Hemoglobin content prediction method and equipment

Abstract

The invention relates to the technical field of biological detection, in particular to a hemoglobin content prediction method and equipment, and the method comprises the following steps: processing multiple groups of light absorption signals obtained based on a hemoglobin solution to obtain time sequence data; inputting the time sequence data into a one-dimensional convolutional neural network to extract multi-scale local features, and inputting the multi-scale local features into an attention transformation network to extract global information and local information to obtain comprehensive attention features; extracting a relationship between different peaks in the target signal corresponding to the time sequence data through a graph attention network to obtain a peak-to-peak relationship feature; and inputting the comprehensive attention features and the peak-to-peak relationship features into a feature fusion network for fusion, and calculating a hemoglobin content index based on the obtained fusion features. The method comprehensively considers the influence of the overall and local characteristics, the peak-to-peak relationship and the like, so that the hemoglobin content index can be accurately detected.

Description

Technical Field

[0001] This application relates to the field of biological detection technology, and in particular to a method and device for predicting hemoglobin content. Background Technology

[0002] Fetal hemoglobin (HbF) is the main type of hemoglobin in the fetus, composed of two α chains and two γ chains (α2γ2). In adults, HbF levels are usually low, but under certain pathological conditions (such as β-thalassemia, sickle cell disease, etc.) or in hereditary persistent hemoglobinemia of fetus (HPFH), HbF levels can be significantly elevated. Therefore, accurate detection and quantification of HbF are of great significance for disease diagnosis, treatment monitoring, and prognostic assessment.

[0003] In existing technologies, high-performance liquid chromatography (HPLC) is the most common method for HbF detection. HPLC is one of the gold standard methods for HbF detection. It separates hemoglobin variants using a chromatographic column and performs quantitative analysis using a UV or visible light detector. This method features high sensitivity and high resolution. However, this method relies on experience in peak identification. If there are significant impurity peaks or baseline drift in the chromatographic peaks, the identification of HbF will be highly subjective, resulting in inaccurate HbF detection.

[0004] In liquid chromatography (LC), peak identification and quantification are crucial steps. Traditional rule-based peak identification methods identify target peaks by setting rules such as retention time windows, peak height thresholds, and peak area thresholds. However, this method suffers from low accuracy in complex samples (such as hemoglobin variants) and is susceptible to baseline drift and peak overlap. Deep learning can automatically extract complex features from chromatographic data, and studies have attempted to use convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for peak identification. However, their application in complex samples (such as hemoglobin variants) remains exploratory and faces numerous challenges. For example, under certain conditions, the HbF peak resolution is low, overlapping peaks occur, and column aging can cause severe tailing or baseline drift, resulting in low detection accuracy. Summary of the Invention

[0005] In view of this, the present application provides a method and device for predicting hemoglobin content, which can effectively solve the problem of inaccurate hemoglobin detection in the prior art.

[0006] In a first aspect, embodiments of this application provide a method for predicting hemoglobin content, applicable to hybrid architecture prediction models; the hybrid architecture prediction model includes a one-dimensional convolutional neural network, an attention transformation network, a graph attention network, and a feature fusion network; the method includes: Multiple sets of optical absorption signals obtained from hemoglobin solution were processed to obtain time-series data; The time-series data is input into the one-dimensional convolutional neural network to extract multi-scale local features, and the multi-scale local features are input into the attention transformation network to extract global and local information to obtain comprehensive attention features. The graph attention network is used to extract the relationship between different peaks in the target signal corresponding to the time series data to obtain the inter-peak relationship features. The integrated attention feature and the inter-peak relationship feature are input into the feature fusion network for fusion, and the hemoglobin content index is calculated based on the obtained fusion feature.

[0007] In some embodiments, the hybrid architecture prediction model further includes a global coefficient embedding network; The method further includes: inputting the constructed global coefficient features into the global coefficient embedding network to extract the change information of the target signal and obtain global trend features; The step of inputting the integrated attention features and the inter-peak relationship features into the feature fusion network for fusion includes: The integrated attention feature, the inter-peak relationship feature, and the global trend feature are input into the feature fusion network for fusion to obtain the fused feature.

[0008] In some embodiments, the global coefficient embedding network includes: a fully connected layer, a hidden layer, a nonlinear activation layer, and a multi-head attention mechanism layer; The step of inputting the constructed global coefficient features into the global coefficient embedding network to extract the change information of the target signal and obtain global trend features includes: performing high-dimensional mapping on the global coefficient features through the fully connected layer to obtain a first feature; performing dimensionality reduction processing on the first feature through the hidden layer to obtain a second feature; performing nonlinear expression enhancement processing on the second feature through the nonlinear activation layer to obtain a third feature; and dynamically adjusting the third feature through the multi-head attention mechanism layer to obtain the global trend features. The global coefficient features are constructed based on parameters including the average tailing factor, blank signal trend, and baseline drift standard deviation.

[0009] In some embodiments, the processing of multiple sets of optical absorption signals obtained based on hemoglobin solution to obtain time-series data includes: The first wavelength light absorption signal and the second wavelength light absorption signal were obtained from the photodetector after the hemoglobin solution was separated by a liquid chromatography separator. The target signal is calculated based on the first wavelength light absorption signal and the second wavelength light absorption signal; The target signal is standardized to obtain the time-series data.

[0010] In some embodiments, the hybrid architecture prediction model further includes an output network; the hemoglobin content index includes hemoglobin area; The calculation of hemoglobin content indicators based on the obtained fusion features includes: The fused features are input into the output network to predict the hemoglobin area, and the hemoglobin content is estimated based on the hemoglobin area.

[0011] In some embodiments, the output network includes: multiple fully connected layers, an activation function located between every two adjacent fully connected layers, and a linear output layer; The step of inputting the fused features into the output network to predict hemoglobin area includes: The input fused features are nonlinearly transformed by the multi-layer fully connected layers to gradually reduce the feature dimension. Multiple activation functions are used to enhance the nonlinear expressive power of the model. The reduced and enhanced features are mapped to the target output space by the linear output layer to obtain the predicted hemoglobin area.

[0012] In some embodiments, the method includes at least one of the following three: First item: The one-dimensional convolutional neural network includes: a feature fusion layer, a convolutional layer, and multiple sets of parallel processing modules; each processing module includes a convolutional kernel, a batch normalization layer, an activation function, and a max pooling layer connected in sequence; the step of inputting the temporal data into the one-dimensional convolutional neural network to extract multi-scale local features includes: performing convolution operations on the temporal data through the convolutional kernels of each group to extract local information of the target signal; batch normalizing the corresponding local information through the batch normalization layers of each group; introducing nonlinearity into the batch normalized features through the activation functions of each group; downsampling the features after introducing nonlinearity through the max pooling; concatenating the downsampled features of each group in the channel dimension through the feature fusion layer; and fusing the concatenated features through the convolutional layer to generate the multi-scale local features; Second item: The attention transformation network includes: a global attention mechanism layer, a local attention mechanism layer, a feature fusion layer, and a fully connected layer connected in sequence; The step of inputting the multi-scale local features into the attention transformation network to extract global and local information to obtain comprehensive attention features includes: The global attention mechanism layer performs global attention calculation on the multi-scale local features to obtain global features; The local attention mechanism layer divides the multi-scale local features into multiple local windows, calculates the local information within each local window, and then aggregates them into local features. The global and local features are concatenated along the channel dimension by a feature fusion layer, and the concatenated features are then subjected to a non-linear transformation by a fully connected layer to generate the comprehensive attention feature. Third item: The feature fusion network includes: a feature splicing layer, an attention mechanism layer, and a weighted summation layer; The integrated attention features and the inter-peak relationship features are input into the feature fusion network for fusion, including: The comprehensive attention feature, the inter-peak relationship feature, and the global trend feature are concatenated along the channel dimension through the feature concatenation layer. The attention mechanism layer performs a linear transformation on the concatenated features and calculates attention weights to generate multiple attention weights. The concatenated features are then weighted and summed according to the multiple attention weights to obtain the fused feature.

[0013] In some embodiments, the step of extracting the relationship between different peaks in the target signal corresponding to the time-series data through the graph attention network to obtain inter-peak relationship features includes: Multiple peak features are extracted from the time-series data based on the peak information of the target signal; The multiple peak features are input into the graph attention network to extract the relationship between different peaks in the target signal, thereby obtaining the inter-peak relationship features; The multiple peak characteristics include at least one of the following: retention time, peak height, peak tail, half-peak tail, left half-peak width, right half-peak width, left peak separation, right peak separation, fitted peak area, and total area of ​​each peak.

[0014] In some embodiments, the graph attention network includes: a multi-head attention mechanism, a feature aggregation layer, a batch normalization layer, and a non-linear activation function; The step of inputting the multiple peak features into the graph attention network to extract the relationship between different peaks in the target signal, and obtaining the inter-peak relationship features, includes: A graph structure containing multiple nodes is constructed based on the aforementioned multiple peak features; The attention weight between each node and its neighboring nodes is calculated through the multi-head attention mechanism layer; For each node, the features of all its corresponding neighboring nodes are weighted and summed according to their respective attention weights, and the features of the current node are updated using the weighted summed features. The updated node features are batch normalized, and then a nonlinear relationship is introduced through the nonlinear activation function to obtain the inter-peak relationship features; Optionally, constructing a graph structure containing multiple nodes based on the multiple peak features includes: The peaks in the target signal and multiple neighboring peaks are used as nodes; wherein, the features of each node include at least one of the following: retention time, peak height, Gaussian fitted peak area, tailing factor, half-peak tailing factor, left Gaussian similarity, and right Gaussian similarity. The edges of the graph are constructed based on the relationship characteristics between the nodes, wherein the relationship characteristics include the distance between peaks, the ratio of peak heights, and the ratio of peak fitting areas. Calculate the target features for each edge and store them as an edge attribute tensor; wherein the target features include at least one of peak distance, peak shape similarity, peak height ratio, and peak area ratio; The graph structure is obtained based on each of the nodes and each of the edges.

[0015] Secondly, embodiments of this application provide a terminal device, the terminal device including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement a hemoglobin content prediction method provided in the first aspect of this application.

[0016] The embodiments of this application have the following beneficial effects: The hybrid architecture prediction model in this application includes a one-dimensional convolutional neural network, an attention transformation network, a graph attention network, and a feature fusion network. The method includes: processing multiple sets of light absorption signals obtained from a hemoglobin solution to obtain time-series data; inputting the time-series data into the one-dimensional convolutional neural network to extract multi-scale local features; inputting the multi-scale local features into the attention transformation network to extract global and local information to obtain comprehensive attention features; extracting the relationship between different peaks in the target signal corresponding to the time-series data through a graph attention network to obtain inter-peak relationship features; inputting the comprehensive attention features and the inter-peak relationship features into the feature fusion network for fusion; and calculating the hemoglobin content index based on the obtained fused features. Because this application comprehensively considers the influence of overall and local features, inter-peak relationships, etc., it can accurately detect the hemoglobin content index. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a hemoglobin content prediction method according to an embodiment of this application is shown; Figure 2 This paper illustrates a schematic diagram of a hybrid architecture prediction model used in the hemoglobin content prediction method according to an embodiment of this application. Figure 3 This paper illustrates another structural schematic diagram of the hybrid architecture prediction model used in the hemoglobin content prediction method of this application embodiment; Figure 4 This paper shows a schematic diagram of the global coefficient embedding network of the hybrid architecture prediction model used in the hemoglobin content prediction method of this application embodiment; Figure 5 This paper shows a schematic diagram of the structure of a one-dimensional convolutional neural network of the hybrid architecture prediction model used in the hemoglobin content prediction method of this application embodiment; Figure 6 This paper shows a schematic diagram of the attention transformation network structure of the hybrid architecture prediction model used in the hemoglobin content prediction method of this application embodiment; Figure 7 This paper shows a schematic diagram of the feature fusion network structure of the hybrid architecture prediction model used in the hemoglobin content prediction method of this application embodiment; Figure 8 This paper illustrates another structural schematic diagram of the hybrid architecture prediction model used in the hemoglobin content prediction method of this application embodiment; Figure 9 A schematic diagram of the output network of the hybrid architecture prediction model used in the hemoglobin content prediction method of this application is shown.

[0019] Explanation of key component symbols: 110 - One-dimensional convolutional neural network; 120 - Attention transformation network; 130 - Graph attention network; 140 - Feature fusion network; 150 - Global coefficient embedding network; 160 - Output network. Detailed Implementation

[0020] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0021] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] In the following text, the terms "comprising," "having," and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more combinations thereof. Furthermore, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0023] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0024] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0025] The following examples illustrate the method for predicting hemoglobin levels.

[0026] Figure 1 A flowchart of a hemoglobin content prediction method according to an embodiment of this application is shown.

[0027] Exemplary, this hemoglobin content prediction method is applicable to hybrid architecture prediction models; such as Figure 2As shown, the hybrid architecture prediction model includes a one-dimensional convolutional neural network 110, an attention transformation network 120, a graph attention network 130, and a feature fusion network 140. The output of the one-dimensional convolutional neural network 110 is connected to the attention transformation network 120, and the outputs of both the attention transformation network 120 and the graph attention network 130 are connected to the feature fusion network 140. The one-dimensional convolutional neural network 110 is also known as 1D-CNN, the graph attention network 130 is also known as GAT, and the attention transformation network 120 includes, but is not limited to, a Transformer.

[0028] The method for predicting hemoglobin levels includes the following steps: S100 processes multiple sets of light absorption signals obtained from hemoglobin solution to obtain time-series data.

[0029] Whole blood was anticoagulated with EDTA, and hemoglobin solution was obtained by centrifugation after lysing red blood cells. In this embodiment, a detection system was used to process the hemoglobin solution to obtain multiple sets of light absorption signals. These multiple sets of light absorption signals were then processed to obtain a target signal, and time-series data was obtained based on the target signal.

[0030] As an example, multiple sets of optical absorption signals obtained based on hemoglobin solutions were processed to obtain time-series data, including: S110, acquiring a first wavelength light absorption signal and a second wavelength light absorption signal from a photodetector after separating the hemoglobin solution using a liquid chromatography separator. The hemoglobin solution includes, but is not limited to, a solution containing fetal hemoglobin.

[0031] In this embodiment, a detection system is used to process the hemoglobin solution, obtaining two sets of light absorption signals. Exemplarily, the detection system in this embodiment includes a liquid chromatography separator and a light absorption detector. During the liquid chromatography separation process based on the liquid chromatography separator, the light absorption detector detects the absorbance of the eluted sample (hemoglobin solution) at a predetermined wavelength and generates a light absorption signal. A chromatogram is then plotted based on the light absorption signal. The light absorption signal reflects the concentration level of each component in the sample. The predetermined wavelength includes, but is not limited to, a first wavelength and a second wavelength. For example, the first wavelength is 415 nm, and the second wavelength is 500 nm. In liquid chromatography, the sample is separated in the chromatographic column by the mobile phase, and different components elute sequentially due to differences in retention time (RT). A commonly used ultraviolet-visible (UV-Vis) detector measures the absorption of a component by light at a specific wavelength (following Lambert-Beer's law). The efflux time series of each component is obtained by real-time detection of the output signal intensity of the optical sensor (horizontal axis is time, vertical axis is detector response). Essentially, it is a two-dimensional data stream of time and signal intensity, with each peak corresponding to a component, and peak area / height reflecting concentration. The target signal is obtained by subtracting the light absorption signal corresponding to the first wavelength from the light absorption signal corresponding to the second wavelength. In other words, the target signal is obtained by subtracting the two sets of two-dimensional time-signal intensity data streams.

[0032] S120, the target signal is calculated based on the first wavelength light absorption signal and the second wavelength light absorption signal.

[0033] For example, the target signal can be obtained by subtracting the two sets of light absorption signals.

[0034] S130: Standardize the target signal to obtain time-series data.

[0035] For example, using the retention time of the HbA0 peak as a benchmark, batch-to-batch retention time drift is corrected, and a Savitzky-Golay filter is used to remove noise before sampling to obtain time-series data. HbA is a major form of hemoglobin in normal adults, including non-glycated hemoglobin (HbA0) and glycated hemoglobin (HbA1). In normal adults, HbA accounts for approximately 95%–97% of total hemoglobin, with HbA0 accounting for about 90% and HbA1 accounting for about 5%–8%. In high-performance liquid chromatography (HPLC) analysis, HbA0 forms a dominant peak, called the HbA0 peak. The Savitzky-Golay filter is a commonly used signal smoothing and denoising method.

[0036] S200, the time series data is input into a one-dimensional convolutional neural network 110 to extract multi-scale local features, and the multi-scale local features are input into an attention transformation network 120 to extract global and local information to obtain comprehensive attention features.

[0037] S300 uses a graph attention network 130 to extract the relationship between different peaks in the target signal corresponding to the time series data, and obtains the inter-peak relationship features.

[0038] S400 inputs the comprehensive attention features and inter-peak relationship features into the feature fusion network 140 for fusion, and calculates the hemoglobin content index based on the obtained fusion features.

[0039] This application provides a novel hybrid structure prediction model that considers the influence of global features corresponding to global information, local features corresponding to local information, and inter-peak relationships. Specifically, it combines a multi-scale one-dimensional convolutional neural network 110 (1D-CNN) with a local Transformer and a graph attention network 130 (GAT) to simultaneously capture local details and global information of the target signal. This model extracts features at different scales through multi-scale convolution and enhances the modeling ability of local details through the local attention mechanism of the Transformer, thereby improving the accuracy and efficiency of signal processing, i.e., improving the accuracy of hemoglobin content prediction.

[0040] In one implementation, to further improve prediction accuracy, embodiments of this application also consider the signal change trend caused by system variations by embedding all coefficients. It can be understood that, as Figure 3 As shown, the hybrid architecture prediction model also includes a global coefficient embedding network 150. The output of the one-dimensional convolutional neural network 110 is connected to the attention transformation network 120, and the outputs of the attention transformation network 120, the graph attention network 130, and the global coefficient embedding network 150 are all connected to the feature fusion network 140.

[0041] The method also includes: inputting the constructed global coefficient features into the global coefficient embedding network 150 to extract the change information of the target signal and obtain global trend features.

[0042] Exemplary, global coefficient features are constructed based on the mean tailing factor, blank signal trend, and baseline drift standard deviation. For example, global coefficient features (global system parameters) are obtained by Z-score normalization of the mean tailing factor, blank signal trend, and baseline drift standard.

[0043] The average tailing factor is calculated based on the tailing factors of the main peaks from the most recent batches. The main peak is the peak corresponding to HbA0.

[0044] The blank signal trend is calculated based on the timing, signal strength, and signal length of the blank signal. The blank signal trend reflects the overall linear change of the target signal, such as sensor baseline drift.

[0045] The blank signal refers to the background signal generated by the detection system itself in the absence of target hemoglobin solution (i.e., "zero concentration"). It is the basis for assessing the system's noise level and baseline stability. In other words, the blank signal is the environmental background signal collected before each measurement. Furthermore, the blank signal is the light absorption signal in a blank test, which is a test performed without a sample, only with reagents.

[0046] The overall trend of blank signals in the characteristic segment ,in Indicates timing, This represents the signal strength of the blank signal, where n is the signal length of the blank signal. i Indicates the index of blank signal data points.

[0047] The baseline drift standard deviation is calculated based on the target signal strength, the baseline fit value, and the target signal length. The baseline fit value refers to the fitted value output at a specific time point or data point in the chromatogram, modeled using a mathematical model (such as polynomial fitting, spline interpolation, wavelet transform, etc.) to represent the background signal strength at that point.

[0048] Baseline drift standard deviation: ,in It is the signal strength of the target signal. This is the baseline fit value. n is the signal length of the target signal. i Indicates the index of the target data point.

[0049] Furthermore, such as Figure 4 As shown, the global coefficient embedding network 150 includes: a fully connected layer, a hidden layer, a non-linear activation layer, and a multi-head attention mechanism layer.

[0050] Understandably, the constructed global coefficient features are input into the global coefficient embedding network 150 to extract the change information of the target signal, resulting in global trend features, including: The global coefficient features are mapped to a high-dimensional feature space through a fully connected layer to obtain the first feature. The first feature is then reduced in dimensionality through a hidden layer to obtain the second feature. The second feature is further enhanced with a non-linear activation layer to improve its non-linear expressive power, resulting in the third feature. Finally, a multi-head attention mechanism layer dynamically adjusts the third feature to capture the importance of global information, yielding the global trend feature. The non-linear activation layer includes, but is not limited to, the ReLU activation function.

[0051] As an example, global coefficients are mapped to 128 dimensions in a fully connected layer, and then the 128-dimensional features are mapped to 64 dimensions in a hidden layer. The ReLU activation function is used to enhance the non-linear expressive power, and then a multi-head attention mechanism layer is used to dynamically adjust the features that enhance the non-linear expressive power.

[0052] Understandably, the embedded global coefficient features include: mean tailing factor, blank signal trend, and baseline drift standard deviation.

[0053] Average tailing factor: The average tailing factor of the HbA0 peak in this batch of samples, reflecting the overall characteristics of the peak shape. The tailing factor is an important parameter for measuring peak symmetry and evaluating instrument condition.

[0054] Overall signal trend (blank signal trend): The overall trend of the signal, reflecting the global changes in the signal.

[0055] Baseline drift (expressed as baseline drift standard deviation): The degree of baseline drift of the target signal, which measures the dispersion of the data and reflects the stability of the target signal.

[0056] Understandably, in step S400, the integrated attention features and inter-peak relationship features are input into the feature fusion network 140 for fusion, including: The comprehensive attention features, inter-peak relationship features, and global trend features are input into the feature fusion network 140 for fusion to obtain fused features.

[0057] The hybrid architecture model in this application comprehensively considers the influence of overall and local features, inter-peak relationships, and signal trends, further improving the accuracy of hemoglobin content prediction.

[0058] In one implementation, such as Figure 5 As shown, the one-dimensional convolutional neural network 110 includes: a feature fusion layer, convolutional layers, and multiple sets of parallel processing modules; each processing module includes a convolutional kernel, a batch normalization layer, an activation function, and a max pooling layer connected in sequence. Exemplarily, this embodiment includes three sets of processing modules. Temporal data is input into the three sets of processing modules respectively. The three sets of processing modules each employ convolutional kernels of different sizes: large, medium, and small. Small convolutional kernel: Uses a smaller convolutional kernel to extract local details of the target signal. Medium convolutional kernel branch: Uses a medium-sized convolutional kernel 11 to extract local features of the target signal. Large convolutional kernel branch: Uses a larger convolutional kernel 51 to extract the overall trend of the target signal. For example, the first set uses a 3×3 convolutional kernel, the second set uses an 11×11 convolutional kernel, and the third set uses a 51×51 convolutional kernel.

[0059] Understandably, in step S200, the temporal data is input into a one-dimensional convolutional neural network 110 to extract multi-scale local features, including: The time-series data is convolved using convolutional kernels in each group to extract local information of the target signal. Batch normalization layers in each group normalize the corresponding local information to improve stability. Activation functions in each group introduce nonlinearity into the batch-normalized features. Max pooling is then used to downsample the nonlinearized features to retain key features. A feature fusion layer concatenates the downsampled features along the channel dimension, and a convolutional layer fuses the concatenated features to generate multi-scale local features. Downsampling reduces the spatial dimension (e.g., height, width) or resolution of the data while retaining the most important information, thereby reducing computational complexity, increasing model stability, and preventing overfitting.

[0060] Demonstratively, convolutional kernels of sizes 3×3, 11×11, and 51×51 are used to convolve the input temporal data to extract local details of the target signal. For each group of convolutional kernels, batch normalization is performed on the convolutional results to accelerate training and improve model stability. Then, the ReLU activation function is used to introduce nonlinearity, followed by adaptive max pooling to downsample the features introducing nonlinearity, preserving key features. Finally, in the feature fusion layer, the output features of the corresponding groups of small, medium, and large convolutional kernels are concatenated along the channel dimension, and a 1×1 convolution is used to fuse the concatenated features, generating the final multi-scale local features.

[0061] In one implementation, such as Figure 6 As shown, the attention transformation network 120 includes: a global attention mechanism layer, a local attention mechanism layer, a feature fusion layer, and a fully connected layer connected in sequence. The global attention mechanism uses a standard Transformer encoder to capture global information of the target signal. The local attention mechanism divides the input data into multiple local windows, calculates attention weights within each window, and captures local details. The feature fusion layer concatenates the output features of the global and local attention along the channel dimension, and uses a fully connected layer to perform a non-linear transformation to generate the final feature representation.

[0062] Understandably, in step S200, the multi-scale local features are input into the attention transformation network 120 to extract global and local information, resulting in comprehensive attention features, including: S210, global features are obtained by performing global attention calculation on multi-scale local features through a global attention mechanism layer; S220 divides multi-scale local features into multiple local windows through a local attention mechanism layer, calculates local information within each local window, and then aggregates them into local features. S230 uses a feature fusion layer to concatenate global and local features along the channel dimension, and then uses a fully connected layer to perform a non-linear transformation on the concatenated features to generate comprehensive attention features.

[0063] Furthermore, the global attention mechanism layer includes a multi-head self-attention mechanism and a feedforward neural network. The local attention mechanism layer includes a local window partitioning unit, a feature aggregation unit, and a feature fusion unit.

[0064] Understandably, the multi-head self-attention mechanism is used to compute the relationship between each position (temporal sequence) and other positions (temporal sequences) in the multi-scale local features of the input, capture global information (global features), and use a feedforward neural network to perform a nonlinear transformation on the output of the multi-head self-attention to enhance the feature representation.

[0065] The local attention mechanism's local window partitioning unit divides multi-scale local features into multiple local windows, each containing a fixed number of locations (e.g., 5 locations). Attention weights are calculated within each window to capture local details. Specifically, the attention weights between each location and other locations within the window are calculated. Finally, based on the attention weights of each location, the features are aggregated by a feature aggregation unit (mean pooling), and the aggregated features are fused by a feature fusion unit.

[0066] Finally, through feature fusion, the output features of the global attention mechanism layer and the local attention mechanism layer are concatenated along the channel dimension, and the concatenated features are subjected to nonlinear transformation using a fully connected layer to generate the final feature representation.

[0067] In one implementation, step S300 involves extracting the relationship between different peaks in the target signal corresponding to the time-series data using a graph attention network 130 to obtain inter-peak relationship features, including: S310 extracts multiple peak features from time-series data based on the peak information of the target signal.

[0068] In this embodiment of the application, the peak information of the target signal is obtained through preliminary peak detection. For example, the first derivative method is used to locate all peaks (including miscellaneous peaks) to obtain peak information.

[0069] The root peak information extracts multiple peak features from time series data, including at least one of the following: retention time, peak height, peak tail, half-peak tail, left half-peak width, right half-peak width, left peak separation, right peak separation, fitted peak area, and total area of ​​each peak.

[0070] S320, multiple peak features are input to the graph attention network 130 to extract the relationship between different peaks in the target signal and obtain the inter-peak relationship features.

[0071] Furthermore, the graph attention network 130 includes: a multi-head attention mechanism, a feature aggregation layer, a batch normalization layer, and a non-linear activation function.

[0072] Understandably, multiple peak features are input into the graph attention network 130 to extract the relationships between different peaks in the target signal, resulting in inter-peak relationship features, including: S321, constructs a graph structure containing multiple nodes based on multiple peak features.

[0073] Furthermore, a graph structure containing multiple nodes is constructed based on multiple peak features, including: The peaks in the target signal and multiple neighboring peaks are used as nodes; the features of each node include at least one of the following: retention time, peak height, Gaussian fitted peak area, tailing factor, half-peak tailing factor, left Gaussian similarity, and right Gaussian similarity; that is, the node features include retention time, peak height, Gaussian fitted peak area, tailing factor, half-peak tailing factor, left Gaussian similarity, and right Gaussian similarity.

[0074] The edges of the graph are constructed based on the relationship characteristics between the nodes; where the relationship characteristics include the distance between peaks, the ratio of peak heights, and the ratio of peak fitting areas. The target features of each edge are calculated and stored as an edge attribute tensor. The target features include at least one of the following: peak distance, peak shape similarity, peak height ratio, and peak area ratio. That is, edge features include peak distance, peak shape similarity, peak height ratio, and peak area ratio. The main purpose of storing these as edge attribute tensors is to provide additional feature information for the edges in the graph structure. This feature information helps the hybrid architecture model of this application embodiment better capture the relationships between nodes, thereby improving the model's performance.

[0075] The graph structure is obtained based on each node and each edge.

[0076] S322 calculates the attention weight between each node and its neighboring nodes through a multi-head attention mechanism layer.

[0077] The multi-head attention mechanism layer includes linear transformation, attention computation, multi-head concatenation, and linear transformation. Specifically, it calculates the attention weights of each node with its neighboring nodes, and when calculating the attention weights, it introduces edge features (inter-peak distance, peak similarity, etc.) to enhance the information content of the graph structure.

[0078] S323. For each node, the features of all its corresponding neighboring nodes are weighted and summed according to the corresponding attention weights, and the features of the current node are updated using the weighted summed features.

[0079] S324 performs batch normalization on the updated node features, and then introduces non-linear relationships through a non-linear activation function to obtain inter-peak relationship features. Batch normalization of the updated node features accelerates training and improves model stability, and the ReLU activation function is used to introduce non-linear relationships.

[0080] In one implementation, such as Figure 7 As shown, the feature fusion network 140 includes: a feature concatenation layer, an attention mechanism layer, and a weighted summation layer. The feature concatenation layer (feature fusion) concatenates features from different modules along the channel dimension. Attention weight calculation: The attention mechanism layer calculates the weights of different features. Feature weighting: The features are weighted and summed according to the attention weights to generate the fused feature representation.

[0081] Understandably, in step S400, the integrated attention features and inter-peak relationship features are input into the feature fusion network 140 for fusion, including: The feature concatenation layer concatenates comprehensive attention features, inter-peak relationship features, and global trend features along the channel dimension. The attention mechanism layer performs a linear transformation on the concatenated features and calculates attention weights to generate multiple attention weights. The concatenated features are then weighted and summed based on these multiple attention weights to obtain the fused features.

[0082] In one implementation, such as Figure 8 As shown, the hybrid architecture prediction model also includes an output network 160; the hemoglobin content index includes hemoglobin area.

[0083] In step S400, the hemoglobin content index is calculated based on the obtained fusion features, including: The fused features are input into the output network 160 to predict the hemoglobin area, and the hemoglobin content is estimated based on the hemoglobin area.

[0084] Furthermore, such as Figure 9 The output network 160 includes: multiple fully connected layers, activation functions located between every two adjacent fully connected layers, and a linear output layer. For example, in this embodiment, it includes three sequentially connected fully connected layers, with an activation function set between every two fully connected layers. The activation function includes, but is not limited to, the ReLU activation function.

[0085] Understandably, the fused features are input into the output network 160 to predict hemoglobin area, including: The fused features of the input are nonlinearly transformed through multiple fully connected layers to progressively reduce the feature dimensionality. Multiple activation functions are used to enhance the model's nonlinear expressive power. A linear output layer maps the reduced-dimensional and enhanced features to the target output space to obtain the predicted hemoglobin area. The reduced-dimensional and enhanced features include those obtained after processing by multiple fully connected layers and those processed by multiple activation functions. Specifically, the hybrid architecture prediction model enhances its ability to capture complex features and patterns by introducing nonlinear transformations using activation functions in each layer. The use of activation functions allows the model to progressively extract and combine more complex features, thereby improving its nonlinear expressive power.

[0086] The training and validation of the hybrid architecture prediction model in this application are described below: The training to validation data ratio is 4:1, and all data comes from IFCC sample tests.

[0087] Data preparation: Input data includes preprocessed waveform data, a constructed peak characteristic table, and global coefficients. The target data is the HbF peak area. The gold standard for HbF peak area is derived from the IFCC sample target value sheet.

[0088] Forward calculation: Based on the peak feature table and the relationship between peaks, a graph structure is constructed to generate node features, edge indices and edge features. The preprocessed waveform signal, graph structure and global coefficients are then input into the model to obtain the predicted HbF area.

[0089] Loss function: The Huber Loss (a parametric loss function used for regression tasks) is used to calculate the loss between the predicted HbF peak area and the target HbF peak area.

[0090]

[0091] Where Loss is the loss value. To predict the peak area, This represents the actual peak area.

[0092] Backpropagation: Calculate the gradient of the loss function with respect to the model parameters and update the model parameters using the Adam optimizer.

[0093] This application extracts local features at different scales from the input signal using a multi-scale 1D-CNN; captures local details and global information of the target signal using a local attention Transformer; captures the relationships between different features in the modeling signal using a graph attention network (GAT); fuses the features extracted by the multi-scale 1D-CNN, Transformer, and GAT using a feature fusion module; and finally, makes predictions based on the fused features using an output module.

[0094] This application provides a hemoglobin content prediction device. It is suitable for hybrid architecture prediction models; the hybrid architecture prediction model includes a one-dimensional convolutional neural network 110, an attention transformation network 120, a graph attention network 130, and a feature fusion network 140. Exemplarily, the hemoglobin content prediction device includes: The data processing module is used to process multiple sets of light absorption signals obtained based on hemoglobin solution to obtain time-series data; The attention information extraction module is used to input time series data into a one-dimensional convolutional neural network 110 to extract multi-scale local features, and input the multi-scale local features into an attention transformation network 120 to extract global and local information to obtain comprehensive attention features; The peak-to-peak relationship extraction module is used to extract the relationship between different peaks in the target signal corresponding to the time series data through the graph attention network 130, and obtain the peak-to-peak relationship features. The fusion processing module is used to input the comprehensive attention features and inter-peak relationship features into the feature fusion network 140 for fusion, and calculate the hemoglobin content index based on the obtained fusion features.

[0095] It is understood that the device in this embodiment corresponds to the hemoglobin content prediction method in the above embodiments, and the options in the above embodiments are also applicable to this embodiment, so they will not be described again here.

[0096] This application also provides a terminal device, exemplary of which includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to enable the terminal device to perform the functions of the various modules in the above-described hemoglobin content prediction method or the above-described hemoglobin content prediction device. Exemplarily, this terminal device is a hemoglobin detection instrument.

[0097] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0098] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0099] This application also provides a computer-readable storage medium for storing the computer program used in the aforementioned terminal device. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0101] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0102] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting hemoglobin content, characterized in that, Applicable to hybrid architecture prediction models; the hybrid architecture prediction model includes a one-dimensional convolutional neural network, an attention transformation network, a graph attention network, and a feature fusion network; the method includes: Multiple sets of optical absorption signals obtained from hemoglobin solution were processed to obtain time-series data; The time-series data is input into the one-dimensional convolutional neural network to extract multi-scale local features, and the multi-scale local features are input into the attention transformation network to extract global and local information to obtain comprehensive attention features. The graph attention network is used to extract the relationship between different peaks in the target signal corresponding to the time series data to obtain the inter-peak relationship features. The integrated attention feature and the inter-peak relationship feature are input into the feature fusion network for fusion, and the hemoglobin content index is calculated based on the obtained fusion feature.

2. The method for predicting hemoglobin content according to claim 1, characterized in that, The hybrid architecture prediction model also includes a global coefficient embedding network; The method further includes: inputting the constructed global coefficient features into the global coefficient embedding network to extract the change information of the target signal and obtain global trend features; The step of inputting the integrated attention features and the inter-peak relationship features into the feature fusion network for fusion includes: The integrated attention feature, the inter-peak relationship feature, and the global trend feature are input into the feature fusion network for fusion to obtain the fused feature.

3. The method for predicting hemoglobin content according to claim 2, characterized in that, The global coefficient embedding network includes: a fully connected layer, a hidden layer, a nonlinear activation layer, and a multi-head attention mechanism layer; The step of inputting the constructed global coefficient features into the global coefficient embedding network to extract the change information of the target signal and obtain global trend features includes: The global coefficient features are mapped to a high dimension through the fully connected layer to obtain a first feature; the first feature is reduced in dimension through the hidden layer to obtain a second feature; the second feature is enhanced in non-linear expression through the non-linear activation layer to obtain a third feature; and the third feature is dynamically adjusted through the multi-head attention mechanism layer to obtain the global trend feature. The global coefficient features are constructed based on parameters including the average tailing factor, blank signal trend, and baseline drift standard deviation.

4. The method for predicting hemoglobin content according to claim 1, characterized in that, The process of processing multiple sets of optical absorption signals obtained based on hemoglobin solution to obtain time-series data includes: The first wavelength light absorption signal and the second wavelength light absorption signal were obtained from the photodetector after the hemoglobin solution was separated by a liquid chromatography separator. The target signal is calculated based on the first wavelength light absorption signal and the second wavelength light absorption signal; The target signal is standardized to obtain the time-series data.

5. The method for predicting hemoglobin content according to claim 1, characterized in that, The hybrid architecture prediction model also includes an output network; the hemoglobin content index includes hemoglobin area; The calculation of hemoglobin content indicators based on the obtained fusion features includes: The fused features are input into the output network to predict the hemoglobin area, and the hemoglobin content is estimated based on the hemoglobin area.

6. The method for predicting hemoglobin content according to claim 5, characterized in that, The output network includes: multiple fully connected layers, activation functions located between every two adjacent fully connected layers, and a linear output layer; The step of inputting the fused features into the output network to predict hemoglobin area includes: The input fused features are nonlinearly transformed by the multi-layer fully connected layers to gradually reduce the feature dimension. Multiple activation functions are used to enhance the nonlinear expressive power of the model. The reduced and enhanced features are mapped to the target output space by the linear output layer to obtain the predicted hemoglobin area.

7. The method for predicting hemoglobin content according to claim 5, characterized in that, The method includes at least one of the following three: First item: The one-dimensional convolutional neural network includes: a feature fusion layer, a convolutional layer, and multiple sets of parallel processing modules; each processing module includes a convolutional kernel, a batch normalization layer, an activation function, and a max pooling layer connected in sequence. The step of inputting the time-series data into the one-dimensional convolutional neural network to extract multi-scale local features includes: performing convolution operations on the time-series data through convolutional kernels of each group to extract local information of the target signal; batch normalizing the corresponding local information through batch normalization layers of each group; introducing nonlinearity into the batch-normalized features through activation functions of each group; downsampling the features after introducing nonlinearity through max pooling; concatenating the downsampled features of each group in the channel dimension through the feature fusion layer; and fusing the concatenated features through the convolutional layer to generate the multi-scale local features. Second item: The attention transformation network includes: a global attention mechanism layer, a local attention mechanism layer, a feature fusion layer, and a fully connected layer connected in sequence; The step of inputting the multi-scale local features into the attention transformation network to extract global and local information to obtain comprehensive attention features includes: The global attention mechanism layer performs global attention calculation on the multi-scale local features to obtain global features; The local attention mechanism layer divides the multi-scale local features into multiple local windows, calculates the local information within each local window, and then aggregates them into local features. The global and local features are concatenated along the channel dimension by a feature fusion layer, and the concatenated features are then subjected to a non-linear transformation by a fully connected layer to generate the comprehensive attention feature. Third item: The feature fusion network includes: a feature splicing layer, an attention mechanism layer, and a weighted summation layer; The integrated attention features and the inter-peak relationship features are input into the feature fusion network for fusion, including: The comprehensive attention feature, the inter-peak relationship feature, and the global trend feature are concatenated along the channel dimension through the feature concatenation layer. The attention mechanism layer performs a linear transformation on the concatenated features and calculates attention weights to generate multiple attention weights. The concatenated features are then weighted and summed according to the multiple attention weights to obtain the fused feature.

8. The method for predicting hemoglobin content according to claim 1, characterized in that, The step of extracting the relationship between different peaks in the target signal corresponding to the time-series data through the graph attention network to obtain the inter-peak relationship features includes: Multiple peak features are extracted from the time-series data based on the peak information of the target signal; The multiple peak features are input into the graph attention network to extract the relationship between different peaks in the target signal, thereby obtaining the inter-peak relationship features; The multiple peak characteristics include at least one of the following: retention time, peak height, peak tail, half-peak tail, left half-peak width, right half-peak width, left peak separation, right peak separation, fitted peak area, and total area of ​​each peak.

9. The method for predicting hemoglobin content according to claim 8, characterized in that, The graph attention network includes: a multi-head attention mechanism, a feature aggregation layer, a batch normalization layer, and a non-linear activation function; The step of inputting the multiple peak features into the graph attention network to extract the relationship between different peaks in the target signal and obtain the inter-peak relationship features includes: constructing a graph structure containing multiple nodes based on the multiple peak features; calculating the attention weight between each node and its neighboring nodes through the multi-head attention mechanism layer; for each node, calculating the weighted sum of the features of all its corresponding neighboring nodes according to the corresponding attention weights, and updating the features of the current node using the weighted summed features; batch normalizing the updated node features, and then introducing a non-linear relationship through the non-linear activation function to obtain the inter-peak relationship features. Optionally, constructing a graph structure containing multiple nodes based on the multiple peak features includes: The peaks in the target signal and their neighboring peaks are used as nodes. Each node is characterized by at least one of the following: retention time, peak height, Gaussian fitted peak area, tailing factor, half-peak tailing factor, left Gaussian similarity, and right Gaussian similarity. Edges of the graph are constructed based on the relationship features between the nodes, including peak-to-peak distance, peak height ratio, and peak fitted area ratio. The target features of each edge are calculated and stored as an edge attribute tensor, including at least one of peak-to-peak distance, peak shape similarity, peak height ratio, and peak area ratio. The graph structure is obtained based on each node and each edge.

10. A terminal device, characterized in that, The terminal device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the hemoglobin content prediction method according to any one of claims 1-9.

Images