Early-stage lung cancer risk prediction model based on meta analysis results

Abstract

The invention provides a lung cancer prediction model based on meta analysis, and the lung cancer prediction model is optimized and fitted by using a machine learning method and a small amount of clinical data so as to achieve an optimal prediction solution. And the feasibility of the provided model for predicting on a small-scale data set and serving as an enhancement model for improving the accuracy of other models is verified. The method is composed of a model linear contribution formed by meta analysis results of VOCs and an offset term for fitting by using a small amount of clinical data. After the model is established, accurate prediction on a small-scale data set can be completed, and the prediction effect of the fusion model can also be enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of cancer diagnosis, and in particular relates to an early lung cancer risk prediction model based on meta-analysis results. Background Art

[0002] Breath analysis has been proven to be effective in identifying early lung cancer by identifying the concentration of volatile organic compounds (VOCs) in the exhaled breath of subjects through mass spectrometry. VOCs are small molecular weight compounds with high vapor pressure and low boiling point. The measurement and analysis of VOCs provides a non-invasive, rapid, easy to implement, low-cost, widely available, small, portable, and highly sensitive method for early cancer screening. The biggest advantage of the breath analysis method over LDCT is its timely diagnostic characteristics, which is its threshold for perceiving disease.

[0003] Existing lung cancer screening usually uses an invasive form of LDCT, but LDCT has high false positive rates and the risk of radiation-induced cancer for continuous screening, especially because there is no unified and accurate standard for defining high-risk populations. In addition, in terms of non-invasive gas detection, given the complexity and diversity of human physiology and pathology, the development and validation of clinical prediction models for lung cancer screening often requires the use of extensive clinical data sets. However, due to privacy and security issues, it is usually not easy to obtain a large number of clinical data sets, and models trained and fitted using public data sets generally do not have good generalization capabilities. Therefore, in the absence of a large number of own clinical data sets, the performance of clinical prediction models is weak and lacks good robustness.

[0004] The present invention designs a method to integrate, summarize and model the gas markers of lung cancer through meta-analysis, so as to provide assistance in the field of early screening of lung cancer. This method aims to integrate all relevant studies, accurately estimate the effects of lung cancer gas markers, and is conducive to exploring the consistency of evidence from various studies and the differences between studies. This method only requires a small amount of clinical data for correction to obtain a high-accuracy lung cancer early screening model, which can not only make accurate predictions on small-scale data sets, but also serve as an enhanced model to improve the predictive ability of its existing lung cancer prediction model, effectively overcoming the limitations of the above-mentioned clinical model in development and verification. Summary of the invention

[0005] The present invention provides an early lung cancer risk prediction model based on meta-analysis results, which is characterized by comprising three stages: data screening and processing, linear contribution calculation, and model fitting.

[0006] Furthermore, in the data screening and processing stage, the exclusion criteria for inclusion in the study must first be set, and then the screened data must be cleaned and standardized.

[0007] Furthermore, in the stage of calculating the linear contribution, the processed data is used as input and passed into the metafor meta-analysis package to calculate the effect size to obtain the linear contribution.

[0008] Furthermore, in the model fitting stage, the model is fitted using clinical data and activation functions to obtain the bias term of the model. The linear contribution and the bias term are combined to obtain the lung cancer early screening model.

[0009] Furthermore, the exclusion criteria for inclusion in the study were as follows, to see whether the following VOCs factors were present:

[0010] [1] Early screening for lung cancer;

[0011] [2] The substance being tested is the patient’s gas, not urine or other substances;

[0012] [3] The detection method used is mass spectrometry;

[0013] [4] The study must include specific concentration values ​​or a list of variances, standard deviations, medians, and thresholds for each type of VOCs tested;

[0014] [5] The VOCs concentration unit must be ppb, or relevant properties that can be converted to ppb units must be provided, such as the pressure and temperature of the substance during measurement;

[0015] [6] The data quality in the study was too poor to meet the expected calculation standards;

[0016] If any of the above factors exist, the data will be screened out for further preprocessing; if none of the above factors exist, it will be excluded.

[0017] Further, the preprocessing includes:

[0018] 1. When the data presents a right-skewed or left-skewed distribution, logarithmic transformation is performed;

[0019] 2. When the data appears in the form of thresholds and medians, the mean and standard deviation are converted. The calculation formula is as follows,

[0020]

[0021] Where a and b are the threshold values ​​of the data, m is the median of the data, and n is the sample size of the data;

[0022] The standard deviation S is calculated as follows:

[0023]

[0024] where Φ is the cumulative distribution function of the standard normal distribution

[0025] 3. When encountering continuous data, convert it into binary OR value form; first calculate the Cohen effect size using the Cohen formula, the formula is as follows:

[0026]

[0027]

[0028]

[0029] Secondly, after calculating the Cohen effect size, the corresponding OR value is calculated. The formula is as follows:

[0030]

[0031] Where SMD is the standardized mean difference, n0 and n1 are the sample sizes of the two study groups.

[0032] Furthermore, rma in the metafor meta-analysis package was selected as the function of the fitting model, and the OR value and SE value were used as the yi and vi input values ​​of the function. The effect size value of each VOCs was obtained by calculation as the linear contribution of the model. The obtained model is as follows, where A, B, …M, N are the effect sizes of each VOCs:

[0033] y=Ax1+Bx2+Cx3+Dx4+…+Nx n .

[0034] Furthermore, after obtaining the linear contribution, we need to use a small amount of clinical data to calculate the bias term of the model, and use a small amount of clinical data fitting process such as:

[0035] First, input the model into the activation function sigmoid. The sigmoid function is as follows:

[0036]

[0037] The σ(y) of the healthy population is 0, and the σ(y) of the lung cancer population is 1. From the sigmoid function graph, we can see that the critical value of the healthy population and the lung cancer population is 0.5, that is, the critical y value is 0. Set the bias term of the model to β, and the current model is as follows:

[0038] y=Ax1+Bx2+Cx3+Dx4+…+Nx n +β

[0039] Then input the VOCs data of healthy people into the above formula to obtain:

[0040] y=α+β

[0041] Where α is the constant of linear contribution, β is the bias term required, and in the healthy population, y < 0. Inputting all the data of the healthy population to obtain the maximum value of the bias term of all the healthy population can ensure the prediction of all the healthy population.

[0042] Similarly, input the data of all lung cancer populations to obtain another maximum value of the bias term, take the middle intersection of the two maximum values, re-predict all data with an increase interval of 0.01 to 0.1, and test the bias term whose predicted value is closest to the true value. This bias term is what the model is looking for.

[0043] The present invention also proposes an application of an early lung cancer risk prediction model based on the meta-analysis results. In a neural network model, the output of the model is added as an additional feature. The specific steps are:

[0044] Step 1: Calculate y = Ax1 + Bx2 + Cx3 + Dx4 + ... + Nx n +β;

[0045] Step 2: Convert the y value into a probability value through the sigmoid function;

[0046] Step 3: Add the probability value as a new feature to the input feature set of the neural network model;

[0047] Step 4: Train and validate the integrated model and evaluate its prediction performance.

[0048] Compared with the prior art, the present invention has the following technical effects:

[0049] (1) Through meta-analysis of VOCs, an efficient early screening model for lung cancer was established. After the model was established, only a small amount of clinical data was needed for fitting to achieve good prediction results. This greatly reduced the cost and difficulty of data acquisition, making the model more practical.

[0050] (2) In view of the fact that VOCs are not yet fully used in the current early screening methods for lung cancer, the model proposed in this invention can not only accurately identify the specific metabolites of lung cancer patients, but also effectively combine data from different sources, reduce the random errors of a single study, and provide more accurate effect estimates.

[0051] (3) The potential of VOCs in early screening of lung cancer can be fully explored, and the prediction effect of the model can be further optimized by fitting a small amount of clinical data. At the same time, the model can be used as an enhancement module and combined with other existing lung cancer early screening algorithms to effectively improve the prediction accuracy and stability of the model, thereby providing strong technical support for the early diagnosis and treatment of lung cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Shown is a flow chart of the present invention;

[0053] Figure 2 Shown is a model fitting bias term flow chart of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The inventive point of the present invention is to establish an early screening model for lung cancer through a meta-analysis of volatile organic compounds (VOCs), and use a small amount of clinical data to fit and optimize the model. The present invention uses a meta-analysis method to integrate data from multiple studies, and establishes a more robust lung cancer prediction model through a comprehensive analysis of VOCs. Meta-analysis can effectively reduce the deviation caused by a single study and improve the accuracy and stability of the model. Many lung cancer prediction models in the prior art mainly rely on data sets from a single study, or use traditional statistical methods to establish models. These models may be affected by the limitations of the data set and the heterogeneity of the research, resulting in poor stability and generalization of the model.

[0056] like Figure 1-2 As shown: The present invention provides an early lung cancer risk prediction model based on meta-analysis results, which includes three stages: data screening and processing, calculation of linear contribution, and model fitting.

[0057] In the data screening and processing stage, we first need to set the exclusion criteria for inclusion in the study, and then clean and standardize the screened data, so that the data characteristics can be balanced and the robustness of the model can be enhanced;

[0058] In the linear contribution calculation stage, the processed data is used as input and passed into the metafor meta-analysis package to calculate the effect size and obtain the linear contribution;

[0059] In the model fitting stage, we use clinical data and activation functions to fit the model and obtain the bias term of the model.

[0060] Combining the linear contribution with the bias term gives us the lung cancer early screening model.

[0061] The specific technical details are as follows:

[0062] First, taking the PubMed database as an example, a total of 352,920 articles were retrieved under the main keyword lung cancer, and a total of 15,666 articles were retrieved under the main keyword VOCs. A total of 251 articles were found by combining lung cancer and VOCs papers. The criteria for inclusion of articles are as follows: [1] Early screening for lung cancer; [2] The substance being detected is patient gas, not urine, etc.; [3] The detection method used is mass spectrometry; [4] The study must include specific values ​​of the concentration of various VOCs detected or a list of variances, standard deviations, medians, and thresholds; [5] The unit of the queried VOCs concentration must be ppb, or relevant attributes that can be converted to ppb units must be provided, such as the pressure and temperature of the substance during measurement; [6] The data quality in the study is too poor to meet the expected calculation standards.

[0063] After screening the VOCs data in the included literature, we analyzed the number of times each VOCs appeared in the literature and the degree of correlation between each VOCs and cancer in each literature. We can also explore VOCs factors according to our own needs. We screened out the linear contribution factors VOCs of the lung cancer early screening model.

[0064] The exclusion criteria for the included studies were as follows: whether the following VOCs factors were present:

[0065] [1] Early screening for lung cancer;

[0066] [2] The substance being tested is the patient’s gas, not urine or other substances;

[0067] [3] The detection method used is mass spectrometry;

[0068] [3] The detection method used is mass spectrometry;

[0069] [4] The study must include specific concentration values ​​or a list of variances, standard deviations, medians, and thresholds for each type of VOCs tested;

[0070] [5] The VOCs concentration unit must be ppb, or relevant properties that can be converted to ppb units must be provided, such as the pressure and temperature of the substance during measurement;

[0071] If any of the above factors exists, the data will be screened out for further preprocessing; if none of the above factors exist, or the data quality in the study is too poor to meet the expected calculation standards, it will be excluded.

[0072] Secondly, pre-process the VOCs data screened from each document:

[0073] 1. Perform logarithmic transformation. Logarithmic transformation helps reduce the skewness of the data. When the data is right-skewed (positively skewed) or left-skewed (negatively skewed) distributed, logarithmic transformation can make the data closer to normal distribution. And it can not only reduce the impact of outliers on the analysis. Outliers are usually far away from most data points. Logarithmic transformation shrinks large values, thereby reducing the interference of outliers on the overall analysis results. It can also help to balance variables of different magnitudes, which is very helpful for optimizing the performance of the algorithm;

[0074] 2. Convert to mean and standard deviation. This method uses a method to convert the sample size, threshold, median or quartile of the data into the mean of the data. and standard deviation S. When the data is converted in the form of threshold and median, the mean The calculation formula is as follows, where a and b are the thresholds of the data, m is the median of the data, and n is the sample size of the data.

[0075]

[0076] The standard deviation S is calculated as follows, where Φ is the cumulative distribution function of the standard normal distribution.

[0077]

[0078] 3. Continuous data are converted into binary OR values. This method first calculates the Cohen effect size using the Cohen formula, which is as follows:

[0079]

[0080]

[0081]

[0082] After calculating the Cohen effect size, the corresponding OR value is calculated as follows:

[0083]

[0084] SMD is the standardized mean difference, and n0 and n1 are the sample sizes of the two study groups.

[0085] Then, after obtaining the OR values ​​corresponding to VOCs in healthy people and cancer people in each article, we chose to use R software to calculate the effect size of the meta-analysis. We chose rma in the metafor meta-analysis package as the function of the fitting model, and used the OR value and SE value as the yi and vi input values ​​of the function. The effect size value of each VOCs was calculated as the linear contribution of the model. The resulting model is as follows, where A, B, ...M, N are the effect sizes of each VOCs:

[0086] y=Ax1+Bx2+Cx3+Dx4+…+Nx n .

[0087] Finally, after obtaining the linear contribution, we need to use a small amount of clinical data to calculate the bias term of the model. Figure 2 As shown:

[0088] First, input the model into the activation function sigmoid. The sigmoid function is as follows:

[0089]

[0090] The σ(y) of the healthy population is 0, and the σ(y) of the lung cancer population is 1. From the sigmoid function graph, we can see that the critical value of the healthy population and the lung cancer population is 0.5, that is, the critical y value is 0. Set the bias term of the model to β, and the current model is as follows:

[0091] y=Ax1+Bx2+Cx3+Dx4+…+Nx n +β

[0092] Then input the VOCs data of healthy people into the above formula to obtain:

[0093] y=α+β

[0094] Among them, α is the constant of linear contribution, β is the bias term required, in the healthy population, y < 0, all the data of the healthy population are input to obtain the maximum value of the bias term of all healthy populations, which can ensure the prediction of all healthy populations;

[0095] Similarly, input the data of all lung cancer populations to obtain another maximum value of the bias term, take the middle intersection of the two maximum values, re-predict all data with an increase interval of 0.01 to 0.1, and test the bias term whose predicted value is closest to the true value. This bias term is what the model is looking for.

[0096] The model created using the method proposed in the present invention can also enhance the accuracy of other existing models. As a breath VOCs concentration model, the model is simple in form, contains limited parameters and bias terms, has low computational complexity, and is easy to understand and implement. Therefore, the model can be easily applied to a variety of different prediction scenarios and is particularly suitable for being added as an enhancement module to existing complex prediction models.

[0097] This model is designed as an independent prediction module that outputs the probability of lung cancer. This output result can be added as a new feature to the input feature set of the existing model to improve the prediction ability of the existing model.

[0098] The specific implementation method is as follows:

[0099] In a neural network model, you only need to add the output of this model as an additional feature to achieve seamless integration. The specific steps are:

[0100] Step 1: Calculate y = Ax1 + Bx2 + Cx3 + Dx4 + ... + Nx n +β.

[0101] Step 2: Convert the y value into a probability value through the sigmoid function.

[0102] Step 3: Add the probability value as a new feature to the input feature set of the existing model.

[0103] Step 4: Train and validate the integrated model and evaluate its prediction performance.

[0104] By integrating this model into the existing model, the overall prediction performance can be improved. In theory, this model provides additional valuable information by capturing the relationship between VOCs concentration and lung cancer, allowing the integrated model to more comprehensively consider relevant factors when making predictions, thereby improving prediction accuracy.

[0105] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0106] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is limited by the appended claims and their equivalents.

Claims

1. An early lung cancer risk prediction model based on meta-analysis results, characterized by: It includes three stages: data screening and processing, calculation of linear contribution, and model fitting.

2. The early lung cancer risk prediction model based on meta-analysis results according to claim 1, characterized in that: In the data screening and processing stage, the exclusion criteria for inclusion in the study must first be set, and then the screened data must be cleaned and standardized.

3. The early lung cancer risk prediction model based on meta-analysis results according to claim 2, characterized in that: In the linear contribution calculation stage, the processed data is used as input and passed into the metafor meta-analysis package to calculate the effect size to obtain the linear contribution.

4. The early lung cancer risk prediction model based on meta-analysis results according to claim 3, characterized in that: In the model fitting stage, the model is fitted using clinical data and an activation function to obtain the bias term of the model, and the early screening model for lung cancer is obtained by combining the linear contribution and the bias term.

5. The early lung cancer risk prediction model based on meta-analysis results according to claim 2, characterized in that: The exclusion criteria for the included studies were as follows: whether the following VOCs factors were present: [1] Early screening for lung cancer; [2] The substance being tested is the patient’s gas, not urine or other substances; [3] The detection method used is mass spectrometry; [4] The study must include specific concentration values ​​or a list of variances, standard deviations, medians, and thresholds for each type of VOCs tested; [5] The VOCs concentration unit must be ppb, or relevant properties that can be converted to ppb units must be provided, such as the pressure and temperature of the substance during measurement; [6] The data quality in the study was too poor to meet the expected calculation standards; If any of the above factors exist, the data will be screened out for further preprocessing; if none of the above factors exist, it will be excluded.

6. The early lung cancer risk prediction model based on meta-analysis results according to claim 5, characterized in that: The pre-processing comprises:

1. When the data presents a right-skewed or left-skewed distribution, logarithmic transformation is performed; 2. When the data appears in the form of thresholds and medians, the mean and standard deviation are converted. The calculation formula is as follows, Where a and b are the threshold values ​​of the data, m is the median of the data, and n is the sample size of the data; The standard deviation S is calculated as follows: Where Φ is the cumulative distribution function of the standard normal distribution; 3. When encountering continuous data, convert it into binary OR value form; first calculate the Cohen effect size using the Cohen formula, the formula is as follows: Secondly, the Cohen effect size is calculated and the corresponding OR value is calculated. The formula is as follows: Where SMD is the standardized mean difference, n0 and n1 are the sample sizes of the two study groups.

7. The early lung cancer risk prediction model based on meta-analysis results according to claim 3, characterized in that: Select rma in the metafor meta-analysis package as the function of the fitting model, use the OR value and SE value as the yi and vi input values ​​of the function, and calculate the effect size of each VOCs as the linear contribution of the model. The obtained model is as follows, where A, B, ... M, N are the effect sizes of each VOCs: y=Ax1+Bx2+Cx3+Dx4+…+Nx n 。 8. The early lung cancer risk prediction model based on meta-analysis results according to claim 4, characterized in that: After obtaining the linear contribution, a small amount of clinical data is used to calculate the bias term of the model. The fitting process using a small amount of clinical data is as follows: First, input the model into the activation function sigmoid. The sigmoid function is as follows: The σ(y) of the healthy population is 0, and the σ(y) of the lung cancer population is 1. From the sigmoid function graph, we can see that the critical value of the healthy population and the lung cancer population is 0.5, that is, the critical y value is 0. Set the bias term of the model to β, and the current model is as follows: y=Ax1+Bx2+Cx3+Dx4+…+Nx n +β Then input the VOCs data of healthy people into the above formula to obtain: y＝α+β Among them, α is the constant of linear contribution, β is the bias term required, in the healthy population, y < 0, all the data of the healthy population are input to obtain the maximum value of the bias term of all healthy populations, which can ensure the prediction of all healthy populations; Similarly, input the data of all lung cancer populations to obtain another maximum value of the bias term, take the middle intersection of the two maximum values, re-predict all data with an increase interval of 0.01 to 0.1, and test the bias term whose predicted value is closest to the true value. This bias term is what the model is looking for.

9. An application of an early lung cancer risk prediction model based on meta-analysis results, according to any one of claims 1 to 8, wherein: In a neural network model, the output of the model is added as an additional feature. The specific steps are: Step 1: Calculate y = Ax1 + Bx2 + Cx3 + Dx4 + ... + Nx n +β; Step 2: Convert the y value into a probability value through the sigmoid function; Step 3: Add the probability value as a new feature to the input feature set of the neural network model; Step 4: Train and validate the integrated model and evaluate its prediction performance.

Images