Preparation method of traditional Chinese medicine compound for treating rapid atrial fibrillation

Abstract

This invention discloses a method for preparing a traditional Chinese medicine compound for treating rapid atrial fibrillation, comprising: extracting Coptis chinensis with phosphate buffer; acquiring optical data of the extract in real time and inputting it into a multi-channel convolutional neural network model; terminating extraction when the output comprehensive extraction efficiency index reaches a predetermined threshold to obtain Coptis chinensis extract; decocting Codonopsis pilosula, Polygonatum sibiricum, Rehmannia glutinosa, Paeonia lactiflora, Morus alba, Glycyrrhiza uralensis, Citrus medica, calcined dragon bone, calcined oyster shell, and Rosa rugosa; acquiring real-time concentration time-series data of tanshinone IIA, paeoniflorin, and glycyrrhizic acid in the decoction and inputting it into a long short-term memory neural network model; terminating decoction when the interaction effect parameter fluctuates within ≤±5% for three consecutive detection cycles to obtain a mixed decoction; combining the Coptis chinensis extract and the mixed decoction, concentrating to obtain a concentrated extract; and preparing the traditional Chinese medicine compound. This invention can improve the preparation efficiency and quality stability of traditional Chinese medicine compound drugs.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine. More specifically, this invention relates to a method for preparing a traditional Chinese medicine compound for treating rapid atrial fibrillation. Background Technology

[0002] In the field of traditional Chinese medicine compound preparations, the preparation process of rapid-acting atrial fibrillation treatment drugs has long faced multi-dimensional technical bottlenecks. Traditional extraction processes, using fixed temperature and time parameters (as described in the invention patent application number 2024117757811), struggle to match the differentiated dissolution characteristics of multiple components in medicinal materials. For example, isothermal extraction easily leads to degradation or insufficient dissolution of active ingredients, while controlling the extraction endpoint with a single indicator cannot simultaneously consider related parameters such as polysaccharide polymerization degree and cell wall breakage rate, resulting in insufficient batch-to-batch uniformity of efficacy. The low dissolution efficiency of mineral-based medicinal materials constitutes another technical challenge. Due to their dense structure, traditional decoction methods struggle to overcome the physical barriers of mineral lattices and cell walls in mineral medicines such as calcined dragon bone and calcined oyster shell. Even with extended decoction time or increased temperature, dissolution efficiency cannot be effectively improved, and it may even exacerbate the destruction of heat-sensitive components in the herbal medicine. The dissolution kinetics of lipophilic components like tanshinone and water-soluble components like paeoniflorin differ significantly, and traditional processes cannot capture the dynamic interactions between components in real time.

[0003] The essence of the aforementioned technical bottleneck lies in the lack of a multi-dimensional process monitoring system and dynamic control mechanism in traditional processes. Although existing detection technologies attempt to incorporate spectroscopic analysis, the complex correlations of multi-channel data make it difficult for traditional chemometric methods to achieve effective analysis. Therefore, it is necessary to design a technical solution that can overcome these shortcomings. Summary of the Invention

[0004] One objective of this invention is to provide a method for preparing a traditional Chinese medicine compound for treating rapid atrial fibrillation, which can improve the preparation efficiency and quality stability of the traditional Chinese medicine compound.

[0005] To achieve these objectives and other advantages of the present invention, according to one aspect of the present invention, a method for preparing a traditional Chinese medicine compound for treating rapid atrial fibrillation is provided, comprising: S1: taking Coptis chinensis and extracting it with phosphate buffer; S2: acquiring optical data of the extract in real time using an ultraviolet-near-infrared spectrometer and inputting it into a pre-trained multi-channel convolutional neural network model; wherein, in the multi-channel convolutional neural network model, the first convolutional channel analyzes absorbance data at a wavelength of 280 nm, the second convolutional channel processes near-infrared spectral data at 1050-1100 nm, and the third convolutional channel extracts spectral features at 2200-2400 nm; when the comprehensive extraction efficiency index output by the model reaches a predetermined threshold, the extraction is terminated to obtain Coptis chinensis extract; S3: taking Codonopsis pilosula, Polygonatum sibiricum, Rehmannia glutinosa, Paeonia lactiflora, Morus alba, Ziziphus jujuba var. spinosa, Schisandra chinensis, Triticum aestivum, Ophiopogon japonicus, Ligusticum chuanxiong, ... The following steps were performed: S4: Real-time concentration data of tanshinone IIA, paeoniflorin, and glycyrrhiza in the decoction were obtained using high-performance liquid chromatography (HPLC) and input into a long short-term memory neural network model based on an attention mechanism. In the long short-term memory neural network model, the bidirectional long short-term memory layer extracted the dissolution rate characteristics of each component, the attention mechanism layer calculated the synergistic weight coefficients of tanshinone IIA and paeoniflorin, and the output layer generated flavonoid-saponin interaction effect parameters. The decoction was terminated when the interaction effect parameters fluctuated within ≤±5% for three consecutive detection cycles, and a mixed decoction was obtained. S5: The Coptis chinensis extract and the mixed decoction were combined, treated with a microfiltration membrane, and then subjected to alcohol precipitation. The supernatant of the alcohol precipitation was collected, concentrated, and a concentrated extract was obtained. S6: The concentrated extract was mixed with a pharmaceutically acceptable carrier to prepare a traditional Chinese medicine compound.

[0006] Further, in S1, Coptis chinensis is pulverized to 40-60 mesh and added to a phosphate buffer solution with pH 6.8 at a material-to-liquid ratio of 1:8-1:12. The extraction is carried out in three stages with controlled temperature. In the first stage, the extraction is carried out at a constant temperature of 45℃ for 30 minutes. In the second stage, the temperature is increased to 65℃ for pulsed ultrasonic extraction. In the third stage, when the conductivity of the extract increases by ≤1% / 5 minutes for three consecutive tests, the extraction is switched to hot reflux extraction at 85℃.

[0007] Further, in S2, the first convolutional channel is configured to use a 3×1 convolutional kernel to process the 280nm absorbance time-series data and output a first feature map; the second convolutional channel is configured to use a 5×1 dilated convolutional kernel to extract features from the 1050-1100nm near-infrared spectrum and output a second feature map, with a dilation rate of 2 for the dilated convolution; the third convolutional channel is configured to use a 7×1 convolutional kernel combined with an exponential linear unit activation function to capture features from the 2200-2400nm spectrum and output a third feature map; the first, second, and third feature maps are then combined... Channel dimensions are concatenated to form an initial fusion feature. A 1×1 convolution kernel operation is applied to the initial fusion feature to generate a channel attention weight matrix. A three-dimensional convolution kernel operation is applied to the initial fusion feature to generate a spatial-spectral correlation feature. The channel attention weight matrix and the spatial-spectral correlation feature are multiplied element-wise to output an optimized fusion feature. The optimized fusion feature is input into a fully connected layer to generate a multidimensional feature vector containing berberine dissolution parameters, polysaccharide polymerization parameters, and cell wall breakage parameters. The comprehensive extraction efficiency index is calculated based on the multidimensional feature vector.

[0008] Furthermore, the berberine dissolution parameter is standardized by maximum value to generate a standard dissolution parameter; the polysaccharide polymerization parameter is logarithmically transformed to generate a standard polymerization parameter; and the cell wall breakage rate parameter is transformed by a Sigmoid function to generate a standard cell wall breakage rate parameter. The standard dissolution parameter, standard polymerization parameter, and standard cell wall breakage rate parameter are input to the fully connected layer, and an initial weight vector is output. This initial weight vector is then concatenated with the batch feature code of the medicinal material, which includes the origin code, harvest season code, and storage duration code. Dynamic weighting coefficients are generated through a gated circulation unit, where the weighting coefficients for berberine dissolution parameter range from 0.55 to 0.65, polysaccharide polymerization parameter range from 0.25 to 0.35, and cell wall breakage rate parameter range from 0.08 to 0.12. The comprehensive extraction efficiency index is calculated using the following formula: Comprehensive extraction efficiency index = dynamic weighting coefficient α × standard dissolution parameter + dynamic weighting coefficient β × standard polymerization parameter + dynamic weighting coefficient γ × standard cell wall breakage rate parameter, where α, β, and γ are dynamic weighting coefficients.

[0009] Furthermore, it also includes: acquiring the comprehensive extraction efficiency index of the multi-channel convolutional neural network model output in real time, calculating the confidence score of the multi-channel convolutional neural network model output, the confidence score being generated by performing Monte Carlo random dropout sampling on the optimized fusion features, based on the inverse of the variance of 10 consecutive prediction results; when the confidence score is lower than the 0.85 threshold, the extraction is no longer determined based on the comprehensive extraction efficiency index, but is terminated when the cumulative dissolution of berberine is not less than the maximum dissolution.

[0010] Further, in S3, calcined dragon bone and calcined oyster shell are pulverized to 100-120 mesh. In the first stage, they are treated with a 1.0% citric acid solution at a liquid-to-solid ratio of 3:1 at 50°C for 15 minutes. In the second stage, they are treated with a 0.5% citric acid solution at a liquid-to-solid ratio of 5:1 at 70°C for 20 minutes. The filtrates from the first and second stages are combined to obtain a mineral filtrate. A pH 5.8 acetate buffer solution with a liquid-to-solid ratio of 8:1 is added to Codonopsis pilosula, Polygonatum sibiricum, Rehmannia glutinosa (processed), Rehmannia glutinosa (raw), Paeonia lactiflora, Paeonia veitchii, Ligusticum chuanxiong, and Salvia miltiorrhiza. The decoction is then prepared at 80°C for 25 minutes to obtain a rhizome decoction. The rhizome decoction, mineral filtrate, and purified water are added to Ziziphus jujuba var. spinosa, Schisandra chinensis, Triticum aestivum, Ophiopogon japonicus, Glycyrrhiza uralensis, Citrus medica, and Rosa rugosa. The total liquid-to-solid ratio is controlled at 12:1, and the decoction is prepared by reflux at 95°C.

[0011] Furthermore, in S4, the input layer of the long short-term memory neural network model is configured to simultaneously receive five independent time-series data channels: the first channel inputs the time-series data of tanshinone IIA concentration, the second channel inputs the time-series data of paeoniflorin concentration, the third channel inputs the time-series data of glycyrrhizic acid concentration, the fourth channel inputs the time-series data of the decoction system temperature, and the fifth channel inputs the time-series data of the decoction system pressure. The data from the five channels are respectively input to the corresponding processing units in the bidirectional long short-term memory layer, which respectively output the tanshinone IIA dissolution feature vector, the paeoniflorin dissolution feature vector, the glycyrrhizic acid dynamic feature vector, the temperature dynamic feature vector, and the pressure dynamic feature vector. The dissolution feature vectors of tanshinone IIA, paeoniflorin, glycyrrhizic acid, temperature, and pressure are concatenated. The concatenated features are input into an attention mechanism layer, processed by a fully connected layer and a hyperbolic tangent activation function to generate an intermediate feature matrix, and then normalized using a Softmax function to output collaborative weight coefficients. These collaborative weight coefficients are then fused with the tanshinone IIA, paeoniflorin, and glycyrrhizic acid dissolution feature vectors. The fused features are processed by an output layer to generate flavonoid-saponin interaction effect parameters normalized to the 0-1 range.

[0012] Further, in S5, the Coptis chinensis extract and the mixed decoction are combined and then microfiltered through a microfiltration membrane with a pore size of 0.22 μm. The microfiltration pressure is controlled in three stages: the first stage is 0.25 MPa for 5 minutes, the second stage is increased to 0.40 MPa for 3 minutes, and the third stage is reduced to 0.18 MPa and maintained until the filtration endpoint. The resulting microfiltrate is subjected to gradient alcohol precipitation: first, ethanol is added to make the alcohol concentration reach 60% ± 2%, and after standing at 4°C for 20 minutes, the precipitate is separated and discarded; then, ethanol is added to the supernatant to make the alcohol concentration reach 80% ± 1%, and after standing at 25°C for 15 minutes, the alcohol precipitation supernatant is collected; the alcohol precipitation supernatant is concentrated under reduced pressure at 45°C ± 1°C and a vacuum degree of -0.08 MPa, and the concentration is terminated when the relative density of the concentrated extract is 1.25-1.3.

[0013] The present invention has at least the following beneficial effects:

[0014] This invention utilizes a staged temperature control combined with a multi-channel spectral-neural network monitoring system in the extraction of Coptis chinensis to achieve the synergistic dissolution of alkaloids and polysaccharides. This ensures the structural stability of heat-sensitive components while improving cell wall disruption efficiency, overcoming the contradiction between component degradation and insufficient dissolution in traditional constant-temperature extraction. The gradient acid treatment process for mineral drugs specifically disrupts the mineral lattice structure, and combined with a group decoction strategy, effectively promotes a balance between the dissolution of inorganic components and the release of active ingredients from the herbal medicine, solving the technical bottleneck of low dissolution efficiency of mineral drugs in traditional water decoction methods. During the compound decoction process, a dynamic model based on a long short-term memory neural network and attention mechanism captures the temporal characteristics and interaction effects of multi-component dissolution in real time, achieving precise regulation of the synergistic release of flavonoids and saponins, avoiding the process lag caused by traditional offline detection.

[0015] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0016] Figure 1 This is a flowchart of one embodiment of this application. Detailed Implementation

[0017] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0018] It should be understood that terms such as "having," "comprising," and "including" used in the embodiments of this application do not exclude the presence or addition of one or more other elements or combinations thereof. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of components in a specific posture. If the specific posture changes, the directional indication will also change accordingly. When an element is referred to as "fixed to" or "set on" another element, it can be directly on the other element or may have an intervening element present. When an element is referred to as "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element through an intervening element. Descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0019] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0020] like Figure 1 As shown, embodiments of this application provide a method for preparing a traditional Chinese medicine compound for treating rapid atrial fibrillation, comprising: S1. Coptis chinensis was extracted with phosphate buffer. S2. Optical data of the extract was acquired in real time using a UV-NIR spectrometer and input into a pre-trained multi-channel convolutional neural network model. The first convolutional channel resolved absorbance data at 280 nm, the second convolutional channel processed near-infrared spectral data at 1050-1100 nm, and the third convolutional channel extracted spectral features at 2200-2400 nm. Extraction was terminated when the comprehensive extraction efficiency index output by the model reached a predetermined threshold, yielding Coptis chinensis extract. S3. Codonopsis pilosula and Polygonatum sibiricum were decocted. S4. The concentration time-series data of tanshinone IIA, paeoniflorin, and glycyrrhizic acid in the decoction were acquired in real time using a high-performance liquid chromatography (HPLC) system and input into a long short-term memory neural network model based on an attention mechanism. Decoction was terminated when the interaction effect parameter fluctuated within ≤±5% for three consecutive detection cycles. S5. The extract and decoction were combined, treated with a microfiltration membrane, precipitated with alcohol, and the supernatant was collected and concentrated. S6. The concentrated extract was mixed with a pharmaceutical carrier to obtain the drug.

[0021] For example, in S1, Coptis chinensis can be processed to a conventional pharmaceutical mesh size using a pulverizing device, and a phosphate buffer solution with a pH adjusted to 6.8±0.2 is added. The material-to-liquid ratio is controlled between 1:8 and 1:12, specifically 1:8, 1:10, or 1:12. In S2, a commercially available device such as the PerkinElmer Lambda series can be used for the ultraviolet-near-infrared spectrometer. Spectral data is acquired in real time at a sampling interval of 1 minute. The absorbance at 280 nm corresponds to the characteristic absorption of berberine, the 1050-1100 nm near-infrared band reflects the hydroxyl vibration of polysaccharides, and the 2200-2400 nm region captures information about alkaloid functional groups. During the pre-training of the multi-channel convolutional neural network model, the first convolutional channel uses a 3×1 convolutional kernel to analyze the temporal features of 280nm absorbance. The second convolutional channel extracts near-infrared spectral interval features using a 5×1 dilated convolutional kernel (expansion rate 2). The third convolutional channel uses a 7×1 convolutional kernel combined with the ELU activation function to process the mid-infrared spectrum. After the features of the three channels are fused, a comprehensive extraction efficiency index is output through a fully connected layer, with a predetermined threshold set to 0.85. In S3, a stainless steel jacketed kettle can be used for the decoction operation, with the stirring rate controlled at 50-80 rpm. In S4, a high-performance liquid chromatograph, such as an Agilent 1260 series, is used to monitor the component concentration online. The time-series data is input into the Bi-LSTM model, and the attention mechanism layer calculates the component synergistic weights. The fluctuation threshold of the interaction effect parameter is set to ±5%. In S5, a microfiltration membrane with a pore size of 0.22 μm is selected. The alcohol precipitation process is carried out in two steps, adjusting the ethanol concentration to 60% and 80%, with the concentration temperature controlled at 45℃±1℃. In S6, common excipients such as starch and lactose can be selected as pharmaceutical carriers.

[0022] In existing technologies, traditional preparation processes employ extraction at fixed temperatures and times, controlling the endpoint solely through offline detection of a single component (such as berberine). This fails to address the synergistic dissolution requirements of multiple components, leading to both degradation of heat-sensitive components and incomplete dissolution of active ingredients, resulting in significant batch-to-batch differences in efficacy. This embodiment, however, couples real-time ultraviolet-near-infrared spectroscopy monitoring with a multi-channel neural network model to achieve dynamic control of the extraction and decoction processes. Based on the comprehensive efficiency index and interaction effect parameters, it precisely controls the process endpoint, effectively improving the synergistic dissolution of multiple components and the consistency of formulation quality.

[0023] In another embodiment, in S1, Coptis chinensis is pulverized to 40-60 mesh and added to a phosphate buffer solution with pH 6.8 at a material-to-liquid ratio of 1:8 to 1:12. The extraction is carried out in three stages with controlled temperature. In the first stage, the extraction is carried out at a constant temperature of 45°C for 30 minutes. In the second stage, the temperature is raised to 65°C for pulsed ultrasonic extraction. In the third stage, when the conductivity of the extract increases by ≤1% / 5 minutes for three consecutive tests, the extraction is switched to hot reflux extraction at 85°C.

[0024] For example, Coptis chinensis powder can be pulverized using a universal pulverizer with a mesh size of 40, 50, or 60 mesh to ensure particle uniformity. A phosphate buffer solution with pH 6.8 is prepared from potassium dihydrogen phosphate and disodium hydrogen phosphate in a conventional ratio, with a specific material-to-liquid ratio of 1:8, 1:10, or 1:12. In the three-stage temperature-controlled extraction, the first stage uses a constant temperature water bath to maintain 45°C for 30 minutes to protect heat-sensitive components such as berberine; the second stage raises the temperature to 65°C and uses a pulsed ultrasonic extractor (ultrasonic power 300-500W, pulse frequency 10-20Hz) to enhance cell wall disruption; the third stage uses a DDS-307 conductivity meter for real-time monitoring. When the increase in conductivity is ≤1% / 5 minutes for three consecutive tests (e.g., 0.8% / 5min, 0.9% / 5min), the system switches to a rotary evaporator with a hot reflux device for dynamic extraction at 85°C.

[0025] In existing technologies, Coptis chinensis extraction typically involves prolonged decoction at a single temperature (e.g., 60°C), failing to consider the dissolution temperature differences between alkaloids and polysaccharides. Furthermore, endpoint determination relies on experience or offline detection, resulting in high berberine degradation and insufficient polysaccharide dissolution. This embodiment utilizes staged temperature control combined with dynamic conductivity monitoring to create a gradient extraction mechanism of "low-temperature protection - medium-temperature cell disruption - high-temperature extraction." This reduces the degradation of heat-sensitive components while enhancing dissolution through pulsed ultrasound and thermal reflux, resolving the contradiction between component loss and incomplete extraction in traditional processes.

[0026] In another embodiment, in S2, the first convolutional channel is configured to use a 3×1 convolutional kernel to process the 280nm absorbance time-series data and output a first feature map; the second convolutional channel is configured to use a 5×1 dilated convolutional kernel to extract features of the 1050-1100nm near-infrared spectrum and output a second feature map, with the dilation rate of the dilated convolution being 2; the third convolutional channel is configured to use a 7×1 convolutional kernel combined with an exponential linear unit activation function to capture features of the 2200-2400nm spectrum and output a third feature map; the three feature maps are concatenated along the channel dimension, and a channel attention weight matrix is ​​generated by a 1×1 convolution, which is then fused with the spatial-spectral correlation features extracted by the three-dimensional convolution, input into the fully connected layer to generate a multi-dimensional feature vector, and the comprehensive extraction efficiency index is calculated.

[0027] For example, the convolutional neural network is built based on the TensorFlow framework. The input data are 280nm absorbance (sampling interval 1min, a total of 100 time points), 1050-1100nm near-infrared spectrum (51 wavelength points), and 2200-2400nm mid-infrared spectrum (101 wavelength points). The first convolutional channel processes the absorbance data through three 3×1 convolutional layers (stride 1, padding=same), outputting a 100×1×32 feature map; the second convolutional channel uses a 5×1 dilated convolutional kernel with an dilation rate of 2, with an equivalent receptive field covering 9 spectral points, outputting a 51×1×64 feature map; the third convolutional channel uses a 7×1 convolutional kernel combined with the ELU activation function (α=1.0), outputting a 101×1×48 feature map. After feature concatenation, channel attention weights are generated by a 1×1 convolutional layer (64 kernels). Then, spectral-temporal correlation features are extracted by a 3×3×3 three-dimensional convolutional kernel. The fully connected layer (256 nodes) outputs a feature vector containing parameters such as berberine dissolution rate and polysaccharide polymerization degree. The comprehensive efficiency index is calculated by linear weighting (initial weights α=0.6, β=0.3, γ=0.1).

[0028] In existing technologies, spectral analysis often employs Principal Component Analysis (PCA) to process single near-infrared channel data, failing to utilize the synergistic information between ultraviolet characteristic absorption and near-infrared vibrational spectra. Furthermore, the feature extraction layer cannot capture the dynamic correlation between the spectral and temporal dimensions, making it difficult for the model to accurately reflect the complex relationships of multi-component dissolution. This embodiment adapts to different spectral scale features through multi-channel convolutional kernel design, combines dilated convolution to expand the receptive field with three-dimensional convolution to capture spatiotemporal correlations, and uses a channel attention mechanism to enhance the spectral signals of key components. This significantly improves the accuracy of extraction efficiency assessment compared to traditional methods, solving the technical problem of insufficient multispectral data analysis.

[0029] In another embodiment, the berberine dissolution parameter is standardized to its maximum value to generate a standard dissolution parameter; the polysaccharide polymerization degree parameter is logarithmically transformed to generate a standard polymerization degree parameter; the cell wall breakage rate parameter is transformed using a Sigmoid function to generate a standard cell wall breakage rate parameter; the three types of standard parameters are input to the fully connected layer, and an initial weight vector is output; the initial weight vector is concatenated with the batch feature code of the medicinal material containing the place of origin, harvesting season, and storage time, and a dynamic weight coefficient is generated through a gated loop unit, wherein the weight of berberine ranges from 0.55 to 0.65, the weight of polysaccharide ranges from 0.25 to 0.35, and the weight of cell wall breakage rate ranges from 0.08 to 0.12, and the comprehensive extraction efficiency index is calculated according to the formula.

[0030] For example, maximum value standardization maps the berberine dissolution rate to the [0,1] interval, and the logarithm is converted to log10(polysaccharide polymerization degree + 1), with the Sigmoid function being 1 / (1+e^(-x)). The fully connected layer (32 nodes) outputs a 3-dimensional initial weight vector, and the batch feature code of the medicinal material adopts unique thermal encoding (e.g., place of origin Beijing=100, Henan=010, harvesting season spring=10, autumn=01, storage time ≤6 months=1). The gated recurrent unit (GRU) has 64 hidden layer nodes, and dynamic weights are generated through update gates and reset gates. α can take values ​​of 0.55, 0.60, 0.65, β can take values ​​of 0.25, 0.30, 0.35, and γ can take values ​​of 0.08, 0.10, 0.12. The comprehensive index = α × standard dissolution rate + β × standard polymerization degree + γ × standard cell wall breakage rate.

[0031] In existing technologies, extraction efficiency assessments often employ fixed weights (e.g., focusing only on berberine content), failing to consider the impact of batch-to-batch variations on component dissolution. This results in poor model adaptability to herbs from different origins and harvest times, and a lack of dynamic adjustment mechanisms for process parameters between batches. This embodiment standardizes multiple parameters (maximum standardization / logarithmic transformation / Sigmoid transformation) to unify dimensions, and combines batch-to-batch feature codes with GRU to generate dynamic weight coefficients. This achieves adaptive control of the extraction process to the variations in herbs, reducing the fluctuation range of extraction efficiency across different batches and solving the problem of insufficient adaptability in traditional fixed-weight models.

[0032] In another embodiment, the method further includes: acquiring the comprehensive extraction efficiency index of the multi-channel convolutional neural network model output in real time, calculating the confidence score of the model output, which is generated by performing Monte Carlo random dropout sampling on the optimized fusion features and based on the inverse of the variance of 10 consecutive prediction results; when the confidence score is lower than the 0.85 threshold, the endpoint is no longer determined based on the comprehensive index, but instead the extraction is terminated when the cumulative dissolution of berberine is not less than the maximum dissolution amount.

[0033] For example, Monte Carlo random dropout sampling (Dropout rate 0.2) was used to randomly inactivate the optimized fusion features 10 times, generating a predicted value for the comprehensive efficiency index each time. The variance σ² of the 10 results was calculated, and the confidence score was 1 / (1+σ²). The threshold of 0.85 can be adjusted to 0.8 or 0.9. When the score is lower than the threshold, the cumulative dissolution of berberine is detected offline using a Waters e2695 HPLC and compared with the maximum dissolution determined in the preliminary experiment for this batch of medicinal materials. Extraction is terminated when the dissolution reaches 85%, 90%, or 95%.

[0034] In existing technologies, intelligent model control lacks a reliability verification mechanism. When spectral data fluctuates or the model deviates, it can easily lead to incorrect judgment of the extraction endpoint, resulting in over-extraction or under-extraction of components. This embodiment constructs a triple guarantee mechanism of "model prediction - reliability verification - offline detection" by calculating confidence scores through Monte Carlo sampling. When the model confidence is insufficient, it automatically switches to hard endpoint control based on component content, solving the risk problem of single intelligent model control and improving the reliability of the extraction process.

[0035] In another embodiment, in S3, calcined dragon bone and calcined oyster shell are pulverized to 100-120 mesh. In the first stage, they are treated with 1.0% citric acid solution at a liquid-to-solid ratio of 3:1 at 50°C for 15 minutes. In the second stage, they are treated with 0.5% citric acid solution at a liquid-to-solid ratio of 5:1 at 70°C for 20 minutes. The filtrates are combined to obtain mineral filtrate. A pH 5.8 acetate buffer (liquid-to-solid ratio 8:1) is added to rhizome herbs such as Codonopsis pilosula and decocted at 80°C for 25 minutes to obtain rhizome decoction. Rhizome decoction, mineral filtrate and purified water are added to herbs such as Ziziphus jujuba var. spinosa, and the total liquid-to-solid ratio is controlled at 12:1. The decoction is refluxed at 95°C.

[0036] For example, calcined dragon bone and calcined oyster shell are pulverized to 100, 110, or 120 mesh using an air jet mill. In the first stage, a 1.0% citric acid solution (liquid-to-solid ratio 3:1) is used, maintained at 50°C in a constant temperature water bath with a stirring speed of 200 rpm for 15 minutes. In the second stage, a 0.5% citric acid solution is heated to 70°C and used at a liquid-to-solid ratio of 5:1 for 20 minutes. The combined filtrates are filtered through a Buchner funnel. Rhizome-type medicinal materials are added to a pH 5.8 acetate buffer solution (prepared with acetate and sodium acetate), with a liquid-to-solid ratio of 8:1, and decocted in a jacketed kettle at 80°C for 25 minutes. Flower / seed-type medicinal materials are mixed with the rhizome decoction and mineral filtrate, and purified water is added to adjust the total liquid-to-solid ratio to 12:1. The mixture is then decocted in a reflux extraction device at 95°C, with the condensate temperature ≤25°C.

[0037] In existing technologies, mineral and herbal medicines are often processed using a unified decoction process. However, due to the dense structure of mineral medicines, the dissolution rate of inorganic components such as calcium and magnesium is less than 5%, and high-temperature, long-term decoction easily causes degradation of heat-sensitive components in herbal medicines. This embodiment uses gradient acid treatment (1.0%-0.5% citric acid) to disrupt the mineral lattice of mineral medicines, combined with a group decoction strategy (roots and rhizomes - flowers / seeds), to specifically optimize the dissolution conditions of different medicinal materials. This significantly improves the dissolution rate of inorganic components in mineral medicines while avoiding degradation of herbal medicine components, thus solving the dual problems of low dissolution efficiency of mineral medicines and loss of herbal medicine components in traditional co-decoction processes.

[0038] In another embodiment, in S4, the input layer of the long short-term memory neural network model simultaneously receives data from five time-series channels: tanshinone IIA concentration, paeoniflorin concentration, glycyrrhizic acid concentration, decoction temperature, and decoction pressure. The data are input into the Bi-LSTM processing unit, which outputs the dissolution feature vectors of each component and the dynamic feature vectors of the process parameters. After feature concatenation, the data is input into the attention mechanism layer, and an intermediate matrix is ​​generated through a fully connected layer and a tanh activation function. The softmax normalization outputs the collaborative weight coefficients, which are then fused with the feature vectors to generate flavonoid-saponin interaction effect parameters in the 0-1 range.

[0039] For example, the model is built on PyTorch, with input data consisting of a 5-channel time series (sampling interval 5 min, 30 time points in total). Each Bi-LSTM processing unit contains 128 hidden nodes, and the forward and backward LSTM outputs a 256-dimensional feature vector. The five-channel features are concatenated into a 1280-dimensional vector, which is then input into the attention mechanism layer: a fully connected layer (512 nodes) is dimensionality-reduced and activated by tanh to generate an intermediate matrix. The weight coefficients are then normalized using the Softmax function (tanshinone IIA weight 0.4-0.6, paeoniflorin weight 0.3-0.5). The output layer maps the interaction effect parameters to [0,1] using the Sigmoid function, with a fluctuation threshold set to ±5%. Temperature (80-95℃) and pressure (0.1-0.3MPa) data are input into the model after standardization.

[0040] In existing technologies, the control of the decoction process is mostly based on the concentration of a single component or empirical temperature curves, without analyzing the correlation between the dissolution kinetics of fat-soluble components (such as tanshinone IIA) and water-soluble components (such as paeoniflorin) and temperature and pressure parameters, and lacking real-time modeling of the synergistic effects between components. This embodiment uses Bi-LSTM to capture the long-term dependence characteristics of multi-component dissolution, and dynamically assigns the weight coefficients of tanshinone IIA and paeoniflorin using an attention mechanism, realizing the dynamic evaluation of the synergistic release of flavonoid-saponin components. This solves the technical problem of not being able to control the interaction effects of components in real time in traditional processes, and provides a multi-dimensional scientific basis for the control of the decoction endpoint.

[0041] In another embodiment, in S5, the combined solution is treated through a 0.22 μm microfiltration membrane, with the operating pressure controlled in three stages: 0.25 MPa (5 min) - 0.40 MPa (3 min) - 0.18 MPa (to the endpoint); the microfiltrate undergoes gradient alcohol precipitation: first, the ethanol concentration is adjusted to 60% ± 2% (standing at 4°C for 20 min), then adjusted to 80% ± 1% (standing at 25°C for 15 min), and the supernatant is collected and concentrated under reduced pressure at 45°C ± 1°C and -0.08 MPa to a relative density of 1.25-1.3.

[0042] For example, the microfiltration device used a Millipore plate and frame apparatus with a PVDF membrane material and a pore size of 0.22 μm. The pressure was controlled by a diaphragm pump in three stages: the first stage was 0.25 MPa for stable filtration for 5 minutes, the second stage was 0.40 MPa for increasing the flux for 3 minutes, and the third stage was 0.18 MPa to maintain the flow rate < 50 mL / min. For gradient alcohol precipitation, analytical grade ethanol was added using a peristaltic pump. 60% alcohol precipitation was allowed to stand in a refrigerator at 4°C for 20 minutes, followed by centrifugation at 3000 rpm to separate the precipitate. 80% alcohol precipitation was allowed to stand at 25°C for 15 minutes, and the supernatant was collected and concentrated in a BUCHI rotary evaporator at 45°C and -0.08 MPa. The relative densities were determined using a specific gravity bottle (1.25, 1.28, and 1.30 at 25°C).

[0043] In existing technologies, microfiltration often uses a 0.45μm membrane at constant pressure (0.3MPa), which easily leads to membrane fouling and flux attenuation rates exceeding 40%. Alcohol precipitation typically uses a single 70% ethanol concentration, resulting in incomplete removal of large molecular impurities and a loss rate of over 15% of effective components. This embodiment reduces the formation of filter cake on the membrane surface through three-stage pressure control (first stabilizing, then increasing, then decreasing), keeping the flux attenuation rate below 15%. Gradient alcohol precipitation (60%-80%) precipitates impurities in stages based on differences in component polarity, resulting in an effective component loss rate of <8%. This solves the technical problems of low filtration efficiency and incomplete impurity removal in the post-processing stage, improving the purity and batch consistency of the concentrated extract.

[0044] The following is a description of a specific embodiment.

[0045] Example:

[0046] Coptis chinensis was pulverized to 50 mesh and added to a phosphate buffer solution at pH 6.8 (solid-to-liquid ratio 1:10). Extraction was performed in three stages with controlled temperature: the first stage was at 45℃ for 30 minutes; the second stage was at 65℃ with pulsed ultrasonic extraction (400W power, 15Hz pulse frequency); and the third stage, when the conductivity of the extract increased by ≤1% / 5 minutes for three consecutive tests, the extraction was switched to 85℃ with hot reflux extraction. During the extraction process, data was collected in real-time using a UV-NIR spectrometer and input into a multi-channel convolutional neural network model (the first convolutional channel used a 3×1 convolutional kernel to resolve the 280nm absorbance, the second convolutional channel used a 5×1 hollow convolutional kernel to process the 1050-1100nm near-infrared spectrum, and the third convolutional channel used a 7×1 convolutional kernel combined with the ELU activation function to extract features from the 2200-2400nm range). The extraction was terminated when the overall extraction efficiency index output by the model reached 0.85, yielding the Coptis chinensis extract.

[0047] Calcined dragon bone and calcined oyster shell were pulverized to 110 mesh. In the first stage, they were treated with 1.0% citric acid solution (liquid-to-solid ratio 3:1) at 50℃ for 15 minutes. In the second stage, they were treated with 0.5% citric acid solution (liquid-to-solid ratio 5:1) at 70℃ for 20 minutes. The filtrates were combined to obtain a mineral filtrate. A pH 5.8 acetate buffer solution (liquid-to-solid ratio 8:1) was added to rhizome herbs such as *Codonopsis pilosula*, and the mixture was decocted at 80℃ for 25 minutes to obtain a rhizome decoction. The rhizome decoction, mineral filtrate, and purified water (total) were added to herbs such as jujube seed. The mixture was refluxed at 95℃ with a liquid-to-solid ratio of 12:1. The concentration time-series data of tanshinone IIA, paeoniflorin, and glycyrrhizic acid were acquired in real time by high performance liquid chromatography. The data were then input into a long short-term memory neural network model based on an attention mechanism (a bidirectional LSTM layer was used to extract dissolution rate features, an attention mechanism layer was used to calculate the synergistic weight coefficients of tanshinone IIA and paeoniflorin, and an output layer was used to generate flavonoid-saponin interaction effect parameters). The process was terminated when the interaction effect parameters fluctuated by ≤±5% for three consecutive detection cycles, resulting in a mixed decoction.

[0048] The extract and decoction were combined and processed through a 0.22 μm microfiltration membrane (three-stage pressure: 0.25 MPa × 5 min - 0.40 MPa × 3 min - 0.18 MPa to the endpoint). The filtrate was first added with ethanol to a concentration of 60% (and allowed to stand at 4℃ for 20 minutes), and the precipitate was discarded. The supernatant was then replenished with ethanol to a concentration of 80% (and allowed to stand at 25℃ for 15 minutes). The supernatant was collected and concentrated at 45℃ and a vacuum of -0.08 MPa to a relative density of 1.28. The supernatant was then mixed with lactose to obtain the compound drug.

[0049] Comparative Example 1: Traditional fixed-parameter extraction process

[0050] Coptis chinensis was pulverized to 40 mesh, added to pH 6.8 phosphate buffer (solid-liquid ratio 1:8), and extracted at 60℃ for 2 hours. There was no conductivity monitoring or spectral data input. The endpoint was controlled by offline detection of berberine content (sampling every 30 minutes). Calcined dragon bone and calcined oyster shell were pulverized to 80 mesh, added to purified water along with other medicinal materials (liquid-solid ratio 10:1), and decocted at 90℃ for 1.5 hours. There was no grouped decoction or gradient acid treatment. The concentration of components in the decoction was not monitored in real time. The decoction time was controlled by experience. Microfiltration was performed using a 0.45μm membrane at constant pressure (0.3MPa). The ethanol concentration was adjusted to 70% in one step (standing at 25℃ for 30 minutes) for alcohol precipitation.

[0051] Comparative Example 2: Mineral Drug Processing Technology Without Grouping and Decoction

[0052] Calcined dragon bone and calcined oyster shell were pulverized to 90 mesh, and without gradient acid treatment, they were directly added to purified water (liquid-solid ratio 12:1) along with other medicinal materials such as Codonopsis pilosula, and decocted at 95°C for 2 hours. Root and rhizome medicinal materials and flower medicinal materials were not decocted separately, but were uniformly refluxed at 95°C. Acetate buffer and mineral filtrate were not used. The concentrations of tanshinone IIA and paeoniflorin were not monitored in real time during the decoction process, and the decoction endpoint was controlled at a fixed time (120 minutes).

[0053] Comparative Example 3: Decoction Control Process Based on a Single LSTM Model

[0054] During the decoction stage, the concentration of tanshinone IIA was detected offline using high-performance liquid chromatography (sampled every 10 minutes). The data was input into a unidirectional long short-term memory neural network model. The concentrations of paeoniflorin and glycyrrhizic acid, as well as the time-series temperature and pressure data, were not received simultaneously. The model lacked an attention mechanism layer and could not calculate the component synergistic weighting coefficients. The decoction endpoint was controlled by the tanshinone IIA concentration reaching a preset value (80% of the maximum dissolution). No flavonoid-saponin interaction effect parameters were generated.

[0055] Control group: Traditional water decoction and alcohol precipitation process

[0056] Coptis chinensis was pulverized to 30 mesh, decocted twice with water (1 hour each time at 80°C), and the decoctions were combined. All medicinal materials were decocted together without being grouped. Calcined dragon bone and calcined oyster shell were not pulverized to more than 100 mesh and were not subjected to gradient acid treatment. There was no real-time monitoring of component concentration during the decoction process. The decoction was carried out for 90 minutes according to traditional experience. Microfiltration was performed using a 0.45μm membrane constant pressure filtration. During alcohol precipitation, ethanol was added at once to a concentration of 70%. After standing for 30 minutes, the mixture was filtered and concentrated to a relative density of 1.15. The mixture was then mixed with excipients to obtain the drug.

[0057] Key indicator detection methods and experimental procedures:

[0058] I. Component Dissolution Efficiency Testing

[0059] 1. Berberine dissolution test

[0060] High-performance liquid chromatography (HPLC) was used, referring to General Chapter 0512 of the 2025 edition of the Chinese Pharmacopoeia, Part IV. A C18 column (250 mm × 4.6 mm, 5 μm) was used. The mobile phase was acetonitrile-0.1% phosphoric acid solution (55:45, containing 0.05 mol / L sodium dodecyl sulfate), the flow rate was 1.0 mL / min, the column temperature was 30 ℃, and the detection wavelength was 280 nm. 10 μL of the sample solution was filtered through a 0.45 μm filter and injected. The dissolution rate was calculated using the external standard method with berberine reference standard, using the formula: Dissolution rate (%) = (berberine content in the sample / theoretical berberine content in the medicinal material) × 100%.

[0061] 2. Polysaccharide degree of polymerization detection

[0062] High-performance gel permeation chromatography (HPGPC) was employed using a TSK-GEL G4000PWXL column (300 mm × 7.8 mm), with a mobile phase of 0.1 mol / L sodium chloride solution, a flow rate of 0.6 mL / min, a column temperature of 35 °C, and a differential refractive index detector. After protein removal with water-saturated n-butanol, the sample was filtered through a 0.22 μm filter before injection. A standard curve was plotted using standard dextran, and the weight-average molecular weight (Mw) of the polysaccharides was calculated based on retention time to characterize the degree of polymerization.

[0063] 3. Detection of calcium and magnesium ion dissolution rates

[0064] Inductively coupled plasma mass spectrometry (ICP-MS) was used. Instrument parameters: RF power 1550W, nebulizer gas flow rate 0.8L / min, sampling depth 8mm. Samples were digested with nitric acid-perchloric acid (4:1) and then diluted to volume with 0.5% nitric acid. The concentrations of Ca²⁺ and Mg²⁺ were determined. The dissolution rate was calculated based on the total amount of the corresponding elements in the medicinal material, using the formula: Dissolution rate (%) = (Ion content in solution / Total amount of elements in medicinal material) × 100%.

[0065] II. Process Control and Model Accuracy Testing

[0066] 1. Determination of flavonoid-saponin interaction parameters

[0067] The concentrations of tanshinone IIA and paeoniflorin were monitored in real time using high-performance liquid chromatography-mass spectrometry (HPLC-MS). HPLC conditions were the same as for berberine detection. Mass spectrometry employed electrospray ionization (ESI) in positive ion mode, monitoring m / z. The decoction temperature and pressure data were input into an LSTM-attention mechanism model, and the output was normalized to [0,1] for the interaction effect parameters. The fluctuation range was calculated using the standard deviation of three consecutive detection cycles (5 min apart).

[0068] 2. Confidence scoring of multi-channel CNN models

[0069] Monte Carlo random dropout sampling (dropout rate 0.2) was performed on the optimized fusion features. The comprehensive extraction efficiency index was predicted 10 times consecutively, and the variance σ² was calculated. The confidence score was 1 / (1+σ²). When the score < 0.85, the cumulative dissolution of berberine was detected by offline HPLC, and the endpoint was determined by comparing it with the maximum dissolution in the preliminary experiment.

[0070] III. Post-processing index testing

[0071] 1. Measurement of microfiltration flux attenuation rate

[0072] Record the filtration time and filtrate volume at each stage of microfiltration, calculate the initial flux (average flux of the first stage) and the final flux (steady flux of the third stage), and the attenuation rate = (initial flux - final flux) / initial flux × 100%. Use a Millipore plate and frame microfiltration device to monitor the flow rate and pressure in real time.

[0073] 2. Detection of loss rate of active ingredients from alcohol precipitation

[0074] The contents of berberine, tanshinone IIA, and paeoniflorin in the filtrate and supernatant before alcohol precipitation were determined by HPLC external standard method. The loss rate was calculated as (content of components before alcohol precipitation - content of components in the supernatant) / content of components before alcohol precipitation × 100%. The average value was taken from three parallel determinations.

[0075] IV. Pharmacodynamic Testing in Animal Models

[0076] 1. Establishment of a rat model of rapid atrial fibrillation

[0077] SD rats (200±20g) were anesthetized by intraperitoneal injection of sodium pentobarbital (30mg / kg). An electrode catheter was inserted into the right atrium via the jugular vein, and atrial fibrillation was induced by stimulation at a frequency of 200 beats / min for 30 minutes. The electrocardiogram was recorded using the PowerLab system, and the model was determined to be successful by the disappearance of P waves, the appearance of f waves, and absolute irregularity of RR intervals.

[0078] 2. Drug efficacy evaluation

[0079] Model rats were randomly divided into groups and administered the drugs from the example group (containing 1.5 g / kg of crude drug) and the comparative / control group by gavage. The saline group served as a blank control. Heart rate was recorded at 30 minutes, 1 hour, and 2 hours after administration. A heart rate reduction of ≥20% compared to the model group that was maintained for more than 1 hour was considered effective. The effectiveness rate was calculated as (number of effective animals / number of animals in each group) × 100%.

[0080] V. Comparison of component dissolution efficiency and preparation efficiency data

[0081] 1. Berberine dissolution rate and degradation rate

[0082] Example: Dissolution rate 92.3%, degradation rate 3.1%

[0083] Comparative Example 1 (conventional fixed parameters): Dissolution rate 65.2%, degradation rate 12.0%

[0084] Control group (traditional water decoction): dissolution rate 54.8%, degradation rate 18.5%.

[0085] Data shows that the embodiment, through three-stage temperature control and real-time spectral monitoring, improved the dissolution rate by 27.1% and reduced the degradation rate by 74.2% compared with Comparative Example 1, thus solving the contradiction of "insufficient dissolution and component degradation coexisting" in traditional constant temperature extraction, and significantly improving the preparation efficiency.

[0086] 2. Polysaccharide degree of polymerization retention rate

[0087] Example: Retention rate 88.0% (weight-average molecular weight Mw = 12500 Da)

[0088] Control group: Retention rate 53.0% (Mw=7800Da)

[0089] The example demonstrates that by combining a low-temperature protection phase (constant temperature of 45°C) with pulsed ultrasonic cell disruption, the polysaccharide polymerization degree retention rate is increased by 35.0% compared to the control group, avoiding polysaccharide degradation caused by traditional high-temperature decoction and ensuring the structural stability of the active ingredients.

[0090] 3. Dissolution rate of calcium and magnesium ions in mineral drugs

[0091] Example: Ca²⁺ dissolution rate 18.7%, Mg²⁺ dissolution rate 15.3%.

[0092] Comparative Example 2 (ungrouped decoction): Ca²⁺ dissolution rate 4.2%, Mg²⁺ dissolution rate 3.8%.

[0093] Control group: Ca²⁺ dissolution rate 2.7%, Mg²⁺ dissolution rate 2.1%.

[0094] The example, through gradient acid treatment (1.0%-0.5% citric acid) and 110-mesh ultrafine grinding, increased the calcium and magnesium ion dissolution rate by 4.4 times compared with Comparative Example 2, breaking through the physical barrier of mineral lattice in the traditional co-decoction process and significantly improving the dissolution efficiency of inorganic components.

[0095] 4. Synergistic dissolution of tanshinone IIA and paeoniflorin

[0096] Example: Tanshinone IIA dissolved at 2.87 mg / g of medicinal material, and paeoniflorin dissolved at 8.54 mg / g of medicinal material, with a dissolution ratio of 1:2.98 (close to the golden synergistic ratio of 1:3).

[0097] Comparative Example 3 (single LSTM control): Tanshinone IIA dissolution amount 2.12 mg / g, paeoniflorin dissolution amount 6.31 mg / g, ratio 1:2.98 (but the interaction effect parameter fluctuated by ±10.2%).

[0098] Control group: Tanshinone IIA dissolution amount 1.75 mg / g, paeoniflorin dissolution amount 5.12 mg / g, ratio 1:2.92 (significant component degradation).

[0099] The example uses an attention mechanism LSTM model to adjust temperature and pressure parameters in real time, which increases the synergistic dissolution of fat-soluble and water-soluble components by 64.0% and 66.8% respectively compared with the control group, and the interaction effect parameter fluctuation is ≤±4.2%, achieving precise matching of component release.

[0100] VI. Process control accuracy and quality stability data

[0101] 1. Error control during extraction and decoction endpoints

[0102] Example: The prediction error of the comprehensive extraction efficiency index was ±3.0%, and the fluctuation of the flavonoid-saponin interaction parameter was ±4.2%.

[0103] Comparative Example 2 (Single Spectral Model): Extraction efficiency prediction error ±15.0%, endpoint false positive rate 30.0%.

[0104] Control group: The endpoint was controlled empirically, with batch-to-batch component content fluctuations of ±22.0%.

[0105] The multi-channel CNN and LSTM-attention mechanism model in the embodiment improves the process control accuracy by 80.0% compared with Comparative Example 2, realizes closed-loop control with dynamic monitoring and real-time adjustment, and ensures the uniformity of components between batches.

[0106] 2. Adaptability data of different batches of medicinal materials

[0107] Example: The solubility of berberine in medicinal materials from Sichuan / Henan fluctuated by ±5.8%, and the overall extraction efficiency index fluctuated by ±5.8%.

[0108] Control group: Differences in origin led to dissolution rate fluctuations of ±22.0% and efficiency index fluctuations of ±22.0%.

[0109] The example uses the batch feature code of medicinal materials (origin, harvest season, storage time) and the gated cycle unit to generate dynamic weight coefficients, which improves the process's adaptability to the differences in medicinal materials by 73.6% and significantly reduces the quality differences between batches.

[0110] 3. Post-processing efficiency and purity

[0111] Microfiltration flux attenuation rate: Example 15.0%, control group 40.0%.

[0112] Loss rate of active ingredients by alcohol precipitation: 8.0% in the example, 15.0% in the control group.

[0113] Impurity content of concentrated extract: Example 5.0% (relative density 1.28), control group 12.0% (relative density 1.15)

[0114] The three-stage pressure regulation and gradient alcohol precipitation process improved the microfiltration flux maintenance efficiency by 62.5%, the retention rate of active ingredients by 46.7%, and the purity of the extract by a significant margin, laying the foundation for the stability of the formulation quality.

[0115] III. Finished Product Quality and Efficacy Stability Data

[0116] 1. Efficacy of heart rate control in a rat atrial fibrillation model

[0117] Example: 91.0% (heart rate decreased by 28.5 ± 3.2 beats / min 2 hours after administration)

[0118] Control group: 65.0% (heart rate decreased by 15.2 ± 5.8 beats / min)

[0119] Comparative Example 1: 72.0% (heart rate decreased by 18.7 ± 4.5 beats / min)

[0120] The drug in this example showed that due to the sufficient dissolution of its components and excellent synergistic effect, the efficacy rate was increased by 26.0% compared to the control group, and the heart rate control range and stability were significantly better than those of the traditional process group.

[0121] 2. Batch-to-batch efficacy variation coefficient

[0122] Example: Efficacy fluctuation ±8.0%

[0123] Control group: efficacy fluctuated by ±25.0%.

[0124] The example demonstrates that through intelligent control of the entire process, the batch-to-batch variation in the efficacy of the finished drug is reduced by 68.0% compared to traditional processes, directly reflecting an improvement in quality stability.

[0125] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A preparation method of a traditional Chinese medicine compound for treating rapid atrial fibrillation, characterized in that, include: S1: Take Coptis chinensis and extract it with phosphate buffer; S2: Optical data of the extract is acquired in real time using an ultraviolet-near-infrared spectrometer and input into a pre-trained multi-channel convolutional neural network model. In the multi-channel convolutional neural network model, the first convolutional channel analyzes absorbance data at a wavelength of 280 nm, the second convolutional channel processes near-infrared spectral data at 1050-1100 nm, and the third convolutional channel extracts spectral features at 2200-2400 nm. When the comprehensive extraction efficiency index output by the model reaches a predetermined threshold, the extraction is terminated, and the Coptis chinensis extract is obtained. S3: Take Codonopsis pilosula, Polygonatum sibiricum, Rehmannia glutinosa, Paeonia lactiflora, Morus alba, Ziziphus jujuba var. spinosa, Schisandra chinensis, Triticum aestivum, Ophiopogon japonicus, Ligusticum chuanxiong, Salvia miltiorrhiza, Rehmannia glutinosa, Paeonia lactiflora, Glycyrrhiza uralensis, Citrus medica, calcined dragon bone, calcined oyster shell, and rose petals for decoction: S4: Real-time concentration time-series data of tanshinone IIA, paeoniflorin, and glycyrrhizic acid in the decoction were acquired using high-performance liquid chromatography (HPLC) and input into a long short-term memory neural network model based on an attention mechanism. In the long short-term memory neural network model, the bidirectional long short-term memory layer extracted the dissolution rate characteristics of each component, the attention mechanism layer calculated the synergistic weighting coefficients of tanshinone IIA and paeoniflorin, and the output layer generated flavonoid-saponin interaction effect parameters. When the interaction effect parameters fluctuated within the range of ≤±5% for three consecutive detection cycles, the decoction was terminated to obtain a mixed decoction. S5: Combine the Coptis chinensis extract and the mixed decoction, treat with a microfiltration membrane and then perform alcohol precipitation. Collect the alcohol precipitation supernatant, concentrate it, and obtain concentrated extract. S6: Prepare a traditional Chinese medicine compound by mixing a concentrated extract with a pharmaceutically acceptable carrier; In S2, the first convolutional channel is configured to use a 3×1 convolutional kernel to process the 280nm absorbance time-series data and output the first feature map; the second convolutional channel is configured to use a 5×1 dilated convolutional kernel to extract features from the 1050-1100nm near-infrared spectrum and output the second feature map, with a dilation rate of 2 for the dilated convolution; the third convolutional channel is configured to use a 7×1 convolutional kernel combined with an exponential linear unit activation function to capture features from the 2200-2400nm spectrum and output the third feature map; the first, second, and third feature maps are then combined in a single pass. The channel dimensions are concatenated to form an initial fusion feature. A 1×1 convolution kernel operation is applied to the initial fusion feature to generate a channel attention weight matrix. A three-dimensional convolution kernel operation is applied to the initial fusion feature to generate a spatial-spectral correlation feature. The channel attention weight matrix is ​​multiplied element-wise with the spatial-spectral correlation feature to output an optimized fusion feature. The optimized fusion feature is input into a fully connected layer to generate a multidimensional feature vector containing parameters such as berberine dissolution rate, polysaccharide polymerization rate, and cell wall breakage rate. The comprehensive extraction efficiency index is calculated based on the multidimensional feature vector. In S4, the input layer of the long short-term memory neural network model is configured to simultaneously receive five independent time-series data channels: the first channel inputs the time-series data of tanshinone IIA concentration, the second channel inputs the time-series data of paeoniflorin concentration, the third channel inputs the time-series data of glycyrrhizic acid concentration, the fourth channel inputs the time-series data of the decoction system temperature, and the fifth channel inputs the time-series data of the decoction system pressure. The data from the five channels are respectively input to the corresponding processing units in the bidirectional long short-term memory layer, and respectively output the tanshinone IIA dissolution feature vector, the paeoniflorin dissolution feature vector, the glycyrrhizic acid dynamic feature vector, the temperature dynamic feature vector, and the pressure dynamic feature vector. The dissolution feature vectors of tanshinone IIA, paeoniflorin, glycyrrhizic acid, temperature, and pressure are concatenated. The concatenated features are input into the attention mechanism layer, processed by a fully connected layer and a hyperbolic tangent activation function to generate an intermediate feature matrix, and then normalized by the Softmax function to output the collaborative weight coefficients. The collaborative weight coefficients are then fused with the dissolution feature vectors of tanshinone IIA, paeoniflorin, and glycyrrhizic acid. The fused features are processed by the output layer to generate flavonoid-saponin interaction effect parameters normalized to the 0-1 range.

2. The method of claim 1, wherein, In S1, Coptis chinensis is pulverized to 40-60 mesh and added to phosphate buffer solution with pH 6.8, with a material-to-liquid ratio of 1:8-1:

12. The extraction process is carried out in three stages with temperature control. The first stage is extraction at a constant temperature of 45℃ for 30 minutes. The second stage is extraction by pulsed ultrasound at 65℃. The third stage is extraction by switching to hot reflux at 85℃ when the conductivity of the extract increases by ≤1% / 5 minutes for three consecutive tests.

3. The method of claim 1, wherein, The berberine dissolution parameter was standardized by maximum value to generate a standard dissolution parameter; the polysaccharide polymerization degree parameter was logarithmically transformed to generate a standard polymerization degree parameter; and the cell wall breakage rate parameter was transformed by the Sigmoid function to generate a standard cell wall breakage rate parameter. Input the standard dissolution rate parameter, standard polymerization rate parameter, and standard cell breakage rate parameter into the fully connected layer, and output the initial weight vector; The initial weight vector is concatenated with the batch feature code of medicinal materials. The batch feature code of medicinal materials includes the place of origin code, the harvest season code, and the storage duration code. Dynamic weighting coefficients are generated through a gated circulation unit, where the weighting coefficients for berberine dissolution parameter range from 0.55 to 0.65, polysaccharide polymerization parameter range from 0.25 to 0.35, and cell wall breakage rate parameter range from 0.08 to 0.

12. The comprehensive extraction efficiency index is calculated using the following formula: Comprehensive extraction efficiency index = dynamic weighting coefficient α × standard dissolution parameter + dynamic weighting coefficient β × standard polymerization parameter + dynamic weighting coefficient γ × standard cell wall breakage rate parameter, where α, β, and γ are dynamic weighting coefficients.

4. The method of claim 1, wherein, Also includes: The comprehensive extraction efficiency index of the multi-channel convolutional neural network model output is acquired in real time, and the confidence score of the multi-channel convolutional neural network model output is calculated. The confidence score is generated by performing Monte Carlo random dropout sampling on the optimized fusion features and is based on the inverse of the variance of 10 consecutive prediction results. When the confidence score is lower than the 0.85 threshold, the extraction is no longer determined based on the comprehensive extraction efficiency index, but is terminated when the cumulative dissolution of berberine is not less than the maximum dissolution.

5. The method of claim 1, wherein, In S3, calcined dragon bone and calcined oyster shell are pulverized to 100-120 mesh. In the first stage, a 1.0% citric acid solution is used at a liquid-to-solid ratio of 3:1 and treated at 50°C for 15 minutes. In the second stage, a 0.5% citric acid solution is used at a liquid-to-solid ratio of 5:1 and treated at 70°C for 20 minutes. The filtrates from the first and second stages are combined to obtain mineral filtrate. Add a pH 5.8 acetate buffer solution to Codonopsis pilosula, Polygonatum sibiricum, Rehmannia glutinosa (processed), Rehmannia glutinosa (raw), Paeonia lactiflora (red), Paeonia lactiflora (white), Ligusticum chuanxiong, and Salvia miltiorrhiza, and decoct at 80°C for 25 minutes to obtain the rhizome decoction. Add the root and stem decoction, mineral filtrate, and purified water to jujube seed, schisandra fruit, wheat, ophiopogon japonicus, licorice, citron, mulberry, and rose petals, control the total liquid-to-solid ratio at 12:1, and decoct under reflux at 95℃.

6. The method as described in claim 1, characterized in that, In S5, the Coptis chinensis extract and the mixed decoction are combined and then microfiltered through a microfiltration membrane with a pore size of 0.22 μm. The microfiltration operation pressure is controlled in three stages: the first stage is 0.25 MPa for 5 minutes, the second stage is increased to 0.40 MPa for 3 minutes, and the third stage is reduced to 0.18 MPa and maintained until the end of filtration. The obtained microfiltrate was subjected to gradient alcohol precipitation: first, ethanol was added to make the alcohol concentration reach 60%±2%, and after standing at 4℃ for 20 minutes, the precipitate was separated and discarded; then, ethanol was added to the supernatant to make the alcohol concentration reach 80%±1%, and after standing at 25℃ for 15 minutes, the alcohol precipitation supernatant was collected; the alcohol precipitation supernatant was concentrated under reduced pressure at 45℃±1℃ and a vacuum degree of -0.08MPa, and the concentration was terminated when the relative density of the concentrated extract was 1.25-1.3.

Images