Food evaluation value prediction method, food evaluation value prediction model generation method, and program

Abstract

This technology improves the accuracy of predicting food quality values. [Solution] The prediction method includes the step of inputting multiple analytical parameters of the food to be predicted into a prediction model that has been pre-trained to take multiple analytical parameters indicating the results of food analysis by an analytical instrument as input and output a food evaluation value that quantifies one of the senses of taste, texture, and smell of the food, in order to obtain the food evaluation value. The multiple analytical parameters indicating the results of food to be predicted include at least two types of analytical parameters obtained from analytical parameters obtained by a taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS).

Description

[Technical Field]

[0001] This disclosure relates to a technology for predicting food evaluation values ​​using machine learning. [Background technology]

[0002] Food evaluation values, which quantify the senses of taste, texture, and smell of food, are defined by sensory evaluation based on human sensory characteristics. Sensory evaluation has the advantage of directly quantifying human senses, but in order to achieve highly reproducible evaluations, it is necessary to train a large number of trained evaluators (panelists), which is costly and time-consuming.

[0003] In contrast, Patent Document 1 mentions that research and development are being conducted to evaluate texture by physical measurement using instruments, and discloses that it is considered in principle difficult to evaluate human texture by measuring the uniaxial reaction force when food is pressed. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2019-207123 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] This disclosure provides a technology to improve the prediction accuracy of food evaluation values. [Means for solving the problem]

[0006] The food evaluation value prediction method of this disclosure includes the steps of: acquiring a plurality of analytical parameters that indicate the results of food analysis using an analytical instrument; and obtaining the food evaluation value by inputting the acquired plurality of analytical parameters into a prediction model that takes the plurality of analytical parameters that indicate the results of food analysis using an analytical instrument as input and outputs a food evaluation value that quantifies one of the senses of taste, texture, and smell of the food. The plurality of analytical parameters that indicate the results of food analysis include at least two types of analytical parameters obtained from analytical parameters obtained by a taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS). [Brief explanation of the drawing]

[0007] [Figure 1] This is a block diagram showing a food evaluation value prediction model generation system and a food evaluation value prediction system. [Figure 2] This figure shows the parameters (food evaluation values) of the average TI curve plotted using the Overbosch method. [Figure 3] This is a schematic diagram illustrating the measurement state of a taste sensor and the response value (mV) obtained from the sensor. [Figure 4] This is an explanatory diagram relating to multiple analytical parameters obtained by the taste sensor of the first embodiment. [Figure 5] This is a flowchart showing the method for generating a predictive model. [Figure 6] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for RMinc of the TI method in the first embodiment of food evaluation values. [Figure 7] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the first embodiment: Tend of the TI method. [Figure 8]This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for RMdec of the TI method in the first embodiment of food evaluation values. [Figure 9] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for Rinc of the TI method in the first embodiment of food evaluation values. [Figure 10] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the Rdec of the TI method in the first embodiment of food evaluation values. [Figure 11] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the DUR of the TI method in the first embodiment of food evaluation values. [Figure 12] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the first embodiment: Dpla of the TI method. [Figure 13] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for Dinc in the first embodiment of food evaluation values ​​using the TI method. [Figure 14] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for Ddec of the TI method in the first embodiment of food evaluation values. [Figure 15] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the AUC of the TI method in the first embodiment of food evaluation values. [Figure 16] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the first embodiment: Apla of the TI method. [Figure 17] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the first embodiment: Ainc of the TI method. [Figure 18]Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of Adec in the TI method. [Figure 19] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of astringency in the QDA method. [Figure 20] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of citrus in the QDA method. [Figure 21] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of herbal medicine in the QDA method. [Figure 22] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of kire in the QDA method. [Figure 23] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of lastingness in the QDA method. [Figure 24] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of rapidness in the QDA method. [Figure 25] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of spread in the QDA method. [Figure 26] Food evaluation value of the first embodiment: A diagram showing the accuracy of the regression equations and the analysis parameters used as inputs to the regression equations for Comparative Example 1 and Examples 1 and 2 of thickness in the QDA method. [Figure 27]This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for Woody, a food evaluation value of the first embodiment using the QDA method. [Figure 28] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Rinc of the TI method. [Figure 29] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: RMinc of the TI method. [Figure 30] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Rdec of the TI method. [Figure 31] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: RMdec of the TI method. [Figure 32] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Tstart of the TI method. [Figure 33] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Dinc of the TI method. [Figure 34] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Tmax of the TI method in the second embodiment of food evaluation values. [Figure 35] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Dpla of the TI method. [Figure 36] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Ddec of the TI method. [Figure 37]This figure shows the food evaluation values ​​of the second embodiment: Comparative Example 1, the accuracy of the regression equations for Examples 1 and 2, and the analytical parameters used as input to the regression equations for the Tend of the TI method. [Figure 38] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: DUR of the TI method. [Figure 39] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Ainc of the TI method. [Figure 40] This figure shows the food evaluation values ​​of the second embodiment: Comparative Example 1, Examples 1 and 2 for Apla of the TI method, the accuracy of the regression equations, and the analytical parameters used as input to the regression equations. [Figure 41] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: Adec of the TI method. [Figure 42] This figure shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and the food evaluation values ​​of the second embodiment: AUC of the TI method. [Figure 43] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the initial intensity of the top of the food evaluation value of the second embodiment: QDA method. [Figure 44] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the second embodiment: lastingness of the QDA method. [Figure 45] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the bitterness (Bitter) of the food evaluation value of the second embodiment using the QDA method. [Figure 46] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the astringency of the food evaluation value of the second embodiment using the QDA method. [Figure 47] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the monotonic (Flat) value of the QDA method in the second embodiment of food evaluation values. [Figure 48] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the second embodiment: stickiness (coating) using the QDA method. [Figure 49] This figure shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the second embodiment: aroma (caramel) using the QDA method. [Figure 50] This diagram illustrates the proportion of each parameter selected as an important variable. [Modes for carrying out the invention]

[0008] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. Figure 1 is a block diagram showing the food evaluation value prediction model generation system 1 and the food evaluation value prediction system 2.

[0009] The food evaluation value prediction model generation method described herein can be executed by the food evaluation value prediction model generation system 1. Furthermore, the food evaluation value prediction method described herein can be executed by the food evaluation value prediction system 2. Both the food evaluation value prediction model generation system 1 and the food evaluation value prediction system 2 can be implemented using a computer.

[0010] As shown in Figure 1, the food evaluation value prediction model generation system 1 constructs a prediction model 20 using machine learning with training data D1. The food evaluation value prediction system 2 uses the prediction model 20 to predict a food evaluation value (estimated value) that quantifies one of the senses of taste, texture, and smell of the food, based on multiple analytical parameters that show the analysis results of the food to be predicted by the analytical instrument 3. The food evaluation value prediction system 2 uses the prediction model 20 that was previously generated by machine learning by the food evaluation value prediction model generation system 1. Details of the food evaluation value prediction model generation system 1 and the food evaluation value prediction system 2 will be described later.

[0011] <First Embodiment> [Food evaluation values ​​(sensory evaluation values, bitterness evaluation values)] The food evaluation values ​​in this disclosure are numerical values ​​that quantify one of the senses of taste, texture, and smell of food. In the first embodiment, bitterness among the senses of taste is used as an example. Nine types of bittering agents and bittering materials were used as food samples to evaluate only bitterness. Specifically, these were Caffeine, Jamaica quassia, Naringin, Wormwood, Kuding tea, Iso alpha acids, Gentiana lutea (root), Quina, and Amur cork tree. A selection panel of 50 people compared each sample with a control (0.04% Caffeine aqueous solution) and serially diluted samples, and selected the one they felt was more bitter (two-point discrimination method). The inflection point of the sigmoid curve plotted based on the selectivity of each concentration was defined as the subjective equivalent concentration of bitterness, and in subsequent evaluations and analyses, samples were prepared and provided at that concentration. Each sample was evaluated using Quantitative Descriptive Analysis (QDA) and Time-Intensity Analysis (TI), and the following 25 food evaluation values ​​(bitterness evaluation values) were obtained.

[0012] [Food evaluation values ​​showing the qualitative characteristics of bitterness based on Quantitative Descriptive Analysis (QDA method)] A panel of 14 selected members determined 10 characteristic parameters and their definitions suitable for the sample. Through trial evaluations, appropriate evaluation frequencies and sample quantities were determined. After sufficient evaluation training, the intensity of each characteristic parameter was evaluated on a 15cm line scale. The evaluation was repeated twice. Each sample was assigned a three-digit random number, and the liquid color was blinded during the test. The following shows the 10 food evaluation values ​​(characteristic items) and their definitions. For example, the definition of the characteristic item "astringency" is "irritation to the oral cavity." • Astringency: Irritation in the oral cavity (the higher the intensity, the stronger the irritation) • Rapidity of expression: The speed of the timing (higher intensity means faster). • Sharp: The gradient of the rising angle (the higher the strength, the greater the gradient) • Kire: The gradient of the angle of the drop (the greater the strength, the steeper the gradient) • Lastingness: The length of the lingering effect (longer for stronger effects) • Thickness: The flavor profile from the middle to the last stage (the higher the thickness, the stronger the flavor profile). • Spread: Spreads throughout the entire mouth (the stronger the intensity, the wider the spread). • Herbal medicine: Herbal, mossy bitterness (the stronger the concentration, the more herbal or mossy the taste). • Woody: A woody, pencil lead-like bitterness (the stronger the intensity, the more woody or pencil lead-like the taste). • Citrus: Grapefruit-like bitterness (the higher the intensity, the stronger the grapefruit-like taste)

[0013] [Food evaluation values ​​showing the time-dependent characteristics of bitterness using the Time-Intensity (TI) method] A selection panel of 13 participants conducted trial evaluations to determine appropriate sample quantities, swallowing times, etc. After sufficient evaluation training, the temporal intensity of "bitterness" was evaluated using a slider bar on a 10cm line scale. The evaluation was repeated three times. Figure 2 shows the parameters (food evaluation values) of the average TI curve plotted using the Overbosch method. In Figure 2, the vertical axis represents intensity, and the horizontal axis represents time. As shown in Figure 2, in the first embodiment, there are 15 food evaluation values ​​that indicate the temporal characteristics of bitterness using the TI method. The 15 food evaluation values ​​and their definitions are as follows. • Rinc: Slope of increase • RMinc: Maximum slope of increase • Rdec: Slope of decrease • RMdec: Maximum slope of decrease • Tstart: The first time bitterness was detected. • Dinc: Increased time • Tmax: Time at which the intensity first reaches its maximum. • Dpla: Duration of the plateau of maximum intensity ·Ddec: Decrease time • Tend: The time when the bitterness is no longer noticeable. • DUR: Total time spent feeling bitterness Ainc: Area under the increasing curve • Apla: Area under the curve of the maximum intensity plateau • Adec: Area under the decreasing curve • AUC: Area under the curve of the entire curve

[0014] [Example of analysis parameter: Taste sensor] An example of analytical instrument 3 is a taste sensor. An example of analytical parameters obtained by a taste sensor is described below. A taste sensor is a sensor that obtains response information to a lipid membrane. Each sample was analyzed using eight types of taste sensors. The specific items of the eight types of taste sensors and their identifiers are as follows. ·Acidic bitterness:C00 • Hydrochloride bitter taste: BT0 • Basic salt bitterness: AN0 • Astringency: AE1 • Acidity: CA0 • Saltiness: CT0 • Umami: AAE • Sweetness: GL1 For each taste sensor, a first-stage initial taste measurement, a second-stage first aftertaste measurement, and a third-stage second aftertaste measurement were performed, and the response value over a 30-second measurement period was measured in 1-second increments at each stage. Figure 3 is a schematic explanatory diagram showing the measurement state of the taste sensor and the response value (mV) obtained from the sensor. The measurement method is as schematically shown in Figure 3 for membrane potential (sensor output), first by immersing the taste sensor in a reference solution to obtain a baseline membrane potential (sensor output). Next, when the taste sensor is immersed in the sample, the membrane potential changes due to interaction with the taste substance, and the amount or portion of the change corresponds to the initial taste (sometimes referred to as test_1). Next, the taste sensor is washed with the reference solution. Next, the taste sensor is immersed in the reference solution and the change in membrane potential is measured. The amount or portion of the change corresponds to the first aftertaste (sometimes referred to as CPA1). In the first embodiment, the taste sensor is further washed with the reference solution, and then the taste sensor is immersed in the reference solution and the change in membrane potential is measured. The amount or part of the change corresponds to the second aftertaste (sometimes referred to as CPA2).

[0015] Figure 4 is an explanatory diagram of the multiple analytical parameters obtained by the taste sensor of the first embodiment. As shown in Figure 4, the analytical parameters are obtained based on the sensor intensity at multiple time points (0 seconds, 1 second, 2 seconds, ..., 29 seconds, 30 seconds) measured between the start and end of the measurement (30 seconds). The 10 basic analytical parameters obtained by the taste sensor and their definitions are as follows. ·Imax: Maximum intensity • Tmax: Time to reach maximum intensity • Plateau Intensity: 90% Imax (Maximum intensity plateau) • Plateau_Reach_Time: Time to reach 90% Imax (maximum intensity plateau) • Plateau_Duration: Duration of the plateau of maximum intensity • Intensity_at_0_Sec: Intensity at 0 seconds (start of measurement) • Intensity_at_30_Sec: Intensity at 30 seconds (end of measurement) • AUC_total: Area under the curve of the entire curve ·AUC_pre_plateau: Area under the curve before Plateau_Reach_Time AUC_plateau: Area under the curve of Plateau_Duration The above 10 parameters are basic parameters. Since the initial taste, the first aftertaste, and the second aftertaste each have the above 10 reference parameters, there are a total of 30 parameters. For example, the maximum intensity of the initial taste is denoted as Imax_test_1, the maximum intensity of the first aftertaste is denoted as Imax_CPA_1, and the maximum intensity of the second aftertaste is denoted as Imax_CPA_2. Furthermore, for each basic parameter, combination parameters can be generated that represent the difference (def) and ratio (ratio) of two combinations of the initial taste, first aftertaste, and second aftertaste. For example, the ratio of the initial taste to the second aftertaste in the AUC_total of the taste sensor (C00) can be expressed as TasteSensor_SB2C00_AUC_total_ratio_test_1_vs_CPA_2. The difference between the initial taste and the second aftertaste in the AUC_total of the taste sensor (C00) can be expressed as TasteSensor_SB2C00_AUC_total_diff_test_1_vs_CPA_2. Based on one basic parameter, there are three combination parameters: the initial taste and first aftertaste, the initial taste and second aftertaste, and the first aftertaste and second aftertaste. Each combination has two types: a difference and a ratio. Therefore, based on one basic parameter, 3 × 2 = 6 combination parameters can be generated.

[0016] In summary, the analytical parameters obtained by the taste sensor may include intensities prior to the end of the measurement (30 seconds). An example of this is Intensity_at_0_Sec [intensity at 0 seconds (start of measurement)]. In some cases, Plateau_Intensity [90% Imax (plateau of maximum intensity)] can also be used as an example. The analytical parameters obtained by the taste sensor may include an intensity that is a predetermined percentage of the maximum intensity. In this example, if the predetermined percentage is set to 90%, Plateau_Intensity[90%Imax (plateau of maximum intensity)] can be exemplified. In the first embodiment, the predetermined percentage is 90%, but any value greater than 0% and less than 100%, such as 80%, can be appropriately set. The analytical parameters obtained from the taste sensor can exemplify the time it takes to reach a predetermined percentage of the maximum intensity. An example of this is Plateau_Reach_Time [time to reach 90%Imax (plateau of maximum intensity)]. The analytical parameters obtained by the taste sensor may include the duration at which the intensity is equal to or greater than a predetermined percentage of the maximum intensity. An example of this is Plateau_Duration [Duration of the maximum intensity plateau]. The analytical parameters obtained by the taste sensor may include an area determined by the intensity over a predetermined period. Examples include at least one of AUC_total [area under the curve of the entire curve], AUC_pre_plateau [area under the curve before Plateau_Reach_Time], and AUC_plateau [area under the curve of Plateau_Duration]. Conventionally, a single taste sensor only had the intensity at the end of the measurement (30 seconds), and only the first initial taste and the first aftertaste were recorded. Specifically, Intensity_at_30_Sec_test_1 and Intensity_at_30_Sec_CPA_1 are known.

[0017] [Example of analysis parameters: DART-MS] An example of analytical instrument 3 is a real-time direct mass spectrometer (DART-MS). DART-MS can obtain comprehensive component information without targeting. Each sample was tested directly without pretreatment, and the intensity of each spectrum was obtained as the analytical parameter.

[0018] [Example of analysis parameters: NIRS] An example of analytical instrument 3 is a near-infrared analyzer (NIRS). NIRS can obtain comprehensive component information without targeting. Each sample was tested directly without pretreatment, and the intensity of each spectrum was obtained as an analytical parameter.

[0019] [Example of analytical parameters: NMR] An example of analytical instrument 3 is a nuclear magnetic resonance (NMR) spectrometer. NMR can obtain comprehensive component information without targeting. Each sample was tested directly without pretreatment, and the intensity of each spectrum was obtained as an analytical parameter.

[0020] [Example of analytical parameters: GC-MS] An example of analytical instrument 3 is a gas chromatography-mass spectrometer (GC-MS). Since even trace amounts of aroma components can affect sensory differences, they were analyzed using a highly sensitive GC-MS. After stirring and extraction of the components using a Starbar (Twister, GERSTEL), they were subjected to GC-MS, and the relative content of each component was obtained as an analytical parameter by calculating the peak area ratio with an internal standard.

[0021] <Food Evaluation Value Prediction Model Generation System 1, Food Evaluation Value Prediction System 2> As shown in Figure 1, the food evaluation value prediction system 2 includes a prediction target analysis parameter acquisition unit 21 and a prediction unit 22. The food evaluation value prediction model generation system 1 includes a training data acquisition unit 10, training data D1, and a learning unit 11. The training data D1 is stored in memory 1a. The training data acquisition unit 10 and the learning unit 11 are implemented by processor 1b. The prediction target analysis parameter acquisition unit 21 and the prediction unit 22 are implemented by processor 2b. In the first embodiment, processors 1b and 2b in one device implement each unit, but the system is not limited to this. For example, each process may be distributed using a network, and multiple processors 1b and 2b may be configured to execute the processing of each unit. That is, one or more processors 1b and 2b execute the processing.

[0022] [Predictive model generation] The training data acquisition unit 10 takes multiple analytical parameters representing the analysis results of food by the analytical instrument 3 as input (explanatory variables) and acquires training data D1 with food evaluation values ​​as output (dependent variable). In the first embodiment, the multiple branching parameters in the training data D1 include analytical parameters obtained by a taste sensor (indicated as TS in Figure 1), analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS). The number of analytical parameters obtained by the taste sensor includes 8 types of taste sensors × (10 basic parameters + 6 × 10 combination parameters) = approximately 560 parameters. The analytical parameters obtained by DART-MS include approximately 2000 parameters. The analytical parameters obtained by NIRS include approximately 1000 parameters. The analytical parameters obtained by NMR include approximately 350 parameters. The analytical parameters obtained by GC-MS include approximately 50 parameters. In a single model, there is one food evaluation value per food item as the output (dependent variable) for each input (group of explanatory variables). In Figure 1, if the food evaluation value X1 is, for example, Tstart, then the food evaluation value (Tstart) for food 1 is X 1_1 This is written as follows, and the food evaluation value (Tstart) of food 2 is X 1_2 Similarly, the food evaluation value (Tstart) of food 9 is expressed as X 1_9 This is how it is written. In other words, in the first embodiment, in order to generate a predictive model 20 that predicts food evaluation values ​​(Tstart), there are nine rows of data, each row taking the above-mentioned multiple analysis parameters as input and outputting one food evaluation value. Although not shown in Figure 1, there are 25 types of food evaluation values, so training data D1 is obtained similarly for each food evaluation value, and one predictive model 20 is generated for each type of food evaluation value. There are 25 types of predictive models 20 generated.

[0023] The learning unit 11 shown in Figure 1 generates a predictive model 20 that outputs food evaluation values ​​by performing machine learning using the training data D1. In the first embodiment, the predictive model 20 uses a regression equation based on PLS (partial least squares) regression analysis, but it is not limited to this as long as it is a supervised machine learning model. For example, various models such as regression trees, random forests, support vector machines, neural networks, and ensembles can be used. In the first embodiment, since there are only nine samples for a large number of explanatory variables, PLS regression analysis, which has a high predictive value even with a small number of samples for a large number of explanatory variables, is used. However, if a large number of samples can be prepared, it is not limited to PLS regression analysis, and other regression methods can also be used.

[0024] Furthermore, in the first embodiment, PLS regression analysis, which is robust against multicollinearity, is used. However, in order to reduce multicollinearity as much as possible, a first removal processing unit 12 that performs a first removal process and a second removal processing unit 13 that performs a second removal process are provided. The first removal process is a process of determining which analytical parameters to be removed are selected from among a plurality of parameters obtained by the analytical instrument 3, and whose correlation coefficient with the food evaluation value is lower than a predetermined threshold. The second removal process is a process of determining which of the analytical parameters obtained by the analytical instrument 3, which have a relatively high correlation coefficient with the food evaluation value, will be excluded. The generation flow of the prediction model 20, including the first removal process, the second removal process, and PLS regression analysis, will be explained with reference to Figure 5.

[0025] Figure 5 is a flowchart illustrating the method for generating a predictive model. As shown in Figure 5, the first elimination process is initiated. Specifically, in step ST1, the correlation coefficient between each of the multiple analytical parameters (approximately 4000) obtained by multiple analytical instruments 3 and the food evaluation value is calculated. In the next step ST2, analytical parameters whose absolute value of the correlation coefficient with the food evaluation value is lower than a predetermined threshold (0.5) are determined to be excluded and deleted. This completes the first elimination process. In one example, approximately 2000 explanatory variables remained at the end of the first elimination process.

[0026] Next, the second elimination process begins. Specifically, in step ST3, all possible combinations (pairs) of multiple analysis parameters are generated. In the next step, ST4, the pairs of analysis parameters generated in step ST3 are sorted in descending order of their correlation coefficient with the food evaluation value. In the next step, ST5, one of the first or second analysis parameters constituting a pair is determined to be excluded and deleted. More specifically, the correlation index with respect to the food evaluation value of the group of analysis parameters to which the first analysis parameter constituting the pair belongs is compared with the correlation index with respect to the food evaluation value of the group of analysis parameters to which the second analysis parameter constituting the pair belongs. The analysis parameter belonging to the group with the lower correlation index is determined to be excluded and deleted. More specifically, the correlation index is the RV coefficient with respect to the food evaluation value obtained by MFA analysis. The reason for performing the second elimination process is, for example, when a first analytical parameter obtained by a taste sensor and a second analytical parameter obtained by NMR are paired (when both the first and second analytical parameters have relatively high correlation coefficients with respect to the food evaluation value), the correlation index (RV coefficient) of all analytical parameters obtained by the taste sensor with respect to the food evaluation value is compared with the correlation index (RV coefficient) of all analytical parameters obtained by NMR. The analytical parameters with high group correlations with the food evaluation value are retained, and those with low group correlations are eliminated. This makes it possible to consider the correlation of all types of analytical parameters with respect to the food evaluation value, not just individual values. The process in step ST5 is repeated until the number of remaining analytical parameters (explanatory variables) is 30 or less (step ST6: NO). The second elimination process makes it possible to reduce the number of explanatory variables from approximately 2000 to 30 or less.

[0027] In step ST6, if the number of remaining analysis parameters (explanatory variables) is 30 or less (step ST6: YES), then in the next step ST7, PLS regression analysis is performed. A detailed explanation of PLS ​​regression analysis is omitted. Briefly, it involves repeatedly performing PLS mapping based on the remaining explanatory variables to select the optimal factors and generate a regression equation, and then generating a regression equation using only explanatory variables with high variable weights (VIP) (e.g., thresholds of 0.9 or higher). The termination conditions for the iteration include either the number of explanatory variables being less than or equal to a predetermined value (5 in the first embodiment) or the number of explanatory variables being greater than or equal to the number of selected factors. This generates a predictive model 20 (regression equation) with 5 or fewer analysis parameters as input. In some cases, the number of analysis parameters may be greater than 5.

[0028] [Food Evaluation Value Prediction] The prediction target analysis parameter acquisition unit 21 acquires multiple analysis parameters obtained by the analysis instrument 3. The analysis parameters to be acquired are five or fewer analysis parameters selected during the generation stage of the prediction model 20, and these differ for each prediction model 20. The analysis parameters are measured by the analysis instrument 3 in the same manner as the analysis parameters included in the training data D1 acquired by the training data acquisition unit 10.

[0029] As shown in Figure 1, the prediction model 20 constituting the prediction unit 22 is a model pre-trained to take multiple analysis parameters representing the analysis results of food by the analytical instrument 3 as input and output food evaluation values, as described above. The prediction unit 22 inputs the multiple analysis parameters acquired by the prediction target analysis parameter acquisition unit 21 into the prediction model 20 and outputs food evaluation values. The multiple analysis parameters input into the prediction model 20 do not include analysis parameters that were determined to be excluded in the first exclusion process and the second exclusion process. Furthermore, the number of multiple analysis parameters input into the prediction model 20 in the prediction stage is smaller than the number of multiple analysis parameters obtained by the multiple analytical instruments 3 in the prediction model 20 generation stage.

[0030] [Effects of this disclosure] Comparative Example 1 To demonstrate the effects of the above disclosure, Comparative Example 1, Example 1, and Example 2 are shown below. In Comparative Example 1, the multiple analytical parameters obtained by the analytical instrument 3 used in the generation stage of the prediction model 20 are only analytical parameters obtained from taste sensors, and for each taste sensor there are only two types: Intensity_at_30_Sec_test_1 and Intensity_at_30_Sec_CPA_1. Therefore, the analytical parameters of Comparative Example 1 are 2 types of parameters × 8 types of taste sensors = 16 analytical parameters.

[0031] Example 1 In Example 1, the analytical parameters obtained by the analytical instrument 3 used in the generation stage of the prediction model 20 include analytical parameters obtained by other types of analytical instruments 3, in addition to the analytical parameters obtained by the taste sensor in Comparative Example 1. Specifically, analytical parameters obtained by the taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS) are used in the generation stage of the prediction model 20.

[0032] Example 2 In Example 2, the analytical parameters obtained by the analytical instrument 3 used in the generation stage of the prediction model 20 are, in addition to the analytical parameters from the taste sensor in Example 1, a total of approximately 560 analytical parameters are input from the eight types of taste sensors as described above. Otherwise, it is the same as Example 1.

[0033] Comparative Example 1 and Examples 1 and 2 for some of the food evaluation values ​​among the 25 types will be explained.

[0034] Figure 6 shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the first embodiment: RMinc of the TI method. Regarding the accuracy of the regression equations, Q 2 and cumulative R2 Y and R 2 A value closer to 1 indicates higher prediction accuracy, while a value closer to 0 for RMSE (error) indicates smaller error.

[0035] Figure 7 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the first embodiment:Tend of the TI method.

[0036] Figure 8 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the RMdec of the TI method for food evaluation values ​​in the first embodiment.

[0037] Figure 9 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Rinc of the food evaluation value of the first embodiment using the TI method.

[0038] Figure 10 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Rdec of the food evaluation value of the first embodiment using the TI method.

[0039] Figure 11 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the first embodiment: DUR of the TI method.

[0040] Figure 12 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Dpla of the TI method in the first embodiment of food evaluation values.

[0041] Figure 13 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Dinc of the food evaluation value of the first embodiment using the TI method.

[0042] Figure 14 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the first embodiment: Ddec of the TI method.

[0043] Figure 15 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the AUC of the TI method in the first embodiment of food evaluation values.

[0044] Figure 16 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Apla of the TI method, according to the first embodiment of food evaluation values.

[0045] Figure 17 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Ainc of the food evaluation value of the first embodiment using the TI method.

[0046] Figure 18 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Adec of the food evaluation value of the first embodiment using the TI method.

[0047] Figure 19 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, regarding the astringency of the food evaluation value of the first embodiment using the QDA method.

[0048] Figure 20 shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for citrus fruits using the QDA method for food evaluation values ​​in the first embodiment.

[0049] Figure 21 shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for herbal medicine using the QDA method for food evaluation values ​​in the first embodiment.

[0050] Figure 22 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, regarding the food evaluation value: kire (sharpness) of the QDA method in the first embodiment.

[0051] FIG. 23 is a diagram showing the accuracy of the regression equations and the analysis parameters serving as the inputs of the regression equations for Comparative Example 1 and Examples 1 and 2 regarding the lastingness of the food evaluation value: QDA method of the first embodiment.

[0052] FIG. 24 is a diagram showing the accuracy of the regression equations and the analysis parameters serving as the inputs of the regression equations for Comparative Example 1 and Examples 1 and 2 regarding the rapidness of the manifestation of the food evaluation value: QDA method of the first embodiment.

[0053] FIG. 25 is a diagram showing the accuracy of the regression equations and the analysis parameters serving as the inputs of the regression equations for Comparative Example 1 and Examples 1 and 2 regarding the spread of the food evaluation value: QDA method of the first embodiment.

[0054] FIG. 26 is a diagram showing the accuracy of the regression equations and the analysis parameters serving as the inputs of the regression equations for Comparative Example 1 and Examples 1 and 2 regarding the thickness of the food evaluation value: QDA method of the first embodiment.

[0055] FIG. 27 is a diagram showing the accuracy of the regression equations and the analysis parameters serving as the inputs of the regression equations for Comparative Example 1 and Examples 1 and 2 regarding the woody of the food evaluation value: QDA method of the first embodiment.

[0056] Regarding Comparative Example 1 and Examples 1 and 2 shown in FIGS. 6 to 27, Q 2 and the cumulative R 2 Y and R 2 all show that Examples 1 and 2 are greater than Comparative Example 1, and the RMSE shows that Examples 1 and 2 are smaller than Comparative Example 1. That is, it can be seen that when using the analysis parameters obtained by two or more types of analysis instruments 3 (Examples 1 and 2), the prediction accuracy can be improved compared to when using the analysis parameters obtained by one type of analysis instrument 3 (Comparative Example 1).

[0057] <Second Embodiment> [Food Evaluation Value (Sensory Evaluation Value, Sweetness Evaluation Value)] In the second embodiment, sweetness among the tastes is illustrated. Ten types of sweeteners and sweetening materials were used as food samples to evaluate only sweetness. Specifically, these were sucrose, sucralose, acesulfame K, advantame, thaumatin, stevia extract (A), stevia extract (B), stevia extract (C), monk fruit extract (A), and monk fruit extract (B). Twenty selected panelists compared each sample with a control (6% sucrose aqueous solution) and serially diluted samples, selecting the one they perceived as sweeter (two-point discrimination method). The concentration at which the selectivity rate was 50% was defined as the subjective equivalent concentration of sweetness, and samples were prepared to this concentration for subsequent evaluations and analyses. Each sample was evaluated using Quantitative Descriptive Analysis (QDA) and Time-Intensity Analysis (TI), and the following 22 food evaluation values ​​(sweetness evaluation values) were obtained.

[0058] [Food evaluation values ​​showing the qualitative characteristics of sweetness based on Quantitative Descriptive Analysis (QDA method)] A panel of 21 selected members determined seven characteristic parameters and their definitions that were suitable for the sample. Through trial evaluations, appropriate evaluation frequencies and sample quantities were determined. After sufficient evaluation training, the intensity of each characteristic parameter was evaluated on a 15cm line scale. The evaluation was repeated twice. Each sample was assigned a three-digit random number, and the liquid color was blinded during the testing. The following shows the 10 food evaluation values ​​(characteristic items) and their definitions. For example, the definition of the characteristic item "Top Intensity" is shown to be "Top Taste Sensation of Sweetness." • Initial Intensity: The initial taste sensation of sweetness (higher intensity indicates a stronger taste sensation). • Lastingness: A sweet aftertaste lingers, the longer the lingering sensation (the higher the intensity, the longer the aftertaste). • Bitterness: Maximum bitterness intensity (higher intensity indicates stronger bitterness) • Astringency: Irritation, numbness, or bitterness in the mouth (the higher the intensity, the stronger the irritation). • Flat: Mild, light, bland (the higher the intensity, the milder it is) • Stickiness (Coating): Sticky, clinging, rough, persistent (the stronger the stickiness, the more it feels like it's sticking to the inside of the mouth) • Aroma (Caramel): Darkness of brown sugar, caramel, savory, soy sauce, soybeans, cooked flavor (the higher the intensity, the stronger the brown sugar flavor)

[0059] [Food evaluation values ​​showing the time-dependent characteristics of sweetness using the Time-Intensity (TI) method] A selection panel of 21 participants conducted trial evaluations to determine appropriate sample quantities, swallowing times, etc. Furthermore, after sufficient evaluation training, the temporal intensity of "sweetness" was evaluated using a slider bar on a 10cm line scale. The evaluation was repeated three times. As shown in Figure 2, in the second embodiment, there are 15 food evaluation values ​​that indicate the temporal characteristics of sweetness using the TI method. The 15 food evaluation values ​​and their definitions are as follows. • Rinc: Slope of increase • RMinc: Maximum slope of increase • Rdec: Slope of decrease • RMdec: Maximum slope of decrease • Tstart: The first time sweetness was detected. • Dinc: Increased time • Tmax: Time at which the intensity first reaches its maximum. • Dpla: Duration of the plateau of maximum intensity ·Ddec: Decrease time • Tendency: The time when you no longer taste sweetness. • DUR: Total time spent feeling sweetness Ainc: Area under the increasing curve • Apla: Area under the curve of the maximum intensity plateau • Adec: Area under the decreasing curve • AUC: Area under the curve of the entire curve

[0060] [Example of analysis parameter: Taste sensor] An example of analytical instrument 3 is a taste sensor. An example of analytical parameters obtained by the taste sensor will be described. The taste sensor is a sensor that obtains response information to the lipid membrane. Each sample was analyzed with nine types of taste sensors. The specific items of the nine types of taste sensors and the sensor identifiers are as follows. In the second embodiment, UM2 is added to the eight types of taste sensors in the first embodiment. The measurement method for each taste sensor is the same as in the first embodiment. ·Acidic bitterness:C00 • Hydrochloride bitter taste: BT0 • Basic salt bitterness: AN0 • Astringency: AE1 • Acidity: CA0 • Saltiness: CT0 Umami: AAE, UM2 • Sweetness: GL1 For each taste sensor, a first-stage initial taste measurement, a second-stage first aftertaste measurement, and a third-stage second aftertaste measurement were performed, and the response value over a 30-second measurement period was measured in 1-second increments at each stage. Figure 3 is a schematic explanatory diagram showing the measurement state of the taste sensor and the response value (mV) obtained from the sensor. The measurement method is as schematically shown in Figure 3 for membrane potential (sensor output), first by immersing the taste sensor in a reference solution to obtain a baseline membrane potential (sensor output). Next, when the taste sensor is immersed in the sample, the membrane potential changes due to interaction with the taste substance, and the amount or portion of the change corresponds to the initial taste (sometimes referred to as test_1). Next, the taste sensor is washed with the reference solution. Next, the taste sensor is immersed in the reference solution and the change in membrane potential is measured. The amount or portion of the change corresponds to the first aftertaste (sometimes referred to as CPA1). In the first embodiment, the taste sensor is further washed with the reference solution, and then the taste sensor is immersed in the reference solution and the change in membrane potential is measured. The amount or part of the change corresponds to the second aftertaste (sometimes referred to as CPA2).

[0061] The following describes the multiple analytical parameters obtained by the taste sensor in the second embodiment. Similar to the first embodiment, the analytical parameters are obtained based on the sensor intensity at multiple time points (0 seconds, 1 second, 2 seconds, ..., 29 seconds, 30 seconds) measured between the start and end of the measurement (30 seconds). The 13 basic analytical parameters obtained by the taste sensor and their definitions are as follows. ·Imax: Maximum intensity • Tmax: Time to reach maximum intensity • Plateau Intensity: 90% Imax (Maximum intensity plateau) • Plateau_Reach_Time: Time to reach 90% Imax (maximum intensity plateau) • Plateau_Duration: Duration of the plateau of maximum intensity • Intensity_at_1_Sec: Intensity at 1 second (start of measurement) • Intensity_at_3_Sec: Intensity at 3 seconds (start of measurement) • Intensity_at_30_Sec: Intensity at 30 seconds (end of measurement) • Slope_1_to_3: Slope from 1 second to 3 seconds. • Slope_3_to_Tmax: Slope from 3 seconds to Tmax • Slope_Tmax_to_30: Slope from Tmax to 30 seconds • Slope_3_to_Plateau: Slope from 3 seconds to 90% Imax · Slope_Plateau_to_Tmax: Slope from 90% Imax to Tmax In the first embodiment, a parameter relating to area is used, but in the second embodiment, the parameter relating to area is replaced with a parameter relating to slope. Since the parameter relating to area is the product of intensity and time, it is a parameter that includes intensity. However, a parameter relating to the slope of the line connecting the intensity at the first time point and the intensity at the second time point is a parameter that does not include intensity, thus differentiating it from parameters that include intensity, such as Imax. The 13 parameters listed above are basic parameters. Since the initial taste, the first aftertaste, and the second aftertaste each have the 13 reference parameters listed above, there are a total of 39 parameters. For example, the maximum intensity of the initial taste is denoted as Imax_test_1, the maximum intensity of the first aftertaste is denoted as Imax_CPA_1, and the maximum intensity of the second aftertaste is denoted as Imax_CPA_2. Furthermore, for each basic parameter, it becomes possible to generate combination parameters that represent the difference (def) between two combinations of the initial taste, first aftertaste, and second aftertaste. For example, the difference between the initial taste and second aftertaste in the AUC_total of the taste sensor (C00) can be expressed as TasteSensor_SB2C00_AUC_total_diff_test_1_vs_CPA_2. Based on one basic parameter, there are three combination parameters: the initial taste and first aftertaste, the initial taste and second aftertaste, and the first aftertaste and second aftertaste. Therefore, it becomes possible to generate three combination parameters based on one basic parameter.

[0062] In taste sensor analysis, measurement error correction is performed using interpolation and difference. In this correction, a regression equation is created with the measured value of sample A, designated as a control, as the dependent variable and the measurement order as the independent variable. An interpolated value is calculated that takes into account the state change of the waiting sample, and the difference is calculated from the test data. In situations where a curve is drawn from measured values, as in this method, a smooth curve may not be drawn when the difference is calculated due to the slope and inflection point of the control. Therefore, in this method, the curve drawn using raw data without corrected interpolation and difference processing was used to calculate parameters (Tmax, ...Time, Slope...) based on the curve shape. For the curve, the response values ​​at each time point were arithmetic mean of the data from three trials that were smoothed using Savizchy-Golay, and each parameter was calculated from the resulting average curve.

[0063] [Example of analysis parameters] DART-MS, NIRS, and NMR are the same as in the first embodiment. Regarding GC-MS, it is the same as in the first embodiment except that the sample was extracted with dichloromethane, concentrated, and then subjected to GC-MS.

[0064] Regarding the method for generating the predictive model, the analysis parameters obtained from the taste sensors included in the training data D1 consist of 9 types of taste sensors × (13 basic parameters + 3 × 13 combination parameters) = approximately 468 parameters. Regarding the method for generating the predictive model, before performing the first elimination process in the first embodiment, the coefficient of variation (CV) of each explanatory variable was calculated, and variables with a coefficient of variation (CV) smaller than a predetermined threshold (e.g., 0.05) were removed from the explanatory variables. Next, for the feature groups classified as DARTnega, DARTposi, and NMR, only the top 500 variables with high F-values ​​calculated by the f_classif function of the scikit-learn library were retained, and all others were excluded.

[0065] [Effects of this disclosure] To demonstrate the effects of the present disclosure described above, Comparative Example 1, Example 1, and Example 2 are shown below. In Comparative Example 1, the analytical parameters obtained by the analytical instrument 3 used in the generation stage of the prediction model 20 are only those obtained from the taste sensor, and for each taste sensor, there are only two types: Intensity_at_30_Sec_test_1 and Intensity_at_30_Sec_CPA_1. Therefore, the analytical parameters in Comparative Example 1 are 2 types of parameters × 9 types of taste sensors = 18 analytical parameters.

[0066] In Example 1, the analytical parameters obtained by the analytical instrument 3 used in the generation stage of the prediction model 20 include analytical parameters obtained by other types of analytical instruments 3, in addition to the analytical parameters obtained by the taste sensor in Comparative Example 1. Specifically, analytical parameters obtained by the taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS) are used in the generation stage of the prediction model 20.

[0067] In Example 2, the analytical parameters obtained by the analytical instrument 3 used in the generation stage of the prediction model 20 are, in addition to the analytical parameters from the taste sensor in Example 1, a total of approximately 468 analytical parameters are input from the nine types of taste sensors as described above. Otherwise, it is the same as Example 1.

[0068] Comparative Example 1 and Examples 1 and 2 for some of the food evaluation values ​​among the 22 types will be explained.

[0069] Figure 28 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Rinc of the food evaluation value of the second embodiment:TI method.

[0070] Figure 29 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for RMinc of the food evaluation value of the second embodiment using the TI method.

[0071] Figure 30 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Rdec of the food evaluation value of the second embodiment using the TI method.

[0072] Figure 31 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the RMdec of the TI method in the second embodiment of food evaluation values.

[0073] Figure 32 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Tstart of the TI method, according to the second embodiment of food evaluation values.

[0074] Figure 33 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Dinc of the food evaluation value of the second embodiment: TI method.

[0075] Figure 34 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Tmax of the food evaluation value of the second embodiment: TI method.

[0076] Figure 35 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Dpla of the TI method in the second embodiment of food evaluation values.

[0077] Figure 36 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Ddec of the food evaluation value of the second embodiment using the TI method.

[0078] Figure 37 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the second embodiment:Tend of the TI method.

[0079] Figure 38 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the second embodiment: DUR of the TI method.

[0080] Figure 39 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Ainc of the food evaluation value of the second embodiment: TI method.

[0081] Figure 40 shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for Apla of the TI method in the second embodiment of food evaluation values.

[0082] Figure 41 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for Adec of the food evaluation value of the second embodiment: TI method.

[0083] Figure 42 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the AUC of the TI method in the second embodiment of food evaluation values.

[0084] Figure 43 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, regarding the initial intensity of the top food evaluation value of the QDA method in the second embodiment.

[0085] Figure 44 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, regarding the food evaluation value of the second embodiment: lastingness of the QDA method.

[0086] Figure 45 shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, and for the food evaluation value of the second embodiment: bitterness using the QDA method.

[0087] Figure 46 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the astringency of the food evaluation value of the second embodiment using the QDA method.

[0088] Figure 47 shows the accuracy of the regression equations and the analytical parameters used as inputs to the regression equations for Comparative Example 1, Examples 1 and 2, and for the monotonic (Flat) food evaluation value of the second embodiment of the QDA method.

[0089] Figure 48 shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, regarding the food evaluation value of the second embodiment: stickiness (coating) using the QDA method.

[0090] Figure 49 shows the accuracy of the regression equations and the analytical parameters used as input to the regression equations for Comparative Example 1, Examples 1 and 2, regarding the food evaluation value of the second embodiment: aroma (Caramel) using the QDA method.

[0091] In Comparative Example 1 and Examples 1 and 2 shown in Figures 28 to 49, Q2, cumulative R2Y, and R2 are all greater in Examples 1 and 2 than in Comparative Example 1, while the RMSE is smaller in Examples 1 and 2 than in Comparative Example 1. In other words, it can be seen that using analytical parameters obtained from two or more analytical instruments 3 (Examples 1 and 2) can improve prediction accuracy compared to using analytical parameters obtained from one type of analytical instrument 3 (Comparative Example 1).

[0092] Therefore, it can be understood that the food evaluation value prediction method of this disclosure is effective whether the food evaluation value for bitterness is as shown in the first embodiment or for sweetness as shown in the second embodiment.

[0093] [Selection rate for each parameter] Figure 50 is an explanatory diagram showing the proportion of each parameter selected as an important variable. As shown in Figure 50, since the contributing components and receptor properties differ for each of the five tastes, the types of important variables selected by regression analysis showed characteristics specific to each taste.

[0094] The types of instruments used to select key variables showed different trends for sweetness and bitterness. In bitterness prediction, all analytical methods except GC-MS contributed relatively equally to the prediction, but in sweetness prediction, there was a bias towards NIR. This is thought to be due to differences in the properties of taste substances. Sweet substances often contain OH groups and have a higher threshold than bitter substances, so it is thought that clearer response values ​​were obtained with NIR. In DART as well, in sweetness prediction, measurement results from the negative charge mode were selected more often as key variables, which is also thought to be due to differences in the properties of contributing substances. If the food evaluation value is a numerical value representing bitterness, the multiple analytical parameters indicating the analysis results of the food may include analytical parameters obtained by a taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), and analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR). If the food evaluation value is a numerical value representing sweetness, the multiple analytical parameters indicating the analysis results of the food may include analytical parameters obtained by a near-infrared analyzer (NIRS).

[0095] In sweetness prediction and bitterness prediction, the types of sensors selected as important variables differed in some respects. In Example 2, where the taste sensor parameters included parameters related to time, intensity, and slope, the sensor response values ​​for sweetness, bitterness, and saltiness were effective in predicting bitterness, while in sweetness prediction, the sensor response values ​​for sweetness, umami, and bitterness tended to be effective. This is thought to be a tendency dependent on the receptor properties of each taste quality. Among the five basic tastes, bitterness, sweetness, and umami are tastes perceived through the response of G protein-coupled receptors, but since sweetness and umami are perceived by dimers containing T1R3, it is thought that sensors corresponding to umami were selected as important variables more often in sweetness prediction. On the other hand, bitterness is known to perceive a wide range of components through 25 T2R receptors, and there have also been reports on mineral responses, so it is thought that sensors corresponding to saltiness were selected as important variables more often in bitterness prediction. If the food evaluation value is a numerical representation of bitterness, and the parameters of the taste sensor include parameters related to time, intensity, and slope, then the taste sensor may also include sensors for sweetness, bitterness, and saltiness. If the food evaluation value is a numerical representation of sweetness, and the parameters of the taste sensor include parameters related to time, intensity, and slope, then the taste sensor may include sensors for sweetness, umami, and bitterness.

[0096] On the other hand, in Example 1, where the taste sensor parameters included only parameters related to intensity, the sensor response values ​​for bitterness and sourness were effective in predicting bitterness, while the sensor response value for sweetness tended to be more effective in predicting sweetness.

[0097] In sweetness prediction, the usage rate of CPA_2 (second aftertaste) tended to be lower than in bitterness prediction. Since many bitter substances are lipid-soluble and their adsorption time to receptors and oral lipid membranes contributes to sensory persistence, the value of CPA_2 was selected more often as an important variable. On the other hand, in the sensory evaluation results for sweetness, "initial intensity" and "persistence" showed a high correlation, suggesting that the intensity of the sensation contributes to the aftertaste, and therefore the value of CPA_2, which indicates membrane adsorption capacity, was not effective in prediction. Furthermore, the usage rate of Time tended to be lower than that of bitterness prediction, suggesting that the curve parameter information disclosed herein was more effective for bitterness prediction. However, even in sweetness prediction, Example 2, which used curve parameter information, showed a much higher accuracy than Example 1, suggesting that it is a sufficiently effective method for predicting tastes other than bitterness as well.

[0098] [1] As in the embodiment described above, the food evaluation value prediction method includes the steps of: acquiring a plurality of analytical parameters that indicate the analysis results of the food to be predicted by the analytical instrument 3; and obtaining a food evaluation value by inputting the acquired plurality of analytical parameters into a prediction model 20 that has been pre-trained to take the plurality of analytical parameters that indicate the analysis results of the food by the analytical instrument 3 as input and output a food evaluation value that quantifies the bitterness, which is the taste of the food. The plurality of analytical parameters that indicate the analysis results of the food may include at least two types of analytical parameters from among analytical parameters obtained by a taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS). As shown in Comparative Example 1 or the comparison between Example 1 and Example 2, using multiple analytical parameters obtained by two or more types of analytical instruments 3 improves the prediction accuracy of food evaluation values ​​compared to using multiple analytical parameters obtained by one type of analytical instrument 3.

[0099] [2] The food evaluation value prediction method described in [1] above may be configured such that, in the generation stage of the prediction model 20, a first exclusion process is performed to determine which analytical parameters to be excluded from among a plurality of analytical parameters obtained by the analytical instrument 3 have a correlation coefficient with the food evaluation value that is lower than a predetermined threshold, and the plurality of analytical parameters to be input to the prediction model 20 do not include the analytical parameters that were determined to be excluded in the first exclusion process. By reducing the number of explanatory variables, it becomes possible to reduce or avoid multicollinearity and improve the prediction accuracy of food evaluation values.

[0100] [3] The food evaluation value prediction method described in [1] or [2] above may be such that, in the generation stage of the prediction model 20, a second exclusion process is performed in which one of the analytical parameters with a relatively high correlation coefficient with the food evaluation value from among the multiple analytical parameters obtained by the analytical instrument 3 is selected as an analytical parameter to be excluded, and the multiple analytical parameters input to the prediction model 20 in the prediction stage do not include the analytical parameters that were selected as to be excluded in the second exclusion process. By reducing the number of explanatory variables, it becomes possible to reduce or avoid multicollinearity and improve the prediction accuracy of food evaluation values.

[0101] [4] As in the embodiment described above, the food evaluation value prediction model generation method includes the step of generating a prediction model 20 that outputs a texture evaluation value, using training data D1 which takes as input multiple analytical parameters that show the results of food analysis by the analytical instrument 3 and outputs a food evaluation value that quantifies one of the senses of taste, texture, and smell of the food. The multiple analytical parameters that show the results of food analysis may include at least two types of analytical parameters from among analytical parameters obtained by a taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS).

[0102] [5] As in the above embodiment, the program may cause one or more processors 1b, 2b to execute the method described in any of [1] to [4] above.

[0103] The computer-readable temporary recording medium according to the above embodiment stores the above program.

[0104] Although embodiments of this disclosure have been described above with reference to the drawings, it should be understood that the specific configurations are not limited to these embodiments. The scope of this disclosure is indicated not only by the description of the embodiments above but also by the claims, and further includes all modifications within the meaning and scope equivalent to the claims.

[0105] The structures adopted in each of the above embodiments can be adopted in any other embodiment. The specific configuration of each part is not limited to the embodiments described above, and various modifications are possible without departing from the spirit of this disclosure.

[0106] <Variation> (A) In the above embodiments, bitterness is given as an example of a food evaluation value that quantifies one of the senses of taste, texture, and smell of food, but the invention is not limited to this. For example, this disclosure can also be applied to food evaluation values ​​that quantify tastes such as sweetness, umami, depth of flavor, saltiness, and sourness as tastes of food. Furthermore, since the analytical method using analytical instrument 3 mentioned in the examples can analyze aroma-contributing components, it can also be applied to food evaluation values ​​that quantify smell. Furthermore, it can also be applied to food evaluation values ​​that quantify textures such as chewiness and biteability of food. Since the analytical method using analytical instrument 3 mentioned in the examples can analyze components that contribute to viscosity and hardness, it can also be applied to food evaluation values ​​that quantify texture.

[0107] (B) In the above embodiment, the number of types of food evaluation values ​​is 25, but is not limited to this. If one type of food evaluation value is sufficient, the number of types of food evaluation values ​​may be just one, and the number of types of food evaluation values ​​can also be changed in various ways.

[0108] (C) The sensory evaluation method for food evaluation values ​​is not limited to the above. For example, the number of evaluators can be changed as appropriate.

[0109] (D) In ​​the above embodiment, both the first removal process and the second removal process are performed, but the embodiment is not limited to this. Only the first removal process may be performed, or only the second removal process may be performed, and the prediction accuracy can be improved compared to when no removal process is performed. [Explanation of symbols]

[0110] 1b: Processor 2b: Processor 3:Analytical equipment 20: Predictive Models D1: Training data

Claims

1. The steps include obtaining multiple analytical parameters that represent the analysis results of the food to be predicted using an analytical instrument, and The process involves taking multiple analytical parameters representing the results of food analysis by an analytical instrument as input, and inputting the acquired multiple analytical parameters into a pre-trained predictive model that outputs a food evaluation value that quantifies one of the senses of taste, texture, and smell of the food, in order to obtain the food evaluation value. Includes, A method for predicting food evaluation values, wherein the plurality of analytical parameters indicating the analysis results of food include at least two analytical parameters from among analytical parameters obtained by a taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS).

2. In the generation stage of the prediction model, a first exclusion process is performed to determine which analytical parameters to be excluded from among a plurality of analytical parameters obtained by an analytical instrument have a correlation coefficient with the food evaluation value that is lower than a predetermined threshold. The method for predicting food evaluation values ​​according to claim 1, wherein the plurality of analytical parameters input to the prediction model in the prediction stage do not include analytical parameters that have been determined to be excluded in the first exclusion process.

3. In the generation stage of the prediction model, a second exclusion process is performed in which, among a plurality of analytical parameters obtained by the analytical instrument, one of the analytical parameters with a relatively high correlation coefficient with the food evaluation value is selected as the analytical parameter to be excluded. The method for predicting food evaluation values ​​according to claim 1 or 2, wherein the plurality of analytical parameters input to the prediction model in the prediction stage do not include analytical parameters that have been determined to be excluded in the second exclusion process.

4. The process includes the step of generating a predictive model that outputs a texture evaluation value, using training data that takes multiple analytical parameters representing the results of food analysis by analytical instruments as input and outputs a food evaluation value that quantifies one of the senses of taste, texture, and smell of the food. A method for generating a food evaluation value prediction model, wherein the multiple analytical parameters indicating the analysis results of the food include at least two analytical parameters from among analytical parameters obtained by a taste sensor, analytical parameters obtained by a real-time direct mass spectrometer (DART-MS), analytical parameters obtained by a near-infrared analyzer (NIRS), analytical parameters obtained by a nuclear magnetic resonance spectrometer (NMR), and analytical parameters obtained by a gas chromatography-mass spectrometer (GC-MS).

5. A program that causes one or more processors to execute the method according to claim 1, 2, or 4.

Images