Raw material plasma protein content determination model training method, system, medium and equipment

Abstract

The invention provides a raw material plasma protein content determination model training method, system, medium and equipment, and relates to the field of protein content determination, and the method comprises the following steps: obtaining a raw material plasma sample and a real protein content label thereof; the method comprises the following steps: acquiring near infrared spectrums of raw material plasma samples based on different temperatures by adopting double sensors to obtain near infrared spectrums containing temperature data; training a CNN-LSTM model, and inputting the near infrared spectrums containing the temperature data into the CNN-LSTM model to obtain a predicted protein content label of each near infrared spectrum; and calculating a loss function of the model according to the predicted protein content label and the real protein content label of each raw material plasma sample, and when the loss of the model is minimum, obtaining a trained CNN-LSTM model. Based on a near infrared spectrum analysis technology and a data fusion strategy, a model is utilized to learn a relationship between spectrum data containing temperature information and real protein content, and the influence of temperature on a protein content detection result is eliminated.

Description

Technical Field

[0001] The present invention relates to the technical field of protein content determination, and in particular to a raw material plasma protein content determination model training method, system, medium and equipment. Background Art

[0002] Raw plasma is the plasma of healthy people collected by the single plasma collection method. It is the only raw material for the production of blood products. The total protein content of raw plasma is a key quality parameter to characterize the quality of plasma. Conventional protein content determination methods such as Kjeldahl nitrogen determination method and biuret reagent method have the disadvantage of long determination time and cannot meet the large testing needs such as re-inspection after raw plasma collection and before production in a timely manner.

[0003] Portable miniature near-infrared spectral sensors have extremely high practical value for large-scale on-site detection of plasma samples, but due to the narrow spectral range of a single sensor and less effective information, multi-sensor data fusion is one of the current research hotspots. Through multi-sensor data fusion, the dual goals of increasing effective spectral information and reducing the cost of using NIR instruments can be achieved.

[0004] In addition, since temperature affects the vibration state and energy level transition of molecules, the increase in temperature causes the absorption peak of the spectrum to move toward the short-wave direction, which has a greater impact on the peak height and peak width of the spectrum, thereby affecting the prediction accuracy of the near-infrared spectroscopy quantitative analysis model. In order to alleviate the problem of temperature interference, existing technologies such as the Chinese invention patent with application number 202110361714.5, entitled Spectral Analysis Method and System Based on Multi-level Partial Least Squares Algorithm, separate the influence of temperature on the spectrum by constructing a model, but the computational complexity is high, which affects the efficiency of protein content determination. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a raw material plasma protein content determination model training method, system, medium and equipment. Based on near-infrared spectral analysis technology and data fusion strategy, the real-time temperature during spectral acquisition is fused with the spectral data to obtain spectral data containing temperature information. The model is used to learn the relationship between the spectral data containing temperature information and the actual protein content, thereby eliminating the influence of temperature on the protein content detection results.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for training a raw material plasma protein content determination model, comprising: Obtain raw plasma samples and labels of their true protein content; Using dual sensors to collect near infrared spectra of raw plasma samples based on different temperatures to obtain near infrared spectra containing temperature data; the near infrared spectra include a first spectrum and a second spectrum; The CNN-LSTM model was trained and the near-infrared spectra containing temperature data were input into the CNN-LSTM model to obtain the predicted protein content label of each near-infrared spectrum. The loss function of the model was calculated based on the predicted protein content label and the true protein content label of each raw plasma sample. When the loss of the model was minimized, a trained CNN-LSTM model was obtained.

[0007] Preferably, the protein content includes total protein content, albumin content or globulin content.

[0008] Preferably, the method of using dual sensors to collect near infrared spectra of raw plasma samples based on different temperatures to obtain near infrared spectra containing temperature data specifically includes: A 1 mm optical path cuvette was used, with air as the background, and the background spectrum was collected once an hour. The scanning ranges of the dual sensors were 1350 nm-1650 nm and 1550 nm-1950 nm, respectively. The dual sensors included a first near-infrared sensor and a second near-infrared sensor. The dual sensors were connected to the sample pool via a Y-shaped optical fiber and irradiated by an external light source, so as to build a dual micro sensor detection path. The temperature control accessory was used to set the collection temperature of each sample to 5-45°C with an interval of 5°C. The spectra were collected three times and averaged to obtain the near-infrared spectrum containing temperature data.

[0009] Preferably, after obtaining the near infrared spectrum containing temperature data, the near infrared spectrum image is preprocessed by Hotelling T 2 The Mahalanobis distance method was used to eliminate abnormal spectra.

[0010] Preferably, the CNN-LSTM model includes a connected CNN network and an LSTM network; The CNN network is used to extract shallow and deep features of the infrared spectrum; it consists of a convolution layer, a pooling layer, and a fully connected layer, including two one-dimensional convolution layers, the number of convolution kernels in the first layer is 32, the convolution kernel size is 3×N, the moving step is 2, and the activation function is selected as the Relu function; The LSTM network is used to fuse the shallow features and the deep features to obtain the final features; the LSTM model includes an input gate, an output gate and a forget gate, and the forget gate includes multiple memory cells, at least one sigmoid function and a point multiplication operation.

[0011] Preferably, the specific process of training the CNN-LSTM model is: The near-infrared spectrum containing temperature data and the true protein content label are input into the CNN-LSTM model for training; specifically, the temperature data, the first spectrum and the second spectrum are input into the model in a certain order according to an N×N matrix, and the feature extraction from 1×1 to N×N is performed through the model convolution layer to obtain a predicted protein content label; the loss function of the model is calculated according to the predicted protein content label and the true protein content label, and the convolution kernel size, the number of convolution kernels and the number of fully connected neurons are optimized according to the loss function to obtain a trained CNN-LSTM model.

[0012] Preferably, the certain sequence includes at least two different arrangement sequences of the temperature data, the first spectrum and the second spectrum.

[0013] In a second aspect, the present invention provides a raw material plasma protein content determination model training system, comprising: An initial data acquisition module, used to obtain raw plasma samples and their true protein content labels; A near infrared spectrum acquisition module, used for collecting near infrared spectra of raw plasma samples based on different temperatures using dual sensors to obtain near infrared spectra containing temperature data; the near infrared spectrum includes a first spectrum and a second spectrum; The training module is used to train the CNN-LSTM model. The near-infrared spectrum containing temperature data is input into the CNN-LSTM model to obtain the predicted protein content label of each near-infrared spectrum; the loss function of the model is calculated according to the predicted protein content label and the true protein content label of each raw plasma sample. When the loss of the model is minimized, a trained CNN-LSTM model is obtained.

[0014] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for training a model for determining raw material plasma protein content described in the first aspect.

[0015] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for training a model for determining raw material plasma protein content described in the first aspect are implemented.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The present invention fuses the real-time temperature during spectrum acquisition with the spectrum data to obtain spectrum data containing temperature information, fuses the temperature and spectrum information to participate in model training, utilizes the close coupling and mutual dependence between temperature and spectrum, does not specifically model and analyze the temperature, and can also reduce the interference of temperature on the infrared spectrum on the basis of saving computational complexity. At the same time, a high-performance convolutional neural network-long short-term memory (CNN-LSTM) algorithm is used to train the samples, and the total protein content of the raw material plasma to be tested is predicted, thereby improving the accuracy and efficiency of protein content prediction.

[0017] The data fusion strategy of the present invention is used to collect spectra of samples using dual micro-near-infrared spectral sensors to solve the problem of narrow wavelength range covered by a single micro-near-infrared sensor. In addition, the temperature data and the spectral data are synchronously fused and trained through the CNN-LSTM deep learning algorithm to achieve better prediction accuracy of spectral data under temperature disturbance.

[0018] The present invention realizes rapid characterization of total protein in raw plasma based on near-infrared spectroscopy analysis technology. Furthermore, it can realize simultaneous detection of the contents of multiple indicators such as albumin and globulin in raw plasma under temperature disturbance, providing means for rapid on-site detection of raw plasma, and providing a reference method for quality grading of raw plasma based on multi-protein content. It has important application value and provides a reference for further promotion and application of this technology.

[0019] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention but do not constitute a limitation of the present invention.

[0021] Figure 1 A main flow chart of a raw material plasma protein content determination model training method provided in Example 1 of the present invention; Figure 2 The original spectra of raw plasma provided in Example 1 of the present invention, (a) the original spectrum collected by NIRONE 1.7; (b) the original spectrum collected by NIRONE 2.0; Figure 3 This is a diagram of the CNN-LSTM architecture provided in Example 1 of the present invention; Figure 4This is a graph of the prediction results of the CNN-LSTM model provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0023] Embodiment 1 like Figure 1 As shown, this embodiment discloses a raw material plasma protein content determination model training method, comprising the following steps: S1: Obtain raw plasma samples and their true protein content labels; S2: using dual sensors to collect near infrared spectra of the raw plasma sample at different temperatures to obtain near infrared spectra containing temperature data; the near infrared spectra include a first spectrum and a second spectrum; S3: Train the CNN-LSTM model, input the near-infrared spectrum containing temperature data into the CNN-LSTM model, and obtain the predicted protein content label of each near-infrared spectrum; calculate the loss function of the model according to the predicted protein content label and the true protein content label of each raw plasma sample, and obtain the trained CNN-LSTM model when the loss of the model is minimized.

[0024] (1) Sample collection: Collect a certain amount of raw plasma samples. Raw plasma is collected by a plasmapheresis machine. The amount of raw plasma collected each time shall not exceed 580 ml (including anticoagulant solution, the mass converted by volume ratio shall not exceed 600 g). According to the blood components, the plasma collection equipment uses centrifugation to separate different blood components, which is an intermittent separation type. The plasmapheresis machine uses disposable plasma collection consumables to automatically complete the entire process of blood collection, anticoagulant injection, whole blood separation, plasma collection, and blood cell transfusion in a fully closed system. About 2 ml of sample is retained for each single plasma for subsequent analysis, and a total of 85 samples are obtained.

[0025] (2) Protein content detection: Use a biochemical analyzer to detect the total protein, albumin, and globulin content.

[0026] (3) Spectral acquisition: Near infrared sensor 1 and near infrared sensor 2 are connected to the sample pool through a Y-type optical fiber, irradiated with an external light source, and a dual micro-sensor detection path is built to obtain the original spectrum of the raw plasma, such as Figure 2 NIRONE 1.7 and NIRONE 2.0 are different models of NIRONE Sensor (near infrared spectrum sensor). This embodiment supports different models of NIRONE Sensor to be connected in parallel for simultaneous detection.

[0027] The transmission sampling mode was adopted, and the cuvette with 1 mm optical path was used. With air as the background, the background spectrum was collected once every hour. The scanning ranges were 1350 nm-1650 nm and 1550 nm-1950 nm respectively. The spectrum was collected three times and the average was taken as the original spectrum.

[0028] To simulate the change of ambient temperature, the temperature control accessory was used to set the acquisition temperature of each sample to 5-45°C, with an interval of 5°C. The spectrum was collected three times and averaged as the original spectrum. By repeatedly collecting spectra, the random errors and noise in a single measurement can be averaged out, thus obtaining more stable and accurate spectral data.

[0029] (4) Sample set division: through Hotelling T 2 The Mahalanobis distance method was used to eliminate abnormal spectra, and the KS classification method was used to divide the calibration set and the validation set, including 60 samples in the calibration set and 24 samples in the validation set.

[0030] (5) CNN parameter optimization: CNN is a deep neural network that automatically extracts features from the original data matrix and then performs classification and regression. Its convolution layer has the function of automatically optimizing the extracted features from the original data, and the pooling layer has the function of selecting the most significant features through dimensionality reduction. Set the CNN network structure and perform model training. The original spectrum fusion CNN model is adopted. After the model is configured, the temperature data, dual sensor spectrum data and total protein primary data (real label) are imported into the neural network for training. The temperature data, spectrum data 1 and spectrum data 2 are input in a certain order in the form of N×N matrix. The feature extraction machine optimization from 1×1 to N×N from local to overall is performed through the convolution layer, and the convolution kernel size, the number of convolution kernels and the number of fully connected neurons are optimized to establish the original spectrum fusion CNN model; the CNN model consists of a convolution layer, a pooling layer and a fully connected layer, and is designed as two one-dimensional convolution layers (Conv1D). Because there is no fixed rule and formula for the definition of the convolution kernel, the present invention preferably sets the number of convolution kernels in the first layer to 32, the convolution kernel size to 3×N, the moving step size to 2, and the activation function is selected as the ReLU (Rectified Linear Units, ReLU) function.

[0031] The certain sequence includes at least two different arrangement sequences of temperature data, first spectrum and second spectrum, such as inputting the model into the sequence of temperature data, first spectrum and second spectrum, and the sequence of first spectrum, temperature data and second spectrum.

[0032] In the order one mode, the model processes the temperature data first, which helps the model consider the effect of temperature on spectral features when processing the spectral data later. This is especially important for spectral data with high temperature sensitivity. After the temperature data, the two spectral data are gradually introduced, which allows the model to gradually build a comprehensive understanding of the data, thereby better capturing the complex relationship between spectral features.

[0033] In the order two mode, the model first processes the spectral data, which helps to quickly capture the chemical composition characteristics of the sample, and then learns the impact of temperature changes on the spectral characteristics, which helps the model to more accurately identify the spectral characteristics while taking into account the temperature effect.

[0034] Different input orders have an important impact on the feature extraction and prediction performance of the model. By adopting different orders according to experimental requirements, we can explore which orders are more suitable for specific data sets and experimental requirements. In the feature extraction process from local to global from 1×1 to N×N, different input orders may cause the model to focus on different feature combinations and interactions. For example, a temperature-first order may make the model focus more on the effect of temperature on spectral features; while a spectrum-first order may make the model focus more on quickly capturing chemical composition features. This flexibility makes the CNN model more adaptable and robust when processing complex data.

[0035] In this embodiment, different types of data (temperature data, spectral data 1, spectral data 2) are integrated into an N×N matrix, so that the CNN model can simultaneously consider the potential correlation and mutual influence between them when processing these data. This integration helps the model capture more comprehensive feature information, thereby improving the accuracy of the prediction. By inputting data in matrix form, the CNN model can use the characteristics of the convolutional layer to extract features in local areas and gradually expand to the entire data set. This feature extraction method from local to global helps the model identify key patterns and trends in the data.

[0036] (6) LSTM network parameter optimization: The number of LSTM hidden layer nodes cannot be less than the input feature value. Different numbers of nodes and layers are optimized. Through feature fusion, the shallow and deep features extracted by CNN are input into the LSTM unit to improve the model performance. The LSTM model includes an input gate, an output gate and a forget gate. The forget gate includes multiple memory cells, at least one sigmoid function and a point multiplication operation. Its calculation formula is:

[0037] In the formula, is the output of the current state at time t, which is a vector between [0,1], is the activation function, W is the weight matrix, f is the current forget gate, is the output value of the previous unit, t is the current value, is the input of the current unit, is the bias value.

[0038] The input feature dimension is 17, the hidden state dimension of the LSTM unit is 64, and there are 2 stacked LSTM layers. The absolute mean is constructed as the model loss function, and the Dropout structure d is added to the LSTM neural network. The Dropout value is set to 0.2, and 30% of the node connections are randomly ignored during the network training process. The training learning rate of the model is set to 0.01, the maximum number of iterations is 500, and the weights and bias values ​​of each layer are optimized through layer-by-layer training until the loss function converges to a stable value. The CNN-LSTM architecture diagram is shown in the figure below. Figure 3 As shown, the output TP is total protein, Alb is albumin, and IgG is immunoglobulin G.

[0039] (7) Model evaluation: R p , RMSEP are model evaluation parameters.

[0040] The original spectrum fusion CNN-LSTM model was established, and the temperature data, spectrum data and primary data were imported into the CNN-LSTM network for training. The number of convolution kernels was determined to be 128 and the number of fully connected neurons was 64 through experiments. The convolution kernel size was then optimized and determined to be 12, and the original spectrum fusion CNN-LSTM model was established. The R of the model c ,RMSEC,R p , RMSEP are 0.9906, 0.5324, 0.9911, and 0.6502 respectively. The prediction results of the CNN-LSTM model are shown in the figure below. Figure 4 The CNN-LSTM model is compared with other models, as shown in Table 1. The results show that CNN-LSTM is superior to other models.

[0041] Table 1 Modeling results of different methods

[0042] This specific embodiment fuses the real-time temperature during spectrum acquisition with the spectral data to obtain spectral data containing temperature information, and the fused temperature and spectral information are jointly involved in model training. The close coupling and mutual dependence of temperature and spectrum are utilized, and modeling and analysis are not specifically performed on temperature. On the basis of saving calculation complexity, the interference of temperature on infrared spectrum can also be reduced, the data processing flow is simplified, and the entire system is more concise and clear, which not only improves the operating efficiency of the system, but also reduces the complexity and cost of the system.

[0043] At the same time, the high-performance CNN-LSTM deep learning algorithm is used to train the samples and predict the total protein content of the raw plasma to be tested. The convolution layer operates on the input data through a sliding convolution kernel to extract local features. These local features are crucial for understanding the subtle structure and changes of the data. As the convolution layer goes deeper, the size of the convolution kernel gradually increases (from 1×1 to N×N), and the model can capture a wider range of data features, thereby achieving an understanding of global features. This feature extraction method from local to global helps the model build a more accurate and robust prediction model, improving the accuracy and efficiency of protein content prediction.

[0044] Embodiment 2 This embodiment provides a raw material plasma protein content determination model training system, comprising: An initial data acquisition module, used to obtain raw plasma samples and their true protein content labels; A near infrared spectrum acquisition module, used for collecting near infrared spectra of raw plasma samples based on different temperatures using dual sensors to obtain near infrared spectra containing temperature data; the near infrared spectrum includes a first spectrum and a second spectrum; The training module is used to train the CNN-LSTM model. The near-infrared spectrum containing temperature data is input into the CNN-LSTM model to obtain the predicted protein content label of each near-infrared spectrum; the loss function of the model is calculated according to the predicted protein content label and the true protein content label of each raw plasma sample. When the loss of the model is minimized, a trained CNN-LSTM model is obtained.

[0045] Embodiment 3 This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the method for training a raw material plasma protein content determination model as described in the above-mentioned embodiment 1 are implemented.

[0046] Embodiment 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for training a model for determining the content of raw material plasma protein as described in the first embodiment above are implemented.

[0047] The steps or modules involved in the above embodiments 2 to 4 correspond to those in embodiment 1. For the specific implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A raw material plasma protein content determination model training method, characterized in that: include: Obtain raw plasma samples and labels of their true protein content; Using dual sensors to collect near infrared spectra of raw plasma samples based on different temperatures to obtain near infrared spectra containing temperature data; the near infrared spectra include a first spectrum and a second spectrum; The CNN-LSTM model was trained and the near-infrared spectra containing temperature data were input into the CNN-LSTM model to obtain the predicted protein content label of each near-infrared spectrum. The loss function of the model was calculated based on the predicted protein content label and the true protein content label of each raw plasma sample. When the loss of the model was minimized, a trained CNN-LSTM model was obtained.

2. A raw material plasma protein content determination model training method as claimed in claim 1, characterized in that: The protein content includes total protein content, albumin content or globulin content.

3. A raw material plasma protein content determination model training method as claimed in claim 1, characterized in that: The method of using dual sensors to collect near infrared spectra of raw plasma samples based on different temperatures to obtain near infrared spectra containing temperature data specifically includes: A 1 mm optical path cuvette was used, with air as the background, and the background spectrum was collected once an hour. The scanning ranges of the dual sensors were 1350 nm-1650 nm and 1550 nm-1950 nm, respectively. The dual sensors included a first near-infrared sensor and a second near-infrared sensor. The dual sensors were connected to the sample pool via a Y-shaped optical fiber and irradiated by an external light source, so as to build a dual micro sensor detection path. The temperature control accessory was used to set the collection temperature of each sample to 5-45°C with an interval of 5°C. The spectra were collected three times and averaged to obtain the near-infrared spectrum containing temperature data.

4. A raw material plasma protein content determination model training method as claimed in claim 1, characterized in that: After obtaining the near-infrared spectrum containing temperature data, the near-infrared spectrum image is preprocessed and the temperature is calculated by Hotelling T 2 The Mahalanobis distance method was used to eliminate abnormal spectra.

5. A raw material plasma protein content determination model training method as claimed in claim 1, characterized in that: The CNN-LSTM model includes a connected CNN network and an LSTM network; The CNN network is used to extract shallow and deep features of the infrared spectrum; it consists of a convolution layer, a pooling layer, and a fully connected layer, including two one-dimensional convolution layers, the number of convolution kernels in the first layer is 32, the convolution kernel size is 3×N, the moving step is 2, and the activation function is selected as the Relu function; The LSTM network is used to fuse the shallow features and the deep features to obtain the final features; the LSTM model includes an input gate, an output gate and a forget gate, and the forget gate includes multiple memory cells, at least one sigmoid function and a point multiplication operation.

6. A raw material plasma protein content determination model training method as claimed in claim 1, characterized in that: The specific process of training the CNN-LSTM model is as follows: The near infrared spectrum containing temperature data and the true protein content label are input into the CNN-LSTM model for training; specifically, the temperature data, the first spectrum and the second spectrum are input into the model in a certain order according to an N×N matrix, and the feature extraction from 1×1 to N×N is performed through the model convolution layer to obtain the predicted protein content label; The loss function of the model is calculated based on the predicted protein content labels and the true protein content labels. The convolution kernel size, the number of convolution kernels, and the number of fully connected neurons are optimized based on the loss function to obtain a trained CNN-LSTM model.

7. A raw material plasma protein content determination model training method as claimed in claim 6, characterized in that: The certain sequence includes at least two different arrangement sequences of the temperature data, the first spectrum and the second spectrum.

8. A raw material plasma protein content determination model training system, characterized in that: include: An initial data acquisition module, used to obtain raw plasma samples and their true protein content labels; A near infrared spectrum acquisition module, used for collecting near infrared spectra of raw plasma samples based on different temperatures using dual sensors to obtain near infrared spectra containing temperature data; the near infrared spectrum includes a first spectrum and a second spectrum; The training module is used to train the CNN-LSTM model. The near-infrared spectrum containing temperature data is input into the CNN-LSTM model to obtain the predicted protein content label of each near-infrared spectrum; the loss function of the model is calculated according to the predicted protein content label and the true protein content label of each raw plasma sample. When the loss of the model is minimized, a trained CNN-LSTM model is obtained.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a raw material plasma protein content determination model training method as described in any one of claims 1-7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the raw material plasma protein content determination model training method according to any one of claims 1 to 7 are implemented.

Images