An artificial intelligence system for non-invasive prediction of EGFR / TP53 co-mutated lung cancer patients

Abstract

This invention provides a non-invasive artificial intelligence system for predicting EGFR / TP53 co-mutated lung cancer patients, belonging to the field of lung cancer prediction systems. By collecting patients' clinical and imaging characteristics, and using LASSO feature selection and a random forest classifier to build a model, this invention obtains an artificial intelligence prediction system capable of accurately and non-invasively predicting EGFR / TP53 gene co-mutation status in lung cancer patients. Experimental results show that the artificial intelligence prediction system established in this invention has superior predictive performance for EGFR / TP53 gene co-mutation status in lung cancer patients, with an AUC value as high as 0.746 on the test set. This artificial intelligence prediction system provides a new option for clinical screening of lung cancer patients with EGFR+ / TP53+ co-mutations, as well as screening EGFR-mutant lung cancer patients resistant to TKIs, and has important guiding significance for the precise clinical treatment of EGFR / TP53 co-mutated lung cancer patients.

Description

Technical Field

[0001] This invention belongs to the field of lung cancer prediction systems, and in particular relates to an artificial intelligence system for non-invasive prediction of lung cancer patients with EGFR / TP53 co-mutation. Background Technology

[0002] With economic and technological development, coupled with environmental exposure and unhealthy lifestyles, the incidence of cancer remains high globally, with lung cancer being one of the most common. According to data from the International Agency for Research on Cancer (IARC) GLOBOCAN (Global Cancer Observatory), in 2020, there were approximately 19.3 million new cancer cases and about 10 million cancer deaths worldwide, with lung cancer accounting for approximately 11.1% of new cases and 18.0% of deaths.

[0003] Primary lung cancer is classified into non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC) based on pathological characteristics. NSCLC accounts for approximately 85% of cases, while SCLC accounts for approximately 15%. Research on NSCLC is a key focus in lung cancer management.

[0004] Traditional chemotherapy for NSCLC has limited efficacy, but the development of targeted region sequencing (TRS) technology has brought hope for precision treatment of NSCLC patients with different gene mutations. Studies have found multiple related gene mutation sites in the development and progression of NSCLC, among which the epidermal growth factor receptor (EGFR) gene is one of the most common driver genes in Asian NSCLC patients. EGFR is the expression product of the proto-oncogene ErbB-1. Under normal physiological conditions, EGFR regulates epithelial tissue and maintains homeostasis. When the EGFR gene is mutated or damaged, it drives abnormal cell growth, leading to cancer development and progression. Currently, targeted drugs targeting activating mutations in the EGFR tyrosine kinase region are widely used in lung cancer treatment. For different EGFR mutation sites, including exon 19 deletion, exon 21 p.L858R point mutation, and other atypical mutations, various tyrosine kinase inhibitors (TKIs) are widely used in clinical practice, improving the survival prognosis of NSCLC patients.

[0005] However, clinical experience with TKI treatment in patients with EGFR mutations has revealed varying treatment outcomes among patients with the same TNM stage. Some patients show poorer efficacy and are more prone to TKI resistance. Early identification and intervention for these patients remains a clinical challenge.

[0006] TP53 is one of the most frequently mutated genes in cancer. Donehower et al. analyzed the TP53 mutation status in the whole exon sequences of 10,225 patients from 32 types of cancer in the TCGA Tumor Genome Atlas database. Of these, 3,786 patients had TP53 mutations, and the frequency of TP53 mutations varied across different cancer types. In NSCLC patients, TP53 was found to be one of the most frequently mutated genes. Several studies have found that patients with EGFR mutations who also have TP53 mutations after TKI treatment have a poorer prognosis and are more prone to drug resistance. Therefore, rapid and accurate detection of NSCLC patients with EGFR / TP53 co-mutations is of great significance for precision clinical treatment of NSCLC. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide an artificial intelligence system for non-invasive prediction of lung cancer patients with EGFR / TP53 co-mutations.

[0008] This invention provides an artificial intelligence system for predicting gene mutation status in lung cancer patients, the artificial intelligence system comprising the following five parts:

[0009] Part 1: Data Input Section; Input the imaging and clinical characteristics of lung cancer patients;

[0010] Part Two: Feature Selection; This section selects from the imaging and clinical features input in Part One to obtain the selected features.

[0011] Part 3: Model Building; The features selected in Part 2 are divided into training set data and test set data. The training set data is used to train a random forest classifier to build a prediction model.

[0012] Part Four: Prediction; The prediction model constructed in Part Three is used to process the test set data to determine whether the gene mutation status of lung cancer patients is EGFR / TP53 co-mutation or non-EGFR / TP53 co-mutation.

[0013] Furthermore, in the first part, the imaging features are extracted using the following method:

[0014] (1) The plain CT images of lung cancer patients are segmented into regions of interest to obtain the regions of interest of the lesions;

[0015] (2) The regions of interest in the image are preprocessed by 14 types of filters to obtain the preprocessed image; the 14 types of filters are: additive Gaussian noise filter, binomial image blur filter, box mean filter, box sigma image filter, curvature flow filter, discrete Gaussian filter, Laplacian sharpening filter, mean filter, normalization filter, recursive Gaussian filter, shot noise filter, speckle noise filter, Gaussian Laplacian filter, and wavelet filter.

[0016] (3) Image features were extracted from the region of interest of the unprocessed image and the preprocessed image, respectively, to obtain 7 types of image features: first-order features, shape features, gray-level co-occurrence matrix, gray-level run matrix, gray-level region size matrix, gray-level dependence matrix, and neighboring gray-level difference matrix.

[0017] Furthermore, in Part One, the clinical characteristics are the following 21 dimensions: gender; age (years); smoking history; family history of lung cancer; family history of non-lung cancer; daily number of cigarettes smoked; number of years of smoking; whether or not one has quit smoking; number of years since quitting smoking; whether or not one drinks alcohol; duration of alcohol consumption (years); alcohol consumption (grams / day); whether or not one coughs; whether or not one experiences chest pain; whether or not one coughs up blood; whether or not one expectorates sputum; carcinoembryonic antigen (CEA) (ng / ml); cancer antigen 125 (U / ml); carbohydrate antigen 199 (U / ml); cytokeratin 19 fragment (ng / ml); neuron-specific enolase (ng / ml).

[0018] Furthermore, in the second part, the selection method is to use the lasso algorithm, set the alpha value to 0.0152, and select the top 108 features with the highest feature coefficients among the imaging features and clinical features.

[0019] Furthermore, in the third part, the parameters of the random forest classifier are set as follows: the class weight uses Balance, the metric is Entroy, the maximum depth is 6, the minimum number of samples per leaf node is 3, the minimum number of samples per split is 2, the number of weak classifiers is 2000, and the classification threshold is 0.5.

[0020] Furthermore, the lung cancer in question is non-small cell lung cancer.

[0021] Furthermore, the non-small cell lung cancer is EGFR-mutated non-small cell lung cancer.

[0022] Furthermore, the EGFR-mutated non-small cell lung cancer is resistant to tyrosine kinase inhibitors.

[0023] The present invention also provides a device for predicting gene mutation status in lung cancer patients, wherein the device stores the aforementioned artificial intelligence system.

[0024] The present invention also provides the use of the above-mentioned artificial intelligence system in the preparation of a device for predicting the gene mutation status of lung cancer patients, the device being able to predict whether the gene mutation status of lung cancer patients is EGFR / TP53 co-mutation or non-EGFR / TP53 co-mutation.

[0025] In this invention, EGFR / TP53 co-mutation refers to EGFR gene mutation combined with TP53 gene mutation, denoted as EGFR+ / TP53+.

[0026] Non-EGFR / TP53 comutation refers to cases other than EGFR / TP53 comutation, including EGFR+ / TP53-, EGFR- / TP53+, and EGFR- / TP53-.

[0027] Among them, EGFR+ / TP53- indicates that the EGFR gene is mutated and the TP53 gene is not mutated; EGFR- / TP53+ indicates that the TP53 gene is mutated and the EGFR gene is not mutated; EGFR- / TP53- indicates that the TP53 gene is not mutated and the EGFR gene is not mutated.

[0028] The alpha value is a regularization parameter in LASSO regression, used to adjust the penalty term of LASSO.

[0029] The feature coefficient refers to the weight of a feature.

[0030] This invention, through the collection of clinical, imaging, pathological, genetic, and prognostic information from 2171 patients initially diagnosed with primary NSCLC at West China Hospital of Sichuan University, found that 24.6% of these 2171 primary NSCLC patients had EGFR+ / TP53+ co-mutations. Furthermore, among these primary NSCLC patients, those with EGFR+ / TP53+ co-mutations showed significant differences in survival prognosis compared to patients with other gene mutation states: early-stage (stage I+II) EGFR+ / TP53+ patients had a significantly lower 1-year survival rate than EGRF+ / TP53- patients (97.5% vs. 100%, P = 0.032); late-stage (stage III+IV) EGFR+ / TP53+ patients had a significantly lower 1-year survival rate than EGRF+ / TP53- patients (81.7% vs. 90.2%, P = 0.033). This invention also found that EGFR+ / TP53+ patients are more likely to develop drug resistance and have shorter progression-free survival (PFS) when treated with first-line EGFR-TKIs than EGFR+ / TP53- patients. Therefore, accurately predicting EGFR+ / TP53+ co-mutant patients in NSCLC has important guiding significance for the clinical precision treatment of EGFR / TP53 co-mutant lung cancer patients.

[0031] This invention collects patients' clinical and imaging characteristics, and establishes a model using LASSO feature selection and random forest classifier to obtain an artificial intelligence prediction system that can accurately and non-invasively predict the EGFR / TP53 gene co-mutation status of lung cancer patients.

[0032] Compared with existing prediction systems, the artificial intelligence prediction system of the present invention has the following advantages:

[0033] 1. The LASSO feature selection method used in this invention can filter out redundant feature vector types, thereby selecting important feature vectors and providing corresponding clinical guidance;

[0034] 2. In this invention, LASSO selection is performed on features, and the coefficients of important features are arranged so that the coefficients of unimportant feature vectors are 0. The results show that LASSO feature selection can improve the performance of random forest classifiers in machine learning.

[0035] 3. The clinical and imaging features included in this invention are all information that is easily obtained in actual clinical practice;

[0036] 4. Experiments show that compared with clinical prediction models and imaging prediction models, the non-invasive prediction model (clinical + imaging prediction model) established in this invention for lung cancer patients with EGFR / TP53 co-mutation has a better predictive effect on the EGFR / TP53 gene co-mutation status of lung cancer patients, with an AUC value as high as 0.746 on the test set.

[0037] This invention provides a non-invasive prediction system for screening lung cancer patients with EGFR+ / TP53+ co-mutations and for screening EGFR-mutant lung cancer patients resistant to TKIs. It has important guiding significance for the clinical precision treatment of lung cancer patients with EGFR / TP53 co-mutations.

[0038] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0039] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Attached Figure Description

[0040] Figure 1 LASSO-Alpha parameter characteristics in Example 1.

[0041] Figure 2Trajectory of characteristic coefficient changes.

[0042] Figure 3 The ROC curves of the training and validation sets for each fold in the five-fold cross-validation of the clinical + imaging prediction model of this invention.

[0043] Figure 4 ROC curves of the training and testing sets of the clinical + imaging prediction model of this invention.

[0044] Figure 5 Overall survival curve for NSCLC patients.

[0045] Figure 6 Survival curves of NSCLC patients with different EGFR / TP53 mutation status; where: red represents the EGFR- / TP53- group, blue represents the EGFR+ / TP53- group, green represents the EGFR- / TP53+ group, and yellow represents the EGFR+ / TP53+ group.

[0046] Figure 7 OS curves for stage I and II NSCLC patients with EGFR+ / TP53- and EGFR+ / TP53+ genes; where: gene group 1 (blue curve) is the EGFR+ / TP53+ group; gene group 2 (green curve) is the EGFR+ / TP53- group.

[0047] Figure 8 OS curves for stage III and IV NSCLC patients with EGFR+ / TP53- and EGFR+ / TP53+ genes; where: gene group 1 (blue curve) is the EGFR+ / TP53+ group; gene group 2 (green curve) is the EGFR+ / TP53- group.

[0048] Figure 9 Survival curves for EGFR+ / TP53- and EGFR+ / TP53+ patients treated with TKIs; where: gene group 1 (blue curve) is the EGFR+ / TP53+ group; gene group 2 (green curve) is the EGFR+ / TP53- group.

[0049] Figure 10 Survival curves for EGFR+ / TP53- and EGFR+ / TP53+ patients who did not receive TKI treatment; where: gene group 1 (blue curve) is the EGFR+ / TP53+ group; gene group 2 (green curve) is the EGFR+ / TP53- group.

[0050] Figure 11 PFS curves for patients treated with EGFR-TKIs alone; where: gene group 1 (blue curve) is the EGFR+ / TP53+ group; gene group 2 (green curve) is the EGFR+ / TP53- group.

[0051] Figure 12 LASSO-Alpha parameter characteristics in Experiment Example 2.

[0052] Figure 13 The top ten features with the highest coefficients; Note: The left side shows the feature names, and the right side shows the coefficients obtained after the LASSO algorithm.

[0053] Figure 14 Trajectory of changes in the top ten high-coefficient characteristic coefficients.

[0054] Figure 15 ROC curves for the training and testing sets of the imaging prediction model, the clinical prediction model, and the clinical + imaging model; where: red represents the clinical + imaging prediction model; green represents the imaging prediction model; and blue represents the clinical prediction model. Detailed Implementation

[0055] The raw materials and equipment used in this invention are all known products, obtained by purchasing commercially available products.

[0056] Example 1: Method for establishing a non-invasive model (clinical + imaging prediction model) for predicting lung cancer patients with EGFR / TP53 co-mutations according to the present invention.

[0057] The patient data used in this embodiment are from 1055 patients who had plain CT images with a slice thickness of 1 mm within 60 days before the pathological diagnosis, out of 2171 patients who were first diagnosed with primary NSCLC at West China Hospital of Sichuan University from January 2013 to December 2019.

[0058] All patients underwent chest CT scans using Siemens, Philips Brilliance Big Bore, and GE Discovery CT750 HD equipment. All scans were performed at maximum inspiration, using spiral CT scanning of the upper body from the apex to the base of the lungs. Instrument parameters were: tube voltage: 120V, tube current: 200–500mA, rotation time: 0.4–0.7s, pixel matrix: 512×512. Siemens scanners used soft convolution kernels (B31f, B30f); GE scanners used Standard kernels.

[0059] In the patient data used in this embodiment, all gene mutation statuses were tested. The test results showed that among the 1055 patients, there were 339 in the EGFR+ / TP53+ group, 409 in the EGFR+ / TP53- group, and 307 in the EGFR- / TP53- group.

[0060] Step 1: Extraction of imaging and clinical features

[0061] 1. Image Feature Extraction

[0062] (1) Image Acquisition

[0063] All 1mm thin-slice CT images were exported in DICOM format from the Picture Archiving and Communication System (PACS) station using IQQA software (IQQA-Chest, EDDA Technology, Princeton Junction, NJ, USA).

[0064] (2) Image segmentation

[0065] Because the original image contains too many cluttering factors, it is necessary to perform region of interest (ROI) segmentation. This embodiment uses deep learning for lesion ROI segmentation: the VB-Net model, which is an improvement on the Bézier Curve, is used to segment the lesion and obtain the region of interest (ROI) of the lesion.

[0066] (3) Image filtering and feature extraction

[0067] The image's region of interest (ROI) is preprocessed using 14 types of filters, including 12 filters from the simpleitk library and 2 filters from pyradiomics. The filtering method used in this embodiment is shown below:

[0068] (3.1) Additive Gaussian Noise Filter:

[0069] This filter can alter an image with enhanced Gaussian white noise. Enhanced Gaussian white noise can be modeled as:

[0070] I = I0 + N;

[0071] Where I is the observed image, I0 is the noise-free image, and N represents the mean μ and variance σ. 2 Normally distributed random variables:

[0072] N~N(μ,σ 2 ).

[0073] During the processing, noise and pixel grayscale values ​​are irrelevant, so the important information of the image will not be changed.

[0074] (3.2) Binomial Blur Image Filter:

[0075] This filter performs separable blurring in each dimension of the image. Binomial blurring involves calculating the nearest neighbor average for each image dimension. The final result after n iterations approximates a Gaussian convolution.

[0076] (3.3) Box Mean Filter:

[0077] Implement fast rectangular mean filtering using the accumulator method.

[0078] (3.4) Box Sigma Image Filter:

[0079] Implement fast rectangular sigma filtering using the accumulator method.

[0080] (3.5) Curvature Flow Filter:

[0081] Denoising of the image is achieved using curvature-driven flow, without affecting the image's boundary information; smoothing occurs only within the region. Isoluminance contours in the input CT grayscale image are considered a level set, and the level levels evolve using a curvature-based velocity function.

[0082] I t =κ|▽I|;

[0083] Where k is the curvature.

[0084] Continuous use of this filter will cause each contour to shrink to zero and eventually disappear, removing all edge information. Since the image to be denoised in this invention is already a level set, unlike level set segmentation algorithms, the `SetInput()` method will be used to directly set it as input. This invention requires manual adjustment of two parameters for the filter: the number of update iterations and the time step between each update. To ensure numerical stability, the time step must meet the CFL (Courant-Friedrichs-Levy) condition, making the step size sufficiently small. CFL effectively limits the contour movement of each level set to less than one grid position / one time step.

[0085] This invention utilizes a filter that leverages the hierarchical structure of a multi-threaded finite difference solver. It employs a Curvature Flow Function to compute zero-throughput Neumann boundary conditions for the derivatives near image data boundaries, updating the object computation. To support streaming of this filter, the invention produces a padded output that considers edge effects. The padded size is m_NumberOfIterations on each edge, using a valid center region.

[0086] (3.6) Discrete Gaussian Filter:

[0087] Blur the image through separable convolution with a discrete Gaussian kernel. This filter performs Gaussian blurring by performing a separable convolution of the image and a discrete Gaussian operator (kernel). The Gaussian kernel used here is designed based on Tony Lindeberg's Gaussian operator to compensate for smoothing and derivation operations after smoothing and discretization.

[0088] (3.7) Laplacian Sharpening Filter:

[0089] This filter uses the Laplacian operator to sharpen the image. Laplacian sharpening highlights areas with rapidly changing intensity, thus emphasizing or enhancing edges. The result is a sharper visual effect in the processed image.

[0090] (3.8) Mean Filter:

[0091] The mean filter is applied in image preprocessing. It calculates the mean value for a given pixel as the average of the pixels surrounding the corresponding input pixel. The mean filter is one of a family of linear filters.

[0092] (3.9) Normalize Filter:

[0093] This invention uses this filter to normalize an image by setting its mean to zero and its variance to 1. The Normalize Image Filter can shift and scale an image so that the mean and unit variance of the pixels in the image are zero. This invention uses the Statistics Image Filter to calculate the mean and variance of the input, and then applies the Shift Scale Image Filter to shift and scale the pixels. However, because this filter normalizes the data to between -1 and 1, integer types will generate images with no unit variance.

[0094] (3.10) Recursive Gaussian Filter

[0095] This filter is used to compute the base class of IIR convolutions with Gaussian kernel approximations:

[0096]

[0097] The Recursive Gaussian Image Filter is the base class for recursive filters that approximate convolution with a Gaussian kernel. For multi-component images, the filter acts independently on each component.

[0098] (3.11) Shot Noise Filter

[0099] This filter processes images with shot noise. Shot noise follows a Poisson distribution.

[0100] I = N(I0);

[0101] Where N(I0) is a Poisson distributed random variable with mean I0. Therefore, the noise depends on the pixel intensity in the image. The intensity in the image can be scaled by a user-provided value to map the pixel value to the actual number of particles. The scaling factor can be viewed as the reciprocal of the gain used during acquisition. The noise signal is then scaled back to its input intensity range:

[0102]

[0103] Where s is the scaling factor.

[0104] The Poisson distribution variable λ is calculated using the following algorithm:

[0105] k←0

[0106] p←1

[0107] Repeat

[0108]

[0109] Until p>e λ

[0110] Return(k);

[0111] U() provides a uniformly distributed random variable with intervals [0,1]. For large values ​​of λ, the algorithm is inefficient. However, when λ is sufficiently large, the Poisson distribution can be accurately approximated by a Gaussian distribution with mean and variance λ. This invention uses a value of 50 to make the algorithm run faster.

[0112]

[0113] Where N() is a normally distributed random variable with mean 0 and variance 1.

[0114] (3.12) Speckle Noise Filter

[0115] Iterate over an image with speckle (multiplicative) noise. The speckle noise follows a user-provided gamma distribution with a mean of 1 and a standard deviation of 1. The noise is proportional to the pixel intensity. The algorithm is modeled as follows:

[0116] I = I0 * G;

[0117] Where G is a gamma-distributed random variable with a mean of 1 and a variance proportional to the noise level:

[0118]

[0119] (3.13) LoG Filter (Laplace Gaussian Filter)

[0120] This filter requires calculating the Laplacian Gaussian operator (LoG) of the image. The LoG of the image is calculated by convolving it with the second derivative of the Gaussian operator, implemented using a recursive Gaussian filter.

[0121] (3.14) Wavelet Filter

[0122] This filter employs wavelets, which are short, wave-like oscillations whose amplitude starts at zero, increases, decreases, and then returns to zero. Wavelet filters can be used to extract signal features using convolutional techniques of "inversion, shifting, multiplication, and integration." Wavelets are combined with known portions of the corrupted signal to extract information from the unknown parts. Wavelets can be applied to various types of data to extract the unknown information required by this invention, including audio signals and image information used in this invention. First, a set of wavelets processes and analyzes the input known data; then, a complementary set of sub-wavelets reversibly decomposes the data, ensuring no gaps or overlaps between data points. Using a wavelet-based compression / decompression algorithm, the complementary wavelets recover the original unknown information with minimal loss.

[0123] The original image, after processing with the aforementioned filter, yielded 25 different images. Feature extraction was performed on each image, resulting in 104 imaging features across 7 categories. In other words, 104 imaging features were extracted for each image, ultimately combining to extract a total of 2600 dimensions (2600 = 104 × 25) of imaging features.

[0124] The seven categories of imaging features are first-order features, shape features, gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), gray-level dependence matrix (GLDM), and neigbouring gray tonedifference matrix (NGTDM). The 104 imaging features are shown in Table 1.

[0125] Table 1 104 imaging features

[0126]

[0127]

[0128]

[0129] 2. Clinical feature extraction

[0130] Table 2 shows the 21-dimensional clinical characteristics of patients in three groups with different gene mutation states: gender, age (years), smoking history, family history of lung cancer, family history of non-lung cancer, daily number of cigarettes smoked (cigarettes), number of years of smoking, whether or not smoking has been stopped, number of years since smoking was stopped, whether or not alcohol consumption has occurred, duration of alcohol consumption (years), amount of alcohol consumed (grams / day), presence or absence of cough, presence or absence of chest pain, presence or absence of hemoptysis, presence or absence of sputum production, carcinoembryonic antigen (CEA, ng / ml), cancer antigen 125 (CA125, U / ml), carbohydrate antigen 199 (CA199, U / ml), cytokeratin 19 fragment (CYFRA21-1, ng / ml), and neuron-specific enolase (NSE, ng / ml).

[0131] Table 2 Summary of 21-dimensional clinical characteristics of patients in the three gene groups

[0132]

[0133]

[0134] Step 2: Feature Selection

[0135] The Lasso algorithm (least absolute shrinkage and selection operator, LASSO) was used to select from 2600-dimensional imaging features and 21-dimensional clinical features (a total of 2621 features). The specific operation is as follows:

[0136] The data was trained and tested using a five-fold crossover method, and the alpha parameter was adjusted based on the area under the curve (AUC) value. Figure 1 For each alpha value, the purple dot represents the selected feature, resulting in the AUC value. The coefficients of each feature are selected by alpha. When alpha(0,1) reaches a certain value, further increasing the number of features included in the model, i.e., decreasing the alpha value, does not significantly improve model performance. For 2621-dimensional imaging and clinical features, the final alpha value was adjusted to 0.0152, and the top 108 features with the highest coefficients were selected. Figure 2 The example shows the coefficient generation path for different features. Each curve in the figure represents the trajectory of the feature coefficient. The vertical axis is the value of the coefficient, and the horizontal axis is log(alpha). When the alpha value is specified, the specific coefficient of the feature can be determined. Table 3 lists the specific features with the top ten non-negative coefficients and their corresponding coefficient values.

[0137] Table 3 shows the specific characteristics and corresponding coefficient values ​​of the top ten nonnegative coefficients selected by LASSO.

[0138] feature coefficient cea 0.055265 capacity for liquor 0.05185 discretegaussian_glrlm_runvariance 0.049149 Quit smoking 0.045633 gender 0.043102 wavelet_glcm_wavelet-hlh-correlation 0.042925 log_glszm_log-sigma-0-5-mm-3d-sizezonenonuniformitynormalized 0.037892 discretegaussian_glszm_largearealowgraylevelemphasis 0.03499 wavelet_glcm_wavelet-hlh-idn 0.034847 wavelet_glszm_wavelet-lhh-graylevelvariance 0.031228

[0139] Step 3: Build a model using a random forest classifier and perform five-fold cross-validation.

[0140] First, the data sequence is randomly shuffled. Then, a five-fold crossover method is used for training and testing: the data of three groups of patients with different gene mutation states are divided into five mutually exclusive (no overlap between each pair) parts according to the number of each label. Each time, one part is selected as the test set and the other four parts are selected as the training set. The training set and test set results can be obtained each time. Finally, the five training / test results are averaged.

[0141] Random forest is a supervised learning algorithm based on decision trees. It combines multiple randomly created decision trees to form a "forest." The 108-dimensional features selected from the LASSO dataset above are used to build a model using a random forest classifier, resulting in the non-invasive prediction model for EGFR / TP53 co-mutant lung cancer patients presented in this invention.

[0142] When modeling, the random forest classifier uses the following parameters: class weights are Balanced, the metric is Entroy, maximum depth is 6, minimum number of samples per leaf node is 3, minimum number of samples per split is 2, number of weak classifiers is 2000, and classification threshold is 0.5.

[0143] The model for non-invasive prediction of EGFR / TP53 co-mutated lung cancer patients established above was used to test the test set data, and the results were output.

[0144] The efficacy of the non-invasive model for predicting EGFR / TP53 co-mutant lung cancer patients established in this invention was evaluated using the area under the characteristic curve (ROC) (AUC). Figure 3 This represents the ROC curves for each fold of the five-fold cross-validation test, for both the training and validation sets. The results show that the AUC values ​​for the training and test sets are 0.807 and 0.746, respectively. Figure 4 The accuracy on the training and test sets was 0.602 and 0.544, respectively; the precision was 0.604 and 0.561, respectively; the recall was 0.599 and 0.549, respectively; and the F1 score (H-mean score) was 0.6 and 0.547, respectively.

[0145] The above results demonstrate that the model established in this invention can accurately and non-invasively predict the EGFR / TP53 gene co-mutation status of lung cancer patients. The model has an AUC value as high as 0.746 on the test set. The model established in this invention has important guiding significance for the clinical precision treatment of lung cancer patients with EGFR / TP53 co-mutation.

[0146] Example 2: The Artificial Intelligence System for Non-invasive Prediction of Lung Cancer Patients with EGFR / TP53 Co-mutations of the Present Invention

[0147] The patient data used in this embodiment are from 1055 patients who had plain CT images with a slice thickness of 1 mm within 60 days before the pathological diagnosis, out of 2171 patients who were first diagnosed with primary NSCLC at West China Hospital of Sichuan University from January 2013 to December 2019 (same as in Embodiment 1).

[0148] The gene mutation status of all patients was tested. The results showed that among the 1055 patients, there were 339 in the EGFR+ / TP53+ group, 409 in the EGFR+ / TP53- group, and 307 in the EGFR- / TP53- group.

[0149] Part 1: Data Input Section

[0150] Extract 2600-dimensional imaging features and 21-dimensional clinical features according to the method in step one of Example 1, and input them into the system.

[0151] Part Two: Feature Selection

[0152] Following the method in step 2 of Example 1, LASSO was used to select 2600-dimensional imaging features and 21-dimensional clinical features (a total of 2621 features), and 108 features were selected.

[0153] Part Three: Model Building

[0154] The data from three groups of patients with different gene mutation states were divided into five mutually exclusive sets (no overlap between any two sets) according to the number of each label. One set was selected as the test set each time, and the other four sets were selected as the training set.

[0155] Following the method in step 3 of Example 1, the 108-dimensional features selected by LASSO in the training set were used to build a model using a random forest classifier, resulting in a non-invasive model for predicting lung cancer patients with EGFR / TP53 co-mutations.

[0156] Part Four: Prediction of Lung Cancer Patients with EGFR / TP53 Co-mutations

[0157] The trained model is used to predict the gene mutation status of patients in the test set to determine whether the patient has EGFR / TP53 co-mutated lung cancer.

[0158] The following experimental examples demonstrate the beneficial effects of the present invention.

[0159] Experimental Case 1: Clinicopathological Features and Prognosis of Non-Small Cell Lung Cancer Patients with EGFR / TP53 Co-mutations

[0160] 1. Research Subjects

[0161] This study included 2171 patients who were initially diagnosed with primary NSCLC at West China Hospital of Sichuan University between January 2013 and December 2019, and who visited the outpatient or inpatient departments. The invention involved obtaining informed consent from patients beforehand. After approval from the Ethics Committee of West China Hospital of Sichuan University, tissue samples from the primary or metastatic lesions of the lung cancer were collected and subjected to high-throughput sequencing of 56 target genes using a panel.

[0162] Inclusion criteria: 1. Pathologically confirmed primary NSCLC; 2. Informed consent given by the patient; 3. The patient was newly diagnosed and had not received any lung cancer-related treatment.

[0163] Exclusion criteria: Tumor tissue could not be obtained or high-throughput sequencing could not be performed (tumor cell content <20%).

[0164] 2. Collection of clinical, imaging, pathological, genetic, and prognostic information

[0165] 2.1 Clinical Features

[0166] The patient's basic clinical information was collected and verified through the electronic pathology system of West China Hospital of Sichuan University. Clinical characteristics included: age at the time of initial pathological diagnosis, smoking history, family history of lung cancer, family history of non-lung cancer malignant tumors, TNM stage of lung cancer at the time of initial diagnosis, status of distant metastases, and status of lesions in both lungs.

[0167] 2.2 Image Features

[0168] CT images taken within 30 days prior to the initial pathological diagnosis of patients were retrieved using the Picture Archiving and Communication System (PACS) of West China Hospital, Sichuan University. The maximum slice diameter (mm) of the lesions and gross imaging characteristics were measured. The included CT examinations included 5mm slice thickness plain scans, contrast-enhanced scans, and 0.625mm and 1mm thin-slice CT scans.

[0169] 2.3 Pathological features

[0170] Immunohistochemical staining of pathological tissues was performed according to NCCN guidelines: TTF1, Napsin A, CK7, P63, and P40, to classify lung cancer tissue types, including lung adenocarcinoma, lung squamous cell carcinoma, and other types of NSCLC. For lung adenocarcinoma, further subtyping was performed, with key components including papillary, acinar, lepidic, and micropapillary structures, and the degree of histological differentiation was assessed.

[0171] 2.4 Gene Sequencing

[0172] The target gene panel was designed to include 56 genes that have been shown in existing research to be associated with the development and progression of non-small cell lung cancer: AKT1, ALK, ARAF, ATM, BIM, BRAF, BRCA1, BRCA2, CCND1, CDK4, CDK6, CDKN2A, CTNNB1, CYP2D6, DDR2, DPYD, EGFR, ERBB2, ERBB3, ERBB4, FGF19, FGF3, FGF4, and FGF. R1, FGFR2, FGFR3, FLT3, HRAS, JAK1, JAK2, KDR, KIT, KRAS, MAP2K1, MET, MTOR, MYC, NRAS, NRG1, NTRK1, NTRK2, NTRK3, PDGFRA, PIK3CA, PTCH1, PTEN, RAF1, RB1, RET, ROS1, SMO, STK11, TP53, TSC1, TSC2, UGT1A1. The status of these 56 genes in the patients was tested.

[0173] 2.5 Survival Prognosis

[0174] The patient's medical records were verified through the electronic medical record system of West China Hospital of Sichuan University. Telephone follow-ups ended on December 31, 2020. The primary endpoint for all included patients was overall survival (OS), defined as the time interval from diagnosis of primary lung cancer to death from any cause. OS is the primary outcome measure of this invention.

[0175] For patients who received EGFR-TKI as first-line treatment at West China Hospital of Sichuan University, and who did not concurrently receive immunotherapy, radiotherapy, chemotherapy, or surgery, and who were regularly followed up (with whole-body CT imaging), disease progression was evaluated according to the Response Evaluation Criteria in Solid Tumor (RECIST) 1.1, and progression-free survival (PFS) was calculated. PFS is the time from the start of TKI treatment to the onset of disease progression or death, or to the end of follow-up. PFS is a secondary endpoint of this invention.

[0176] 3. Statistical Analysis

[0177] SPSS 23.0 (SPSS Inc., Chicago, USA) was used for statistical analysis, and chi-square (χ²) was used. 2 Intergroup comparisons were performed using the p-test or Fisher's exact test. A two-tailed test was used, and p < 0.05 was considered statistically significant.

[0178] 4. Experimental Results

[0179] 4.1 Overall Population Characteristics and Survival Prognosis Analysis

[0180] The average age of the 2171 patients was 59.1 ± 11.3 years. The gender distribution was as follows: 1153 males (53.1%) and 1018 females (46.9%). Regarding smoking history, 703 patients (3.4%) were former or current smokers, 1235 patients (56.9%) were non-smokers, and 233 patients (10.7%) had unknown smoking history. Among the total population, 307 patients (14.1%) had a family history of cancer, and 1864 patients (85.9%) had no family history of cancer. Of these, 111 patients (5.1%) had a family history of lung cancer, and 205 patients (9.4%) had a family history of other types of malignant tumors. According to TNM staging, there were 754 patients in stage I (34.7%), 117 patients in stage II (5.4%), 336 patients in stage III (15.5%), and 775 patients in stage IV (35.7%). 189 patients could not be evaluated for TNM staging (8.7%).

[0181] Of all the patients, 1,789 (82.4%) had adenocarcinoma; 244 (11.2%) had squamous cell carcinoma; and 138 (6.4%) had other types of NSCLC.

[0182] The 56-gene analysis results of 2171 NSCLC patients showed that the patients were involved in 13 types of gene mutations: non-frameshifting deletion, non-frameshifting insertion, nonsense mutation, missense mutation, in-frame insertion, in-frame deletion, splice site mutation, frameshift mutation, start codon deletion, stop codon deletion, gene fusion / re-arrangement, copy number amplification, and copy number loss. The results showed that EGFR was the most frequently mutated gene in NSCLC patients, with 1128 patients exhibiting EGFR mutations (52.0%, 1128 / 2171). TP53 was the second most frequently mutated gene, with 1103 patients exhibiting TP53 mutations (50.8%, 1103 / 2171). The remaining eight genes among the top ten most frequently mutated genes were: KRAS (12%), PIK3CA (10%), ERBB2 (7%), MET (7%), ATM (6%), CDKN2A (5%), CDK4 (5%), and ROS1 (5%). In the selected 56-gene panel sequencing, eight genes—NRG1, BIM, CYP2C19, CYP2D6, CYP3A4, DPYD, NF1, and UGT1A1—did not show significant mutations. EGFR and TP53 mutations were predominantly missense mutations. Of the top two mutated genes, 535 patients (24.6%) had EGFR mutations combined with TP53 mutations.

[0183] A summary of lung cancer treatments revealed that 1,133 patients (52.5%) underwent surgery, 234 patients (10.8%) received stereotactic radiotherapy to the lung lesions, 635 patients (29.2%) received chemotherapy, 579 patients (26.7%) received targeted therapy, 105 patients (4.8%) received anti-angiogenic therapy, and 88 patients (4.1%) received immunotherapy.

[0184] Based on outpatient follow-up and telephone follow-up information, relevant information was collected from patients who had not received anti-tumor therapy before targeted testing, and these patients were followed up. A total of 1447 patients were included who were followed up for more than one year or experienced an outcome event (death). Overall survival results are as follows... Figure 5 As shown, the 1-year survival rate for patients was 78.93%.

[0185] 4.2 Relationship between EGFR / TP53 gene mutations and patient survival prognosis

[0186] Univariate / multivariate Cox regression models were used to analyze the impact of clinical, pathological, and genetic factors on one-year overall survival (OS) in NSCLC patients. The results are shown in Tables 4 and 5. Age greater than 65 years, TNM stage III-IV, KARS mutation, and TP53 mutation were associated with shorter OS in NSCLC patients; while EGFR mutation, targeted therapy, and surgical resection were associated with longer OS.

[0187] Table 4 Univariate / multivariate analysis of overall survival (OS) in NSCLC patients

[0188]

[0189]

[0190] Table 5 Cox Regression Assignment Table

[0191]

[0192] Overall survival (OS) for patients in the EGFR- / TP53- group, EGFR+ / TP53- group, EGFR- / TP53+ group, and EGFR+ / TP53+ group is as follows: Figure 6 As shown, the overall survival (OS) differed among different gene groups (P<0.001). The 1-year survival rate was 80.24% for EGFR- / TP53- patients; 96.30% for EGFR+ / TP53- patients; 57.81% for EGFR- / TP53+ patients; and 84.03% for EGFR+ / TP53+ patients.

[0193] Subgroup analysis was performed based on TNM staging to compare 1-year overall survival (OS) between EGFR+ / TP53- and EGFR+ / TP53+ patients with early (I, II) and late (III, IV) clinical stages.

[0194] Figure 7Survival curves for EGFR+ / TP53+ and EGRF+ / TP53- patients with TNM stage I and II are shown. At the end of the follow-up period, only two patients in the EGFR+ / TP53+ group died, with a 1-year survival rate of 97.5%, while the 1-year survival rate in the EGRF+ / TP53- group was 100%. The difference between the two groups was statistically significant (P = 0.032).

[0195] Figure 8 Survival curves for EGFR+ / TP53+ and EGFR+ / TP53- patients with TNM stage III and IV are shown. The 1-year survival rate of the EGFR+ / TP53+ group was 81.7%, which was lower than that of the EGFR+ / TP53- group (90.2%), and the difference was statistically significant (P = 0.033).

[0196] An analysis of overall survival (OS) in EGFR-mutant patients with stage III and IV EGFR-mutant disease treated with TKIs showed that the 1-year survival rate was 85.9% for EGFR+ / TP53+ patients and 95.9% for EGFR+ / TP53- patients, with a statistically significant difference (P = 0.010). Figure 9 ).

[0197] An analysis of overall survival (OS) in stage III and IV EGFR-mutant patients who had not received TKI treatment revealed that the 1-year survival rate was 63.6% for EGFR+ / TP53- patients and 69.2% for EGFR+ / TP53+ patients. The difference between the two groups was not statistically significant (P = 0.660). Figure 10 ).

[0198] This study analyzed 79 patients who received first-line EGFR-TKI therapy at West China Hospital and were followed up regularly every 4-6 weeks (with whole-body CT imaging). At the end of the follow-up period, 4 patients had not experienced disease progression. Results showed that the median progression-free survival (PFS) was 8.4 months (7.3-9.5) for EGFR+ / TP53+ patients and 10.5 months (8.3-12.7) for EGFR+ / TP53- patients. EGFR+ / TP53+ patients were more likely to experience disease progression than EGFR+ / TP53- patients, and the difference in PFS between the two groups was statistically significant (P = 0.015). Figure 11 ).

[0199] The above results indicate that among the 2171 primary NSCLC patients included in this study, 24.6% had EGFR+ / TP53+ co-mutations (535 / 2171). There were significant differences in survival prognosis among patients with different gene mutation statuses: (1) The 1-year survival rates of the EGFR+ / TP53+ group, EGFR- / TP53- group, EGFR+ / TP53- group, and EGFR- / TP53+ group were 84.03%, 80.24%, 96.30%, and 57.81%, respectively, with statistically significant differences (P<0.001); (2) In the early stages (stages I+II): the 1-year survival rate of EGFR+ / TP53+ patients was lower than that of EGFR+ / TP53- patients (97.5% vs. 100%, P=0.00%). 032); (3) Advanced (III+IV stage): EGFR+ / TP53+ patients had a lower 1-year survival rate than EGFR+ / TP53- patients (81.7% vs. 90.2%, P=0.033); (4) Subgroup analysis of EGFR mutation positive patients and first-line EGFR-TKI treatment: The median PFS of EGFR+ / TP53+ patients was significantly shorter than that of EGFR+ / TP53- patients (months, median) [8.4 (7.3-9.5) vs. 10.5 (8.3-12.7), P=0.015].

[0200] Therefore, accurately predicting EGFR+ / TP53+ co-mutated patients in NSCLC has important guiding significance for the clinical precision treatment of EGFR / TP53 co-mutated lung cancer patients.

[0201] Experimental Example 2: Establishing an imaging prediction model for EGFR / TP53 co-mutant lung cancer patients (control model 1)

[0202] The patient data used in this experiment were from 1055 patients who had plain CT images with a slice thickness of 1 mm within 60 days before the pathological diagnosis, out of 2171 patients who were first diagnosed with primary NSCLC at West China Hospital of Sichuan University between January 2013 and December 2019 (same as Example 1).

[0203] Step 1: Image Feature Extraction

[0204] Using the same method as step 1 of the first step in Example 1, 2600-dimensional imaging features were extracted.

[0205] Step 2: Feature Selection

[0206] LASSO was used to select from 2600-dimensional imaging features and 21-dimensional clinical features (a total of 2621 features). The specific steps are as follows:

[0207] Five-fold crossover was used for training and testing, and the alpha parameter was adjusted based on the AUC value. Figure 12 Because of cross-validation, for each alpha value, the purple dots represent the selected features, and the mean of the simplified AUC calculated using the Radscore formula is obtained. It can be seen that once alpha(0,1) reaches a certain value, further increasing the number of features included in the model, i.e., reducing the alpha value, does not significantly improve model performance. For 2600-dimensional imaging features, the alpha value is adjusted to 0.005, and the top 253 features with the highest feature coefficients are selected. Figure 13 The table shows the top ten features with the highest coefficients. Table 6 lists the specific names and coefficient values ​​of the top ten features by absolute value. In the subsequent model building, features with larger coefficients have a greater impact on the model results. Figure 14 An example of determining the coefficients for the top ten features with the highest absolute values ​​is shown in the figure. Each curve in the figure represents the trajectory of the feature coefficient. The vertical axis is the value of the coefficient, and the horizontal axis is log(alpha). When the alpha value is specified, the specific value of the feature coefficient can be determined.

[0208] Table 6. Top 10 image features with high LASSO feature selection coefficients

[0209] feature coefficient laplaciansharpening_ngtdm_busyness -0.101054326 boxsigmaimage_firstorder_uniformity -0.074092664 discretegaussian_glrlm_runlengthnonuniformity -0.058647547 discretegaussian_glszm_graylevelvariance -0.055099413 wavelet_glrlm_wavelet-hhl-runvariance 0.053083416 wavelet_glcm_wavelet-hlh-idn 0.049407207 normalize_glszm_graylevelnonuniformitynormalized 0.037800465 boxsigmaimage_gldm_largedependencehighgraylevelemphasis -0.036993694 wavelet_glrlm_wavelet-lhl-runentropy 0.03462181 log_gldm_log-sigma-1-5-mm-3d-largedependencehighgraylevelemphasis -0.033199564

[0210] Step 3: Build a model using a random forest classifier and perform five-fold cross-validation.

[0211] The 253-dimensional features selected by LASSO were used to build a model using a random forest classifier, resulting in a non-invasive imaging prediction model for lung cancer patients with EGFR / TP53 co-mutations.

[0212] Referring to the method in Example 1, the efficacy of the non-invasive imaging prediction model for predicting EGFR / TP53 co-mutant lung cancer patients was evaluated using the area under the characteristic curve (ROC) (AUC).

[0213] The results show ( Figure 15 The AUC values ​​of the image prediction model on the training and test sets were 0.993 and 0.719, respectively.

[0214] Compared with the image prediction model established in this experimental example, the AUC value of the non-invasive prediction model (clinical + image prediction model) for lung cancer patients with EGFR / TP53 co-mutation established in Example 1 of this invention was significantly improved (AUC value = 0.746), indicating that the non-invasive prediction model for lung cancer patients with EGFR / TP53 co-mutation established in Example 1 of this invention has a better predictive effect on the EGFR / TP53 gene co-mutation status of lung cancer patients.

[0215] Experimental Example 3: Establishing a clinical predictive model for lung cancer patients with EGFR / TP53 co-mutations (control model 2)

[0216] The patient data used in this experiment were from 1055 patients who had plain CT images with a slice thickness of 1 mm within 60 days before the pathological diagnosis, out of 2171 patients who were first diagnosed with primary NSCLC at West China Hospital of Sichuan University between January 2013 and December 2019 (same as Example 1).

[0217] Step 1: Image Feature Extraction

[0218] Using the same method as step 2 of the first step in Example 1, 21-dimensional clinical features were extracted.

[0219] Step 2: Feature Preprocessing

[0220] The L1 / L2 method is used to preprocess the features.

[0221] Step 3: Build a model using a random forest classifier and perform five-fold cross-validation.

[0222] The 21-dimensional clinical features obtained from the preprocessing were used to build a model, resulting in a non-invasive clinical prediction model for lung cancer patients with EGFR / TP53 co-mutations.

[0223] Referring to the method in Example 1, the efficacy of the non-invasive clinical prediction model for predicting EGFR / TP53 co-mutant lung cancer patients was evaluated using the area under the characteristic curve (ROC) (AUC).

[0224] The results show ( Figure 15 The AUC values ​​for the training and test sets of the clinical prediction model were 0.761 and 0.714, respectively.

[0225] Compared with the clinical prediction model established in this experiment, the AUC value of the non-invasive prediction model (clinical + imaging prediction model) for EGFR / TP53 co-mutated lung cancer patients established in Example 1 of this invention was significantly improved (AUC value = 0.746), indicating that the non-invasive prediction model for EGFR / TP53 co-mutated lung cancer patients established in Example 1 of this invention has a better predictive effect on the EGFR / TP53 gene co-mutation status of lung cancer patients.

[0226] In summary, this invention, by collecting patients' clinical and imaging characteristics and establishing a model using LASSO feature selection and a random forest classifier, has developed an artificial intelligence prediction system capable of accurately and non-invasively predicting EGFR / TP53 gene co-mutations in lung cancer patients. Experiments show that compared to clinical and imaging prediction models, the AI ​​prediction system developed in this invention demonstrates superior predictive performance for EGFR / TP53 gene co-mutations in lung cancer patients, achieving an AUC value as high as 0.746 on the test set. This AI prediction system provides a new option for clinical screening of lung cancer patients with EGFR+ / TP53+ co-mutations, as well as for screening EGFR-mutant lung cancer patients resistant to TKIs, and has significant guiding significance for the precise clinical treatment of lung cancer patients with EGFR / TP53 co-mutations.

Claims

1. An artificial intelligence system for predicting genetic mutation status of a lung cancer patient, characterized in that: The artificial intelligence system comprises the following five parts: The first part: data input part; input the imaging features and clinical features of lung cancer patients; The second part: feature selection part; The imaging features and clinical features input in the first part are selected to obtain selected features; The third part: model construction part; the selected features in the second part are divided into training set data and test set data, and a random forest classifier is trained with the training set data to construct a prediction model; The fourth part: prediction part; the prediction model constructed in the third part is used to process the test set data to determine the gene mutation of lung cancer patients as EGFR / TP53 co-mutation or non-EGFR / TP53 co-mutation; In the first part, the imaging features are obtained by the following method: (1) The plain CT image of the lung cancer patient is segmented by the region of interest to obtain the image region of interest of the lesion; (2) The image region of interest is respectively preprocessed by 14 types of filters to obtain preprocessed images; the 14 types of filters are: additive Gaussian noise filter, binomial image blur filter, box mean filter, box sigma image filter, curvature flow filter, discrete Gaussian filter, Laplacian sharpening filter, mean filter, normalization filter, recursive Gaussian filter, shot noise filter, speckle noise filter, Gaussian Laplace filter, wavelet filter; (3) The image features of the image region of interest without preprocessing and the preprocessed image are extracted respectively to obtain 7 types of imaging features: First-order features, shape features, gray level co-occurrence matrix, gray run length matrix, gray region size matrix, gray dependence matrix, and adjacent gray difference matrix; In the first part, the clinical features are the following 21-dimensional clinical features: gender; age, unit: years; smoking history; lung cancer family history; non-lung cancer family history; daily smoking amount, unit: cigarettes; smoking years; whether to quit smoking; years of quitting smoking; whether to drink; drinking time, unit: years; drinking amount, unit: grams / day; whether to cough; whether to have chest pain; whether to have hemoptysis; whether to have sputum; carcinoembryonic antigen, unit ng / ml; cancer antigen 125, unit U / ml; carbohydrate antigen 199, unit U / ml; cytokeratin 19 fragment, unit ng / ml; neuron-specific enolase, unit ng / ml; In the second part, the selection method is to use the lasso algorithm, set the alpha value to 0.0152, and select the top 108 features with the highest feature coefficients from the imaging features and clinical features. In the third part, the parameters of the random forest classifier are set as follows: the class weight uses Balance, the index selects Entroy, the maximum depth is 6, the minimum sample number of leaf node is 3, the minimum split sample number is 2, the number of weak classifiers is 2000, and the classification threshold is 0.

5. 2.The artificial intelligence system of claim 1, wherein: The lung cancer is non-small cell lung cancer. 3.The artificial intelligence system of any one of claims 1-2, wherein: EGFR 4.The artificial intelligence system of claim 3, wherein: The non-small cell lung cancer is EGFR Mutated non-small cell lung cancer.

5. The artificial intelligence system of claim 4, wherein: The The device stores the artificial intelligence system of any one of claims 1-5. Mutant non-small cell lung cancer is resistant to tyrosine kinase inhibitors.

6. An apparatus for predicting a genetic mutation profile of a lung cancer patient, comprising: ​ 7. Use of the artificial intelligence system according to any one of claims 1 to 5 in the preparation of a device for predicting the genetic mutation status of a lung cancer patient, which can predict the genetic mutation status of a lung cancer patient as EGFR / TP53 co-mutation or non-EGFR / TP53 co-mutation.

Images