Waveform analysis methods and waveform analysis devices

Abstract

A waveform analysis device (4) analyzes a target waveform as a chromatogram or spectrum. The waveform analysis device (4) includes: a waveform segmentation unit (54) that segments the target waveform into multiple partial waveforms; a determination unit (55) that uses a learned model to determine whether each of the multiple partial waveforms of the target waveform is a peak portion, the learned model being generated by machine learning using multiple sets of multiple partial waveforms created by segmenting a reference waveform whose peak position is known; and a classification unit (56) that classifies the target waveform into a continuous peak region of the peak portion and a non-peak region other than the peak region based on the determination result of the determination unit.

Description

Technical Field

[0001] This invention relates to a technique for analyzing the waveforms of chromatograms and spectra. Background Technology

[0002] A chromatograph is used to identify or quantify the components contained in a sample. In the chromatograph, the components in the sample are separated using a column, and the components eluting from the column are detected sequentially. Then, a chromatogram is prepared with time set to the horizontal axis and detection intensity set to the vertical axis, and the concentration or content of the compound corresponding to the peak is determined based on the area or height of the chromatogram.

[0003] To determine the area or height of a peak from a chromatogram, it is necessary to identify the start and end points of the peak rising from the chromatographic baseline. The process of identifying the start and end points of a chromatographic peak is called peak selection. Since noise is often present in the chromatographic baseline and peaks, this must be taken into account to properly identify the peak start and end points. Various peak selection algorithms exist that take such noise into account; for example, Patent Document 1 describes a technique for peak selection using filtering with a filter having a bandwidth and a shielding frequency corresponding to the degree of temporal variation in the chromatographic data. Chromatographic analysis software can execute multiple algorithms, allowing users to select any algorithm and set the desired parameters for peak selection (e.g., Non-Patent Document 1).

[0004] Existing technical documents

[0005] Patent documents

[0006] Patent Document 1: Japanese Patent Application Publication No. 2009-8582

[0007] Non-patent literature

[0008] Non-patent document 1: "Efficiency of Waveform Processing Algorithm and Analysis Business of Waveform Processing Algorithm", [online], [Retrieval August 16, 2019], Shimadzu Corporation, Internet <URL:http: / / www.an.shimadzu.co.jp / hplc / support / faq / faq8 / i-peakfinder_introduction.ht m>

[0009] Non-Patent Document 2: Olaf Ronneberger, two others, “U-Net: Convolutional Networks for Biomedical Image Segmentation”, [online], [submitted May 18, 2015], arXiv.org, Internet<URL:https: / / arxiv.org / pdf / 1505.04597.pdf> Summary of the Invention

[0010] The problem the invention aims to solve

[0011] The shape or noise pattern of chromatographic peaks or baselines varies depending on the type of sample or detector. Therefore, peak selection becomes complex, requiring the exploration of suitable algorithms and parameters for the sample being chromatographically processed and the detector used. Furthermore, analyst subjectivity can easily be introduced into parameter settings, potentially leading to less precise peak selection.

[0012] Here, the peak selection of the chromatogram obtained by separating and measuring the components in the sample using a chromatograph is explained. However, the same problem exists when the peak selection of the spectrum obtained by measuring the components in the sample using a spectrophotometer is performed.

[0013] The problem to be solved by the present invention is to provide a technology that can perform peak picking with high precision without complicated operations.

[0014] Solution for solving the problem

[0015] One aspect of the present invention, completed to solve the above-mentioned problems, is a waveform analysis method that analyzes a waveform that is a chromatogram or spectrum. In this waveform analysis method...

[0016] A learned model is created by using multiple sets of partial waveforms, each consisting of segments of a reference waveform whose peak positions are known, to determine the peak portions contained in an input waveform through machine learning.

[0017] The waveform of the object is divided into multiple partial waveforms.

[0018] The learned model is used to determine whether each part of the multiple waveforms of the object waveform is a peak.

[0019] Based on the determination result, the object waveform is classified into a continuous peak region and a non-peak region outside the peak region.

[0020] In addition, another aspect of the present invention, completed to solve the above-mentioned problems, is a waveform analysis apparatus that analyzes a waveform that is a chromatogram or spectrum, the waveform analysis apparatus comprising:

[0021] A waveform segmentation unit that divides the object waveform into multiple partial waveforms;

[0022] The determination unit uses a learned model to determine whether each portion of the multiple partial waveforms of the object waveform is a peak. The learned model is created through machine learning using multiple sets of partial waveforms created by segmenting a reference waveform whose peak positions are known.

[0023] The classification unit, based on the determination result of the determination unit, classifies the object waveform into a continuous peak region and a non-peak region elsewhere.

[0024] The effects of the invention

[0025] In the waveform analysis method and apparatus of this invention, a learned model for determining whether an input waveform is a peak is pre-built by performing machine learning on multiple sets of partial waveforms created by segmenting a reference waveform whose peak position is known. Here, the reference waveform used in building the learned model is typically a waveform obtained by performing the same type of measurement as the target waveform (chromatography or spectroscopy), but waveforms created through theoretical calculations can also be used. Furthermore, the multiple partial waveforms of the reference waveform are created by segmenting them at intervals narrower than the peak widths of the peaks contained in the reference waveform. In this way, a learned model is built that learns various shapes of partial waveforms constituting a part of a peak. Then, similarly to segmenting the reference waveform, the target waveform is segmented at intervals narrower than the peak widths of the peaks expected to be contained in the target waveform to create partial waveforms. Then, using the learned model described above, it is determined whether each partial waveform of the multiple partial waveforms formed by segmenting the target waveform whose peak position is unknown is a peak. Therefore, multiple parts of the target waveform are classified into peak parts and non-peak parts, and based on this, continuous regions of peak parts are classified as peak regions, and continuous regions of non-peak parts are classified as non-peak regions. In the waveform analysis method and waveform analysis apparatus of the present invention, the target waveform is classified into peak regions and non-peak regions by using a learned model that has learned the features of multiple reference waveforms, thus enabling high-precision peak picking without complex operations. Attached Figure Description

[0026] Figure 1This is a structural diagram of the main parts of a liquid chromatography-mass spectrometry analysis system including an embodiment of the waveform analysis device involved in this invention.

[0027] Figure 2 This is a flowchart illustrating the process of creating a learned model in one embodiment of the waveform analysis method involved in this invention.

[0028] Figure 3 This is a flowchart illustrating the chromatographic analysis process in the waveform analysis method of this embodiment.

[0029] Figure 4 This diagram illustrates the use of the waveform analysis apparatus of this embodiment to create a two-stage output learning model.

[0030] Figure 5 This is a diagram illustrating how the waveform analysis device of this embodiment analyzes the results of chromatograms through two-stage output.

[0031] Figure 6 This diagram illustrates the use of the waveform analysis apparatus of this embodiment to create a learning model with five-level output.

[0032] Figure 7 This is a diagram illustrating the results obtained by analyzing chromatograms using the waveform analysis device of this embodiment through five-level output. Detailed Implementation

[0033] Hereinafter, an embodiment of the waveform analysis apparatus and waveform analysis method according to the present invention will be described with reference to the accompanying drawings. In this embodiment, the waveform analysis apparatus is assembled as part of a chromatography-mass spectrometry analysis system. Furthermore, the waveform analysis apparatus of this embodiment does not necessarily need to be integrally constructed with the chromatography-mass spectrometry analysis system; it can also be constructed separately from the chromatography-mass spectrometry analysis system.

[0034] exist Figure 1 The main structural components of the liquid chromatography-mass spectrometry (LC-MS) analysis system of this embodiment are shown. The LC-MS analysis apparatus of this embodiment is broadly composed of a liquid chromatograph 1, a mass spectrometer 2, and a control / processing unit 4 that controls their operation. The liquid chromatograph 1 includes: a mobile phase container 10 containing the mobile phase; a pump 11 that draws in the mobile phase and supplies it at a fixed flow rate (or flow velocity); a syringe 12 that injects a predetermined amount of sample solution into the mobile phase; and a column 13 that separates the components contained in the sample solution over time. The sample solution flowing out of the column 13 is introduced into the electrospray ionization probe 201 of the mass spectrometer 2. Furthermore, the liquid chromatograph 1 is connected to an automated sampler 14 for sequentially introducing multiple liquid samples into the syringe 12.

[0035] The mass spectrometer 2 has a multi-stage differential evacuation system. This system includes a first intermediate vacuum chamber 21 and a second intermediate vacuum chamber 22, with the vacuum level progressively increased, located between the ionization chamber 20 (at approximately atmospheric pressure) and the analysis chamber 23 (high vacuum after being evacuated by a vacuum pump (not shown)). An electrospray ionization probe (ESI probe) 201 is installed in the ionization chamber 20, which sprays the sample solution supplied from the liquid chromatograph 1 while simultaneously charging it. The ionization chamber 20 and the first intermediate vacuum chamber 21 are connected via a narrow-diameter heated capillary 202. The first intermediate vacuum chamber 21 and the second intermediate vacuum chamber 22 are separated by a separator 212 with a small orifice at the top. Ion guides 211 and 221, respectively, are installed in the first and second intermediate vacuum chambers 21 and 22 to concentrate ions and transport them to subsequent stages. A quadrupole mass filter 231 and an ion detector 232 are installed in the analysis chamber 23.

[0036] Mass spectrometer 2 is capable of performing selected ion monitoring (SIM) and MS scanning measurements. In SIM measurements, the mass-to-charge ratio of ions passing through quadrupole mass filter 231 is fixed, and only ions with that mass-to-charge ratio are detected. In MS scanning measurements, the mass-to-charge ratio of ions passing through quadrupole mass filter 231 is scanned within a specified range, and ions within that specified range are detected for each mass-to-charge ratio.

[0037] In addition to the storage unit 41, the control / processing unit 4 also includes a measurement control unit 51, a learned model creation unit 52, an analysis mode selection unit 53, a waveform segmentation unit 54, a judgment unit 55, a classification unit 56, a noise value calculation unit 57, a noise removal unit 58, a peak separation unit 59, and a baseline estimation unit 60 as functional blocks. The storage unit 41 includes a reference waveform storage unit 42, a measurement data storage unit 43, and a learned model storage unit 44. The control / processing unit 4 is essentially a personal computer, connected to an input unit 6 and a display unit 7. Furthermore, a waveform analysis program is pre-installed in the control / processing unit 4. By executing this waveform analysis program, the functions of the learned model creation unit 52, the analysis mode selection unit 53, the waveform segmentation unit 54, the judgment unit 55, the classification unit 56, the noise value calculation unit 57, the noise removal unit 58, the peak separation unit 59, and the baseline estimation unit 60 are realized.

[0038] The reference waveform storage unit 42 stores training data 421 and validation data 422. Both training data 421 and validation data 422 are data of the waveforms (raw waveforms) obtained by measuring samples containing various components using a chromatography-mass spectrometry analysis device. These waveforms (e.g., total ion chromatography showing the change in the total intensity of all detected ions with mass-to-charge ratios over time when the components separated by liquid chromatograph 1 are measured by MS scanning using mass spectrometer 2, or mass chromatography showing the change in the intensity of ions with a specific mass-to-charge ratio over time when SIM or MRM measurements are performed) are pre-selected to determine the peak positions. Furthermore, these raw waveform data are pre-normalized so that the intensity values ​​are within a specified range (e.g., ±1.0). Normalization unifies multiple chromatograms with different intensity scales into a common intensity scale, thereby improving the accuracy of the learned model described later. Here, chromatograms obtained by measuring actual samples are used as training data 421 and validation data 422, but chromatograms created through simulation can also be used.

[0039] The raw waveform of the chromatogram is divided into a predetermined number of segments along the time axis. This predetermined number is, for example, 1,024 or 512, determined such that the width (length along the time axis) of each segment is at least less than the peak width. For example, it is determined based on the peak width and the number of data points required to constitute a peak. As a specific example, we will illustrate the case where the minimum peak width in the chromatogram obtained through actual measurements is 0.2 min and the maximum peak width is 2.0 min. When acquiring this measurement data and the data of the target analysis, the analytical conditions and the condition (degree of degradation) of the column used for component separation may vary. Therefore, the above-mentioned predetermined number is set to take these differences into account, so that even if the minimum peak width narrows to 0.1 min and the maximum peak width widens to 3.0 min, it is still possible to handle the situation. Considering the reproducibility of peak shape, a peak should contain at least 20 points, and more preferably 30 points. Based on these considerations, in order to divide the chromatogram into a 3.0-minute time range and segment the original waveform by including 30 points within a 0.1-minute range, which is the envisioned minimum peak width, the aforementioned specified number needs to be set to 900 or more. In the analytical example described later, the chromatogram is divided into a 5.0-minute time range, and the aforementioned specified number is set to 1,024 based on the calculation that 5.0 min / 0.1 min × 20 points = 1,000 (≈1,024) points, so that it can handle even when the peak width is wider than the envisioned width.

[0040] Each waveform data segment corresponds to information related to the characteristics of that segment of the waveform (characteristic information). The characteristic information corresponding to a segment of the original waveform includes at least information about whether that segment of the waveform belongs to a peak region or a non-peak region. Other characteristic information will be described later.

[0041] The measurement data storage unit 43 stores various measurement conditions used in the measurement of samples using a chromatography-mass spectrometry analysis apparatus. It also stores chromatographic data obtained from the measurement of the samples. The learned model storage unit 44 stores learned models used for peak selection in chromatography. The learned model storage unit 44 also stores learned models corresponding to the resolution modes described later.

[0042] Next, the waveform analysis method for chromatography using the chromatography-mass spectrometry analysis system of this embodiment will be described. In the chromatography-mass spectrometry analysis system of this embodiment, by executing the waveform analysis program, it is possible to select either creating a learned model or analyzing chromatographic data.

[0043] First, refer to Figure 2 The flowchart illustrates the process of creating a learned model.

[0044] When the user selects to create a learned model, the learned model creation unit 52 prepares an unlearned learned model (step 1). This learned model can use various models capable of performing semantic segmentation. Semantic segmentation is generally used to parse images composed of two-dimensionally distributed pixel data; however, in this embodiment, it is suitable for parsing waveforms of a color spectrum composed of data arranged one-dimensionally along the time axis. Examples of learned models capable of performing semantic segmentation include U-Net (see Non-Patent Document 2), SeGNet, and PSPNet. In this embodiment, U-Net is used.

[0045] Next, training data 421 and validation data 422 are read from the reference waveform storage unit 42 (step 2). Then, the learning model creation unit 52 sets the learning epoch (i) to 1 (step 3) and inputs the training data 421 into the learning model (step 4). Based on a portion of the original waveform of the training data 421 and the characteristic information corresponding to that portion of the original waveform, the variables of the learning model are adjusted (step 5). In the U-Net used as the learning model in this embodiment, the weights of the neural network are adjusted to obtain the correct characteristic information from the portion of the original waveform. When the input of the training data 421 and the adjustment of the variables of the learning model are completed, the learning model (i) at the end of the learning epoch (i) = 1 is set and saved in the storage unit 41 (step 6). Furthermore, validation data 422 is input into the learning model (i), and the learning model (i) analyzes a portion of the original waveform of the validation data 422 to confirm the correct answer rate of the assigned characteristic information (step 7).

[0046] After determining the correct answer rate for a portion of the waveform of the validation data 422, it is determined whether i has reached a predetermined value (e.g., 100) (step 8). At this moment, i = 1, so it is determined that the predetermined value has not been reached ("No" in step 8), and i is incremented by 1 (i←i+1) (step 9), returning to step 4. After returning to step 4, the same process is repeated until i reaches the predetermined value. Then, when i reaches the predetermined value ("Yes" in step 8), an appropriate learning model is selected from the multiple learning models stored in storage unit 41, and this learning model (i) is stored as the learned model in the learned model storage unit 44 (step 10), ending the series of processes. The learned model is selected based on, for example, the highest correct answer rate for the validation data and the absence of overlearning. In addition, the predetermined value is set to a value greater than the number of times a suitable learned model can be constructed. This value can be determined, for example, by referring to examples of past learned models or by preparing to construct a learning model. In the example above, the learning model (i) is saved in the storage unit 41 each time it is learned, but the learning model (i) can also be saved every predetermined number of times.

[0047] Next, refer to Figure 3 The flowchart illustrates the process of resolving the waveform of an unresolved chromatogram.

[0048] When the user sets up the sample in the automatic sampler 14 and indicates the start of the measurement, the measurement control unit 51 reads the measurement conditions stored in the measurement data storage unit 43 and displays these conditions on the screen of the display unit 7. When the user selects the measurement conditions to be used (or applies appropriate changes) based on the displayed measurement conditions and indicates the start of the measurement, the measurement control unit 51 performs chromatographic and mass spectrometric analysis of the sample to obtain the chromatogram. The measurement operation of the measurement control unit 51 is the same as before, so detailed explanation is omitted. Here, an example of obtaining the chromatogram by measuring the sample using the measurement control unit 51 has been described, but chromatographic data can also be obtained by reading previously acquired chromatographic data.

[0049] If, after obtaining chromatographic data through sample measurement or data readout (step 11), the user instructs the user to analyze the chromatographic data, the analysis mode selection unit 53 will display a screen on the display unit 7 asking the user for the analysis mode of the chromatographic data. This screen may display, for example, analysis modes such as "two-stage output (peak region / non-peak region classification)," and the user selects the desired mode to determine the analysis mode. Analysis modes other than two-stage output will be described later.

[0050] When the user selects the resolution mode (step 12), the waveform segmentation unit 54 segments the chromatographic waveform (raw waveform) into a predetermined number of partial waveforms (partial raw waveforms) (step 13). The width (length in the time axis direction) of each partial raw waveform is determined based on the waveform length (length of the chromatographic mass spectrometry analysis execution time) to ensure that the width of each partial raw waveform is at least smaller than the width of the peaks predicted to be included in the chromatogram. This number of segments can be the same as the training data 421 or the validation data 422, or it can be different from both. The number of segments can be appropriately determined according to the above considerations, for example, it can be set to 512 or 1,024.

[0051] Next, the determination unit 55 reads the learned model corresponding to the parsing mode selected by the user from the learned model storage unit 44 (step 14). Then, a portion of the original waveform is input into the learned model (step 15). The learned model determines whether the input portion of the original waveform belongs to the peak region (step 16). In this way, characteristic information (information on whether it belongs to the peak region) is added to each portion of the original waveform.

[0052] When the determination unit 55 finishes determining each part of the original waveform, the classification unit 56 classifies the part of the original waveform based on its results and based on characteristic information (information on whether it belongs to the peak region) (step 17).

[0053] After classifying a portion of the original waveform, the noise value calculation unit 57 calculates the noise value based on the waveform in the non-peak region (step 18). The noise value can be set as, for example, the average intensity of the portion of the original waveform belonging to the non-peak region. The noise removal unit 58 removes noise by subtracting the calculated noise value from the original waveform (step 19) and displays it on the screen of the display unit 7.

[0054] Next, an example of creating a fully learned model using actual chromatographic data and performing chromatographic waveform analysis (a two-stage output example) will be explained. In this example, 13,359 chromatographic data points were obtained by performing chromatographic-mass spectrometry analysis on a sample containing primary metabolites from an organism, and peak picking was manually performed on them. Of these, 1,400 chromatographic data points were used as validation data. Additionally, to confirm the correct answer rate of the fully learned model, 1,400 chromatographic data points were used as test data. The remaining chromatographic data points were used as training data. Furthermore, the number of times the input of the training data and the confirmation of the correct answer rate of the validation data (the predetermined number mentioned above) were performed when creating the fully learned model was set to 60, and the learning model with the 60th training iteration was selected as the fully learned model.

[0055] Figure 4 It is a graph showing the change in the correct answer rate during the model creation process after learning. Figure 4 "Main / accuracy" refers to the correct answer rate for training data 421, and "validation / main / accuracy" refers to the correct answer rate for validation data 422. It can be seen that the correct answer rates for both training and validation data 421 gradually increase with the number of inputs. However, it is not the case that more inputs to training and validation data 421 will necessarily lead to a greater increase in the correct answer rate. Excessive inputs can cause a phenomenon known as overlearning, resulting in a decrease in the correct answer rate. Therefore, regarding the inputs to training and validation data 421, it is preferable to monitor the progression of the correct answer rate and terminate the process before overlearning occurs. With this learned model, when using test data to validate the accuracy of the two-level output, the correct answer rate was 98%.

[0056] Figure 5 This is the result obtained by analyzing the original waveform using the model learned above. Figure 5 In the diagram, -1.25 (peak region) and -1.5 (non-peak region) represent the correct answers for the two-stage output, and -1.75 (peak region) and -2.0 (non-peak region) represent the judgment results after using the learned model. Furthermore, these values ​​are displayed along with the original waveform. For example... Figure 5 As shown, the peak and non-peak regions of the original chromatographic waveform are known to be accurately classified.

[0057] The above-described resolution mode (two-stage output) represents the minimum number of classification examples. Next, several preferred additional structures capable of more detailed chromatographic waveform resolution will be described. The structures described below can be appropriately combined with the above embodiments. Furthermore, multiple additional structures can be combined, provided that the processing is not reversed.

[0058] One preferred additional structure is a three-level output obtained by adding peak boundaries to the two-level outputs described above. In this case, a learned model is created by performing machine learning on multiple sets of partial waveforms produced by segmenting training data 421 and validation data 422 (hereinafter collectively referred to as "reference waveforms") whose peak boundaries are known. Then, the peak regions, which are one of the characteristic information in the two-level outputs described above, are classified into two categories: peak boundaries and non-peak boundary regions (others).

[0059] In chromatography, although columns separate the components in a sample, it may not be possible to completely separate all components in a sample containing many components. Therefore, sometimes multiple peaks overlap on the chromatogram (superimposed peaks). Determining the concentration or content of a compound based on peak area or height assumes that the peaks do not overlap. Therefore, peak separation is necessary when multiple peaks overlap. By setting the output to the three levels described above, peak separation can be easily achieved during this process.

[0060] Furthermore, in the aforementioned three-level output structure, a four-level output structure can also be adopted, which classifies the peak boundary, as one of the characteristic information, into two points: the peak start point and the peak end point. In this case, a fully learned model is created by performing machine learning on multiple sets of partial waveforms, each consisting of a reference waveform whose start and end points are known, by segmenting the waveform.

[0061] For separating superimposed peaks, conventional methods have included tailing, which divides the target peak into two peaks by setting the shape from the start point to the end point as one peak and superimposing another peak on that peak; complete separation, which separates the target peak into two peaks by sequentially connecting the start point, minimum point, and end point; and vertical separation, which separates the two peaks by using a perpendicular line passing through the minimum point of the target peak. For example, Patent Document 1 describes software that can automatically select peaks when the analyst selects any of the tailing, complete separation, or vertical separation methods and inputs the required parameters. In addition, peaks are sometimes separated by fitting model functions such as Gaussian functions.

[0062] In the aforementioned tailing process, regarding the start and end points of the two peaks, the first peak start point and the second peak start point exist sequentially from the side with the shorter holding time, followed by the first and second peak end points. On the other hand, in complete separation or vertical segmentation, the first peak start point and the second peak end point exist from the side with the shorter holding time, followed by the second peak start point and the second peak end point. In the above four-level output structure, the peak boundaries are distinguished as peak start points and peak end points, thus allowing for the discussion of appropriate peak segmentation methods after the user confirms the arrangement.

[0063] Furthermore, a structure that classifies peak regions, as one of the characteristic information elements, into superimposed peak regions and single peak regions can also be adopted. In this case, a learned model can be created simply by using multiple sets of reference waveforms, each consisting of a segment created by dividing a reference waveform containing single peaks and superimposed peaks whose positions are known, and performing machine learning on these segments. By employing this structure, users can easily identify the regions requiring peak separation.

[0064] Furthermore, this structure can also be configured as a five-level output structure that classifies superimposed peak regions into three types: vertical segmentation, complete separation, or tailing processing. In this case, a learned model can be created simply by using multiple sets of grouped partial waveforms created by segmenting a reference waveform to perform machine learning. This reference waveform includes superimposed peaks that have been separated by various methods, including tailing processing, complete separation, and vertical segmentation.

[0065] The appropriate method for segmenting (separating) overlapping peaks in the chromatogram of the target object varies depending on factors such as the shape of baseline drift or peak shape, and is easily influenced by user subjectivity. In the five-level output structure described above, the learned model can automatically determine the appropriate method for segmenting overlapping peaks, thus enabling objective and high-precision peak segmentation without incorporating user subjectivity.

[0066] Figure 6 This is an example of a fully learned model used in the aforementioned parsing pattern (five-level output). It is similar to an example of a two-level output model. Figure 4 same, Figure 6 "Main / accuracy" refers to the correct answer rate on training data 421, while "validation / main / accuracy" refers to the correct answer rate on validation data 422. Additionally, with... Figure 4 Similarly, in this example, the correct answer rate for both training data 421 and validation data 422 gradually increases with the number of inputs. Regarding the learned model, the correct answer rate is 97% when using test data to validate the accuracy of the five-level output.

[0067] Figure 7 This is an example of outputting the results obtained by analyzing the raw waveform of the chromatogram using the learned model in a five-level output format. Figure 7 In the diagram, -1.1 (peak end point), -1.2 (peak start point), -1.3 (vertical peak segmentation), -1.4 (single peak), and -1.5 (non-peak region) represent the correct answers for the five-level output, while -1.6 (peak end point), -1.7 (peak start point), -1.8 (vertical peak segmentation), -1.9 (single peak), and -2.0 (non-peak region) represent the outputs of the learned model. These values ​​are also displayed along with the original waveform. Figure 7 As shown, it can be seen that the partial waveforms of the chromatogram were accurately classified.

[0068] In the structure that classifies superimposed peaks and individual peaks, it is also possible to separate peaks by fitting a predetermined model function (e.g., Gaussian fitting, EMG (Exponential Modified Gaussian) fitting). In this case, the pre-used model function is stored in storage unit 41. Additionally, for regions determined by the learned model to be superimposed peak regions, a peak separation unit 59 is used to separate peaks by fitting a predetermined model function.

[0069] In structures that classify superimposed peaks and single peaks, further classifications corresponding to the number of peaks contained in a superimposed peak can be added. In this case, a learned model can be created simply by using multiple sets of partial waveforms—groups created by segmenting reference waveforms containing multiple superimposed peaks of varying numbers—through machine learning. Thus, users can easily determine the number of peaks contained in a superimposed peak.

[0070] The preferred embodiment described above mainly adds a classification method related to peak separation, but it is also possible to add a function to estimate the baseline contained in the chromatogram. In this case, a baseline estimation unit 60 is used to estimate the baseline based on the partial waveform of the non-peak region. The baseline estimation unit 60 estimates the baseline of the target waveform as a whole by fitting the partial waveform of the non-peak region with a model function. This makes it easier to discuss which is more suitable, for example, vertical segmentation, complete separation, or tailing treatment when separating overlapping peaks. In addition, in the structure that uses a learning model to classify overlapping peaks, the user can easily verify the appropriateness of the classification results based on the learning model.

[0071] Furthermore, it is preferable to add a peak area calculation unit (or a peak height calculation unit) to the functional block within the control / processing unit 4 of the chromatography-mass spectrometry analysis apparatus of the above embodiment, which calculates the area of ​​the peak region classified by the learning model. This allows for convenient quantification of the component corresponding to the peak.

[0072] The above embodiments are all examples and can be appropriately modified in accordance with the spirit of the present invention. In the above embodiments, the structure is designed as part of a chromatography-mass spectrometry analysis system, but it can be configured as a waveform analysis device independent of the chromatography-mass spectrometry analysis apparatus. In this case, analysis can be performed simply by reading pre-acquired chromatographic data through the chromatography-mass spectrometry analysis apparatus. Furthermore, the above embodiments have been described using the processing of waveforms of chromatograms obtained by chromatography-mass spectrometry analysis as an example, but the same analysis can also be performed on chromatograms obtained by chromatographs or gas chromatographs equipped with detectors other than mass spectrometers (spectrometers). Moreover, the object of analysis is not limited to chromatograms; it can also be performed on spectra (waveforms representing changes in detection intensity relative to wavelength or wavenumber axis) obtained by measurements performed by a spectrophotometer, for example, in the same way as described above.

[0073] [Way]

[0074] Those skilled in the art will understand that the above-described exemplary embodiments are specific examples of the following methods.

[0075] (First item)

[0076] One aspect of the present invention relates to a waveform analysis method that analyzes a waveform that is an object waveform in chromatography or spectroscopy, wherein,

[0077] A learned model is created by using multiple sets of partial waveforms, each consisting of segments of a reference waveform whose peak positions are known, to determine the peak portions contained in an input waveform through machine learning.

[0078] The waveform of the object is divided into multiple partial waveforms.

[0079] The learned model is used to determine whether each part of the multiple waveforms of the object waveform is a peak.

[0080] Based on the determination result, the object waveform is classified into a continuous peak region and a non-peak region outside the peak region.

[0081] (Item 10)

[0082] Another aspect of the present invention relates to a waveform analysis apparatus for analyzing a waveform that is a chromatogram or spectrum, the waveform analysis apparatus comprising:

[0083] A waveform segmentation unit that divides the object waveform into multiple partial waveforms;

[0084] The determination unit uses a learned model to determine whether each portion of the multiple partial waveforms of the object waveform is a peak. The learned model is created through machine learning using multiple sets of partial waveforms created by segmenting a reference waveform whose peak positions are known.

[0085] The classification unit, based on the determination result of the determination unit, classifies the object waveform into a continuous peak region and a non-peak region elsewhere.

[0086] In the waveform analysis method described in the first item and the waveform analysis apparatus described in the tenth item, a learned model for determining whether an input waveform is a peak is pre-built by performing machine learning on multiple sets of partial waveforms created by segmenting a reference waveform whose peak position is known. Here, the reference waveform used in building the learned model is typically a waveform obtained by performing the same type of measurement as the target waveform (chromatography or spectroscopy), but waveforms created through theoretical calculations can also be used. Furthermore, the multiple partial waveforms of the reference waveform are created by segmenting them at intervals narrower than the peak widths of the peaks contained in the reference waveform. In this way, a learned model is built that learns various shapes of partial waveforms constituting a part of a peak. Then, similarly to segmenting the reference waveform, the target waveform is segmented at intervals narrower than the peak widths of the peaks expected to be contained in the target waveform to create partial waveforms. Then, using the learned model described above, it is determined whether each partial waveform of the multiple partial waveforms formed by segmenting the target waveform whose peak position is unknown is a peak. Therefore, multiple parts of the target waveform are classified into peak parts and non-peak parts, and based on this, continuous regions of peak parts are classified as peak regions, and continuous regions of non-peak parts are classified as non-peak regions. In the waveform analysis method and waveform analysis apparatus of the present invention, the target waveform is classified into peak regions and non-peak regions by using a learned model that has learned the features of multiple reference waveforms, thus enabling high-precision peak picking without complex operations.

[0087] (Second item)

[0088] In the waveform analysis method described in the first item,

[0089] The learned model was created by performing machine learning on multiple sets of partial waveforms, each consisting of a reference waveform whose peak boundaries were known, segmented together.

[0090] In the waveform analysis method,

[0091] The boundaries of peaks contained in the partial waveform that are determined to be part of the peak portion are identified.

[0092] The peak regions are classified into peak boundary regions and non-peak boundary regions.

[0093] (Item 11)

[0094] In the mass spectrometry apparatus described in item ten,

[0095] The learned model was created by performing machine learning on multiple sets of partial waveforms, each consisting of a reference waveform whose peak boundaries were known, segmented together.

[0096] The determination unit identifies the boundaries of peaks contained in the partial waveform that are determined to be part of the peak portion.

[0097] The classification department further classifies the peak region into peak boundary regions and non-peak boundary regions.

[0098] In the waveform analysis method described in the second item and the waveform analysis apparatus described in the eleventh item, superimposed peaks can be easily separated when performing quantitative analysis of sample components.

[0099] (Third item)

[0100] In the waveform analysis method described in the second item,

[0101] The learned model is created by performing machine learning on multiple sets of partial waveforms, each consisting of a reference waveform whose start and end points are known, and segmented from these waveforms.

[0102] In the waveform analysis method,

[0103] Determine whether the boundary of the peak is the start point or the end point of the peak.

[0104] The boundaries of the peaks are further classified into peak start points and peak end points.

[0105] (Item 12)

[0106] In the waveform analysis device described in item eleven,

[0107] Furthermore, the product ions extracted by the product ion extraction unit are the candidate compounds. The learned model is created by machine learning using multiple sets of partial waveforms formed by segmenting a reference waveform whose peak start and end points are known.

[0108] The determination unit determines whether the boundary of the peak is the start point or the end point of the peak.

[0109] The classification department further classifies the peak boundary into peak start point and peak end point.

[0110] In the waveform analysis method described in the third item and the waveform analysis apparatus described in the twelfth item, the peak boundaries are distinguished as peak start points and peak end points, so that an appropriate peak segmentation method can be discussed after the user has confirmed the arrangement.

[0111] (Item 4)

[0112] In any of the waveform analysis methods described in the first to third items,

[0113] The learned model is created through machine learning using multiple sets of partial waveforms, each consisting of a reference waveform containing a single peak and a superimposed peak, both with known positions.

[0114] In the waveform analysis method,

[0115] Determine whether the peaks contained in the peak region, which is a continuous region of the peak portion, are single peaks or superimposed peaks.

[0116] The peak regions are further classified into single peak regions and superimposed peak regions.

[0117] (Item 13)

[0118] In any of the waveform analysis devices described in items ten through twelfth,

[0119] The learned model is created through machine learning using multiple sets of partial waveforms, each consisting of a reference waveform containing a single peak and a superimposed peak, both with known positions.

[0120] The determination unit determines whether the peaks contained in the peak region, which is a continuous region of the peak portion, are single peaks or superimposed peaks.

[0121] The classification department further classifies the peak regions into single peak regions and superimposed peak regions.

[0122] In the waveform analysis method described in item four and the waveform analysis device described in item thirteen, the user can easily identify the region that needs peak separation.

[0123] (Item 5)

[0124] In the waveform analysis method described in the fourth item,

[0125] The learned model is created through machine learning using multiple sets of partial waveforms produced by segmenting a reference waveform. The reference waveform contains superimposed peaks that have undergone peak separation using methods such as tailing, complete separation, and vertical segmentation. Tailing involves separating the peaks into regions where the start and end points of the superimposed peaks are defined as a single peak, and then superimposing other peaks on top of that peak. Complete separation uses lines connecting the start, minimum, and end points of the superimposed peaks to separate them. Vertical segmentation uses perpendicular lines passing through the minimum points of the superimposed peaks to separate the two peaks.

[0126] In the waveform analysis method,

[0127] A method for determining the separation of the plurality of peaks contained in the superimposed peak region.

[0128] (Item Fourteen)

[0129] In the mass spectrometry apparatus described in item thirteen,

[0130] The learned model is created through machine learning using multiple sets of partial waveforms produced by segmenting a reference waveform. The reference waveform contains superimposed peaks that have undergone peak separation using methods such as tailing, complete separation, and vertical segmentation. Tailing involves separating the peaks into regions where the start and end points of the superimposed peaks are defined as a single peak, and then superimposing other peaks on top of that peak. Complete separation uses lines connecting the start, minimum, and end points of the superimposed peaks to separate them. Vertical segmentation uses perpendicular lines passing through the minimum points of the superimposed peaks to separate the two peaks.

[0131] The determination unit determines a method suitable for separating the plurality of peaks contained in the superimposed peak region.

[0132] In the waveform analysis method described in item 5 and the waveform analysis apparatus described in item 14, the learned model can be used to automatically determine the appropriate method for segmenting superimposed peaks. Therefore, peaks can be segmented objectively and with high precision without incorporating the user's subjectivity.

[0133] (Item 6)

[0134] In the waveform analysis method described in the fourth item,

[0135] The multiple peaks contained in the superimposed peak region are separated using a pre-determined model function.

[0136] (Seventh item)

[0137] In the waveform analysis method described in item six,

[0138] Multiple peaks contained in the superimposed peaks are separated by Gaussian fitting or exponentially modified Gaussian fitting.

[0139] (Item 15)

[0140] In the waveform analysis apparatus described in item thirteen,

[0141] It also includes a peak separation section, which uses a predetermined model function to separate the multiple peaks contained in the superimposed peak region.

[0142] (Item 16)

[0143] In the waveform analysis apparatus described in item fifteen,

[0144] The peak separation section separates peaks using Gaussian fitting or exponentially modified Gaussian fitting.

[0145] In the waveform analysis method described in item six or seven and the waveform analysis apparatus described in item fifteen or sixteen, superimposed peaks can be automatically separated using a model function.

[0146] (Item 8)

[0147] In any of the waveform analysis methods described in items one through seven,

[0148] The noise value is determined based on the waveform of the object classified as the non-peak region.

[0149] (Item 17)

[0150] In any of the waveform analysis devices described in items 10 through 16,

[0151] It also includes a noise value calculation unit, which calculates the noise value based on the object waveform classified as the non-peak region.

[0152] In the waveform analysis method described in item 8 and the waveform analysis apparatus described in item 17, the noise value contained in the waveform of the object can be automatically calculated.

[0153] (Item 9)

[0154] In any of the waveform analysis methods described in items one through eight,

[0155] Baseline is estimated based on the waveforms of objects classified as non-peak regions.

[0156] (Item 18)

[0157] In any of the waveform analysis devices described in items 10 through 17,

[0158] It also includes a baseline estimation unit that estimates the baseline based on the waveform of the object classified as the non-peak region.

[0159] In the waveform analysis method described in item 9 and the waveform analysis apparatus described in item 18, it is helpful to discuss which method is suitable, for example, vertical segmentation, complete separation, or tailing processing when separating superimposed peaks. Furthermore, in the structure that uses a learning model to classify superimposed peaks, the user can easily verify the appropriateness of the classification results based on the learning model.

[0160] Explanation of reference numerals in the attached figures

[0161] 1: Liquid Chromatograph; 2: Mass Spectrometer; 4: Control / Processing Unit; 41: Storage Unit; 42: Reference Waveform Storage Unit; 421: Training Data; 422: Validation Data; 43: Measurement Data Storage Unit; 44: Model Storage Unit; 5: Procedure; 51: Measurement Control Unit; 52: Model Creation Unit; 53: Resolution Mode Selection Unit; 54: Waveform Segmentation Unit; 55: Judgment Unit; 56: Classification Unit; 57: Noise Value Calculation Unit; 58: Noise Removal Unit; 59: Peak Separation Unit; 60: Baseline Estimation Unit; 6: Input Unit; 7: Display Unit.

Claims

1. A waveform analysis method for analyzing a waveform that is a chromatogram or spectrum, wherein in the waveform analysis method, By dividing multiple reference waveforms, whose peak positions are known, along with the time direction of the chromatogram or the wavelength direction of the spectrum into groups of partial waveforms smaller than the width of the peak, a fully learned model is created for determining the peak portions contained in the input waveform through machine learning using these groups of partial waveforms created from the multiple reference waveforms. The waveform of the object is divided into multiple partial waveforms in the time direction of the chromatography or the wavelength direction of the spectrum. The learned model is used to determine whether each part of the multiple waveforms of the object waveform is a peak. Based on the determination result, the object waveform is classified into a continuous peak region and a non-peak region outside the peak region.

2. The waveform analysis method according to claim 1, wherein, The learned model is created by dividing multiple reference waveforms, each with a known peak boundary position, into groups of partial waveforms for each reference waveform, and then performing machine learning on these groups of partial waveforms created from the multiple reference waveforms. In the waveform analysis method, The boundaries of peaks contained in the partial waveform that are determined to be part of the peak portion are identified. The peak regions are classified into peak boundary regions and non-peak boundary regions.

3. The waveform analysis method according to claim 2, wherein, The learned model is created by dividing multiple reference waveforms, whose peak start and end points are known, into groups of partial waveforms for each reference waveform, and then using these groups of partial waveforms created from the multiple reference waveforms for machine learning. In the waveform analysis method, Determine whether the boundary of the peak is the start point or the end point of the peak. The boundaries of the peaks are further classified into peak start points and peak end points.

4. The waveform analysis method according to claim 1, wherein, The noise value is determined based on the waveform of the object classified as the non-peak region.

5. The waveform analysis method according to claim 1, wherein, Baseline is estimated based on the waveforms of objects classified as non-peak regions.

6. A waveform analysis method for analyzing a waveform that is a chromatogram or spectrum, wherein in the waveform analysis method, By segmenting multiple reference waveforms whose peak positions are known, creating groups of partial waveforms for each reference waveform, and then performing machine learning on these groups of partial waveforms based on the multiple reference waveforms, a learned model is created to determine the peak portions contained in the input waveform. The waveform of the object is divided into multiple partial waveforms. The learned model is used to determine whether each part of the multiple waveforms of the object waveform is a peak. Based on the determination result, the object waveform is classified into a continuous peak region and a non-peak region outside the peak region. The learned model is created by dividing multiple reference waveforms, each containing a single peak and overlapping peaks with known positions, into groups of partial waveforms for each reference waveform. Machine learning is then performed using these groups of partial waveforms created from the multiple reference waveforms. The superimposed peak is composed of multiple peaks stacked together. In the waveform analysis method, Determine whether the peaks contained in the peak region, which is a continuous region of the peak portion, are single peaks or superimposed peaks. The peak regions are further classified into single peak regions and superimposed peak regions.

7. The waveform analysis method according to claim 6, wherein, The learned model is created by dividing multiple reference waveforms into groups of partial waveforms for each reference waveform, and then performing machine learning on these groups of partial waveforms. The reference waveforms include superimposed peaks that have undergone peak separation using methods such as tailing, complete separation, and vertical segmentation. Tailing involves separating the peaks into regions where the start and end points of the superimposed peaks form a single peak, and then superimposing other peaks on top of this peak. Complete separation uses lines connecting the start, minimum, and end points of the superimposed peaks to separate them. Vertical segmentation uses perpendicular lines passing through the minimum points of the superimposed peaks to separate the two peaks. In the waveform analysis method, A method suitable for separating multiple peaks contained in the superimposed peak region.

8. The waveform analysis method according to claim 6, wherein, A pre-determined model function is used to separate the multiple peaks contained in the superimposed peak region.

9. The waveform analysis method according to claim 8, wherein, Multiple peaks contained in the superimposed peak region are separated by Gaussian fitting or exponentially modified Gaussian fitting.

10. A waveform analysis apparatus for analyzing a waveform that is a chromatogram or spectrum, the waveform analysis apparatus comprising: A waveform segmentation unit that divides the object waveform into multiple partial waveforms; The determination unit uses a learned model to determine whether each portion of the waveform, segmented into multiple partial waveforms along the time direction of the chromatography or the wavelength direction of the spectrum, is a peak. The learned model is created by: dividing multiple reference waveforms (with known peak positions) along the time direction of the chromatography or the wavelength direction of the spectrum into groups of multiple partial waveforms for each reference waveform, with a width smaller than the peak width; and performing machine learning on these groups of partial waveforms created from the multiple reference waveforms. The classification unit, based on the determination result of the determination unit, classifies the object waveform into a continuous peak region and a non-peak region elsewhere.

11. The waveform analysis apparatus according to claim 10, wherein, The learned model is created by dividing multiple reference waveforms, each with a known peak boundary position, into groups of partial waveforms for each reference waveform, and then performing machine learning on these groups of partial waveforms created from the multiple reference waveforms. The determination unit identifies the boundaries of peaks contained in the partial waveform that are determined to be part of the peak portion. The classification department further classifies the peak region into peak boundary regions and non-peak boundary regions.

12. The waveform analysis apparatus according to claim 11, wherein, The learned model is created by dividing multiple reference waveforms, whose peak start and end points are known, into groups of partial waveforms for each reference waveform, and then using these groups of partial waveforms created from the multiple reference waveforms for machine learning. The determination unit determines whether the boundary of the peak is the start point or the end point of the peak. The classification department further classifies the peak boundary into peak start point and peak end point.

13. The waveform analysis apparatus according to claim 10, wherein, It also includes a noise value calculation unit, which calculates the noise value based on the object waveform classified as the non-peak region.

14. The waveform analysis apparatus according to claim 10, wherein, It also includes a baseline estimation unit that estimates the baseline based on the waveform of the object classified as the non-peak region.

15. A waveform analysis apparatus for analyzing a waveform that is a chromatogram or spectrum, the waveform analysis apparatus comprising: A waveform segmentation unit that divides the object waveform into multiple partial waveforms; The determination unit uses a learned model to determine whether each portion of a plurality of partial waveforms of the object waveform is a peak. The learned model is created by: dividing multiple reference waveforms whose peak positions are known into groups of partial waveforms for each reference waveform, and performing machine learning on these groups of partial waveforms created from the multiple reference waveforms; and... The classification unit, based on the determination result of the determination unit, classifies the object waveform into a continuous peak region and non-peak regions outside the peak portion, wherein... The learned model is created by dividing multiple reference waveforms, each containing a single peak and a superimposed peak with known positions, into groups of partial waveforms for each reference waveform. Machine learning is then performed using these groups of partial waveforms created from the multiple reference waveforms, where the superimposed peaks are composed of multiple overlapping peaks. The determination unit determines whether the peaks contained in the peak region, which is a continuous region of the peak portion, are single peaks or superimposed peaks. The classification department further classifies the peak regions into single peak regions and superimposed peak regions.

16. The waveform analysis apparatus according to claim 15, wherein, The learned model is created by dividing multiple reference waveforms into groups of partial waveforms for each reference waveform, and then performing machine learning on these groups of partial waveforms. The reference waveforms include superimposed peaks that have undergone peak separation using methods such as tailing, complete separation, and vertical segmentation. Tailing involves separating the peaks into regions where the start and end points of the superimposed peaks form a single peak, and then superimposing other peaks on top of this peak. Complete separation uses lines connecting the start, minimum, and end points of the superimposed peaks to separate them. Vertical segmentation uses perpendicular lines passing through the minimum points of the superimposed peaks to separate the two peaks. The determination unit determines a method suitable for separating multiple peaks contained in the superimposed peak region.

17. The waveform analysis apparatus according to claim 15, wherein, It also includes a peak separation section, which uses a predetermined model function to separate multiple peaks contained in the superimposed peak region.

18. The waveform analysis apparatus according to claim 17, wherein, The peak separation section separates peaks using Gaussian fitting or exponentially modified Gaussian fitting.

Images