Tumor margin identification method and system based on mass spectrometry data

Abstract

The present application belongs to the technical field of tumor margin identification, and particularly relates to a tumor margin identification method and system based on mass spectrum data, which comprises the following steps: based on initial mass spectrum data, adopting a triple complementary strategy to perform missing value processing, combining principal component analysis to remove abnormal samples, and then obtaining high-confidence candidate markers through multi-mechanism feature screening and intersection; taking the candidate markers as input and the margin state as label, adopting LASSO regression to construct an identification model, and optimizing the regularization parameter through cross-validation; comparing the identification result with the postoperative pathological gold standard, and then arranging and storing it in the database; when the accumulation reaches the threshold, triggering model retraining, and correcting the model parameters by using feedback samples. Through triple complementary missing processing and multi-mechanism feature screening intersection, the present application effectively reduces the false positive rate, combines a closed-loop iteration mechanism to realize continuous evolution of the model, and significantly improves the accuracy and clinical adaptability of tumor margin identification.

Description

Technical Field

[0001] This invention belongs to the field of tumor margin recognition technology, and particularly relates to a method and system for tumor margin recognition based on mass spectrometry data. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] During tumor surgery, the surgical margin status is a key factor affecting postoperative recurrence; negative margins can significantly reduce the recurrence rate of cancer. Although traditional pathological testing methods can provide intraoperative margin information, they require cumbersome steps such as freezing, sectioning, staining, and microscopic observation, resulting in problems such as poor timeliness, easy missed detection, and false negatives.

[0004] Therefore, rapid tumor margin detection technologies based on mass spectrometry have emerged in recent years, such as the iKnife system based on rapid evaporation ionization mass spectrometry. These systems establish a reference database by learning from a large number of known tissue samples before surgery, and use algorithms such as principal component analysis and linear discriminant analysis to extract the differences in mass spectrometric characteristics between normal and cancerous tissues, forming a classification model.

[0005] However, existing mass spectrometry-based intraoperative resection margin identification algorithms have significant shortcomings. The entire process is an open-loop design, and the model is statically frozen after deployment, making it unable to receive feedback from postoperative pathological gold standards. Furthermore, mass spectrometry data itself is characterized by high dimensionality, high noise, and strong collinearity, while existing methods often rely on single-feature screening algorithms, which easily generate false positives, leading to overfitting and decreased generalization ability. Therefore, there is an urgent need for a tumor resection margin identification model that can fully utilize the characteristics of mass spectrometry data and effectively suppress false positives. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a tumor margin identification method and system based on mass spectrometry data. The aim is to use a triple complementary strategy to process missing values ​​and remove abnormal samples, combine multi-mechanism feature screening to obtain high-confidence candidate biomarkers by taking the intersection, use LASSO regression to construct an identification model, and use postoperative pathological gold standard feedback to achieve closed-loop iterative optimization of the identification model.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: Firstly, a method for identifying tumor resection margins based on mass spectrometry data is disclosed, including: Based on the initial mass spectrometry data, a triple complementarity strategy was used to handle missing values. After removing outlier samples by principal component analysis, high-confidence candidate biomarkers were obtained by taking the intersection of the triple complementarity strategy. Using candidate biomarkers as input and cutting edge states as labels, a recognition model is constructed using LASSO regression to output recognition results, and the regularization parameters are optimized through cross-validation. The identification results are compared with the postoperative pathological gold standard and then stored in the database. When the accumulation reaches the threshold, the model is retrained. The identification model parameters are corrected using feedback samples and then used for subsequent identification after evaluation and verification.

[0008] Furthermore, it also includes preprocessing steps: sequentially performing asymmetric least squares baseline correction, continuous wavelet transform peak detection, peak alignment based on the reference spectrum, and total ion current normalization on the mass spectrometry signal to obtain a standardized mass spectrometry feature vector.

[0009] Furthermore, the triple complementarity strategy for missing value processing includes: calculating the missing rate of each feature in all samples and deleting features with a missing rate exceeding a threshold; for features with a missing rate not exceeding the threshold and whose missing values ​​are randomly distributed, using the k-nearest neighbor algorithm to fill in the missing values; and for missing values ​​caused by values ​​below the detection limit, using the minimum value supplementation method.

[0010] Furthermore, the method of obtaining high-confidence candidate markers through a triple complementary strategy and intersection includes: using orthogonal partial least squares discriminant analysis to screen features with variable projection importance greater than 1, using a decision tree algorithm based on the Gini coefficient to screen key splitting features, and using the rank-sum test to screen features with significant differences between the negative and positive cutting edge groups.

[0011] Furthermore, the regularization parameter for the LASSO regression is selected using a combination of 10-fold cross-validation and the 1-SE rule.

[0012] Furthermore, the process of comparing the identification results with the postoperative pathological gold standard and then organizing and storing them includes: recording the mass spectrometry characteristics, identification probability, identification status and true label of the sample, and marking and storing the correctly identified samples and the incorrectly identified samples respectively.

[0013] Furthermore, after the triggered model is retrained, the performance of the new model is evaluated through cross-validation. If the accuracy, sensitivity, or specificity is improved by more than a set threshold compared to the original model, the optimized model replaces the original model; otherwise, the original model is retained to continue accumulating samples.

[0014] Secondly, a tumor margin identification system based on mass spectrometry data is disclosed, including: The data preprocessing module is used to handle missing values ​​based on the initial mass spectrometry data using a triple complementary strategy, and to remove outlier samples by combining principal component analysis. The feature selection module is used to obtain high-confidence candidate markers by taking the intersection of a triple complementary strategy; The model building module is used to construct a recognition model with candidate markers as input and edge states as labels, using LASSO regression to output recognition results, and optimizing the regularization parameters through cross-validation. The closed-loop iteration module is used to compare the recognition results with the postoperative pathological gold standard and then organize and store them in the database. When the accumulation reaches the threshold, the model is retrained. The recognition model parameters are corrected using feedback samples and then used for subsequent recognition after evaluation and verification.

[0015] Thirdly, a computer device is disclosed, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the steps of the method described above.

[0016] Fourthly, a computer-readable storage medium is disclosed having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0017] The above one or more technical solutions have the following beneficial effects: This invention employs a triple complementary strategy combining deletion, k-nearest neighbor imputation, and minimum value supplementation to overcome the shortcomings of single imputation methods, which are prone to introducing bias or masking true information. This strategy can adaptively select the processing method based on the missing pattern, effectively preserving the biological differences in the original data and providing a high-quality data foundation for subsequent feature selection and model construction.

[0018] This invention combines three feature selection methods with vastly different principles: orthogonal partial least squares discriminant analysis, decision trees, and rank-sum tests. By taking the intersection of their selection results, it avoids the problem of selecting false positive biomarkers that is common with single methods. This "voting" selection mechanism can eliminate spurious associations caused by random noise, batch effects, and collinearity, obtaining candidate biomarkers with clear biological significance and high cross-batch repeatability. This significantly reduces the risk of overfitting in subsequent identification models and improves generalization ability.

[0019] This invention establishes a closed-loop iterative mechanism based on postoperative pathological gold standard feedback. After model deployment, intraoperative identification results are compared with final pathological results, and tagged samples are stored in a database. When the accumulated data reaches a set threshold, model retraining is automatically triggered, using newly added real feedback data to correct model parameters. Compared to existing open-loop systems, this invention can continuously adapt to changes in data distribution caused by different hospitals, different mass spectrometry platforms, and different patient groups, allowing model performance to continuously evolve with clinical use and avoiding the lag and inefficiency of manual periodic optimization.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of the tumor margin identification method based on mass spectrometry data according to Embodiment 1 of the present invention; Figure 2 This is a block diagram of a tumor margin identification system based on mass spectrometry data according to Embodiment 1 of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Example 1 like Figure 1 As shown, this embodiment discloses a tumor margin identification method based on mass spectrometry data, including: Step 1: Based on the initial mass spectrometry data, a triple complementarity strategy is used to handle missing values, and principal component analysis is combined to remove outlier samples, resulting in a sample feature set. Specifically: Raw mass spectrometry data of tumor tissue samples were acquired using a mass spectrometry acquisition device. Based on the initial mass spectrometry data, baseline correction, peak detection, peak alignment, and normalization preprocessing were performed sequentially to obtain a standardized mass spectrometry feature matrix. Specifically, baseline correction involves subtracting the baseline for each mass spectrum signal using the asymmetric least squares (AsLS) method, with the parameters set as follows: smoothing factor λ = 10. 5 The asymmetric weight p=0.01 and the number of iterations=10. This step can effectively eliminate mass spectrometry baseline drift.

[0027] Peak detection specifically employs the Continuous Wavelet Transform (CWT) algorithm, with the second-order Ricker wavelet as the wavelet basis function. The scale range is set to 2~30, and the step size is 1.

[0028] Wavelet coefficients are calculated at each scale, local maxima are traced along the scale direction to form ridges, and the ridges are inversely calculated to the original mass axis. The screening criteria are set as signal-to-noise ratio (SNR) ≥ 3 and full width at half maximum (FWHM) between 3 and 15 data points. The mass-to-charge ratio (m / z) and intensity of each peak are output.

[0029] Peak alignment was set to use a peak matching method based on the reference spectrum. The reference spectrum was constructed as follows: 10 typical samples were randomly selected from the training set, and the Gaussian kernel density (bandwidth = 0.1 Da) of the peak list of each sample was estimated. Local maxima were taken as reference peaks, resulting in approximately 125 reference peaks.

[0030] The alignment strategy involves matching the peak of each measured sample with the reference peak within a quality error window of ±0.2 Da (or ±50 ppm). Unmatched peaks are considered missing and their intensities are padded with 0. The output is an intensity vector with a dimension of 125×1.

[0031] Normalization specifically involves using total ion current (TIC) normalization, which divides the intensity of all matched peaks in each sample by the total ion current intensity of that sample (i.e., the sum of the intensities of all matched peaks). The calculation formula is as follows: ,in Let be the original intensity of the i-th matched peak, and P=125 be the total number of reference peaks. This normalization can correct for signal intensity differences caused by laser energy fluctuations.

[0032] To address the issue of missing values ​​in the preprocessed mass spectrometry feature matrix, this embodiment first performs a missing pattern analysis on each feature.

[0033] Missing values ​​in mass spectrometry data mainly come from three situations: (1) A certain feature is not detected in most samples, which is a high missing rate feature, usually related to the instrument detection limit or sample matrix effect; (2) A certain feature is randomly missing in some samples, which may be related to uneven laser sampling or instantaneous signal fluctuations; (3) A certain feature was detected in all samples, but in a few samples the signal intensity was lower than the instrument detection limit, which is a low-value truncation type of missing.

[0034] Different types of missing data have different effects on the data distribution. A single imputation method, such as imputation using only the mean or only the minimum, can lead to the following problems: imputation of features with high missing rates will introduce a lot of artificial noise; imputation using the minimum for random missing data will underestimate the true signal strength; and imputation using the mean for low-value truncated missing data will mask the important biological information of "below the detection limit".

[0035] Based on this, this embodiment proposes a triple complementary imputation strategy based on missing pattern recognition, specifically including: Calculate the missing rate of each feature in all samples, and directly delete features with missing rates exceeding a threshold. This threshold is determined by statistical analysis of the missing rate distribution of each feature in the historical mass spectrometry dataset. Such features are too missing to be reliably recovered by the algorithm, and forcibly filling them in will introduce a large number of false signals, interfering with subsequent differential feature screening.

[0036] For features with a missing value rate not exceeding a threshold and where the missing values ​​are randomly distributed, the k-nearest neighbor (kNN) algorithm is used for imputation. Specifically, k=5, and the five most similar samples are found based on the Euclidean distance between samples on other non-missing features. The mean of these five samples on that feature is then used as the imputation value. The advantage of this method is that it infers missing values ​​using the overall similarity of the samples, thus better preserving the original distribution structure of the data.

[0037] For missing values ​​due to being below the detection limit, the signal intensity is truncated at the detection threshold. The minimum value imputation method is used, which uses the minimum value of the feature in all samples as the imputation value.

[0038] This strategy ensures that the true signal of samples below the detection limit lies between zero and the detection limit, and fills in the gap with the minimum observed value of this feature. This preserves the semi-quantitative information that "the content of this metabolite in this sample is extremely low" while avoiding distribution bias caused by arbitrary assignment.

[0039] After processing missing values ​​using three strategies working in tandem, a reconstructed feature matrix was obtained. This matrix effectively eliminated data biases introduced by different types of missing values ​​while preserving the original biological differences, providing a high-quality data foundation for subsequent outlier removal and differential feature selection.

[0040] To further improve data quality, this embodiment uses principal component analysis to reduce the dimensionality of the reconstructed feature matrix, and identifies and removes abnormal samples that deviate from the main sample group based on the dimensionality reduction results, thus obtaining a quality-controlled sample feature set.

[0041] Principal Probe Analysis (PCA) is a classic unsupervised linear dimensionality reduction technique. Its basic principle is to project samples from the original high-dimensional feature space onto several uncorrelated directions through orthogonal transformation. These directions are arranged in order of sample variance, called the first principal component, second principal component, etc. The larger the variance of the principal component, the more original data information it retains. PCA can compress high-dimensional mass spectrometry data into a low-dimensional space while preserving the differential structure between samples to the greatest extent, facilitating visualization and the detection of anomalous samples.

[0042] Step 2: Obtain high-confidence candidate biomarkers by combining the sample feature sets through a triple complementarity strategy and taking the intersection. To identify the most relevant differential features to the tumor resection margin status, this embodiment employs a triple complementary strategy, using three methods with distinct principles for joint screening, and then taking the intersection of the screening results.

[0043] The reason for adopting this strategy is that mass spectrometry data has characteristics such as high dimensionality, strong collinearity, and large noise interference. If only a single screening method is used, it is easy to select "false positive" features that are accidentally related to the cutting edge state due to random noise, batch effects, or collinearity. These features are difficult to replicate in different sample sets or different experimental batches, resulting in poor generalization ability of subsequent identification models.

[0044] Specifically, this embodiment first employs orthogonal partial least squares discriminant analysis (OPLS-DA). This method can effectively separate metabolic differences between the negative and positive cut edge groups by extracting components that are orthogonal to and related to the classification label, while suppressing intra-group noise interference.

[0045] The Variable Projection Importance (VIP) value output by OPLS-DA quantifies the contribution of each mass spectrometry feature to the classification model. In this embodiment, features with a VIP greater than 1 are selected to obtain feature set A. The advantage of this method is that it can capture the covariant relationships among multiple variables and reflect the overall changes at the metabolite network level, but its disadvantage is that it is not sensitive enough to anomalies in univariate distributions.

[0046] This embodiment further employs the CART decision tree algorithm, using the Gini coefficient to evaluate the contribution of each feature to sample partitioning. This method can automatically capture nonlinear interactions and threshold effects between features, does not rely on data distribution assumptions, and is particularly sensitive to discrete features or features with complex boundary relationships.

[0047] In this embodiment, features whose Gini coefficient decreases by no less than a set threshold are selected to obtain feature set B. The advantage of this method is that it can discover non-linearly dependent metabolic features, but its disadvantage is that it is easily affected by noisy data and may select random splitting features that are irrelevant to classification.

[0048] This embodiment also employs the Mann-Whitney U rank-sum test, a non-parametric test method that does not require data to follow a normal distribution and exhibits good robustness to outliers and irregularities. This embodiment screens features that show significant differences between the negative and positive cut edge groups, obtaining feature binding C. The advantage of this method is its robust assessment of inter-group differences for each feature from a univariate perspective, capable of identifying simple mean changes. However, its limitation lies in its inability to capture collaborative relationships between features.

[0049] In this embodiment, the intersection of feature sets A, B, and C is calculated, i.e., candidate marker = A ∩ B ∩ C. Only features that appear in all three sets simultaneously are retained as the final high-confidence candidate markers.

[0050] A feature is considered to have high authenticity and repeatability only if it shows a strong correlation with the edge state in three aspects: multivariate covariant structure, nonlinear splitting ability, and univariate distribution difference.

[0051] This approach significantly reduces false positives introduced by the specific limitations of a single method. For example, a feature with a high VIP value but no significant difference between groups will be excluded by the rank-sum test; a feature with strong splitting ability in a decision tree but weak contribution in a multivariate model will be excluded by OPLS-DA. Candidate biomarkers obtained by taking the intersection exhibit stronger biological robustness and cross-batch reproducibility, laying a solid foundation for building a well-generalized identification model using LASSO regression.

[0052] Step 3: Using candidate markers as input and cutting edge states as labels, construct a recognition model using LASSO regression to output recognition results, and optimize the regularization parameters through cross-validation. Based on the high-confidence candidate markers obtained, this embodiment further employs LASSO logistic regression to construct a cutting edge state recognition model.

[0053] Although candidate biomarkers have undergone screening using a triple complementarity strategy, collinearity among features may still exist (i.e., multiple mass spectrometry features are highly correlated due to metabolic pathway associations). Ordinary logistic regression is unstable in estimation when collinearity exists and is prone to overfitting.

[0054] In this embodiment, LASSO regression introduces an L1 regularization term, which can impose sparsity constraints on the regression coefficients while fitting the model, automatically compressing the coefficients of irrelevant or redundant features to zero, thereby achieving simultaneous completion of feature selection and model training.

[0055] Use the selected differential features as input variables X∈R N×p Where N is the number of samples, p is the number of candidate biomarkers, and the cutting edge state is used as the output label yi∈{0,1}, where 0=negative cutting edge, 1=positive cutting edge. A LASSO logistic regression model is used to construct the identification model, with the objective function being:

[0056] in, β represents the positive cut edge recognition probability output by the model; β is the regression coefficient vector. λ is the L1 regularization term; λ≥0 is the regularization parameter, which controls sparsity. The larger λ is, the sparser the model.

[0057] Regularization parameters The choice of [aspect / component] directly affects model performance. This embodiment uses ten-fold cross-validation to select the optimal [aspect / component]. value.

[0058] The specific steps are as follows: randomly divide the training set into 10 equal parts, take one part as the validation set and the remaining 9 parts as the training set, and then... (The sentence is incomplete and requires more context to translate accurately.) For each value within the range, calculate the average deviation on the validation set. The candidate range is , taking 100 values ​​uniformly in the logarithmic space. After all folds are completed, each For each of the 10 deviation values, take the average and standard deviation.

[0059] Traditional methods select the one that minimizes the average deviation. (Right now However, this method is prone to overfitting. This embodiment employs the "1-SE rule," which selects a value that ensures the average deviation does not exceed one standard error of the minimum deviation. This rule sacrifices a small amount of fitting accuracy in exchange for a sparser model and stronger generalization stability, making it particularly suitable for clinical deployment scenarios. The model exhibits less fluctuation and greater reliability in recognizing new patient data.

[0060] Through ten-fold cross-validation calculations, the optimal method selected in this embodiment is... For samples with unknown cutting edge status, the preprocessed mass spectrometry features are input into the trained LASSO model. The model outputs the positive cutting edge probability. If the positive cutting edge probability is greater than or equal to the preset threshold, preferably 0.5, it is determined to be a positive cutting edge; otherwise, it is a negative cutting edge.

[0061] Step 4: Compare the identification results with the postoperative pathological gold standard and organize them into the database. When the accumulation reaches the threshold, the model is retrained and the identification model parameters are corrected using feedback samples.

[0062] After completing the identification model construction and preliminary verification, this embodiment further establishes a closed-loop feedback iteration mechanism. This mechanism is designed to address the inherent defects of existing technologies: existing intraoperative margin judgment algorithms are usually "open-loop" processes, with the model statically frozen after deployment, unable to receive feedback from the postoperative pathological "gold standard" to correct its own parameters.

[0063] When encountering new cases in clinical practice, such as rare tumor metabolic subtypes or fibrotic tissue after neoadjuvant chemotherapy, the model cannot learn automatically from misjudged or uncertain cases, causing its generalization ability to decline over time or as the distribution of clinical data changes.

[0064] To address the above issues, this embodiment designs the following closed-loop iterative process: For mass spectrometry data of samples with unknown cutting margin states, after the same preprocessing and feature selection as the training set, the data is input into the trained LASSO model. The model outputs the positive cutting margin probability. And make a judgment based on a preset threshold (default 0.5): If Then it is judged as a negative cutting edge, if If the result is positive, the sample is considered a positive cut edge. Simultaneously, record the sample's mass spectrometry characteristics, recognition probability, recognition status, and sample number.

[0065] The model's recognition results are compared sample-by-sample with the postoperative pathological resection margin status, which serves as the gold standard, to determine the accuracy of the recognition. Complete information for each sample, including original mass spectrometry characteristics, recognition probability, true label, and a marker indicating whether the recognition was correct or not, is structurally stored in the database. Samples with incorrect recognition are specially marked for subsequent targeted learning.

[0066] A threshold for accumulating new samples is set; in this embodiment, it is set to 50 samples. When the number of newly added samples with the gold standard label in the database reaches this threshold, the model retraining process is automatically triggered. During retraining, the newly added samples are merged with the original training set, and the same data preprocessing, missing value imputation, anomaly removal, feature selection, and LASSO regression process as the initial modeling is used to recalculate the model parameters, including the regression coefficients and the regularization parameter λ.

[0067] After training, the performance of the new model is evaluated through cross-validation. If the accuracy, sensitivity, or specificity is significantly improved compared to the original model, the optimized model is automatically deployed to the detection device to replace the old model. If there is no significant improvement, the original model is retained and samples continue to be accumulated.

[0068] The above process is repeated cyclically. As clinical data accumulates, the recognition model can continuously adapt to changes in data distribution caused by different hospitals, different mass spectrometry platforms, and different patient groups. The performance of the recognition model is no longer limited by the initial training set, but evolves in sync with the depth of clinical use.

[0069] Compared to existing open-loop methods, this implementation method has the ability to learn online and continuously evolve, and can continuously improve itself with clinical use, providing computational support with clinical practical value for real-time, accurate and reliable determination of tumor margins.

[0070] Example 2 Based on the method described in Implementation 1, the purpose of this embodiment is to provide a tumor margin identification system based on mass spectrometry data, including: The data preprocessing module is used to handle missing values ​​based on the initial mass spectrometry data using a triple complementary strategy, and to remove outlier samples by combining principal component analysis. The feature selection module is used to obtain high-confidence candidate markers by taking the intersection of a triple complementary strategy; The model building module is used to construct a recognition model with candidate markers as input and edge states as labels, using LASSO regression to output recognition results, and optimizing the regularization parameters through cross-validation. The closed-loop iteration module is used to compare the recognition results with the postoperative pathological gold standard and then organize and store them in the database. When the accumulation reaches the threshold, the model is retrained. The recognition model parameters are corrected using feedback samples and then used for subsequent recognition after evaluation and verification.

[0071] Example 3 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0072] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium.

[0073] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0074] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0075] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0076] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for tumor margin identification based on mass spectrometry data, characterized in that, include: Based on the initial mass spectrometry data, a triple complementarity strategy was used to handle missing values. After removing outlier samples by principal component analysis, high-confidence candidate biomarkers were obtained by taking the intersection of the triple complementarity strategy. The triple complementarity strategy for missing value processing includes: calculating the missing rate of each feature in all samples and deleting features with a missing rate exceeding a threshold; for features with a missing rate not exceeding the threshold and whose missing values ​​are randomly distributed, using the k-nearest neighbor algorithm to fill in the missing values; and for missing values ​​caused by values ​​below the detection limit, using the minimum value filling method. The method of obtaining high-confidence candidate markers through a triple complementary strategy and taking the intersection includes: using orthogonal partial least squares discriminant analysis to screen features with variable projection importance greater than 1, using decision tree algorithm based on Gini coefficient to screen key splitting features, and using rank sum test to screen features with significant differences between the negative and positive cutting edge groups. Using candidate biomarkers as input and cutting edge states as labels, a recognition model is constructed using LASSO regression to output recognition results, and the regularization parameters are optimized through cross-validation. The identification results are compared with the postoperative pathological gold standard and then stored in the database. When the accumulation reaches the threshold, the model is retrained. The identification model parameters are corrected using feedback samples and then used for subsequent identification after evaluation and verification.

2. The tumor margin identification method based on mass spectrometry data as described in claim 1, characterized in that, It also includes preprocessing steps: sequentially performing asymmetric least squares baseline correction, continuous wavelet transform peak detection, peak alignment based on the reference spectrum, and total ion current normalization on the mass spectrometry signal to obtain a standardized mass spectrometry feature vector.

3. The tumor margin identification method based on mass spectrometry data as described in claim 1, characterized in that, The regularization parameters for the LASSO regression were selected using a combination of 10-fold cross-validation and the 1-SE rule.

4. The tumor margin identification method based on mass spectrometry data as described in claim 1, characterized in that, The process of comparing the identification results with the postoperative pathological gold standard and then organizing and storing them includes: recording the mass spectrometry characteristics, identification probability, identification status and true label of the sample, and marking and storing the correctly identified samples and the incorrectly identified samples respectively.

5. The tumor margin identification method based on mass spectrometry data as described in claim 1, characterized in that, After the triggered model is retrained, the performance of the new model is evaluated through cross-validation. If the accuracy, sensitivity, or specificity of the new model is improved by more than a set threshold compared with the original model, the optimized model will replace the original model; otherwise, the original model will be retained to continue accumulating samples.

6. A tumor margin identification system based on mass spectrometry data, characterized in that, include: The data preprocessing module is used to handle missing values ​​based on the initial mass spectrometry data using a triple complementary strategy, and to remove outlier samples by combining principal component analysis. The feature selection module is used to obtain high-confidence candidate markers by taking the intersection of a triple complementary strategy; The triple complementarity strategy for missing value processing includes: calculating the missing rate of each feature in all samples and deleting features with a missing rate exceeding a threshold; for features with a missing rate not exceeding the threshold and whose missing values ​​are randomly distributed, using the k-nearest neighbor algorithm to fill in the missing values; and for missing values ​​caused by values ​​below the detection limit, using the minimum value filling method. The method of obtaining high-confidence candidate markers through a triple complementary strategy and taking the intersection includes: using orthogonal partial least squares discriminant analysis to screen features with variable projection importance greater than 1, using decision tree algorithm based on Gini coefficient to screen key splitting features, and using rank sum test to screen features with significant differences between the negative and positive cutting edge groups. The model building module is used to construct a recognition model with candidate markers as input and edge states as labels, using LASSO regression to output recognition results, and optimizing the regularization parameters through cross-validation. The closed-loop iteration module is used to compare the recognition results with the postoperative pathological gold standard and then organize and store them in the database. When the accumulation reaches the threshold, the model is retrained. The recognition model parameters are corrected using feedback samples and then used for subsequent recognition after evaluation and verification.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the tumor margin identification method based on mass spectrometry data as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the tumor margin identification method based on mass spectrometry data as described in any one of claims 1-5.

Images