A TDLAS venous drug concentration prediction method based on a one-dimensional convolutional neural network

Abstract

The present invention provides a TDLAS venous drug concentration prediction method based on a one-dimensional convolutional neural network, comprising the following steps: obtaining second-harmonic data of the venous drug concentration; preprocessing the second-harmonic data of the venous drug and dividing it into a training set, a validation set, and a test set; constructing a 1D-CNN deep learning model based on an adaptive efficient channel attention module and a multi-level residual module; inputting the training set and the validation set into the deep learning model and setting hyperparameters to train the model to obtain an optimal parameter model; inputting the test set into the trained optimal parameter detection model to obtain a prediction result; in the 1D-CNN convolutional neural network model of the present invention, an adaptive channel attention module and a multi-level residual module are combined, and an adaptive convolution kernel and a multi-scale convolution structure are introduced to enhance the feature extraction ability, significantly improving the learning ability and generalization ability of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of spectral analysis, and particularly relates to a method for predicting the concentration of intravenous drugs based on a one-dimensional convolutional neural network and TDLAS. Background Art

[0002] With the continuous progress of medical technology, the pharmacy intravenous admixture service (PIVAS), as a key link in providing intravenous admixture services, has become increasingly important. The facilities and equipment of PIVAS are the basis for ensuring the quality of drug dispensing. Modern equipment and advanced auxiliary equipment can improve the accuracy and efficiency of drug dispensing, reduce the occurrence of human errors, not only reduce the risk of drug cross-contamination, but also improve the quality and safety of drug dispensing. Therefore, the control of intravenous drug dosage is of great significance for the construction of PIVAS.

[0003] In recent years, significant progress has been made in the application of spectral analysis technology in the field of liquid detection. By analyzing the absorption characteristics within the spectral range of liquid samples, the component information of the liquid can be obtained quickly and non-destructively. For example, in the quantitative analysis methods of liquid concentration, Raman spectroscopy detection technology, near-infrared spectroscopy detection technology, and ultraviolet-visible spectroscopy (UV-Vis) are relatively mature. However, the research on the detection of the concentration of intravenous drug solutes by combining the above spectral methods with deep learning is relatively scarce, and the discrimination accuracy for low-concentration liquid medicines is insufficient. Moreover, this method requires a complex optical system. How to provide a method that does not rely on a complex optical system to achieve accurate prediction of the low concentration of intravenous drug solutes is a problem worthy of research. Summary of the Invention

[0004] Object of the Invention: The technical problem to be solved by the present invention is to propose a method for predicting the concentration of intravenous drugs based on a one-dimensional convolutional neural network and TDLAS, which can accurately predict the concentration of low-concentration intravenous drugs in view of the deficiencies of the prior art.

[0005] The present invention specifically provides a method for predicting the concentration of intravenous drugs based on a one-dimensional convolutional neural network and TDLAS, including the following steps:

[0006] Step 1, obtaining the second harmonic data of the concentration of intravenous drugs;

[0007] Step 2, preprocessing the second harmonic data of the intravenous drugs, and dividing the training set, validation set, and test set;

[0008] Step 3, constructing a one-dimensional convolutional neural network 1D-CNN deep learning model;

[0009] Step 4, inputting the training set and the validation set into the deep learning model, and setting hyperparameters to train the model to obtain an optimal parameter model;

[0010] Step 5: Input the test set into the trained optimal parameter model to obtain the prediction result.

[0011] Step 1 includes:

[0012] Step 1-1: In tunable diode laser absorption spectroscopy (TDLAS) technology, adopt wavelength modulation spectroscopy technology. Set the LD laser to irradiate two paths respectively. One path passes through the liquid medicine, and according to the Beer-Lambert law, calculate the light intensity of the liquid concentration. The calculation formula is:

[0013]

[0014] Where, is the incident light wavelength, is the light intensity after transmission, is the incident light intensity, is the absorption coefficient, is the concentration of the liquid to be measured, is the optical path absorption distance, and e is the natural constant;

[0015] The other path does not pass through the liquid and is emitted through the optical attenuator. The light intensity is , is the optical attenuation coefficient; Let the optical attenuation coefficient , and take the difference between the two light intensity signals and to obtain the differential signal ;

[0016] Step 1-2: Perform Taylor series expansion on the obtained differential signal to obtain the second harmonic signal . The calculation formula is:

[0017]

[0018] Where is the modulation amplitude, represents the differential symbol for taking the derivative;

[0019] Step 1-3: Demodulate the second harmonic signal by lock-in amplification, and collect the second harmonic data of intravenous medications at each concentration through the upper computer.

[0020] Step 2 includes the following steps:

[0021] Step 2-1: Adopt the Savitzky-Golay smoothing filter method to smooth the second harmonic data of intravenous medications with a length of N to obtain the output sequence ; i = 0, 1, 2, 3,..., N-1; For the output sequence For each data point, a sliding window with a window size of 2m + 1 is set. The sliding window includes the current data point and m neighboring points on each side of the current data point. Within the sliding window, an nth-order polynomial is used to fit the current data point.

[0022] The Savitzky-Golay smoothing filter formula is:

[0023]

[0024] Where, is the filter coefficient, refers to the second harmonic data of intravenous medication with a length of N in which the element with index is within 2m + 1 data points of the sliding window;

[0025] Step 2-2, obtain the filter coefficient through the following formula :

[0026]

[0027] Where, is an N × (2m + 1) matrix, and each row corresponds to a data point within the sliding window; is a column vector with a length of N, representing the vector of the second harmonic data of intravenous medication ;

[0028] Step 2-3, by solving the filter coefficient in Step 2-2 and substituting it into the Savitzky-Golay smoothing filter formula in Step 2-1, calculate the smoothed second harmonic data of intravenous medication to obtain the second harmonic data set of intravenous medication;

[0029] Step 2-4, divide the second harmonic data set of intravenous medication into a training set, a validation set, and a test set according to 8:1:1.

[0030] In Step 3, the 1D-CNN deep learning model includes an adaptive efficient channel attention module, a multi-level residual module, and a regression prediction module.

[0031] In Step 3, construct the adaptive efficient channel attention module through the following steps:

[0032] Step 3-1, calculate the mean value of the input feature map in the channel dimension to obtain the mean representation of each channel. The channel mean is calculated by the formula:

[0033]

[0034] Among them represents the th element of the th channel input feature map, is the length of the feature map.

[0035] Step 3-2, adjust the sizes of the convolutional kernels and through the learnable dynamic factors and . The calculation formula is:

[0036]

[0037]

[0038] Among them, , is the learnable weight matrix optimized during the training process, is the activation function . Ensure that the generated attention weights are within the range of , and are bias terms. Similar to the weights of the convolutional kernels, they are automatically learned through backpropagation. The update formula for the bias terms is:

[0039]

[0040]

[0041] Among them is the learning rate, and respectively represent the gradients of the loss function with respect to the bias terms and . The initial settings of and are 0.1; and respectively represent the updated values of the bias term and the updated value of the bias term ;

[0042] Step 3-3, adjust the sizes of the convolutional kernels and through the generated dynamic factors and . The calculation formula is:

[0043]

[0044]

[0045] Among them and are the initial convolutional kernel sizes, and are the new convolutional kernel sizes after dynamic factor adjustment; is the rounding function;

[0046] Step 3-4, through the dynamically generated convolutional kernels and , generate the attention weights for each channel, and the calculation formula is:

[0047]

[0048] Among them, the activation function ReLU is responsible for introducing non-linear characteristics, and the activation function scales the output range between 0 and 1; represents the attention weight of the i-th channel;

[0049] Step 3-5, the input feature map obtains the output feature map through the generated dynamic attention weights , and the calculation formula is as follows:

[0050]

[0051] In Step 3, construct a multi-level residual module through the following steps: Set 4 convolutional layers, the size of the first convolutional layer is , the sizes of the second and third convolutional layers are , and the size of the fourth convolutional layer is , and add 1 BN batch normalization layer and ReLU activation function to the output end of each convolutional layer. The input feature obtains the first feature map after passing through the first convolutional layer; Input the first feature map into the second and third convolutional layers to obtain the second feature map ; Add the first feature map and the second feature map through a skip connection to obtain the third feature map ; Pass the third feature map through the fourth convolutional layer to obtain the fourth feature map and perform a skip connection with the input feature to obtain the output feature map .

[0052] In step 3, the regression prediction module includes three fully connected layers, a ReLU activation function layer, and a Dropout random dropout layer. The output of the first fully connected layer is 128, and its output end is connected to a ReLU activation function layer and a Dropout random dropout layer. The output of the second fully connected layer is 32, and its output end is connected to a ReLU activation function layer and a Dropout random dropout layer. The output of the third fully connected layer is 1, which outputs the prediction result.

[0053] In step 4, the mean squared error (MSE) loss function is used for model training, and gradient descent is performed according to the difference between the true value and the predicted value in the training set to optimize the neural network parameters. Among them, the loss function The formula is:

[0054]

[0055] Among them represents the true value of the output sequence , represents the predicted value of the output sequence , and n represents the number of training sets.

[0056] In step 4, the hyperparameter batch size is set to 16, the number of training epochs Epoch is set to 800, the stochastic gradient descent (SGD) optimizer is used, and the learning rate is set to 0.0001. The training set and the validation set are input into the model for training, and the optimal parameter model is saved during the training process.

[0057] The present invention also provides an electronic device, including a processor and a memory. The memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0058] The present invention has the following beneficial effects: (1) Before feature extraction of the data, the data is first preprocessed by SG smoothing filtering, effectively reducing noise, enhancing data quality, and ensuring the accuracy of subsequent feature extraction.

[0059] (2) A dynamic adaptive mechanism is introduced into the original ECA module. By calculating the channel mean and attention weights, the contribution of each channel in the feature map is automatically adjusted, improving the performance of the model.

[0060] (3) By using an improved residual block to replace the traditional convolutional layer to capture multi-level feature information, the risk of gradient disappearance is reduced, and the network is allowed to perform deeper learning, improving the model's ability to process complex data. Description of the Drawings

[0061] Figure 1 This is the schematic diagram of the present invention.

[0062] Figure 2 This is the network structure diagram of the improved one-dimensional convolutional neural network deep learning model of the present invention.

[0063] Figure 3 This is the network structure diagram of the adaptive and efficient channel attention module of the present invention.

[0064] Figure 4 This is the architecture diagram of the improved one-dimensional convolutional neural network deep learning model of the present invention. Specific implementation manners

[0065] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0066] As Figure 1 shown, an embodiment of the present invention provides a TDLAS intravenous drug concentration prediction method based on a one-dimensional convolutional neural network, including the following steps:

[0067] Step 1, obtaining second harmonic data of the intravenous drug concentration, specifically including the following steps:

[0068] Step 1-1, in the tunable diode laser absorption spectroscopy (TDLAS) technology, the wavelength modulation spectroscopy technology is adopted, and the LD laser is set to irradiate two paths respectively. One path passes through the liquid medicine, and the light intensity of the liquid concentration is calculated according to the Beer-Lambert law. The calculation formula is as follows:

[0069]

[0070] Among them, is the incident light wavelength, is the light intensity after transmission, is the incident light intensity, is the absorption coefficient, is the concentration of the liquid to be measured, is the optical path absorption distance.

[0071] The other path does not pass through the liquid and is emitted through the optical attenuator, and its light intensity is , is the optical attenuation coefficient. Let the optical attenuation coefficient , and the difference is made between these two light intensity signals and to obtain the differential signal ;

[0072] Step 1-2, performing Taylor series expansion on the obtained differential signal to obtain the second harmonic signal. The calculation formula is as follows:

[0073]

[0074] Steps 1 - 4, finally, the signal is demodulated by lock - in amplification, and the second - harmonic data of intravenous drugs at each concentration is collected through the host computer.

[0075] Step 2, pre - process the second - harmonic data of intravenous drugs, and divide it into a training set, a validation set, and a test set, specifically as follows:

[0076] Step 2 - 1, adopt the Savitzky - Golay smoothing filter method to smooth the second - harmonic data of intravenous drugs with a length of N to obtain an output sequence ; i = 0, 1, 2, 3, …, N - 1; for each data point in the output sequence , set a sliding window with a window size of 2m + 1, the sliding window includes the current data point and m adjacent points on the left and right of the current data point respectively. Within the sliding window, use an n - order polynomial to fit the current data point;

[0077] The Savitzky - Golay smoothing filter formula is:

[0078]

[0079] where is the filter coefficient, refers to the second - harmonic data of intravenous drugs with a length of N the element with index in

[0080] Step 2 - 2, obtain the filter coefficient through the following formula:

[0081]

[0082] where is an N × (2m + 1) matrix, and each row corresponds to a data point within the sliding window; is a column vector with a length of N, representing the vector of the second - harmonic data of intravenous drugs ;

[0083] Step 2 - 3, solve the filter coefficient in Step 2 - 2, substitute it into the Savitzky - Golay smoothing filter formula in Step 2 - 1, calculate the smoothed second - harmonic data of intravenous drugs , and obtain the second - harmonic data set of intravenous drugs;

[0084] Step 2-4: Divide the intravenous drug second harmonic data set into a training set, a validation set, and a test set according to a ratio of 8:1:1.

[0085] Step 3: Construct a 1D-CNN deep learning model based on an adaptive efficient channel attention module and a multi-level residual module, as Figure 2 shown. The specific steps are as follows:

[0086] As Figure 3 shown, construct an adaptive efficient channel attention module through the following steps:

[0087] Step 3-1: Calculate the mean value of the input feature map in the channel dimension to obtain the mean representation of each channel. The channel mean is calculated as follows:

[0088]

[0089] Where represents the th element of the th channel input feature map, and is the length of the feature map.

[0090] Step 3-2: Adjust the sizes of the convolutional kernels and through the learnable dynamic factors and The calculation formula is as follows:

[0091]

[0092]

[0093] Where, , is the learnable weight matrix optimized during the training process, is the activation function , ensuring that the generated attention weights are within the range, and are the bias terms, which, like the weights of the convolutional kernel, are automatically learned through backpropagation. The update formula for the bias terms is:

[0094]

[0095]

[0096] Where is the learning rate, and are the loss functions for the bias terms and The gradient of and The initialization is set to 0.1;

[0097] Step 3-3, adjust the convolutional kernels and through the generated dynamic factors and The calculation formula is:

[0098]

[0099]

[0100] where and are the initial convolutional kernel sizes, and are the new convolutional kernel sizes after being adjusted by the dynamic factors; is the rounding function;

[0101] Step 3-4, generate the attention weights and for each channel through the dynamically generated convolutional kernels The calculation formula is as follows:

[0102]

[0103] where the activation function ReLU is responsible for introducing non-linear characteristics, and the activation function scales the output range between 0 and 1; represents the attention weight of the i-th channel;

[0104] In Step 3, a multi-level residual module is constructed through the following steps: As Figure 4 shown, specifically: Set 4 convolutional layers. The size of the first convolutional layer is The sizes of the second and third convolutional layers are The size of the fourth convolutional layer is Add 1 BN batch normalization layer and ReLU activation function to the output end of each convolutional layer. The input feature passes through the first convolutional layer to obtain the first feature map ; The first feature map is input into the second and third convolutional layers to obtain the second feature map ; The first feature map is added to the second feature map through the skip connection to obtain the third feature map ; The third feature map passes through the fourth convolutional layer to obtain the fourth feature map And perform skip connections with the input features to obtain the output feature map .

[0105] In step 3, as Figure 4 shown, the regression prediction module includes three fully connected layers, a ReLU activation function layer, and a Dropout random dropout layer; the output of the first fully connected layer is 128, and the output end is connected to a ReLU activation function layer and a Dropout random dropout layer; the output of the second fully connected layer is 32, and the output end is connected to a ReLU activation function layer and a Dropout random dropout layer; the output of the third fully connected layer is 1, and the prediction result is output.

[0106] Step 4, input the training set and the validation set into the deep learning model, and set hyperparameters to train the model to obtain the optimal parameter model, specifically as follows:

[0107] Use the mean squared error (MSE) loss function to train the model, and perform gradient descent according to the difference between the true value and the predicted value in the training set to optimize the neural network parameters; among them, the loss function formula is:

[0108]

[0109] where represents the true value of the output sequence , represents the predicted value of the output sequence , and n represents the number of the training set.

[0110] Set the hyperparameter batch size to 16, the number of training epochs Epoch to 800, use the Stochastic Gradient Descent (SGD) optimizer, and set the learning rate to 0.0001; input the training set and the validation set into the model for training, and save the optimal parameter model during the training process.

[0111] Step 5, input the test set into the trained optimal parameter model to obtain the prediction result.

[0112] To better illustrate the effect of the present invention, using 5% glucose solution as the solvent and ambroxol hydrochloride as the solute, mix them according to the specified ratio and method to prepare the corresponding pharmaceutical dosage form, and obtain 600 data sets with 5 concentrations. Among them, the data set is divided into a training set, a validation set, and a test set in the ratio of 8:1:1, and several existing machine learning models are used to compare with the present invention, as shown in Table 1.

[0113] Table 1

[0114] Model <![CDATA[R 2 > MAE RMSE PCR 0.792 2.881 3.594 SVR 0.827 2.740 3.379 PLSR 0.852 2.382 3.032 The present invention 0.993 0.541 0.710

[0115] In Table 1, R 2 is the coefficient of determination, MAE is the mean absolute error, RMSE is the root mean square error as evaluation indicators, and PCR, SVR, and PLSR are commonly used regression prediction models in machine learning. Compared with these three machine learning models, the present invention has stronger representation learning ability and generalization ability, and is superior to the compared prior art in terms of R 2 , MAE, and RMSE indicators. The present invention has certain advantages in aspects such as signal extraction, feature learning, and complex relationship modeling of intravenous drug-induced second harmonic.

[0116] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A TDLAS intravenous drug concentration prediction method based on a one-dimensional convolutional neural network, characterized in that: The following steps are involved: Step 1, obtaining second harmonic data of intravenous drug concentration; Step 2, preprocessing the second harmonic data of intravenous medication and dividing it into a training set, a validation set, and a test set; Step 3, construct a one-dimensional convolutional neural network 1D-CNN deep learning model; Step 4: Input the training set and validation set into the deep learning model, and set the hyperparameters to train the model to obtain the optimal parameter model; Step 5: Input the test set into the trained optimal parameter model to obtain the prediction result; Step 1 includes: Step 1-1, using wavelength modulation spectroscopy technology in tunable laser absorption spectroscopy TDLAS technology, setting LD laser to irradiate two paths respectively, one of which passes through the liquid, and calculating the light intensity of the liquid concentration according to the Beer-Lambert law. The calculation formula is: I(λ)=I0(λ)e -φ(λ)cL , Among them, λ is the wavelength of incident light, I(λ) is the intensity of light after transmission, I0(λ) is the incident light intensity, φ(λ) is the absorption coefficient, c is the concentration of the liquid to be measured, L is the optical absorption distance, and e is a natural constant; The other path does not pass through the liquid and is emitted through the optical attenuator, with a light intensity of I ′ (λ)=I0(λ)A, A is the light attenuation coefficient; let the light attenuation coefficient A=1, make the difference between the two light intensity signals I(λ) and I′(λ), and get the differential signal ΔI(λ)=I′(λ)-I(λ)=I0(λ)cφ(λ)L; Step 1-2, perform Taylor series expansion on the obtained differential signal ΔI(λ) to obtain the second harmonic signal A2(λ), and the calculation formula is: in is the modulation amplitude, d represents the differential sign; Step 1-3, demodulate the second harmonic signal A2(λ) by phase-locked amplification, and collect the second harmonic data of intravenous medication of various concentrations through the host computer; Step 2 includes the following steps: Step 2-1, using the Savitzky-Golay smoothing filter method, the second harmonic data y of intravenous medication with length N is i Perform smoothing to obtain the output sequence i=0,1,2,3,…,N-1;for the output sequence For each data point in , a sliding window with a window size of 2m+1 is set, and the sliding window includes the current data point and m neighboring points on the left and right of the current data point. In the sliding window, an n-order polynomial is used to fit the current data point; The Savitzky-Golay smoothing filter formula is: Among them, c j is the filter coefficient, y i+j Refers to the second harmonic data y of intravenous medication with length N i The element with index i+j in the sliding window is within 2m+1 data points; Step 2-2, the filter coefficient c is obtained by the following formula j : c j =(B T B) -1 B T y, Where B is an N×(2m+1) matrix, each row corresponds to a data point in the sliding window; y is a column vector of length N, representing the second harmonic data y of intravenous medication i A vector of Step 2-3, by solving the filter coefficient c in step 2-2 j , substitute into the Savitzky-Golay smoothing filter formula in step 2-1, and calculate the smoothed second harmonic data of intravenous medication Get the second harmonic data set of intravenous medication; Step 2-4, divide the intravenous medication second harmonic data set into a training set, a validation set, and a test set; In step 3, the 1D-CNN deep learning model includes an adaptive and efficient channel attention module, a multi-level residual module, and a regression prediction module; In step 3, an adaptive and efficient channel attention module is constructed through the following steps: Step 3-1, calculate the mean of the input feature map in the channel dimension to obtain the mean representation of each channel, where the channel mean m i The calculation formula is: where x i [l] represents the lth element of the input feature map of the i-th channel, and H is the length of the feature map; Step 3-2: Through learnable dynamic factors and Adjust the size of convolution kernels w1 and w2, the calculation formula is: Among them, W1 and W2 are the learnable weight matrices optimized during the training process, σ and Sigmoid are activation functions, b1 and b2 are bias terms, which are automatically learned through back propagation. The update formula of the bias term is: Where η is the learning rate, and Respectively represent the gradient of the loss function with respect to the bias terms b1 and b2; b ′ 1 and b ′ 2 represent the updated value of bias term b1 and the updated value of bias term b2 respectively; Step 3-3, through the generated dynamic factor and Adjust the size of convolution kernels w1 and w2, the calculation formula is: in and is the initial convolution kernel size, and is the new convolution kernel size after dynamic factor adjustment; Step 3-4, generate the attention weight of each channel through the dynamically generated convolution kernels w1 and w2, and the calculation formula is: from i =Sigmoid(w2 ReLU(w1 m i )), The activation function ReLU is responsible for introducing nonlinear characteristics, and the activation function Sigmoid scales the output range between 0 and 1; i represents the attention weight of the i-th channel; Step 3-5, input feature map x i By generating the dynamic attention weight z i Get the output feature map x i ′ , the calculation formula is as follows: x i ′ =x i ×z i ; In step 3, a multi-level residual module is constructed by the following steps: 4 convolutional layers are set, the size of the first convolutional layer is 1×3, the size of the second and third convolutional layers is 1×5, and the size of the fourth convolutional layer is 1×7, and a BN batch normalization layer and a ReLU activation function are added to the output of each convolutional layer. The input feature p0 passes through the first convolutional layer to obtain the first feature map p1; the first feature map p1 is input to the second and third convolutional layers to obtain the second feature map p2; the first feature map p1 is added to the second feature map p2 through a jump link to obtain the third feature map p3; the third feature map p3 passes through the fourth convolutional layer to obtain the fourth feature map p4 and performs a jump link with the input feature to obtain the output feature map p5; In step 3, the regression prediction module includes three fully connected layers, a ReLU activation function layer, and a Dropout random loss layer; the output of the first fully connected layer is 128, and the output end is connected to a ReLU activation function layer and a Dropout random loss layer; the output of the second fully connected layer is 32, and the output end is connected to a ReLU activation function layer and a Dropout random loss layer; the output of the third fully connected layer is 1, and the prediction result is output; In step 4, the mean square error (MSE) loss function is used for model training, and gradient descent is performed according to the difference between the true value and the predicted value in the training set to optimize the neural network parameters; the loss function Loss formula is: where y true，k Represents the output sequence The true value of y pred，k Represents the output sequence The predicted value of , n represents the number of training sets; In step 4, set the hyperparameter batch size to 16, the number of training Epochs to 800, use the stochastic gradient descent SGD optimizer, and the learning rate to 0.0001; input the training set and the validation set into the model training, and save the optimal parameter model during the training process.

2. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to claim 1.

Images