accuracy 69.90％

[0145] like Figures 11-12 According to Table 7, in this embodiment, a lung lesion identification model was constructed using random forest. The performance of the RF model is as follows: accuracy: 88.35%; precision: 82.36%; recall: 85.62%; area under the ROC curve: 0.95; sensitivity: 88.35%; specificity: 95.63%.

[0146] Table 7 Summary of RF Model Data

[0147] actual predict False negative False positive Sensitivity Specificity HV 48 47 1 3 97.92％ 94.55％ BPD 15 15 0 1 100％ 98.86％ ADC 17 14 3 5 82.35％ 94.86％ SCC 13 8 5 1 61.54％ 98.89％ SCLC 10 7 3 2 70％ 97.85％ Sum 103 91 12 12 Accuracy 88.35％

[0148] The results of the various algorithms described above are statistically analyzed, as shown in Table 8:

[0149] Table 8 Comparison of the effects of different models

[0150]

[0151] As shown in Table 8, RF performed better in identifying different lung cancer types based on lectin chip data, achieving an accuracy of 88.35% in the test set. It correctly distinguished 47 out of 48 HV cases, 15 out of 15 BPD cases, 14 out of 17 ADC cases, 8 out of 13 SCC cases, and 7 out of 10 SCLC cases. The other four models performed worse, with accuracy below 80% in the test set.

[0152] Therefore, the lung disease identification system provided by this invention can detect the type of lung disease, especially with high accuracy in identifying lung adenocarcinoma, lung squamous cell carcinoma and small cell lung cancer. Moreover, since it uses saliva for detection, it has the advantage of convenient sampling.

[0153] The above descriptions only cover a few specific embodiments of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention. The above-mentioned serial numbers are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

Claims

1. A lung disease identification system based on salivary lectin chip data, characterized in that, include: The data acquisition module is used to acquire lectin chip data based on saliva samples; The model building module interacts with the data acquisition module to form classification labels and features using the lectin chip data, thereby building and training a lung lesion identification model. The lung disease identification tool based on salivary lectin chip data interacts with the data acquisition module to identify the saliva of the patient to be diagnosed using a trained lung lesion identification model, and to determine whether the patient to be diagnosed belongs to one of the following categories: healthy volunteers, patients with benign lung lesions, or lung cancer patients. The model building module includes: The data preprocessing unit interacts with the data acquisition module to normalize the lectin chip data and obtain normalized lectin chip data. The data classification unit interacts with the data preprocessing unit to randomly sort the normalized lectin chip data, obtain the features and labels of the normalized lectin chip data, and randomly select a portion of the normalized lectin chip data as the training set of the lung lesion identification model according to the labels, and another portion of the normalized lectin chip data as the test set of the lung lesion identification model. A construction unit, connected to the lung disease identification tool, is used to obtain the trained lung lesion identification model and store it within the lung disease identification tool; the construction unit is connected to the data classification unit, and is used to construct the lung lesion identification model based on the lectin chip data using the features and labels through a random forest algorithm; the lung lesion identification model is used to identify healthy individuals, benign lung diseases, small cell lung cancer, lung adenocarcinoma, and lung squamous cell carcinoma. The parameter selection and optimization unit interacts with the construction unit to adjust the lung lesion identification model.

2. The lung disease identification system based on salivary lectin chip data according to claim 1, characterized in that: The construction unit receives the tags and features and uses them to construct the lung lesion identification model based on the lectin chip data through the following steps: Using the Bootstraping method, m samples are randomly selected, and a total of n_tree sampling is performed to generate n_tree training sets; For each of the n_tree training sets, train n_tree decision tree models respectively; For a single decision tree model, the number of training sample features is set to n. When splitting at each node, features are selected for splitting based on information gain, information gain ratio, or Gini index until all training samples of that node belong to the same class. The generated decision trees are combined into a random forest, and the final classification result is determined by voting among the multiple tree classifiers to obtain the lung lesion identification model.

3. The lung disease identification system based on salivary lectin chip data according to claim 2, characterized in that, Parameter selection and optimization unit; The parameter selection and optimization unit adjusts the lung lesion identification model through the following steps: Adjust the parameter n_estimators that has the greatest impact on the lung lesion identification model, and use Gridsearch to explore n_estimators to find the optimal value; The maximum depth of the decision tree is adjusted, and without changing the effect of the lung lesion identification model, a smaller max_depth is searched by performing a grid search to obtain the optimal max_depth parameter. The minimum number of samples required to further split the internal nodes of the decision tree (min_samples_split) and the minimum number of samples required to split the leaf nodes (min_samples_leaf) are simultaneously tuned to obtain the optimal values ​​of min_samples_split and min_samples_leaf. Tune the parameter max_features to obtain the optimal value of max_features.

4. The lung disease identification system based on salivary lectin chip data according to claim 1, characterized in that: The data acquisition module includes: Collection unit, used to collect saliva samples; A labeling unit, connected to the acquisition unit, is used to fluorescently label the saliva sample. A lectin chip unit, connected to the labeling unit, is used to hold fluorescently labeled saliva samples; An incubation unit, connected to the lectin chip unit, is used to incubate the lectin chip unit containing a fluorescently labeled saliva sample. A data acquisition unit, connected to the incubation unit, is used to scan the incubated lectin chip unit and then perform image analysis to acquire the lectin chip data. The data acquisition unit is connected to the data preprocessing unit and is used to acquire the normalized lectin chip data.

5. The lung disease identification system based on salivary lectin chip data according to claim 1, characterized in that, The types of lung cancer patients include: small cell lung cancer patients, lung adenocarcinoma patients, or lung squamous cell carcinoma patients.

6. The lung disease identification system based on salivary lectin chip data according to claim 1, characterized in that, The lung disease identification tool also includes: a data acquisition module; The acquisition module acquires external samples and, after performing the following steps, sends the external samples to the storage medium: After centrifuging the collected whole saliva, the supernatant was collected and the precipitate was discarded. The supernatant was then filtered, labeled with a fluorescent dye, and the free fluorescence was removed to obtain a fluorescently labeled pretreated saliva sample.

7. The lung disease identification system based on salivary lectin chip data according to claim 6, characterized in that, The lung disease identification tool also includes a preprocessing module: The preprocessing module interacts with the acquisition unit and the storage medium to send the lectin chip data to the storage medium after performing the following steps: The glass slide of the prepared lectin chip is cleaned to obtain a cleaned lectin chip. The cleaned lectin chip was incubated with blocking buffer in a chip hybridization box and a rotation reaction was performed to obtain a blocked lectin chip. The fluorescently labeled pretreated saliva sample was mixed with incubation buffer to prepare a loading system, which was then uniformly loaded onto a coverslip. The sealed lectin chip was placed on the cover glass slide and incubated by rotation in the dark. After incubation, the slide was cleaned, centrifuged and dried to obtain the chip to be scanned. The chip to be scanned is scanned, and the chip scanning results are normalized after image analysis to obtain the lectin chip data.

Images