Method for predicting content of phthalate in table grapes

Abstract

The invention belongs to the technical field of food safety early warning, and particularly relates to a method for predicting the content of phthalate in table grapes. Comprising the following steps: detecting a plurality of physicochemical indexes of soil and the phthalate content of grapes to obtain an acquisition data set, the plurality of physicochemical indexes being pH, organic matters, cation exchange capacity, total nitrogen and the phthalate content; dividing the collected data set into a test set and a training set; training the basic random forest model to obtain a first prediction model for predicting grapes through the plurality of physicochemical indexes; and analyzing the importance of the plurality of physicochemical indexes, determining a new feature subset, and training the first prediction model by using the feature subset as input to obtain a final prediction model. By analyzing the PAEs content and physicochemical properties in the grape rhizosphere soil, the PAEs content of the grape fruits can be predicted, and technical support is provided for grape quality safety monitoring and risk prevention and control.

Description

Technical Field

[0001] This invention belongs to the field of food safety early warning technology, and specifically relates to a method for predicting the phthalate content in fresh grapes. Background Technology

[0002] Plasticizers, scientifically known as phthalic acid esters (PAEs), are typical environmental pollutants whose bioaccumulation characteristics pose a persistent risk to agricultural product safety and ecosystems. Table grapes, a widely cultivated and popular fruit, are inevitably contaminated with PAEs during their growth. Among numerous PAE pollution sources, soil is a major source for grape PAEs. As the material basis of agricultural production, soil not only provides the medium and nutrients for crop growth but also serves as a direct receptor for PAEs. Since PAEs in the soil can enter grapevines through root absorption and other pathways, they affect the quality and safety of the grapes. Studies have found that soil physicochemical properties, such as organic matter content and soil pH, are believed to influence the migration behavior of hydrophobic organic pollutants in the soil. Although existing research has revealed the pollution characteristics of PAEs in grapevines, efficient and accurate prediction methods are still lacking for PAE content in grapes, especially for quantifying the complex nonlinear relationship between environmental factors such as soil physicochemical properties and PAE content. Summary of the Invention

[0003] To address the above problems, this invention provides a method for predicting the phthalate content of fresh grapes.

[0004] A method for predicting phthalate content in fresh grapes includes the following steps: Multiple physicochemical indicators of the soil and phthalic acid ester content of grapes were tested to obtain a dataset. The multiple physicochemical indicators are pH, organic matter, cation exchange capacity, total nitrogen, and phthalic acid ester content. The collected dataset was divided into a test set and a training set by random sampling; the basic random forest model was trained based on grid search and cross-validation to obtain the first prediction model for predicting grapes using multiple physicochemical indicators. The importance of multiple physicochemical indicators is analyzed by applying the Gini coefficient of the random forest interpreter to determine a new feature subset. The feature subset is then used as input to train the first prediction model to obtain the final prediction model. Data on the sample to be tested is collected based on the input variables of the final prediction model and input into the final prediction model. The output is the phthalate content in the sample to be tested.

[0005] In another preferred embodiment, the input variables of the final prediction model are the phthalate content of the soil, the organic matter of the soil, the total nitrogen of the soil, and the cation exchange capacity of the soil.

[0006] In another preferred embodiment, a new training set is constructed according to the new feature subset based on the collected dataset, and 85% of the collected data is used to train the new random forest model with different feature subsets, while the remaining 15% of the collected data is used to verify and evaluate the accuracy of the final prediction model.

[0007] In another preferred embodiment, the phthalate specifically refers to butyl benzyl phthalate, dimethyl phthalate, and di-n-octyl phthalate.

[0008] In another preferred embodiment, the accuracy result utilizes the coefficient of determination R. 2 judge.

[0009] In another preferred embodiment, the organic matter content is obtained by potassium dichromate titration method-external heating method; the cation exchange capacity is obtained by hexaamminecobalt trichloride extraction-spectrophotometry; and the phthalate content is obtained according to the "Soil Agricultural Chemical Analysis Methods".

[0010] Compared with the prior art, the present invention has the following beneficial effects: This invention presents a complete solution, from identifying key influencing factors to constructing a robust prediction model. The method first uses a random forest model and variable importance analysis to identify key soil driving factors affecting phthalate uptake in grapes, clearly defining soil PAE content, organic matter, total nitrogen, and cation exchange capacity as core predictive variables. This solves the problem of difficulty in quantifying and analyzing the complex nonlinear relationship between environmental factors and pollutant content in previous studies. Based on these identified key variables, this invention constructs a random forest model capable of accurately predicting the content of three specific PAEs—butyl benzyl phthalate, dimethyl phthalate, and di-n-octyl phthalate—in grape berries. This achieves a direct and quantitative assessment of PAE contamination risk in table grapes, filling a gap in predictive methodologies in this field. This invention comprehensively utilizes grid search, cross-validation, and stepwise feature selection, not only optimizing model parameters and preventing overfitting but also enhancing the model's generalization ability and interpretability. It provides a reliable and scalable methodological paradigm for solving similar environmental-agricultural pollutant prediction problems. This predictive model can provide early warnings about grape safety based on easily monitored soil indicators and provide direct scientific basis for implementing source control through agronomic measures. It has important application value for ensuring agricultural product safety and guiding agricultural production practices.

[0011] This invention enables precise prediction of phthalate content in grapes by detecting phthalate content, organic matter, total nitrogen, cation exchange capacity, and pH in the soil, providing a non-destructive testing method for phthalate content detection in grapes. Due to the bunch-like structure and thin-skinned, juicy biological characteristics of grapes, they are exceptionally fragile in the post-harvest stage. Even minor wounds caused by destructive testing not only affect individual berries but can also rapidly lead to catastrophic rot of entire bunches or boxes of grapes through microbial infection, ethylene transmission, and physical contact, resulting in significant economic losses. Therefore, the method of this invention, used in post-harvest grading and quality testing of grapes, does not damage the grapes and allows for large-scale testing. Compared to destructive testing used for very small sample sizes, the method of this invention has broader applicability. Attached Figure Description

[0012] Figure 1 The following are feature importance ranking graphs for PAEs based on random forest: (a) is the feature importance ranking graph for di-n-octyl phthalate (DNOP); (b) is the feature importance ranking graph for butyl benzyl phthalate (BBP); and (c) is the feature importance ranking graph for dimethyl phthalate (DMP).

[0013] Figure 2 This is for evaluating the stability of the model during repeated training. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0015] Unless otherwise specified, the experimental methods described in the following embodiments are conventional methods; unless otherwise specified, the reagents and materials are commercially available.

[0016] Research on predicting agricultural product quality indicators and heavy metal enrichment factors based on machine learning methods has mainly focused on major crops such as nuts, rice, and peanuts. However, there are no reports on the prediction of PAE content in the soil-grape system using machine learning methods.

[0017] This invention takes fresh grape fruits and soil samples as research objects, integrates the PAEs content and soil factors (soil physicochemical indicators) of the soil-grape system, applies machine learning methods to analyze the main PAEs pollution characteristics of grapes in the region, systematically constructs a random forest (RF) model, predicts accuracy and evaluates its uncertainty, and aims to provide innovative and effective technical support for grape quality and safety monitoring and risk prevention and control.

[0018] 1. Test materials Collection of grape berries and soil samples: From July to September 2023, during the grape ripening period, samples were collected from the northern foothills of the Tianshan Mountains, the Turpan-Hami Basin, the Ili River Valley, and the area surrounding the Tarim Basin. Sixty-nine representative samples were collected. Samples were collected using a grid sampling method, with sampling points evenly covering the entire grape-growing area. Grape fruit samples were placed in cloth bags, numbered, and quickly brought back to the laboratory. They were homogenized using a stainless steel grinder, stored in food-grade silicone bags, and kept at -18℃. Grape fruit samples were matched one-to-one with soil sampling points. Topsoil samples (0-20cm layer) weighing approximately 1kg were collected from the rhizosphere. The collected soil samples were placed in clean cloth bags and sealed for preservation. The collected soil samples were then air-dried, ground (coarse and fine), sieved (60 mesh), and sealed for later use.

[0019] 2. Model the predictive variables and obtain the collected dataset; the predictive variables include soil pH, organic matter, cation exchange capacity, total nitrogen, soil phthalate content and grape phthalate content.

[0020] 2.1 The physical and chemical properties of grape soil were determined according to the methods in "Soil Agricultural Chemical Analysis Methods" and national standards and literature.

[0021] 2.2 Methods for determining soil pH value Accurately weigh 5g of air-dried soil sample (passed through a 2mm sieve) and place it in a 50mL beaker that has been rinsed three times with deionized water and then dried. Use a pipette to accurately add 25mL of deionized water to the beaker, achieving a water-to-soil ratio of 5:1. Place the beaker on a magnetic stirrer and stir at a constant speed for 5 minutes, then remove and let stand for 30 minutes. Before measuring the pH of the soil extract, calibrate the pH meter using standard buffer solutions of pH 4.4 and pH 7.0 at least three times. After calibration, use the calibrated pH meter to measure the pH of the soil extract.

[0022] 2.3 Methods for determining soil organic matter (SOM) To accurately determine the soil organic matter content, the potassium dichromate titration method with external heating was used. The specific operating procedure is as follows: Accurately weigh 0.2500g of air-dried soil sample (passed through a 100-mesh sieve) into a dry, hard test tube. Precisely add 10mL of 0.4mol / L 1 / 6 K₂Cr₂O₇ sulfuric acid solution to the test tube, and place a small funnel at the mouth of the test tube to prevent moisture loss. Set the paraffin oil bath temperature to 180℃. Once it reaches 180℃, place the test tube in a wire cage. Set 1-2 blank controls for each batch of samples, and ensure the oil bath temperature is maintained between 170-180℃. Start timing when the solution in the test tube begins to gently boil and bubble. Remove the test tube immediately after 5 minutes, allow it to cool slightly, and wipe off any oil stains from the outside.

[0023] After cooling, transfer the contents of the test tube to a 250mL Erlenmeyer flask. Wash the test tube and funnel, and control the total volume of the solution to 60-70mL. Add 3-4 drops of o-phenanthroline indicator; the solution will turn orange-yellow. Then titrate with 0.2mol / L ferrous sulfate standard solution, shaking continuously. The titration endpoint is reached when the solution changes from orange-yellow to blue-green and then to brick-red.

[0024] Formula 1; In Formula 1, V0 is the volume of FeSO4 standard solution consumed in the blank test (mL), and V is the volume consumed in the sample determination (mL); c is the concentration of FeSO4 standard solution (mol / L); 0.003 is the millimolecular mass of 1 / 4 carbon atom (g); 1.724 is the organic carbon to organic matter conversion factor, and 1.10 is the oxidation correction factor; m is the mass of the air-dried sample (g); 1000 is used to convert to content per kilogram. The average value of parallel determinations is taken and three significant figures are retained.

[0025] 2.4 Methods for determining soil cation exchange capacity (CEC) Soil cation exchange capacity was determined using a cobalt hexachloroethylene chloride extraction-spectrophotometric method. The specific steps are as follows: Add 0.00 mL, 1.00 mL, 3.00 mL, 5.00 mL, 7.00 mL, and 9.00 mL of 1.66 cmol+ / L hexaamminecobalt trichloride solution to colorimetric tubes, respectively, and dilute with water to prepare a series of standard solutions with concentrations of 0.000, 0.166, 0.498, 0.830, 1.16, and 1.49 cmol+ / L. Using pure water as a reference, measure the absorbance at 475 nm using a 10 mm cuvette, and plot a standard curve with concentration on the x-axis and absorbance on the y-axis. Pass an air-dried soil sample through a 10-mesh nylon sieve and mix thoroughly. Weigh 1.75 g and place it in a 50 mL centrifuge tube, add 25.0 mL of hexaamminecobalt trichloride solution, seal, and shake at 20±2℃ for 60±5 min (adjusting the frequency to maintain suspension). After shaking, centrifuge at 4000 r / min for 10 min, collect the supernatant in a colorimetric tube, and measure the absorbance at 475 nm wavelength within 24 h.

[0026] 2.5 Methods for determining total nitrogen (TN) in soil The Kjeldahl method was used to determine the total nitrogen content in soil. The specific steps are as follows: The air-dried soil sample was passed through a 60-mesh sieve. Approximately 0.5g of the soil sample was accurately weighed using a 0.01g balance, and the precise mass was recorded. The sample was then placed into the bottom of a dry, clean digestion tube using a strip of waxed paper. A small amount of water was added to moisten the sample, followed by the addition of 2.5g of catalyst and 5mL of concentrated sulfuric acid, and the mixture was quickly mixed.

[0027] Add a small funnel to the digestion tube opening and let it sit at room temperature overnight for cold digestion. Then place it in a 330°C digestion oven and heat for 3 hours. Once the solution is clear blue-green and the residue is grayish-white, continue digestion for 30 minutes.

[0028] After the digestion process is completed and cooled, 20-30 mL of water is added for dilution, and the total nitrogen content of the soil is determined using a Kjeldahl nitrogen analyzer.

[0029] 3. Model Construction To further illustrate that the random forest model performs better in detection, we will use the multiple linear regression (MLR) model as a comparison below.

[0030] Multiple linear regression (MLR) is a commonly used model that uses multiple independent variables to explain changes in the dependent variable. It explains the linear relationship with the dependent variable by optimizing the combination of independent variables, and this model can also avoid multicollinearity among the independent variables. Based on soil content and soil physicochemical properties data, the model is established by Equation 2:

[0031] Formula 2; In Equation 2, y is the dependent variable, representing the content of grape PAEs monomers in this study; xi (i=1, 2, 3, ..., n) is the independent variable, namely parameters such as soil content and soil physicochemical properties; bi (i=1, 2, 3, ..., n) is the regression coefficient of variable xi, reflecting the relationship between xi and y, and a is a constant term.

[0032] Random Forest (RF) is an ensemble machine learning algorithm based on classification and regression. It can be used to analyze complex nonlinear relationships between independent and dependent variables. Its key feature is the use of Bootstrap sampling, which randomly selects a portion of the sample data as the training set. Decision trees are then constructed based on these selected data. By repeatedly executing the sampling and training process, numerous decision trees are built, forming the Random Forest model. Due to the randomness of data selection, the decision tree parameters are adjustable, and the RF model does not require testing assumptions such as normality, independence, or multicollinearity, it performs exceptionally well in classification and regression problems. The RF model uses random sampling to divide the dataset into a test set and a training set at a ratio of 15% and 85%, respectively. The training set uses 10-fold cross-validation, dividing it into 10 groups, with one group used as the validation set each time, for a total of 10 training iterations. The random forest has 300 decision trees and a maximum depth of 7.

[0033] The specific construction method using the above approach is as follows: 0) Establishment of the first prediction model Multiple physicochemical indicators of the soil and phthalic acid ester content of grapes were measured to obtain a dataset. These indicators included pH, organic matter, cation exchange capacity, total nitrogen, and phthalic acid ester content. The dataset was divided into a training set and a test set, with 85% for training and 15% for testing. The training set served as the base random forest model, trained using a pre-defined grid search and cross-validation method. Specifically, 10-fold cross-validation was used, dividing the model into 10 groups, with one group selected as the validation set each time, for a total of 10 training iterations. The base random forest had 300 decision trees with a maximum depth of 7, resulting in a first prediction model for predicting grape growth based on multiple physicochemical indicators. The importance of the predictive variables in the first prediction model was analyzed based on variable importance scores.

[0034] 1) Establishment of the final prediction model The Gini coefficient of the random forest interpreter is applied to analyze the importance of multiple physicochemical indicators to determine a new feature subset. The feature subset is then used as input to train the first prediction model to obtain the final prediction model. Specifically, a new training set is constructed according to the new feature subset based on the collected dataset. 85% of the collected data is selected for training the new feature subset random forest model to obtain the final prediction model. The remaining 15% of the collected data is used to verify and evaluate the accuracy of the final prediction model.

[0035] Based on the accuracy results, the key predictor variables for the final prediction model were obtained and used as input variables. Soil phthalate content, organic matter, total nitrogen, and cation exchange capacity were identified as input variables for predicting grape phthalate content. The RF model was further trained, and the R-values ​​of the three monomers DNOP, BBP, and DMP on the test set after each training iteration were obtained. 2 The RF models for all three monomers achieved good predictive stability.

[0036] 4. Results 4.1 Analysis of soil physicochemical properties, grape berries, and soil PAE content characteristics The soil physicochemical properties and soil-grape system PAEs data characteristics are shown in Table 1. The soil pH of the samples ranged from 6.34 to 7.49, with an average of 7.02. Among them, 12.5% ​​were acidic (pH between 5.5 and 6.5), 81.25% were neutral (pH between 6.5 and 7.5), and 6.25% were alkaline (pH between 7.5 and 8.5). The majority of the soil samples were neutral.

[0037] The average values ​​of cation exchange capacity, organic matter, and total nitrogen in the soil samples were 67.63 cmol / kg, 29.30 g / kg, and 1.42 g / kg, respectively. The soil DNOP, BBP, and DMP contents ranged from 30.45 to 462.30 μg / kg, 71.01 to 329.29 μg / kg, and 57.38 to 117.83 μg / kg, respectively. The grape berries contained DNOP, BBP, and DMP in the samples, ranging from 27.37 to 361.93 μg / kg, 80.66 to 325.01 μg / kg, and 59.36 to 156.02 μg / kg, respectively. The average enrichment coefficients of DNOP, BBP, and DMP in the grape berries were 0.93, 1.52, and 1.02, respectively. The enrichment coefficients of BBP and DMP were both above 1, indicating that PAEs have strong mobility and are easily absorbed and accumulated by grapes.

[0038] Table 1. Characteristics of PAE content in the soil-grape system 4.2 Final Prediction Model Using soil pH, organic matter (SOM), cation exchange capacity (CEC), total nitrogen (TN), and soil PAEs content (SC) (n=5), stepwise regression analysis was used to establish multiple linear regression models and random forest models to predict the content of PAE monomers in grapes. Different models showed different results for PAE monomers, as shown in Table 4. In the multiple linear regression model, the predictive effect of different PAE monomers was: BBP(R 2 =0.819) > DNOP(R 2=0.296)>DMP(R 2 =0.179); In the random forest model, the individual prediction performance of different PAEs is: DNOP (R 2 =0.972)>BBP(R 2 =0.926) > DMP(R 2 =0.925). PAEs monomers showed significant differences in predictive performance between the multiple linear regression model and the random forest model, while DNOP monomers showed better prediction performance in the random forest model (R² = 0.925). 2 =0.972) Prediction accuracy is higher than that of the multiple linear regression model (R² = 0.972). 2 =0.296) higher (Table 2); BBP single-unit random forest model (R 2 =0.926) and the multiple linear regression model (R 2 =0.819) all achieved relatively high predictive performance (Table 2); the multiple linear regression model of DMP individual (R 2 Compared to the random forest model (R²=0.179), the random forest model (R²=0.925) significantly achieved better prediction results, and its root mean square error (RMSE=25.324) was also significantly lower than that of the multiple linear regression model (RMSE=68.557) (Table 2). These results indicate that, compared to the multiple linear regression model (R²=0.179), the random forest model significantly outperformed the multiple linear regression model (R²=0.925). 2 =0.179~0.819, RMSE=27.050~68.557), Random Forest model (R 2 =0.925~0.972, RMSE=10.993~25.324) are more suitable for predicting grape PAEs content. The random forest model has achieved better prediction results and lower root mean square error.

[0039] Table 2 Prediction Model for PAEs Content in Grapes 4.4 Analysis of the importance of different predictor variables The RF interpreter was used to rank the importance of feature variables, see [link to RF interpretation]. Figure 1 Organic matter and soil PAEs content are the two most important characteristic variables affecting the prediction of grape DNOP content, with relative importance of 0.5 and 0.3, respectively. The relative importance of the remaining characteristic variables is all below 0.25. The two most important characteristic variables affecting the prediction of grape BBP content are soil PAEs content and total nitrogen. The most important characteristic variables affecting the prediction of grape DMP content are total nitrogen and soil PAEs content, with relative importance of 0.5 and 0.3, respectively. The relative importance of the remaining characteristic variables is all below 0.25. Among them, soil pH has the lowest impact on the prediction of grape DNOP, BBP, and DMP content.

[0040] 4.5 Selection of New Feature Subsets and Determination of Final Model Input Variables Based on the importance analysis of the predictor variables, five new feature subsets were formed, representing five possible combinations of predictor variables, as shown in Table 3. 85% of the different predictor variables were selected for training random forest models with different predictor variable combinations to obtain the final prediction model; the remaining 15% of the different predictor variables were used to validate and evaluate the accuracy of the final prediction model. As the number of input variables increases, the coefficient of determination (R²) of the final prediction model... 2 As the concentration of phthalates increases, the accuracy of the final prediction model gradually improves, and the fitting effect becomes better and better, as shown in Table 4. Based on the model accuracy, soil phthalate content, organic matter, total nitrogen, and cation exchange capacity were selected as the final model input variables.

[0041] Table 3. Five new feature subsets Table 4. Model performance of five new feature subset input schemes Based on model accuracy, soil phthalate content, organic matter, total nitrogen, and cation exchange capacity were determined as the final variable input scheme for predicting phthalate content in grapes. The RF model was further trained, and the R-values ​​of the three monomers DNOP, BBP, and DMP on the test set were obtained after each training iteration. 2 ( Figure 2 The final prediction model was obtained. All three individual RF models achieved good prediction stability.

[0042] The phthalate content in actual grapes is predicted using the final prediction model obtained above. The specific process is as follows.

[0043] From June to September 2024, 13 soil and grape samples were collected from the main grape-producing areas of Turpan, Kashgar, Hotan, and Urumqi County. The specific test results for phthalate content, soil organic matter, total nitrogen, and cation exchange capacity of the soil and grape samples are shown in Table 5. Using the final prediction model, the above test data were input into the model to obtain the predicted results for phthalate content in grapes, as shown in Table 6.

[0044] Table 5. Data Collection Results: Predicted Phthalate Content in Grape Fruits for 2024 Table 6 Predicted Content Results As can be seen from the results in Table 6, the results predicted by the method of the present invention have a small relative error with the actual results, indicating that the method of the present invention has high accuracy in predicting the phthalate content in grapes. This provides a non-destructive testing method with high accuracy for detecting the phthalate content in grapes, effectively avoiding the problems of microbial infection, ethylene transmission, physical contact and rot caused by the tiny wounds on grapes by destructive testing.

[0045] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the phthalate content in fresh grapes, characterized in that, Includes the following steps: Multiple physicochemical indicators of the soil and phthalic acid ester content of grapes were detected to obtain a dataset. The multiple physicochemical indicators are pH, organic matter, cation exchange capacity, total nitrogen, and phthalic acid ester content. The collected dataset was divided into a test set and a training set by random sampling; the basic random forest model was trained based on grid search and cross-validation to obtain the first prediction model for predicting grapes using multiple physicochemical indicators. The importance of multiple physicochemical indicators is analyzed by applying the Gini coefficient of the random forest interpreter to determine a new feature subset. The feature subset is then used as input to train the first prediction model to obtain the final prediction model. Data on the sample to be tested is collected based on the input variables of the final prediction model and input into the final prediction model. The output is the phthalate content in the sample to be tested.

2. The method for predicting phthalic acid ester content in fresh grapes according to claim 1, characterized in that, The input variables for the final prediction model are soil phthalate content, soil organic matter, soil total nitrogen, and soil cation exchange capacity.

3. The method for predicting phthalate content in fresh grapes according to claim 1, characterized in that, Based on the collected dataset, a new training set is constructed according to the new feature subset, and 85% of the collected data is used to train the random forest model on the new feature subset, while the remaining 15% of the collected data is used to verify and evaluate the accuracy of the final prediction model.

4. The method for predicting the phthalate content in fresh grapes according to claim 1, characterized in that, The phthalates specifically refer to butyl benzyl phthalate, dimethyl phthalate, and di-n-octyl phthalate.

5. The method for predicting the phthalate content in fresh grapes according to claim 1, characterized in that, The accuracy results utilize the coefficient of determination R. 2 judge.

6. The method for predicting phthalate content in fresh grapes according to claim 1, characterized in that, The organic matter content was obtained using the potassium dichromate volumetric method with external heating; the cation exchange capacity was obtained using the hexaamminecobalt trichloride extraction-spectrophotometric method; and the phthalate content was obtained according to the "Methods for Chemical Analysis of Soil".

Images