An AI algorithm-based personalized matching cultivation linkage system

Abstract

The application relates to the field of embryo culture and discloses an individualized matching culture linkage system based on an AI algorithm, which comprises an embryo culture box body, a plurality of microenvironment isolation cavities physically isolated by microporous membranes are arranged in the embryo culture box body, each microenvironment isolation cavity contains a culture dish and is connected with an independent micro gas flow guide channel, a piezoelectric micro valve array is arranged in series at the input end of each micro gas flow guide channel, the piezoelectric micro valve array receives an individualized three-gas parameter curve and a pre-adjustment control instruction, and the three-gas input flow and concentration in each microenvironment isolation cavity are independently adjusted. The patient data acquisition unit further comprises an internet hospital evaluation scale data acquisition sub-module which is used for acquiring the reproductive health evaluation scale data filled by a patient through an internet hospital platform, so as to supplement and improve the patient multi-modal data dimension. The application solves the problem that traditional parameter adjustment depends on the experience of operators, and makes the parameter adjustment more scientific and quantitative.

Description

Technical Field

[0001] This invention relates to the field of embryo culture, and more particularly to a personalized, integrated culture system based on AI algorithms. Background Technology

[0002] Embryo culture, as a core component of assisted reproductive technology, directly determines embryo development quality, clinical pregnancy rate, and live birth rate through the precision of its parameter control. With the continuous development of assisted reproductive technology, the core function of embryo culture incubators has gradually focused on simulating the physiological environment within the human uterus. Precise control of the three gas input parameters (O2, CO2, and N2) is crucial for ensuring normal embryo development. Traditional embryo culture techniques primarily use standardized gas parameters based on population clinical data, lacking dynamic adaptation to individual patient differences and real-time embryo development. Overall technological development is evolving towards "personalization, intelligence, and real-time processing," with the integration of AI algorithms and multimodal recognition technology becoming a core driving force for breakthroughs in this field. Furthermore, with the popularization of internet hospitals, the collection and utilization of patient self-assessment data outside the hospital has become a new dimension for improving patient profiles and enhancing the individualized accuracy of culture protocols. However, current technologies have not yet achieved effective integration of internet hospital assessment scale data with embryo culture control systems.

[0003] In existing embryo culture technologies, key core technologies include the three-gas mixing input system of the embryo incubator, the environmental parameter stabilization control module, and embryo development observation technology. The three-gas input system adjusts the concentration of each gas via valves, typically with preset fixed parameters, relying on manual adjustments based on experience or general standards. The control module primarily uses hardware closed-loop control, only capable of maintaining fixed parameters and unable to respond to individual differences. Embryo development observation mostly relies on periodic manual microscopic examination or simple image acquisition, lacking the ability to continuously monitor and accurately analyze the dynamic development process of the embryo. Furthermore, the existing system's patient data acquisition unit lacks the ability to interface with internet hospital platforms, failing to obtain key assessment data such as the patient's remotely submitted psychological state and lifestyle, resulting in a single dimension of patient characteristic analysis and an inability to comprehensively reflect the overall maternal health status that affects embryo development. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a personalized, coordinated training system based on AI algorithms, comprising: The AI ​​multimodal data acquisition module includes a patient data acquisition unit, an embryo development monitoring unit, an incubator environment sensor, and a microelectrode array sensor. The patient data acquisition unit acquires multimodal patient data and has a built-in Internet hospital assessment scale data acquisition submodule. This submodule acquires reproductive health assessment scale data submitted by patients through the Internet hospital platform and integrates it into the multimodal patient data. The embryo development monitoring unit acquires embryo development video streams. The incubator environment sensor acquires real-time dynamic parameters of the incubator. The microelectrode array sensor is integrated into the culture dish to acquire microenvironmental parameters of the culture medium. The AI ​​multimodal analysis module, which is communicatively connected to the AI ​​multimodal data acquisition module, includes a patient feature analysis submodule and an embryo development recognition submodule. The patient feature analysis submodule calls a pre-trained deep neural network parameter mapping model to map the quantized feature vectors extracted from the patient multimodal data into personalized three-gas parameter curves covering multiple developmental stages. The embryo development recognition submodule calls a dual-stream spatiotemporal fusion network to extract the visual spatiotemporal feature vectors of the embryo development video stream through the visual stream and the biochemical time-series feature vectors of the culture medium microenvironment parameters through the biochemical stream. After dynamically weighting and fusing the visual spatiotemporal feature vectors and the biochemical time-series feature vectors through a joint attention mechanism, a real-time embryo development status score is output. The three-gas parameter linkage control module is communicatively connected to the AI ​​multimodal analysis module and includes a long short-term memory network prediction model and a control command generation unit. The long short-term memory network prediction model takes the real-time development status score sequence of the embryo over a preset time period and the real-time dynamic parameter sequence of the incubator as input, and outputs a predicted development score for a preset time period in the future. When the predicted development score is lower than a set first score threshold, the control command generation unit generates a pre-adjustment control command containing the three-gas concentration parameters and temperature parameters based on the deviation between the predicted development score and the set threshold. The embryo culture chamber body is electrically connected to the three-gas parameter linkage control module. The embryo culture chamber body is provided with multiple microenvironment isolation chambers physically isolated by microporous membranes. Each microenvironment isolation chamber contains one culture dish and is connected to an independent micro gas guide channel. The input end of each micro gas guide channel is connected in series with a piezoelectric microvalve array. The piezoelectric microvalve array receives the personalized three-gas parameter curve and the pre-adjustment control command, and independently adjusts the three-gas input flow rate and concentration in each microenvironment isolation chamber.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: Existing embryo incubators typically use fixed, one-size-fits-all three-gas parameters, which cannot be adapted to the individual differences of patient embryos. Furthermore, traditional systems rely solely on in-hospital HIS data for patient data acquisition, lacking the ability to perceive dynamic changes in patients' psychological state and lifestyle outside the hospital. This invention utilizes a patient feature analysis submodule to call a deep neural network parameter mapping model, combined with random forest feature selection and additive attention models, to dynamically weight the patient's multi-dimensional quantitative characteristics (such as advanced age, carrying specific pathogenic genes, hormonal abnormalities, and mental health and lifestyle scale data extracted by the internet hospital assessment scale data acquisition submodule), directly mapping them into personalized three-gas parameter curves covering multiple developmental stages of the embryo. This mechanism can call a threshold library for numerical truncation correction for specific disease labels and abnormal scale scores, overcoming the shortcomings of standardized parameters in adapting to specific physiological tolerance boundaries. It provides each embryo with a unique initial culture environment deeply aligned with its maternal genetic makeup and overall health status, significantly reducing the early developmental arrest rate caused by environmental mismatch. In particular, by incorporating an Internet hospital assessment scale data collection submodule into the patient data collection unit, the system can identify risk factors such as high anxiety levels and low treatment compliance, and adjust the culture strategy in advance to compensate for the potential negative impact of maternal stress on embryonic development.

[0006] Traditional embryo development monitoring relies heavily on single-sensory microscopic examination, which is highly susceptible to misjudgments due to interference from bubbles and noise in the culture medium. This invention innovatively integrates a three-dimensional dual-layer microelectrode array into the bottom layer of the culture dish, combined with a dual-flow spatiotemporal fusion network, to simultaneously extract macroscopic visual spatiotemporal features and microscopic biochemical temporal features (such as differential oxygen partial pressure / carbon dioxide partial pressure). Furthermore, a joint attention mechanism and a Bayesian multi-source verification unit are introduced. When a posterior probability conflict arises between the developmental level determined by macroscopic vision and the underlying biochemical metabolism, the system automatically suspends and intercepts pre-adjustment control commands. This mechanism completely eliminates catastrophic erroneous gas adjustment operations caused by single sensor failure or severe image occlusion from the logical level, greatly improving the fidelity of state perception and the safety of high-value embryo culture.

[0007] Traditional techniques rely on periodic observation, passively adjusting parameters only after significant deterioration in embryonic morphology is detected, resulting in irreversible time lag. This invention embeds a Long Short-Term Memory (LSTM) predictive model into the control module, enabling it to predict future developmental trends within a preset timeframe based on historical developmental scores and dynamic environmental sequences. Coupled with the hierarchical response rules of a fuzzy logic decision engine, once the predicted score shows a trend of falling below a warning threshold, the system can issue pre-adjustment control commands, including coordinated compensation for the concentrations of the three gases and temperature, before substantial morphological damage occurs. This predictive, proactive intervention mechanism transforms "post-event remediation" into "pre-event defense," resolving potential microenvironmental crises at their nascent stage and effectively curbing the surge in reverse embryonic division and fragmentation rate.

[0008] To address the hardware bottlenecks of gas crosstalk and the inability to independently adjust parameters in traditional single-chamber embryo culture incubators, this invention reconstructs the underlying physical structure of the embryo culture incubator. Multiple independent microenvironmental isolation chambers and directional physical flow channels are constructed using microporous membranes, with a dedicated piezoelectric microvalve array connected in series in each chamber. Coupled with a Multi-Agent Distributed Scheduling and Control Architecture (MADRL) at the software layer, the system allocates independent sub-node computation processes and independent gas path execution channels to each culture dish. This structure enables dozens of embryos within the same device to precisely execute their respective differentiated concentration curves without interference, keeping gas cross-contamination to an extremely low level and achieving truly high-concurrency parallel closed-loop culture.

[0009] Most existing culture control algorithms are fixed at the factory, making it impossible to optimize strategies using long-term clinical feedback. This invention proactively configures a deep reinforcement learning closed-loop update module, constructing a multi-objective weighted reward function with "live birth rate" as the highest weight. After each complete clinical cycle, the system automatically absorbs real follow-up outcome data and loads it into an experience replay pool, using a deep Q-network to continuously perform incremental iterative calculations on the underlying parameter mapping network and decision action library. This long-term closed-loop mechanism frees the system from dependence on manually fixed experience. Through feedback from massive amounts of clinical data, the system's control strategy continuously and autonomously approaches the optimal solution that improves the overall pregnancy success rate, endowing the medical device with the ability to self-evolve.

[0010] To balance control precision with the rational allocation of system resources, this invention incorporates a dynamic timing-based operating frequency scheduling algorithm. This algorithm automatically adjusts the closed-loop refresh frequency based on the embryo's developmental stage and key transformation nodes, achieving efficient and flexible tilting with limited computing power. Simultaneously, the gas path control unit is equipped with a dual error calibration link, including feedforward compensation and feedback calibration, effectively offsetting the mechanical dead zone error of the microvalve and ensuring that the command concentration is physically executed with extremely high precision. Furthermore, the underlying fault-tolerant emergency circuit module can instantly trigger timing parameter overlay "blind control" or physically cut off the gas path in the event of extreme physical failures such as sensor disconnection or microvalve jamming, providing a robust safety net for unattended continuous culture lasting for several days. Attached Figure Description

[0011] Figure 1 This is a block diagram provided for an embodiment of the present invention. Detailed Implementation

[0012] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] like Figure 1 A personalized, integrated training system based on AI algorithms includes: The AI ​​multimodal data acquisition module includes a patient data acquisition unit, an embryo development monitoring unit, an incubator environment sensor, and a microelectrode array sensor. The patient data acquisition unit acquires multimodal patient data, the embryo development monitoring unit acquires embryo development video streams, the incubator environment sensor acquires real-time dynamic parameters of the incubator, and the microelectrode array sensor is integrated into the culture dish to acquire microenvironmental parameters of the culture medium. The AI ​​multimodal analysis module, which is communicatively connected to the AI ​​multimodal data acquisition module, includes a patient feature analysis submodule and an embryo development recognition submodule. The patient feature analysis submodule calls a pre-trained deep neural network parameter mapping model to map the quantized feature vectors extracted from the patient multimodal data into personalized three-gas parameter curves covering multiple developmental stages. The embryo development recognition submodule calls a dual-stream spatiotemporal fusion network to extract the visual spatiotemporal feature vectors of the embryo development video stream through the visual stream and the biochemical time-series feature vectors of the culture medium microenvironment parameters through the biochemical stream. After dynamically weighting and fusing the visual spatiotemporal feature vectors and the biochemical time-series feature vectors through a joint attention mechanism, a real-time embryo development status score is output. The three-gas parameter linkage control module is communicatively connected to the AI ​​multimodal analysis module and includes a long short-term memory network prediction model and a control command generation unit. The long short-term memory network prediction model takes the real-time development status score sequence of the embryo over a preset time period and the real-time dynamic parameter sequence of the incubator as input, and outputs a predicted development score for a preset time period in the future. When the predicted development score is lower than a set first score threshold, the control command generation unit generates a pre-adjustment control command containing the three-gas concentration parameters and temperature parameters based on the deviation between the predicted development score and the set threshold. The embryo culture chamber body is electrically connected to the three-gas parameter linkage control module. The embryo culture chamber body is provided with multiple microenvironment isolation chambers physically isolated by microporous membranes. Each microenvironment isolation chamber contains one culture dish and is connected to an independent micro gas guide channel. The input end of each micro gas guide channel is connected in series with a piezoelectric microvalve array. The piezoelectric microvalve array receives the personalized three-gas parameter curve and the pre-adjustment control command, and independently adjusts the three-gas input flow rate and concentration in each microenvironment isolation chamber.

[0014] The patient feature analysis submodule has a built-in random forest feature selection model and an additive attention model; The patient feature analysis submodule divides the patient multimodal data into core influencing features, secondary influencing features, and environmental correlation features, and then inputs them into the random forest feature screening model after standardization, quantification, and digital encoding. The random forest feature selection model calculates the Gini coefficient of each feature and removes features whose Gini coefficient is lower than the set Gini threshold. The additive attention model includes an attention scoring function, which calculates attention scores for features selected and retained by the random forest feature selection model, and generates a dynamic weight vector after normalization by the Softmax function. The patient feature analysis submodule multiplies the dynamic weight vector with the selected and retained features to generate the quantized feature vector, which is then input into the deep neural network parameter mapping model with a four-layer hidden layer structure. When the patient multimodal data contains set special disease label data, the patient feature analysis submodule calls a preset specific population parameter threshold library to perform numerical truncation correction on the personalized three-gas parameter curve.

[0015] The microelectrode array sensor includes a three-dimensional double-layer microelectrode array; The upper layer of the three-dimensional double-layer microelectrode array is distributed with oxygen microelectrodes for detecting oxygen partial pressure, and the lower layer is distributed with carbon dioxide microelectrodes for detecting carbon dioxide partial pressure. The output signal terminals of the oxygen microelectrode and the carbon dioxide microelectrode are connected to the differential circuit. The microenvironmental parameters of the culture medium include differential oxygen partial pressure data and differential carbon dioxide partial pressure data output by the differential circuit, osmotic pressure data output by the micro osmotic pressure sensor, and the ratio of lactic acid to pyruvate concentrations in the culture medium extracted based on a set sampling interval.

[0016] The visual stream of the dual-stream spatiotemporal fusion network includes a semantic segmentation network, a 3D convolutional neural network, and an instance segmentation network; The semantic segmentation network performs image region segmentation on the embryonic development video stream and extracts the fragment volume ratio value. The three-dimensional convolutional neural network extracts spatiotemporal features from consecutive video frames within a set time window and outputs cell division time node values ​​and direct division determination identifier features. The instance segmentation network extracts instance masks from blastomeres in image frames of a preset developmental stage and outputs the multinucleus rate value of a single cell. The visual flow calculates the weighted sum of the fragment volume ratio, the cell division time node, and the multinucleus rate based on a preset primary index weighting coefficient, thereby generating the visual spatiotemporal feature vector.

[0017] The embryo development recognition submodule is equipped with a Bayesian multi-source verification and determination unit. The joint attention mechanism extracts the deviation of the oxygen partial pressure value in the culture medium microenvironment parameters from the preset normal reference range, and dynamically updates the ratio of the first weight parameter of the visual spatiotemporal feature vector to the second weight parameter of the biochemical time series feature vector based on the deviation calculation result. The Bayesian multi-source verification and determination unit calculates the posterior probabilities of the visual spatiotemporal feature vector and the biochemical time-series feature vector, respectively, pointing independently to each preset developmental level classification. When the developmental level classification result corresponding to the maximum posterior probability calculated by the visual spatiotemporal feature vector and the biochemical time-series feature vector is inconsistent, the Bayesian multi-source verification and determination unit outputs a suspension signal to intercept the issuance of the pre-adjustment control command and generates a manual review trigger command.

[0018] The control command generation unit is embedded with a fuzzy logic decision engine, which is configured with a hierarchical response rule base. The hierarchical response rule base maps the real-time developmental status score of the embryo to four consecutive preset score intervals: When the score falls into the first preset score range, the control instruction generation unit outputs an instruction to maintain the current execution parameters of the personalized three-gas parameter curve; When the score falls into the second preset score range, the control command generation unit outputs a fine-tuning command for oxygen concentration and carbon dioxide concentration, which includes a first set range. When the score falls into the third preset score range, the control command generation unit outputs a coordinated intervention command for oxygen concentration and carbon dioxide concentration containing a second preset amplitude and simultaneously outputs a temperature bias adjustment command. The value of the second preset amplitude is greater than the value of the first preset amplitude. When the score falls into the fourth preset score range, the control command generation unit outputs a low oxygen contingency plan parameter switching command.

[0019] The system is equipped with a deep reinforcement learning closed-loop update module that is communicatively connected to the AI ​​multimodal analysis module; The deep reinforcement learning closed-loop update module includes a deep Q-network model. The state space vector of the deep Q-network model is composed of the quantized feature vector, the real-time embryonic development status score, and the current gas path execution parameter vector. The action space vector of the deep Q-network model is a three-gas concentration adjustment matrix containing discrete adjustment step sizes. The deep Q-network model embeds a multi-objective weighted reward function. The input parameters of the multi-objective weighted reward function include the clinical pregnancy rate, live birth rate, and high-quality embryo blastocyst emergence rate obtained from the clinical data interface. The weight coefficient value corresponding to the live birth rate is greater than the weight coefficient values ​​of other input parameters. The deep reinforcement learning closed-loop update module calculates the sample reward value of the experience replay pool based on the multi-objective weighted reward function, and extracts sample data to perform incremental iterative calculation on the node weights of the deep neural network parameter mapping model.

[0020] The system deploys a multi-agent distributed scheduling and control architecture, which includes a master control process that manages the global control environment and multiple sub-node processes that correspond one-to-one with each of the micro-environment isolation cavities. Each of the sub-node processes extracts the identification mark on the surface of the culture dish in its corresponding microenvironment isolation cavity and associates it with the multimodal feature data bound to the embryo, and independently executes the feature mapping, state recognition and instruction generation thread for the corresponding single embryo; Each of the sub-node processes uses a message passing protocol mechanism to send gas regulation command parameters to the main control process, and executes a proportional-integral-derivative control algorithm to generate a low-level drive level signal to be sent to the corresponding piezoelectric microvalve array.

[0021] Each of the microenvironment isolation chambers is provided with a directional physical flow channel structure to guide the unidirectional flow of air; the output port of the micro gas guiding channel is distributed circumferentially around the periphery of the culture dish, and the input gas diffuses into the microenvironment isolation chamber along the directional physical flow channel structure; The three-gas parameter linkage control module is equipped with an error calibration link, including: a feedforward compensation unit, which extracts the deviation mapping data between the target valve opening of the piezoelectric microvalve array and the concentration detected by the sensor from the pre-stored historical database, and superimposes the reverse level compensation value before the pre-adjustment control command is issued; and a feedback calibration unit, which calculates the difference between the real-time dynamic parameter feedback value of the incubator and the target pre-adjustment control command setting value. When the difference is greater than the preset error tolerance, a secondary fine-tuning command is output to the piezoelectric microvalve array for opening compensation.

[0022] The system has a built-in dynamic timing frequency scheduling algorithm: When the real-time developmental status score of the embryo is higher than the preset excellent / good judgment score, the interactive closed loop consisting of the AI ​​multimodal data acquisition module, the AI ​​multimodal analysis module, and the three-gas parameter linkage control module refreshes the data at a first execution frequency; when the real-time developmental status score of the embryo is lower than the excellent / good judgment score, or when the embryo is determined to be at a preset developmental stage transition time node, the interactive closed loop switches to a second execution frequency greater than the first execution frequency to refresh the data. The system also includes a fault-tolerant emergency circuit module. When the environmental sensor signal in the micro-environment isolation chamber is lost, the fault-tolerant emergency circuit module retrieves the current time node parameter of the personalized three-gas parameter curve to overwrite the pre-adjustment control command. When the actual opening degree of the piezoelectric microvalve array is detected to reach the preset mechanical extreme value, the gas supply branch of the corresponding channel is cut off.

[0023] Specifically, the solution of the present invention is as follows: System overall architecture and microenvironment isolation physical structure This embodiment provides a personalized, integrated culture system based on AI algorithms. From the perspective of physical and logical data flow, the system is mainly divided into four core modules: an AI multimodal data acquisition module, an AI multimodal analysis module, a three-gas parameter linkage control module, and the embryo culture incubator itself.

[0024] In order to achieve independent "one embryo, one policy" culture of multiple embryos from different patients in the same incubator and avoid cross-interference of gases, this embodiment reconstructs the underlying physical isolation of the embryo incubator body.

[0025] The incubator body utilizes MEMS microfabrication technology to construct multiple independent microenvironment isolation chambers. Adjacent microenvironment isolation chambers are physically separated by microporous membranes. This microporous membrane design allows for microscopic heat conduction between chambers to maintain overall temperature equilibrium while effectively blocking direct convection of gas molecules between adjacent culture dishes. Each microenvironment isolation chamber contains only one embryo culture dish and is connected to a dedicated micro-gas flow channel.

[0026] Inside each microenvironment isolation chamber, a directional physical flow channel structure is formed on the bottom or sidewall to guide the unidirectional flow of air. The outlet of the micro gas guiding channel is distributed in a circumferential or semi-circular manner along the periphery of the culture dish, so that the input three gases (O2, CO2, N2) diffuse evenly into the microenvironment isolation chamber in the form of unidirectional airflow along the directional physical flow channel structure, avoiding the formation of local airflow eddies within the chamber.

[0027] To achieve independent gas supply, a piezoelectric microvalve array is connected in series at the input end of each micro gas flow channel. This array receives the underlying control level and uses the piezoelectric effect to independently adjust the flow rate and concentration of the three gases in each microenvironment isolation chamber, with a mechanical adjustment accuracy of up to 0.01%.

[0028] Multimodal data sensing and three-dimensional bilayer microelectrode sensing This embodiment details the hardware configuration of the AI ​​multimodal data acquisition module. This module includes a patient data acquisition unit, an embryonic development monitoring unit (such as a high-definition microscope camera orthogonally mounted above the isolation chamber), incubator environmental sensors (temperature, humidity, and macroscopic gas concentration), and a microelectrode array sensor.

[0029] Specifically, for the microenvironment of the culture medium in which the embryo directly contacts, the system embeds a three-dimensional double-layer microelectrode array at the bottom of each culture dish. The specific structure is as follows: the distance between the upper and lower electrode layers is set at the micrometer level (e.g., 50 μm), the upper array contains oxygen microelectrodes for detecting oxygen partial pressure, and the lower array contains carbon dioxide microelectrodes for detecting carbon dioxide partial pressure.

[0030] To address the interference of free ions and electromagnetic noise in the culture medium on weak electrophysiological signals, the output signal terminals of the oxygen microelectrode and carbon dioxide microelectrode are connected to a hardware differential circuit composed of operational amplifiers. Utilizing the extremely high common-mode rejection ratio of the differential signal, high signal-to-noise ratio differential oxygen partial pressure and differential carbon dioxide partial pressure data are output. Simultaneously, combined with a built-in miniature osmolarity sensor and a high-precision temperature probe, the system extracts and calculates the lactic acid to pyruvate concentration ratio data of the culture medium based on a set sampling interval (e.g., once per hour), collectively constructing high-fidelity microenvironment parameters for the culture medium.

[0031] Personalized initial parameter generation mapping based on patient characteristics Before the embryo is placed in the culture chamber, the patient feature analysis submodule in the AI ​​multimodal analysis module is responsible for generating the initial dynamic culture curve for the embryo.

[0032] Feature Quantification and Stratification: The patient data acquisition unit connects to the HIS system to obtain data, which is then divided into core influencing features (such as age, gamete quality, and pathogenic gene carrier markers), secondary influencing features (such as hormone levels), and environmentally related features. Numerical indicators are normalized using Z-Score, while qualitative indicators are encoded using discrete numbers (such as assigning 0 / 1 values).

[0033] Random Forest Dimensionality Reduction: Standardized data is input into a built-in random forest feature selection model (e.g., containing 100 decision trees). This model calculates the Gini coefficient of each feature node and automatically removes low-relevance features with Gini coefficients below a set Gini threshold (e.g., 0.05).

[0034] Weighted Attention Mechanism: The selected features enter the weighted attention model (Additive Attention). The attention scoring function within the model calculates the score for each feature, which is then normalized using the Softmax function to generate a dynamic weight vector. For example, when "advanced age" and "carrier of a single-gene hereditary disease" are identified together, the model automatically increases the weight of the core feature of this combination. The dynamic weight vector is multiplied by the retained features to generate the final quantized feature vector.

[0035] DNN Nonlinear Mapping: The feature vector is input into a deep neural network parameter mapping model with a four-layer hidden layer structure (neuron nodes are sequentially 256-256-128-128, using ReLU activation function and Dropout layer). The model outputs personalized three-gas parameter curves covering four consecutive developmental stages of the embryo: pronuclear stage, cleavage stage, morula stage, and blastocyst stage. If the patient data contains special disease label data (such as specific metabolic defects), the system calls a preset parameter threshold library for specific populations and uses a numerical truncation algorithm to forcibly down-adjust or up-adjust the extreme values ​​of the generated curves to ensure that the parameters do not exceed the absolute boundary of the embryo's physiological tolerance.

[0036] Dual-stream spatiotemporal network identification and Bayesian multi-source false positive verification After the embryo is placed in the incubator, the embryo development recognition submodule calls the dual-stream spatiotemporal fusion network to accurately perceive and verify the real-time status of the embryo.

[0037] Visual stream processing: The semantic segmentation network (such as Mobile U-Net, with frame processing time <50ms) is called to perform image region segmentation on the video stream, accurately remove extracellular debris, and calculate the output debris volume ratio. The three-dimensional convolutional neural network (3D-CNN) is invoked to extract spatiotemporal features from consecutive video frames within a set time window (e.g., 5 minutes), accurately outputting cell division time node values ​​such as t2 and t4, and using a classification layer to determine whether "direct division" (e.g., 2 cells skipping to 3 cells) has occurred, thus identifying and outputting anomaly identification features. By calling an instance segmentation network (such as Mask R-CNN), when a specific developmental stage is identified, instance mask extraction is performed on the blastomeres to identify the number of nuclei in a single cell and output the multinucleus rate value.

[0038] Visual flow generates a visual spatiotemporal feature vector by weighting and summing the above values ​​based on preset primary index weight coefficients.

[0039] Fusion of biochemical flow and joint attention: The biochemical flow network extracts the aforementioned microenvironment parameters of the culture medium to generate a biochemical time-series feature vector. The system's built-in joint attention mechanism monitors in real time the deviation of the differential oxygen partial pressure value from the preset normal reference range (e.g., 25-35 mmHg). When a severe deviation occurs (e.g., a sudden drop in oxygen partial pressure), this mechanism dynamically increases the second weight parameter of the biochemical time-series feature vector and correspondingly decreases the first weight parameter of the visual spatiotemporal feature vector, shifting the system's focus from morphology to the underlying metabolic crisis.

[0040] Bayesian multi-source verification decision unit: To avoid fatal misoperations caused by single-modality issues (such as image contamination or a single electrode short circuit), the decision unit independently calculates the posterior probability of the visual spatiotemporal feature vector and the biochemical time-series feature vector pointing to a preset developmental grade classification (excellent, good, average, poor). When the grades corresponding to the maximum posterior probabilities calculated by the two are inconsistent (e.g., visual judgment is "good," biochemical judgment is "poor"), a modal conflict occurs in the system decision. The decision unit immediately outputs an electrical suspension signal, logically locking and intercepting the issuance of pre-adjustment control commands. Simultaneously, a manual review trigger command is generated on the operating terminal, forcing the embryologist to intervene manually.

[0041] LSTM predictive advance control and fuzzy logic hierarchical response Unlike the lag feedback of traditional incubators, the three-gas parameter linkage control module enables predictive intervention.

[0042] The module's embedded Long Short-Term Memory (LSTM) prediction model (such as configuring 3 hidden layers with 128 neurons each) uses the real-time embryonic development status score sequence and actual dynamic parameter sequence from a preset time period (such as 6 hours) in the past as input to calculate the predicted development score for a preset time period in the future (such as 2-4 hours).

[0043] When the predicted score shows a downward trend and falls below the set first scoring threshold (e.g., a warning line of 6 points), the fuzzy logic decision engine embedded in the control instruction generation unit intervenes. Its tiered response rule base maps the score to four consecutive preset score intervals and executes corresponding actions: First interval (≥9 points): Output instructions to maintain the currently executed parameters of the personalized three-gas parameter curve; Second interval (7-8 minutes): Output includes fine-tuning instructions with the first set amplitude (e.g., O2±0.2%, CO2±0.1%); The third interval (5-6 points): outputs a coordinated intervention command containing a second set amplitude (e.g., O2 ± 0.5%) (the second set amplitude is greater than the first set amplitude). Since large-scale gas adjustment will cause heat loss in the cavity, the system simultaneously outputs a temperature offset adjustment command (e.g., ± 0.2℃) to the heating module for thermodynamic compensation; Fourth interval (<5 points): Output low oxygen contingency plan parameter switching command and execute the highest level of protection.

[0044] Dynamic frequency scheduling, multi-agent distributed control and error calibration This embodiment details the system's underlying communication scheduling and fault tolerance / correction mechanisms.

[0045] Multi-Agent Distributed Scheduling and Control Architecture (MADRL): To support high-concurrency computing across dozens of isolation chambers, the software platform utilizes a MADRL architecture. This includes a master control process that manages the global air supply and overall temperature and humidity environment, as well as multiple child node processes that correspond one-to-one with the micro-environment isolation chambers.

[0046] Each sub-node process extracts identification markers (such as laser QR codes) from the surface of the culture dish using a driven camera, automatically associating them with the patient's multimodal data bound to that embryo. Each sub-node independently and non-blockingly performs feature mapping, status scoring, and instruction generation calculations for a single embryo. Subsequently, the sub-nodes send gas path adjustment command parameters to the master control process using a message passing protocol mechanism (such as the MQTT protocol), and upon approval, independently execute a proportional-integral-derivative (PID) algorithm to generate underlying drive level signals, which are then sent to their respective piezoelectric microvalve arrays.

[0047] Error calibration link and fault tolerance emergency: During the PID signal issuance process, the feedforward compensation unit consults the nonlinear deviation mapping data between the target opening degree of the piezoelectric microvalve and the measured concentration in the historical database, and pre-superimposes a reverse level compensation value on the instruction to offset the mechanical dead zone error; after the instruction is executed, the feedback calibration unit calculates the difference between the sensor's measured feedback value and the target set value. When the absolute value is greater than the preset error tolerance (such as 0.1%), a secondary fine-tuning instruction is output to the piezoelectric microvalve array in the next control cycle.

[0048] The fault-tolerant emergency circuit module serves as a physical safety base: when the sensor signal loss (disconnection) is detected in the isolation chamber, the theoretical parameters of the current time node of the personalized three-gas parameter curve are directly retrieved to overwrite and execute "blind control"; when the actual opening degree of the micro valve is detected to reach the preset mechanical extreme value (jammed or full load), the relay is triggered to directly cut off the gas supply branch of that channel.

[0049] Dynamic timing-based frequency scheduling algorithm: When the embryo score is higher than the excellent / good score, the system data interaction closed loop adopts the first execution frequency (e.g., refreshed once every 5 minutes); when the score falls below the excellent / good score, or when the embryo is visually identified as being at a set developmental stage transition time node (e.g., transitioning from the cleavage stage to the morula stage), the system automatically switches to a second execution frequency greater than the first execution frequency (e.g., refreshed once every 1 minute), concentrating computing power for high-density monitoring and intervention.

[0050] Long-term model closed-loop iteration based on deep reinforcement learning To enable the strategy to evolve over long-term use, the system is equipped with a deep reinforcement learning closed-loop update module.

[0051] This module includes a Deep Q-Network (DQN) model. The state space vector of the DQN is defined as a high-dimensional tensor, which is composed of a quantized feature vector, a real-time embryonic developmental status score, and a vector of current gas path execution parameters; the action space vector is a three-gas concentration adjustment matrix containing a set of discrete adjustment step sizes (e.g., ±0.1%, ±0.2%).

[0052] The core driving force for model training is a multi-objective weighted reward function. After the clinical cycle ends, the system obtains real follow-up outcomes as input parameters through an interface, including: clinical pregnancy rate, live birth rate, and high-quality embryo emergence rate. In the configuration, the weight coefficient corresponding to the live birth rate is forcibly set to the highest value (e.g., weight 0.6, with the other items equally distributed).

[0053] The system periodically packages historical operational data into an experience replay pool. Sample data is extracted through a random sampling mechanism, and sample gradients are calculated based on the reward function. Incremental iterative calculations are then performed on the node weights of the deep neural network parameter mapping model and the policy network of the decision engine using a backpropagation algorithm. Through this long-term unsupervised closed-loop mechanism, the system's parameter prediction and adjustment strategies, fed with massive amounts of real clinical data, continuously approach the theoretically optimal control solution for achieving extremely high live birth rates.

[0054] Specific Implementation of the Data Collection Submodule for the Internet Hospital Assessment Scale This embodiment details the hardware and software configuration of the Internet Hospital Assessment Scale data collection submodule, which is built into the patient data collection unit. This submodule is deployed on a hospital intranet server or a secure cloud computing node, establishing a two-way authenticated connection with the Internet Hospital platform via a standardized HTTPS encrypted channel.

[0055] The questionnaire interface unit adopts the HL7 FHIR standard protocol and periodically (e.g., 2:00 AM daily) polls the internet hospital platform database to obtain the raw data of the reproductive health assessment questionnaires filled out by patients before entering the cycle and during the culture cycle. For high-priority patients (e.g., those with a history of recurrent implantation failure), the system supports a real-time proactive push mode, triggering the data collection process immediately after the patient completes the questionnaire submission.

[0056] The data cleaning unit performs the following preprocessing operations: missing values ​​are filled using multiple imputation (MICE) based on the patient's historical data; obvious logical errors in the entries (such as negative BMI values) are marked and manual review is triggered; and text-based scores are converted into standardized numerical vectors.

[0057] The feature extraction unit extracts quantitative indicators from four core scales: Mental health dimension: A SAS standard score ≥50 is marked as positive for anxiety, and an SDS standard score ≥53 is marked as positive for depression. The perceived stress of fertility is quantified using a 1-10 point scale. Lifestyle dimension: Exercise frequency is converted into weekly metabolic equivalents (MET-minutes / week), and dietary structure score is calculated based on Mediterranean diet pattern adherence; Treatment adherence dimension: Medication adherence was scored using the Morisky Medication Adherence Questionnaire (MMAS-8), with a score <6 indicating poor adherence; Social support dimension: The Perceived Social Support Scale (PSSS) was used for scoring, with a score <50 indicating a low level of social support.

[0058] The time-series alignment unit establishes a unified time axis index, aligning the scale collection time point with the patient's in-hospital examination time and embryo culture time point, supporting dynamic weight calculation based on time decay factors (e.g., the scale data closer to the egg retrieval day has a higher weight), and outputting the fused complete patient multimodal data to the AI ​​multimodal analysis module.

[0059] In the patient feature analysis submodule, the aforementioned scale features, together with in-hospital physiological indicators, constitute the extended feature vector. The random forest feature selection model incorporates scale features into the Gini coefficient calculation. Clinical validation shows that the Gini importance of "anxiety self-rating score" and "sleep quality score" on embryonic developmental outcomes can reach 0.12-0.15, belonging to high-weight items among secondary influencing features. When the additive attention model identifies a sustained increase in the patient's anxiety score (e.g., an increase of >20% in two consecutive assessments), it automatically increases the corresponding weight coefficient. Based on this, the deep neural network parameter mapping model lowers the incubator oxygen concentration setting (e.g., from 5% to 3%) to reduce the potential damage of oxidative stress to the embryo.

[0060] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A personalized, coordinated training system based on AI algorithms, characterized in that, include: The AI ​​multimodal data acquisition module includes a patient data acquisition unit, an embryo development monitoring unit, an incubator environment sensor, and a microelectrode array sensor. The patient data acquisition unit acquires multimodal patient data and has a built-in Internet hospital assessment scale data acquisition submodule. This submodule acquires reproductive health assessment scale data submitted by patients through the Internet hospital platform and integrates it into the multimodal patient data. The embryo development monitoring unit acquires embryo development video streams. The incubator environment sensor acquires real-time dynamic parameters of the incubator. The microelectrode array sensor is integrated into the culture dish to acquire microenvironmental parameters of the culture medium. The AI ​​multimodal analysis module, which is communicatively connected to the AI ​​multimodal data acquisition module, includes a patient feature analysis submodule and an embryo development recognition submodule. The patient feature analysis submodule calls a pre-trained deep neural network parameter mapping model to map the quantized feature vectors extracted from the patient multimodal data into personalized three-gas parameter curves covering multiple developmental stages. The embryo development recognition submodule calls a dual-stream spatiotemporal fusion network to extract the visual spatiotemporal feature vectors of the embryo development video stream through the visual stream and the biochemical time-series feature vectors of the culture medium microenvironment parameters through the biochemical stream. After dynamically weighting and fusing the visual spatiotemporal feature vectors and the biochemical time-series feature vectors through a joint attention mechanism, a real-time embryo development status score is output. The three-gas parameter linkage control module is communicatively connected to the AI ​​multimodal analysis module and includes a long short-term memory network prediction model and a control command generation unit. The long short-term memory network prediction model takes the real-time development status score sequence of the embryo over a preset time period and the real-time dynamic parameter sequence of the incubator as input, and outputs a predicted development score for a preset time period in the future. When the predicted development score is lower than a set first score threshold, the control command generation unit generates a pre-adjustment control command containing the three-gas concentration parameters and temperature parameters based on the deviation between the predicted development score and the set threshold. The embryo culture chamber body is electrically connected to the three-gas parameter linkage control module. The embryo culture chamber body is provided with multiple microenvironment isolation chambers physically isolated by microporous membranes. Each microenvironment isolation chamber contains one culture dish and is connected to an independent micro gas guide channel. The input end of each micro gas guide channel is connected in series with a piezoelectric microvalve array. The piezoelectric microvalve array receives the personalized three-gas parameter curve and the pre-adjustment control command, and independently adjusts the three-gas input flow rate and concentration in each microenvironment isolation chamber.

2. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, The patient feature analysis submodule has a built-in random forest feature selection model and an additive attention model; The patient feature analysis submodule divides the patient multimodal data into core influencing features, secondary influencing features, and environmental correlation features, and then inputs them into the random forest feature screening model after standardization, quantification, and digital encoding. The random forest feature selection model calculates the Gini coefficient of each feature and removes features whose Gini coefficient is lower than the set Gini threshold. The additive attention model includes an attention scoring function, which calculates attention scores for features selected and retained by the random forest feature selection model, and generates a dynamic weight vector after normalization by the Softmax function. The patient feature analysis submodule multiplies the dynamic weight vector with the selected and retained features to generate the quantized feature vector, which is then input into the deep neural network parameter mapping model with a four-layer hidden layer structure. When the patient multimodal data contains set special disease label data, the patient feature analysis submodule calls a preset specific population parameter threshold library to perform numerical truncation correction on the personalized three-gas parameter curve.

3. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, The microelectrode array sensor includes a three-dimensional double-layer microelectrode array; The upper layer of the three-dimensional double-layer microelectrode array is distributed with oxygen microelectrodes for detecting oxygen partial pressure, and the lower layer is distributed with carbon dioxide microelectrodes for detecting carbon dioxide partial pressure. The output signal terminals of the oxygen microelectrode and the carbon dioxide microelectrode are connected to the differential circuit. The microenvironmental parameters of the culture medium include differential oxygen partial pressure data and differential carbon dioxide partial pressure data output by the differential circuit, osmotic pressure data output by the micro osmotic pressure sensor, and the ratio of lactic acid to pyruvate concentrations in the culture medium extracted based on a set sampling interval.

4. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, The visual stream of the dual-stream spatiotemporal fusion network includes a semantic segmentation network, a 3D convolutional neural network, and an instance segmentation network; The semantic segmentation network performs image region segmentation on the embryonic development video stream and extracts the fragment volume ratio value. The three-dimensional convolutional neural network extracts spatiotemporal features from consecutive video frames within a set time window and outputs cell division time node values ​​and direct division determination identifier features. The instance segmentation network extracts instance masks from blastomeres in image frames of a preset developmental stage and outputs the multinucleus rate value of a single cell. The visual flow calculates the weighted sum of the fragment volume ratio, the cell division time node, and the multinucleus rate based on a preset primary index weighting coefficient, thereby generating the visual spatiotemporal feature vector.

5. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, The embryo development recognition submodule is equipped with a Bayesian multi-source verification and determination unit. The joint attention mechanism extracts the deviation of the oxygen partial pressure value in the culture medium microenvironment parameters from the preset normal reference range, and dynamically updates the ratio of the first weight parameter of the visual spatiotemporal feature vector to the second weight parameter of the biochemical time series feature vector based on the deviation calculation result. The Bayesian multi-source verification and determination unit calculates the posterior probabilities of the visual spatiotemporal feature vector and the biochemical time-series feature vector, respectively, pointing independently to each preset developmental level classification. When the developmental level classification result corresponding to the maximum posterior probability calculated by the visual spatiotemporal feature vector and the biochemical time-series feature vector is inconsistent, the Bayesian multi-source verification and determination unit outputs a suspension signal to intercept the issuance of the pre-adjustment control command and generates a manual review trigger command.

6. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, The control command generation unit is embedded with a fuzzy logic decision engine, which is configured with a hierarchical response rule base. The hierarchical response rule base maps the real-time developmental status score of the embryo to four consecutive preset score intervals: When the score falls into the first preset score range, the control instruction generation unit outputs an instruction to maintain the current execution parameters of the personalized three-gas parameter curve; When the score falls into the second preset score range, the control command generation unit outputs a fine-tuning command for oxygen concentration and carbon dioxide concentration, which includes a first set range. When the score falls into the third preset score range, the control command generation unit outputs a coordinated intervention command for oxygen concentration and carbon dioxide concentration containing a second preset amplitude and simultaneously outputs a temperature bias adjustment command. The value of the second preset amplitude is greater than the value of the first preset amplitude. When the score falls into the fourth preset score range, the control command generation unit outputs a low oxygen contingency plan parameter switching command.

7. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, The system is equipped with a deep reinforcement learning closed-loop update module that is communicatively connected to the AI ​​multimodal analysis module; The deep reinforcement learning closed-loop update module includes a deep Q-network model. The state space vector of the deep Q-network model is composed of the quantized feature vector, the real-time embryonic development status score, and the current gas path execution parameter vector. The action space vector of the deep Q-network model is a three-gas concentration adjustment matrix containing discrete adjustment step sizes. The deep Q-network model embeds a multi-objective weighted reward function. The input parameters of the multi-objective weighted reward function include the clinical pregnancy rate, live birth rate, and high-quality embryo blastocyst emergence rate obtained from the clinical data interface. The weight coefficient value corresponding to the live birth rate is greater than the weight coefficient values ​​of other input parameters. The deep reinforcement learning closed-loop update module calculates the sample reward value of the experience replay pool based on the multi-objective weighted reward function, and extracts sample data to perform incremental iterative calculation on the node weights of the deep neural network parameter mapping model.

8. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, Each of the microenvironment isolation chambers is provided with a directional physical flow channel structure to guide the unidirectional flow of air; the output port of the micro gas guiding channel is distributed circumferentially around the periphery of the culture dish, and the input gas diffuses into the microenvironment isolation chamber along the directional physical flow channel structure; The three-gas parameter linkage control module is equipped with an error calibration link, including: a feedforward compensation unit, which extracts the deviation mapping data between the target valve opening of the piezoelectric microvalve array and the concentration detected by the sensor from the pre-stored historical database, and superimposes the reverse level compensation value before the pre-adjustment control command is issued; and a feedback calibration unit, which calculates the difference between the real-time dynamic parameter feedback value of the incubator and the target pre-adjustment control command setting value. When the difference is greater than the preset error tolerance, a secondary fine-tuning command is output to the piezoelectric microvalve array for opening compensation.

9. The personalized training and linkage system based on AI algorithm according to claim 1, characterized in that, The data collection submodule for the Internet hospital assessment scale includes: The scale interface unit connects with the Internet hospital platform through a standardized API interface to obtain the raw data of the reproductive health assessment scale filled out online by patients; The data cleaning unit performs missing value imputation, outlier removal, and format standardization on the original scale data. The feature extraction unit extracts quantitative indicators of mental health scores, lifestyle scores, past treatment adherence scores, and social support scores from the cleaned scale data. The time-series alignment unit aligns the collection timestamps of the scale data with the collection time axis of the patient's multimodal data, establishes a time-series association index, and merges and outputs the results.

10. A personalized, coordinated training system based on AI algorithms according to claim 9, characterized in that, The data from the reproductive health assessment scale includes: The basic information scale collects self-assessment data on patients' age, BMI, marital and reproductive history, and menstrual cycle regularity. Mental health scales were used to collect scores on the Self-Rating Anxiety Scale (SAS), Self-Rating Depression Scale (SDS), perceived stress of childbirth, and sleep quality. Lifestyle scale, collecting scores on smoking and drinking history, exercise frequency, dietary structure, and work environment stress exposure; Treatment adherence scale: The number of previous assisted reproductive treatment cycles, medication adherence score, regularity of follow-up visits, and doctor-patient communication satisfaction score were collected.

Images