Method for establishing an inflammation serum LIBS spectral diagnosis model

Abstract

The present invention relates to a method for establishing an inflammatory serum LIBS spectral diagnostic model, belonging to the technical field of spectral detection. The present invention solves the problems of slow speed, cumbersome process and high technical requirements in the diagnosis of inflammatory serum, and uses laser-induced breakdown spectroscopy (LIBS) technology to achieve accurate and rapid detection of inflammatory serum for the first time. The LIBS spectra of the training-test sets are divided according to each individual, and the Wilcoxon test is used to exclude the training-test sets with significant differences. Several groups are randomly selected from the remaining training-test sets to establish a backpropagation neural network (BPNN) model, and the recognition effect of inflammatory serum is evaluated. For the training-test set with the best recognition effect, multivariate scatter correction (MSC) is used to preprocess the spectral data. The optimal features are applied to other training-test sets to re-establish the MSC-MIV-BPNN model, and the serum recognition effect of the model is verified.

Description

Technical Field

[0001] The present invention relates to a method for establishing an LIBS spectral diagnosis model for inflamed serum, belonging to the technical field of spectral detection. Background Art

[0002] Inflammation is part of the body's biological defense response to harmful stimuli in vascular tissues. When the immune system malfunctions, the inflammatory response can lead to apoptosis or necrosis of cells, and the products of dead cells further spread inflammation, resulting in tissue death or organ failure. Currently, inflammation is a major cause of many common diseases, including colitis, diabetes, pneumonia, liver abscess, sepsis, cancer, etc., which have high morbidity and mortality rates worldwide. Inflammation detection is crucial for clinical diagnosis and treatment. Traditional preliminary diagnosis of inflammation is usually based on routine blood indicators and biomarkers, such as white blood cell count, neutrophil percentage, lymphocyte percentage, C-reactive protein concentration, procalcitonin concentration, etc. Specific indicators need to be confirmed based on the doctor's experience and detected on multiple devices. With these indicators, the final diagnosis result needs to be determined through comprehensive evaluation and even further examinations (such as blood bacterial culture and ultrasound detection). Although the accuracy of traditional inflammation diagnosis is very high, due to its time-consuming, cumbersome operation, high technical requirements and other disadvantages, emergency patients need to wait for a long time or even delay their condition. Therefore, it is very important to find a fast, accurate, portable and stable inflammation recognition technology.

[0003] Laser-induced breakdown spectroscopy (LIBS) is a reliable atomic emission spectroscopic analysis technology with the advantages of fast speed, simple pretreatment and real-time in-situ detection. LIBS has been applied in the medical field, such as elemental imaging, bacteria detection and tissue recognition. In recent years, LIBS has been used for blood detection to diagnose cancer based on the differences in elemental content. However, the training-test set division in existing research is randomly divided according to the mixed spectra of different individuals. This division method will largely result in spectra from the same person appearing in both the training set and the test set, which is very different from the unknown situation in actual medical applications. However, due to the influence of individual differences, spectral fluctuations and spectral redundant information, the method of dividing the training-test set according to each individual may lead to significant differences between the training set and the test set, resulting in poor recognition accuracy and generalization ability of the established model and unable to achieve the expected diagnostic effect. Therefore, it is of great significance to reasonably divide the training-test set, exclude the training-test sets with significant differences, and establish a model that can accurately identify inflamed serum. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for establishing an LIBS spectral diagnosis model for inflamed serum, aiming to quickly, accurately and stably identify inflamed serum.

[0005] The technical solution of the present invention is as follows:

[0006] A method for establishing an LIBS spectral diagnosis model for inflammatory serum, the steps of which include:

[0007] In the first step, prepare m normal serum samples and n inflammatory serum samples on a glass slide, and the inflammatory serum samples include s types of inflammation;

[0008] In the second step, collect a number of LIBS spectra of the m normal serum samples and a number of LIBS spectra of the n inflammatory serum samples on the glass slide;

[0009] In the third step, select k element characteristic spectral lines with physical meanings whose spectral intensities are higher than the set threshold;

[0010] In the fourth step, divide the LIBS spectra collected in the second step into training-test sets according to normal and inflammatory individuals respectively, and obtain multiple division combinations of training-test sets for normal and inflammatory categories respectively;

[0011] When dividing the training-test sets, the normal and inflammatory serum samples are divided according to a certain ratio respectively, and the training set of the inflammatory serum samples must include s types of inflammation;

[0012] In the fifth step, use the Wilcoxon test to perform a significance test on the training-test set combinations obtained in the fourth step. The input is the average value of the intensities of the k element characteristic spectral lines selected in the third step for each sample, and the output is the p-value of each training-test set combination for the normal or inflammatory category. Exclude the combinations with p-values lower than the set value, and the remaining training-test set combinations of the two categories can be randomly combined as the training-test set;

[0013] In the sixth step, randomly select q groups from the remaining training-test sets in the fifth step, and use these training-test sets to evaluate the recognition effect of the backpropagation neural network (BPNN) model;

[0014] In the seventh step, perform multivariate scatter correction (MSC) preprocessing on the LIBS spectra of the training-test set with the best recognition effect in the sixth step to obtain the LIBS spectra of the preprocessed training-test set;

[0015] In the eighth step, select the k element characteristic spectral lines in the third step from the LIBS spectra of the preprocessed training-test set obtained in the seventh step. Use the mean impact value (MIV) method to evaluate the importance of the characteristic spectral lines in the training set, and select different numbers of characteristic spectral lines in descending order of feature importance to establish a BPNN model. The characteristic spectral lines corresponding to the model with the highest recognition accuracy are the optimal features;

[0016] Step 9: Use the q-1 groups of training-test sets except the one with the best recognition effect in Step 6, repeat Step 7, combine the optimal features determined in Step 8, re-establish the model and verify the recognition effect;

[0017] In the aforementioned Step 7, the method for MSC preprocessing is as follows:

[0018] 1) Calculate the average spectrum of the LIBS spectra collected in Step 2 according to formula (1) and set the average spectrum as the ideal spectrum for correcting the LIBS spectra of normal serum samples or inflamed serum samples;

[0019] 2) Perform linear regression between the LIBS spectrum A i collected in Step 2 and the ideal spectrum using the partial least squares method to calculate the spectral line drift k i and the baseline shift b i ;

[0020] 3) Calculate the corrected spectrum according to formula (3), and then extract the element characteristic spectral lines in Step 3 from the corrected spectrum;

[0021]

[0022]

[0023] A MSCi =(A i -b i ) / k i (3)

[0024] In the aforementioned Step 8, the steps of the MIV method include:

[0025] 1) Use the k selected element characteristic spectral lines after correction to train the BPNN model;

[0026] 2) Increase and decrease a certain proportion of the i-th element characteristic spectral line respectively and input it into the above BPNN model to obtain two outputs; i = 1, 2,... k;

[0027] 3) The average of the absolute differences between the two outputs of the samples in the training set is the average influence value MIV (MIV i ) of the i-th element characteristic spectral line;

[0028] 4) For the k selected element characteristic spectral lines, their average influence values must be calculated;

[0029] 5) Repeat the whole process t times, and use the average value of the t times of MIV i as the final feature importance value of the i-th element characteristic spectral line.

[0030] Beneficial effects

[0031] The present invention divides the training - test set samples according to individuals, uses the Wilcoxon test to exclude the training - test sets with significant differences, randomly selects several groups from the remaining training - test sets, establishes a back - propagation neural network (BPNN) model, and evaluates the recognition effect of inflamed serum. For the training - test set with the best recognition effect, multivariate scatter correction (MSC) is used to pre - process the spectral data. The mean influence value (MIV) method is used to evaluate the importance of the characteristic spectral lines in the training set after MSC correction. Different numbers of characteristic spectral lines are selected in descending order of importance to establish a BPNN model for serum recognition. The characteristic spectral lines corresponding to the model with the highest recognition accuracy are the optimal features. The above - mentioned optimal features are applied to other training - test sets, and an MSC - MIV - BPNN model is re - established to verify the serum recognition effect of the model. This model is closer to real detection, with a significant improvement in accuracy and a faster detection time. Brief description of the drawings

[0032] Figure 1 It is a diagram of the experimental device;

[0033] Figure 2 It is a comparison diagram of important features before and after MSC pre - processing;

[0034] Figure 3 It is a diagram of the MIV feature importance ranking and the recognition results of different numbers of features;

[0035] Figure 4 It is the recognition results of the BPNN model and the MSC - MIV - BPNN model for 10 groups of training - test sets. Detailed implementation manners

[0036] The following further describes the present invention with reference to the drawings and embodiments.

[0037] Embodiment

[0038] First step, prepare 10 normal serum samples and 10 inflamed serum samples on a glass slide. The inflamed serum samples include 3 types of inflammation, including 5 cases of hepatitis, 3 cases of liver abscess, and 2 cases of sepsis;

[0039] Second step, use the Figure 1 device shown to collect 150 laser pulses at different positions for each serum sample. Every 3 pulses are averaged to be 1 LIBS spectrum, and 50 spectra can be obtained for each sample;

[0040] Third step, taking 800 as the threshold of spectral intensity, select the element characteristic spectral lines with physical meanings in combination with the NIST library. The number of element characteristic spectral lines is 43, as shown in Table 1;

[0041] Table 1. 43 characteristic spectral lines

[0042]

[0043]

[0044] Step 4: Divide the LIBS spectra collected in Step 2 into training - test sets according to a ratio of 7:3. For healthy samples, randomly select 3 samples from 10 samples as the test set, and the rest as the training set, with a total of 120 combinations. For inflammatory samples, the samples in the training set and the test set should include all 3 types of inflammation. Therefore, randomly select 1 sample from each of the 3 types of inflammation as the test set, and the remaining 7 samples as the training set, with a total of 30 combinations;

[0045] Step 5: Use the Wilcoxon test to conduct a significance test on the training - test sets of the 120 and 30 combinations obtained in Step 4 respectively. The input is the average value of the intensities of 43 elemental characteristic spectral lines, and the combinations with an output p - value less than 0.05 are excluded;

[0046] Step 6: Randomly select 10 groups of training - test sets from the combinations excluded in Step 5, and use these 10 groups of training - test sets to evaluate the recognition results of the BPNN model. The recognition accuracy is recorded in Figure 4 ;

[0047] Step 7: Perform MSC pre - processing on the LIBS spectra of the training - test set combination with the highest accuracy in Step 6. The comparison of the intensities of the important spectral lines of elements K, Ca, Na, and Mg in the training - test set before and after pre - processing is as Figure 2 shown;

[0048] Step 8: Select the 43 features in Step 3 from the LIBS spectra of the pre - processed training - test set obtained in Step 7. Use the MIV method to evaluate the importance of the characteristic spectral lines in the training set, and sequentially select different numbers of characteristic spectral lines from high to low according to the feature importance to establish a BPNN model. The test set accuracy is the highest when the number of characteristic spectra is 36. The importance ranking results of the characteristic spectral lines and the recognition accuracies of the models using different numbers of features are as Figure 3 shown;

[0049] Step 9: Use the 9 groups of training - test sets in Step 6 except the one with the best recognition effect, repeat Step 7, and combine the 36 features determined in Step 8 to re - establish a BPNN model to identify the serum. The recognition results are as Figure 4 shown;

Claims

1. A method for establishing an inflammation serum LIBS spectral diagnosis model, characterized in that The steps of the method include: First, prepare normal serum samples and inflamed serum samples on glass slides; Second, collect the LIBS spectra of the normal serum samples and the inflamed serum samples on the glass slides; Third, extract the element characteristic spectral lines with physical significance whose spectral intensities are higher than the set threshold; Fourth, divide the LIBS spectra collected in the second step into training - test sets according to normal and inflamed individuals respectively. For normal and inflamed categories, multiple combinations of training - test set divisions are obtained; Fifth, use the Wilcoxon test to conduct a significance test on the training - test set combinations obtained in the fourth step. The input is the average value of the intensities of the element characteristic spectral lines selected in the third step for each sample, and the output is the p - value of each training - test set combination for the normal or inflamed category. Exclude the combinations with p - values lower than the set value, and randomly combine the remaining training - test set combinations of the two categories as the training - test set; Sixth, randomly select multiple groups from the remaining training - test sets in the fifth step, and use these training - test sets to evaluate the recognition effect of the backpropagation neural network (BPNN) model; Seventh, perform multiplicative scatter correction (MSC) pre - processing on the LIBS spectra of the training - test set with the best recognition effect in the sixth step to obtain the pre - processed LIBS spectra of the training - test set; Eighth, select the element characteristic spectral lines in the third step from the pre - processed LIBS spectra of the training - test set obtained in the seventh step, use the mean impact value (MIV) method to evaluate the importance of the characteristic spectral lines in the training set, and sequentially select different numbers of characteristic spectral lines from high to low according to the feature importance to establish a BPNN model. The characteristic spectral lines corresponding to the model with the highest recognition accuracy are the optimal features; Ninth, use the other training - test sets in the sixth step except the one with the best recognition effect, repeat the seventh step, combine the optimal features determined in the eighth step, re - establish the model and verify the recognition effect.

2. A method for establishing an inflamed serum LIBS spectral diagnosis model according to claim 1, characterized in that: In the first step, the numbers of normal serum samples and inflamed serum samples are m and n respectively, and the inflamed serum samples include s types of inflammation; In the third step, the number of element characteristic spectral lines is k.

3. A method for establishing an inflamed serum LIBS spectral diagnosis model according to claim 2, characterized in that: In the fourth step, when dividing the training set and the test set, the normal and inflamed samples are divided into training - test sets according to a certain proportion respectively. The training set of the inflamed serum samples needs to include all types of inflammation.

4. A method for establishing an inflamed serum LIBS spectral diagnosis model according to claim 3, characterized in that: In the fifth step, the p - value is set to 0.05, 0.01 or 0.

001.

5. A method for establishing an inflamed serum LIBS spectral diagnosis model according to claim 4, characterized in that: In the seventh step, the number of randomly selected training - test sets is q; In the ninth step, the number of other training - test sets except the one with the best recognition effect is q - 1.

6. The method for establishing an inflammation serum LIBS spectral diagnosis model according to claim 5, wherein: In the seventh step described above, the MSC preprocessing method is: 1) Calculate the average spectrum of the LIBS spectra collected in the second step according to formula (1). And use the average spectrum as the ideal spectrum for correcting the LIBS spectra of normal serum samples or inflamed serum samples. 2) According to formula (2), perform linear regression between the LIBS spectrum A collected in the second step i and the ideal spectrum , and calculate the spectral line drift k i and the baseline shift b i using partial least squares method; 3) Calculate according to formula (3) to obtain the corrected spectrum; A MSCi = (A i - b i ) / k i (3).

7. The method for establishing an inflammation serum LIBS spectral diagnosis model according to claim 6, wherein: In the eighth step described above, the steps of the MIV method include: 1) Use the k element characteristic spectral lines selected from the corrected spectrum to train the BPNN model; 2) The i-th element characteristic spectral line is increased and decreased by a certain proportion respectively to form two new BP neural network inputs and obtain two outputs; i = 1, 2,... k; 3) The average of the absolute differences between the two outputs of all samples is the mean influence value MIV of the characteristic spectral line of the i-th element i ; 4) For the k selected element characteristic spectral lines, their average influence values need to be calculated.

8. The method for establishing an inflammation serum LIBS spectral diagnosis model according to claim 7, wherein: The whole process is repeated t times, and the average value of t times of MIV i is used as the final feature importance value of the characteristic spectral line of the i-th element.

Images