Machine learning-based intelligent deployment method for hemostatic composition ingredients

Abstract

This invention discloses a machine learning-based intelligent formulation method for hemostatic compositions, relating to the fields of biomedical materials and medical devices. This invention incorporates real-time information on the three-dimensional geometry of the wound and the surgical environment into the decision-making process. A pre-trained machine learning model outputs the component ratios for the target wound, and programmable microfluidics enables continuous, loopable dynamic formulation and immediate application, forming a closed loop of perception-inference-execution-feedback. Compared to solutions relying on fixed formulations or offline optimization, this invention can adaptively adjust the ratios of hydrophilic precursors, crosslinking agents, and optional hydrophobic phases according to local morphology and flow rate on complex wound surfaces with curved surfaces, deep cavities, roughness, and fluid erosion. This improves the uniformity of wet surface spreading and gelation, and reduces the risk of leakage and localized failure.

Description

Technical Field

[0001] This invention relates to the field of biomedical materials and medical device technology, and in particular to a method for intelligent formulation of hemostatic composition components based on machine learning. Background Technology

[0002] In clinical hemostasis, the moist and complex wound surfaces (curvature changes, surface roughness, and tissue movement) combined with blood flow shear on the order of 10²-10³s⁻¹ often make it difficult for tissue adhesive to achieve uniform gelation and continuous sealing on curved surfaces and bleeding interfaces, resulting in leakage and local failure. Although existing high-performance hemostatic adhesives can achieve burst pressures of >200-350 mmHg and second-level curing under standardized conditions, their adaptability to real-time geometric and fluid disturbances is limited.

[0003] Existing intelligent research technologies mostly employ historical data-driven formulation optimization, such as Bayesian optimization / self-driven laboratories. These technologies primarily search for optimal components / ratios in offline environments based on predetermined indicators (bursting pressure, adhesion energy, gel time, etc.), lacking a dynamic formulation mechanism based on real-time intraoperative perception. Therefore, achieving an intelligent formulation method for hemostatic compositions that can achieve a closed loop of perception-decision-ratio-application under irregular and time-varying wound conditions remains a pressing technical problem that needs to be solved. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] This invention provides a machine learning-based intelligent formulation method for hemostatic compositions, which solves the problems of uneven gel distribution and leakage caused by irregular and time-varying wounds, and the difficulty of real-time adaptation due to offline optimization.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a machine learning-based intelligent formulation method for hemostatic composition components, comprising:

[0008] Step S1: Obtain morphological data and environmental data of the target wound. The morphological data includes the three-dimensional geometric features of the wound, and the environmental data includes at least one external condition related to hemostasis.

[0009] Step S2: Input the morphological data and environmental data into a pre-trained machine learning model and output the component formulation parameters of the hemostatic composition;

[0010] Step S3: According to the ingredient mixing parameters, control the preparation device to dynamically adjust the ratio of at least two components in the hemostatic composition, and apply the hemostatic composition to the target wound.

[0011] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the machine learning model is a model for predicting the proportion of components, and the model includes at least one of Gaussian process regression, Bayesian optimization surrogate model and / or neural network.

[0012] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the morphological data is acquired through non-contact three-dimensional imaging, and the three-dimensional data of the wound surface is denoised, detrended, and missing data is completed by a preprocessing module. Morphological features such as principal curvature distribution, depth field, and surface roughness are extracted and then input into the machine learning model.

[0013] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the machine learning model is trained based on historical wound data and corresponding hemostatic effect data. The loss function during training simultaneously characterizes at least two of the hemostatic onset time, leakage rate and interface sealing strength, and imposes constraints on biocompatibility and pH range.

[0014] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the environmental data includes blood flow rate and temperature in the target area, and is updated at a frequency no less than a preset refresh frequency during the application of the hemostatic composition, and the formulation parameters are adjusted accordingly.

[0015] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the method constitutes a closed-loop control: a feedback signal is acquired through a sensing unit at the wound site, the feedback signal including at least a bleeding status indicator or an interface sealing strength indicator; the feedback signal and the latest morphological / environmental data are input into the machine learning model to update the formulation parameters.

[0016] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the formulation device is a microfluidic device, which includes two or more independently meterable feeding channels and a mixing channel, and achieves real-time changes in the component ratio during the application process through valve control and pump control.

[0017] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the hemostatic composition comprises a hydrophilic polymer precursor and a crosslinking agent;

[0018] The hydrophilic polymer precursor includes at least one of alginate, chitosan derivative, gelatin derivative or polyethylene glycol derivative;

[0019] The crosslinking agent includes at least one of polyaldehyde, ionic crosslinking agent or enzyme-catalyzed crosslinking agent;

[0020] The component blending parameters are used to determine the target mass ratio of each component.

[0021] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the hydrophilic polymer precursor is specifically oxidized alginate, the crosslinking agent is specifically carboxymethyl chitosan, and the method dynamically adjusts the mass ratio of the two during the application process according to the output of the machine learning model.

[0022] As a preferred embodiment of the intelligent formulation method for hemostatic composition based on machine learning described in this invention, the following steps are performed: digital twin simulation verification is performed before the application of the steps: the wound morphology data and / or environmental data obtained in step S1 and the candidate component formulation parameters are input into a simulation model containing geometric, fluid and / or mass transfer coupling, and the gelation behavior and sealing performance index under different component formulation parameters are evaluated; and the parameters and / or hyperparameters of the pre-trained machine learning model are updated based on the evaluation results to generate corrected component formulation parameters.

[0023] The beneficial effects of this invention are as follows: This invention incorporates real-time information on the three-dimensional geometry of the wound and the surgical environment into the decision-making process. It outputs component ratios tailored to the target wound through a pre-trained machine learning model and utilizes programmable microfluidics to achieve continuous, loopable dynamic mixing and immediate application, forming a closed loop of perception-inference-execution-feedback. Compared to solutions relying on fixed formulations or offline optimization, this invention can adaptively adjust the ratios of hydrophilic precursors, crosslinking agents, and optional hydrophobic phases according to local morphology and flow velocity on complex wound surfaces with curved surfaces, deep cavities, roughness, and fluid erosion. This improves the uniformity of wet surface spreading and gelation, and reduces the risk of leakage and localized failure.

[0024] This invention combines constraints on multiple indicators such as hemostatic onset time, leakage rate, and sealing strength. The method ensures rapid onset of action while also taking into account interface density and subsequent tissue compatibility. Combined with digital twin simulation, the impact of different parameters on sealing performance can be quickly assessed before surgery / application, and the model can be calibrated to improve transferability and robustness under heterogeneous wounds and individual differences.

[0025] This invention reduces reliance on human experience and repeated adjustments, while balancing immediacy and repeatability. It is suitable for highly dynamic scenarios such as emergency surgery, trauma treatment, and bedside care, significantly improving the consistency and reliability of hemostasis for irregular and time-varying wounds. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0027] Figure 1 This is a flowchart illustrating the intelligent formulation method for hemostatic composition components based on machine learning in the embodiments. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0031] This application proposes a machine learning-based intelligent formulation method for hemostatic compositions, combining... Figure 1 As shown, the method includes:

[0032] Step S1: Acquire morphological and environmental data of the target wound. Morphological data includes the three-dimensional geometric features of the wound, and environmental data includes at least one external condition related to hemostasis. In this embodiment, three-dimensional geometric features refer to the point cloud or equivalent height field of the wound surface generated based on non-contact three-dimensional imaging, and environmental data refers to external conditions directly related to the hemostasis process and capable of being acquired online, including but not limited to local blood flow velocity and wound surface temperature. By default, three-dimensional morphological and environmental data are acquired synchronously at a video-level frequency. The frame rate on the morphological side should preferably be no less than 10 frames per second, and the refresh cycle on the environmental side should preferably be no more than 200 milliseconds. The allowable error for the alignment of the timestamps of the two should not exceed one sampling cycle. Morphological data is acquired through the structured light, time-of-flight, or binocular ranging channel of the inherent device, and environmental data is acquired by a sensing unit arranged near the surgical area. Optionally, for short-term missing morphological data caused by occlusion within the field of view, a time-holding strategy for the most recent valid data is adopted, with a holding time not exceeding three seconds; if the environmental data is temporarily abnormal, it is replaced by the most recent stable value and smoothly restored within the next one to three cycles. If necessary, the coordinates of the shape and the environment are aligned using the same reference system for extrinsic parameter calibration, and unit consistency is performed before entering the preprocessing module.

[0033] Step S2 involves inputting morphological and environmental data into a pre-trained machine learning model, which outputs the component mixing parameters of the hemostatic composition. Specifically, in this embodiment, the component mixing parameters are dimensionless proportional vectors, corresponding to the target mass ratio or equivalent volume fraction of at least two components to be mixed. The inference output is normalized and mapped to the instantaneous setpoint of the preparation device. The default update cycle between inference and distribution is no more than 100 milliseconds, and the end-to-end latency should be less than twice the update cycle. To suppress severe oscillations, the output undergoes first-order smoothing before distribution, with the smoothing time constant set between half a second and one second. For example, when the proportional vector given by the model approaches the feasible region boundary, boundary pruning and re-normalization are performed to ensure that the sum of the proportions remains one and that each component is not less than a preset lower limit. If the inference result undergoes abrupt changes in several consecutive cycles, the output is distributed through a linear transition between the previous stable output and the current output, with a transition time of no less than three update cycles.

[0034] Step S3: According to the component mixing parameters, the preparation device dynamically adjusts the ratio of at least two components in the hemostatic composition and applies the hemostatic composition to the target wound. For example, the preparation device includes two or more independently meterable supply channels and a mixing channel. The upper control system links the flow rates of each channel according to the target ratio and completes thorough mixing in the mixing channel. The default total flow rate is set within the range that satisfies applyability, the channel step resolution meets the minimum step size requirement for proportional adjustment, and the switching and stabilization time should be controlled within 500 milliseconds. Optionally, to avoid air bubble introduction and channel stagnation, dead space residue is removed before application through pre-filling and venting processes, and flushing of at least one channel volume is performed after switching the ratio. In boundary cases, when abnormal back pressure or blockage is detected, the system reduces the total flow rate proportionally and maintains the relative ratio of each channel unchanged until it returns to normal.

[0035] In one embodiment, the machine learning model is a model for component proportion prediction, including at least one of Gaussian process regression, Bayesian optimized surrogate model, and / or neural network. Similarly, the training data consists of historical wound samples that have undergone manual quality inspection, containing morphological and environmental inputs and corresponding component proportion and effect annotations. The dataset is recommended to be divided into training, validation, and test subsets with a ratio of not less than 8:1:1. When using a Gaussian process, the kernel function should preferably be an anisotropic length scale and automatic correlation estimation should be enabled. When using a Bayesian optimized surrogate model, the acquisition strategy should preferably use a greedy-exploration balance method such as improved expectation value. When using a neural network, a two- to four-layer fully connected structure or a lightweight convolutional structure is used by default, with the number of hidden units between 64 and 256. The early stopping strategy is triggered based on the validation set metric without boosting rounds. The single latency on the inference side on the embedded platform should preferably not exceed 50 milliseconds to meet online control requirements.

[0036] In one embodiment, morphological data is acquired through non-contact 3D imaging, and the principal curvature distribution, depth field, and surface roughness index are extracted by a preprocessing module before being input into a machine learning model.

[0037] The calculation steps for principal curvature distribution and roughness index include:

[0038] Step C1: A Monge coordinate plane is established locally on the wound surface using non-contact 3D point cloud, and the data is resampled into a regular height field matrix. Median filtering and small hole interpolation are used to suppress outliers and missing data.

[0039] Step C2: Perform a two-dimensional quadratic polynomial or Savitzky-Golay fitting within the sliding window to obtain the first and second partial derivatives of the height field at the current position.

[0040] Step C3: At each location, calculate the two principal curvature pairs together with the partial derivatives obtained in step C2, and form a spatial distribution:

[0041] ,

[0042] in, Indicates position The two principal curvatures, the subscripts here and These represent the larger and smaller principal curvatures, respectively. For local planar coordinates, For height field pair The first-order partial derivative, For the corresponding second-order partial derivative;

[0043] Step C4: For the detrended height field, calculate the root mean square roughness within the analysis window and use it as the main roughness index.

[0044] ,

[0045] in, The root mean square roughness is... The number of grid rows and columns within the window. For row and column indexes, This is the height of the grid. The arithmetic mean of the heights within the window; a multi-scale strategy can be used to first separate the long-wavelength surfaces and then perform statistical analysis to obtain the values ​​at different scales. ;

[0046] Step C5, will The principal curvature distribution is formed by quantiles or histograms, and... Optional statistics are concatenated as features under the same scale label and input into the machine learning model;

[0047] Step C6: Set a mask for the blood clot, instrument obstruction and cutting edge area, and use window shrinking or weighting when the window crosses the edge to suppress boundary deviation;

[0048] Specifically, stable first and second derivatives are used as a bridge to transform the 3D topography into a principal curvature field with physical meaning, and the surface micro-undulations are quantified in the form of root mean square (RMS). The principal curvature expression comes from the classical differential geometry framework, which is suitable for combination with local polynomial fitting, with controllable computational burden and easy online deployment. Roughness adopts RMS statistics, which has strong resistance to random outliers and can be linked with multi-scale decomposition to capture macro-undulations and micro-textures respectively. Distributed and quantized processing makes the features more robust to local anomalies and naturally matches the input form of machine learning. Masking and edge strategies reduce artifacts caused by occlusion and truncation, improving applicability in real wound environments. In this embodiment, the spatial scale of the sliding window is set according to the typical feature size of the wound, with a default value of 2 to 5 mm. The grid spacing for height field resampling is 0.2 to 0.5 mm. The order of local polynomial fitting is 2, the window length is odd and matches the grid spacing. The number of bins for distributed statistics is set between 8 and 16. Optionally, abnormally high and low points are identified and removed using the median deviation criterion, with the threshold for identification coordinated with the data distribution width. For masked regions, they are marked as missing in the feature output and excluded from prediction on the model side using an explicit missing channel or zero-filling masking. If necessary, for cases where the window crosses the edge of the wound, a window reduction or weighting strategy is used to reduce edge errors, and range restrictions are imposed on the statistics of principal curvature and roughness at the output to prevent numerical instability.

[0049] In one embodiment, the machine learning model is trained using historical wound data and corresponding hemostasis effect data as samples. The optimized loss function characterizes at least two of the following: hemostasis onset time, leakage rate, and interfacial sealing strength, and is subject to constraints on biocompatibility and pH range.

[0050] The loss function and weights are defined as follows:

[0051] Step D1: Construct a training set using historical wound samples, and unify the dimensions and intervals for time-on-effectiveness, leakage rate, and interface sealing strength; the model output is a component adjustment parameter vector, which serves as the input for downstream evaluation.

[0052] Step D2: Input the component formulation parameters into the performance evaluation module to obtain differentiable predictions of onset time, leakage rate, sealing strength, pH and biocompatibility score, which are used to construct training signals and constraint terms.

[0053] Step D3: On a single sample, include the deviations of the three performance metrics along with the two types of constraints in the total loss.

[0054] ,

[0055] in, For the total loss of a single sample, The non-negative weights of the three performance losses are: For the threshold is Huber function, These are the predicted onset time, leakage rate, and sealing strength obtained from the evaluation module, respectively. These are the corresponding labeled values ​​(or expected target values). These are the penalty coefficients for biocompatibility constraints, pH constraints, and regularization terms, respectively. As a biocompatibility penalty, among which For predicting biocompatibility scores, As the lower limit of the score, The penalty is based on the pH range, where For pH prediction, Within the allowable pH range, For component adjustment parameter vector The 2-norm regularity, where For the first The proportion of each component, The number of component types;

[0056] Step D4: To balance the scale differences across multiple tasks, learnable weights based on uncertainty are introduced.

[0057] ,

[0058] in, The uncertainty scaling parameters for the three tasks are used as learnable variables and optimized together with the model parameters. This allows for adaptive adjustment based on data, eliminating the need for manual parameter tuning;

[0059] Step D5: Embed pH and biocompatibility constraints as penalties into the total loss of step D3, and apply them to the inference end. Perform normalized projection and lower bound truncation to keep the proportional vector within the feasible region;

[0060] Step D6, on small batches Calculate the mean or weighted sum, and use an adaptive learning rate optimizer to jointly update the model parameters. Relevant parameters are applied until the validation set metrics converge;

[0061] Specifically, in multi-indicator scenarios, direct linear weighting can provide a clear target structure, but different dimensions and noise levels often make weight selection sensitive. An uncertainty-based adaptive weighting scheme incorporates the task scale into learnable parameters, automatically balancing time, leakage, and intensity in the total loss during training, reducing manual parameter tuning costs. Constraints employ differentiable penalty intervals and thresholds, facilitating gradient propagation and providing continuous penalties for out-of-bounds situations; this projection method with harder constraints is more stable in terms of optimization. L2 regularization is introduced for component allocation parameters, suppressing excessive polarization and providing a smooth inductive bias for the surrogate model in sparse sample regions. Furthermore, a unified scale for multiple indicators is achieved by performing interval normalization or standardization on each evaluation quantity. The segmentation position of the robust loss is set according to the corresponding measurement noise and clinically distinguishable threshold, typically on the same order of magnitude as the indicator resolution. Penalty weights are set to the same order of magnitude initially and adaptively updated during training. Feasible region projection is performed progressively at the inference end, including non-negative pruning and sum normalization. The mini-batch size is set between 32 and 128, and the optimizer uses an adaptive learning rate family, with the learning rate automatically annealing according to the validation metric. Optionally, training uses five-fold cross-validation to evaluate generalization ability; early stopping is triggered when the validation set metric shows no improvement over several rounds, and the optimal weights are retained. If a batch has missing annotations, the supervision signal for that sample is skipped and only its constraint-related terms are used, or the sample is removed from the training set.

[0062] In one embodiment, environmental data includes blood flow rate and temperature in the target area, and is updated at a frequency no less than a preset refresh rate during the application of the hemostatic composition, thereby performing rolling corrections on the component preparation parameters. Optionally, the default refresh rate on the environmental side is set to ten times per second, the rolling correction time window is 1 to 2 seconds, and weighted updates are used to improve noise resistance. When only a single environmental quantity can be acquired, a fine-tuning of the proportion is triggered according to the rate of change of that quantity, with the fine-tuning amplitude limited to a single adjustment not exceeding a certain percentage of the total proportion. If unreliable environmental data is detected for multiple consecutive cycles, the previous stable proportion is maintained and the total flow rate is reduced until the data is recovered. If necessary, the environmental data can be aligned with the nearest neighbor of the morphological data using timestamps to ensure input consistency.

[0063] In one embodiment, the method constitutes closed-loop control: feedback signals are acquired through sensing units at the wound site, including at least a bleeding status indicator or an interface seal strength indicator; the feedback signals, along with the latest morphological / environmental data, are input into a machine learning model to update the component blending parameters; in this embodiment, the feedback signals originate from deployed sensing units and available monitoring signals, the bleeding status indicator can be estimated based on visible flow signs in the wound area or flow readings from local acquisition devices, and the interface seal strength indicator can be estimated based on local pressure changes, leakage signs, or adhesion stability indices. The update cycle is synchronized with inference, defaulting to once every 100 milliseconds; to avoid closed-loop oscillation, continuous adjustments in opposite directions are suppressed and a minimum hold time is enabled. If the feedback noise level exceeds a preset threshold, the influence of that channel on the final output is temporarily downweighted, and exponential recovery is performed over multiple cycles.

[0064] In one embodiment, the preparation device is a microfluidic device, which includes two or more independently meterable supply channels and a mixing channel. The component ratio is changed in real time during application through valve control and pump control. Specifically, the minimum resolvable volume and control step of the supply channels meet the proportional resolution requirements, and the flow rate calibration between channels is completed and stored in the device memory before application. The default total flow rate range covers common spraying or spot coating conditions. The residence time in the mixing channel is not less than the time required to achieve uniform mixing, and channel switching jitter is eliminated by synchronous triggering from the upper control unit. Similarly, to reduce cross-contamination, flushing of at least one to two times the dead volume of the channels is performed after a proportional transition. If an abnormality is detected in a supply channel, the system disables it and re-normalizes the proportions according to the remaining channels to maintain a constant total flow rate.

[0065] In one embodiment, the hemostatic composition comprises a hydrophilic polymer precursor and a crosslinking agent;

[0066] The hydrophilic polymer precursor includes at least one of alginate, chitosan derivative, gelatin derivative or polyethylene glycol derivative;

[0067] Crosslinking agents include at least one of polyaldehydes, ionic crosslinking agents, or enzyme-catalyzed crosslinking agents;

[0068] The component formulation parameters are used to determine the target mass ratio of each component. Further, the hydrophilic polymer precursor and crosslinking agent are separately prepared into stock solutions under aseptic conditions, with concentrations selected within a compromise range between applyability and rapid gelation. The target mass ratio is directly determined by the model output and is instantaneously set proportionally to each channel upon entering the formulation device. To avoid premature reaction, the crosslinking agent channel merges with the hydrophilic polymer precursor near the nozzle outlet. If necessary, a short-distance passive mixing section can be set before entering the nozzle outlet to improve mixing uniformity while ensuring that pressure drop and back pressure are within safe ranges. If an abnormal increase in nozzle outlet resistance is detected, the total flow rate is briefly reduced while maintaining the ratio to prevent component mismatch.

[0069] In one embodiment, the hydrophilic polymer precursor is specifically oxidized alginate, and the crosslinking agent is specifically carboxymethyl chitosan. The method dynamically adjusts the mass ratio of the two during application based on the output of a machine learning model. Optionally, a hydrophobic phase is further introduced to improve the gelation stability of the wet surface. For example, the initial mass ratio of oxidized alginate to carboxymethyl chitosan is set near a neutral point of 1:1, and during online operation, it is finely adjusted bidirectionally within permissible ranges according to changes in wound morphology and environmental conditions. It is recommended that the magnitude of a single adjustment be controlled within a few percentage points of the total ratio, and gradually adjusted over multiple cycles to balance reaction kinetics and coating uniformity. Optionally, when the environment shows strong scouring or high local temperatures, a low-percentage linkage of the hydrophobic phase channel is activated to enhance wet surface stability; once the environment stabilizes, the proportion gradually decreases using the same fading strategy. If necessary, a lower limit is set for the proportion to prevent any channel from accidentally reaching zero, leading to mixing failure.

[0070] In one embodiment, a digital twin simulation verification is performed before the application step: the wound morphology data and / or environmental data obtained in step S1, along with candidate component formulation parameters, are input into a simulation model containing geometric, fluid, and / or mass transfer couplings to evaluate the gelation behavior and sealing performance indicators under different component formulation parameters; and the parameters and / or hyperparameters of the pre-trained machine learning model are updated based on the evaluation results to generate calibrated component formulation parameters; optionally, the geometric domain of the digital twin is obtained by meshing the collected three-dimensional wound data, the fluid and mass transfer boundary conditions are given by the velocity and temperature measured on the environmental side, and the simulation time window covers the application and initial gelation stages. The simulation output includes predicted values ​​of the sealing performance indicators and sensitivity information to component ratios, both of which are used for offline model calibration or updating of its hyperparameters, with the update strategy following the principle of not reducing online stability. If the simulation fails to converge within a predetermined time, the previously verified model parameter set is used and a delayed calibration is prompted, without affecting the current online application process.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0072] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for intelligent formulation of hemostatic composition components based on machine learning, characterized in that, include: Step S1: Obtain morphological data and environmental data of the target wound. The morphological data includes the three-dimensional geometric features of the wound, and the environmental data includes at least one external condition related to hemostasis. Step S2: Input the morphological data and environmental data into a pre-trained machine learning model and output the component formulation parameters of the hemostatic composition; Step S3: According to the ingredient mixing parameters, control the preparation device to dynamically adjust the ratio of at least two components in the hemostatic composition, and apply the hemostatic composition to the target wound. The environmental data includes blood flow rate and temperature in the target area, and is updated at a frequency no less than a preset refresh rate during the application of the hemostatic composition, and the component preparation parameters are adjusted accordingly. This method constitutes a closed-loop control: feedback signals are acquired through a sensing unit at the wound site, the feedback signals including at least a bleeding status indicator or an interface sealing strength indicator; the feedback signals and the latest morphological / environmental data are input into the machine learning model to update the component formulation parameters.

2. The intelligent formulation method for hemostatic composition based on machine learning as described in claim 1, characterized in that, The machine learning model is a model for predicting component proportions, and the model includes at least one of Gaussian process regression, Bayesian optimized surrogate model and / or neural network.

3. The intelligent formulation method for hemostatic composition based on machine learning as described in claim 1, characterized in that, The morphological data is acquired through non-contact 3D imaging, and the 3D data of the wound surface is denoised, detrended, and missing measurements are completed by the preprocessing module. The principal curvature distribution, depth field, and surface roughness morphological features are then extracted and input into the machine learning model.

4. The intelligent formulation method for hemostatic composition based on machine learning as described in claim 1, characterized in that, The machine learning model is trained based on historical wound data and corresponding hemostasis effect data. The loss function during training simultaneously represents at least two of the hemostasis onset time, leakage rate and interface sealing strength, and imposes constraints on biocompatibility and pH range.

5. The intelligent formulation method for hemostatic composition based on machine learning as described in claim 1, characterized in that, The preparation device is a microfluidic device, which includes two or more independently meterable feeding channels and a mixing channel, and achieves real-time changes in the component ratio during the application process through valve control and pump control.

6. The intelligent formulation method for hemostatic composition based on machine learning as described in claim 1, characterized in that, The hemostatic composition comprises a hydrophilic polymer precursor and a crosslinking agent; The hydrophilic polymer precursor includes at least one of alginate, chitosan derivative, gelatin derivative or polyethylene glycol derivative; The crosslinking agent includes at least one of polyaldehyde, ionic crosslinking agent or enzyme-catalyzed crosslinking agent; The component blending parameters are used to determine the target mass ratio of each component.

7. The intelligent formulation method for hemostatic composition based on machine learning as described in claim 6, characterized in that, The hydrophilic polymer precursor is specifically oxidized alginate, and the crosslinking agent is specifically carboxymethyl chitosan. The method dynamically adjusts the mass ratio of the two during the application process based on the output of a machine learning model.

8. The intelligent formulation method for hemostatic composition based on machine learning as described in claim 1, characterized in that, Before applying the steps, perform digital twin simulation verification: input the wound morphology data and / or environmental data obtained in step S1 and the candidate component formulation parameters into a simulation model that includes geometric, fluid and / or mass transfer coupling, evaluate the gelation behavior and sealing performance indicators under different component formulation parameters; and update the parameters and / or hyperparameters of the pre-trained machine learning model based on the evaluation results to generate corrected component formulation parameters.

Images