An intelligent quality release method for fresh stem cell products

Abstract

This invention discloses an intelligent quality release method for fresh stem cell products, belonging to the field of biotechnology. The method first collects and aligns multi-source time-series data from the preparation process to construct a structured process dataset. Then, the dataset is input into a spatiotemporal prediction model based on graph neural networks. By analyzing the dynamic correlations between multi-source data, the method predicts key quality attributes at the planned data return point and generates release recommendations with clear classification conclusions and risk levels. Finally, by comparing the predicted values ​​with actual observed data, the model's online parameter fine-tuning is triggered when deviations exceed thresholds or process changes occur, achieving self-optimization. Key decision-making and optimization records throughout the process are stored on a blockchain, forming an auditable traceability chain. This invention achieves a paradigm shift from post-production detection to process prediction, significantly reducing release time while ensuring quality, and possessing the advantages of continuous learning and transparency.

Description

Technical Field

[0001] This application relates to the field of biotechnology, and in particular to a smart quality release method for fresh stem cell products. Background Technology

[0002] In existing technologies, the quality release decision for stem cell products mainly relies on the subjective experience of technicians and static standard verification, resulting in low efficiency and poor consistency. This problem is particularly prominent in the scenario of "direct reinfusion of fresh stem cells." Since this model requires the product to be released and reinfused within hours of completion of culture, the traditional process relying on fixed-cycle testing and manual verification is completely unable to meet its extremely high timeliness requirements. Furthermore, existing methods can only make static judgments on the cell state at the time of testing and cannot effectively predict the dynamic quality state of the cells at the planned reinfusion point. In addition, the current decision-making process lacks intelligent fusion analysis of real-time data from multiple dimensions such as cell metabolism and culture environment, and the models used are often built based on static historical data, making it difficult to adapt to changes brought about by production process optimization or individual donor differences, leading to a gradual decline in predictive performance. Finally, the lack of deep correlation and causal traceability between culture process data and the final release conclusion creates data silos, making quality traceability and root cause analysis extremely difficult. Summary of the Invention

[0003] This invention addresses the technical problems existing in the background art by proposing an intelligent quality release method for fresh stem cell products. It automatically completes intelligent release decisions within a very short time window by integrating and analyzing multi-source preparation data in real time; it constructs an advanced model capable of dynamically predicting the quality status of cells at the time of reinfusion, exceeding static threshold judgments; it establishes a self-optimizing AI framework with continuous learning capabilities to ensure the accuracy and adaptability of the model in the face of process changes and individual differences; and ultimately achieves full-chain verifiable traceability from process parameters and real-time monitoring data to AI decision characteristics and release conclusions, breaking down data barriers and improving the intelligence level and system transparency of quality control.

[0004] To solve the technical problem, the technical solution of the present invention is as follows:

[0005] A smart quality release method for fresh stem cell products, the method comprising: S1: Collect multi-source data during the stem cell preparation process, align and fuse them in the time dimension to generate a structured process dataset; S2: Input the process dataset into the prediction model. The prediction model analyzes the dynamic correlation between multi-source data and outputs the predicted values ​​of key quality attributes for the planned feedback time point. Based on the predicted values ​​of key quality attributes, a product release recommendation is generated. At the same time, a trend evaluation conclusion on the predicted values ​​of key quality attributes is output. S3: Obtain the actual quality observation data of the stem cell product at the planned reinfusion time point or a nearby time point, compare the actual quality observation data with the predicted value of the key quality attribute, and when the comparison deviation exceeds a preset threshold or the system detects a preset process change event, use new batch data containing the actual quality observation data to fine-tune the parameters of the prediction model to obtain an updated prediction model.

[0006] Furthermore, the method also includes: Record and fix the causal relationship graph on which the release recommendation is based; record and fix the trigger optimization process record; and jointly store the causal relationship graph, the optimization process record, the core feature vector hash value of the process dataset, and the release recommendation on the blockchain to form an auditable full-chain traceable record of decision-making and optimization.

[0007] Furthermore, the multi-source data in S1 specifically includes: time-series data of biochemical indicators of the culture environment obtained through continuous monitoring by embedded biosensors; time-series data of cell dynamic morphological parameters obtained through image analysis of microscopic images captured at time intervals; structured operation event data recorded in the preparation process; and cell viability, purity, and microbial detection data performed at preparation nodes or endpoints.

[0008] Furthermore, the biochemical indicators of the culture environment include pH value, dissolved oxygen, glucose concentration, and lactate concentration; the cell dynamic morphological parameters include cell spreading area, aggregation degree, and pseudopodia activity.

[0009] Furthermore, the prediction model is a spatiotemporal graph model based on a graph neural network (GNN); the specific processing steps include: Spatiotemporal graph construction: Each key observation moment in the structured process dataset is defined as a graph structure; wherein, data features from different sources, including operational events, environmental parameters, metabolic signals and morphological parameters, are modeled as graph nodes, and the connection edges between nodes are defined according to the physical source of the data, temporal adjacency and process logic relationship; Spatiotemporal correlation learning; the spatiotemporal graph model aggregates the feature information of the neighboring nodes of each node at the same time through graph convolutional layers to learn the spatial correlation between multi-source data; and captures the dynamic evolution pattern of node features along the time dimension through embedded temporal attention mechanism or recurrent connection to learn the temporal correlation. Dynamic prediction generation; the spatiotemporal graph model, based on the spatiotemporal correlation, infers the state of key quality attributes at the planned feedback time point through the readout layer, and outputs the predicted value of the key quality attributes.

[0010] Furthermore, the predicted values ​​of the key quality attributes include: predicted cell viability at the planned infusion time and predicted positivity rates of cell-specific surface markers.

[0011] Furthermore, the prediction model is fine-tuned using parameters by employing an online learning algorithm based on gradient descent, and the following steps are performed: Loss calculation; input the new batch of data into the prediction model for forward propagation to obtain the corresponding predicted value; compare the predicted value with the actual quality observation data, and calculate the prediction loss of the model on the current new batch of data through a preset loss function; Constrained backpropagation: When calculating the update gradient of model parameters through the backpropagation algorithm, elastic constraints are applied to the core network layer parameters of the prediction model; the elastic constraints dynamically limit the update magnitude according to the importance weight of the parameters in historical batch data, and output the constrained model parameter gradient. Incremental parameter update: Using a learning rate lower than that of the original model training phase, the model parameter gradients calculated by the constrained backpropagation step are used to perform one or more small-batch iterative updates on the parameters of the current prediction model to obtain the updated prediction model, thus completing the parameter fine-tuning of the prediction model.

[0012] Furthermore, the causal relationship graph is automatically generated during decision-making, and displays the key data nodes, influence paths and weights that affect this release decision in the form of a visual graph.

[0013] Furthermore, the optimization process record specifically includes: the event type that triggers fine-tuning, the deviation data between the predicted value and the actual observed value, the data batch identifier used for fine-tuning, and the version hash of the updated model.

[0014] This application has the following advantages: This application transforms quality decision-making from being based on static rules and human experience to dynamic data-driven intelligent prediction based on the entire cell lifecycle. By integrating multi-dimensional time-series information such as real-time metabolism and morphology, the system can proactively predict cell quality at the reinfusion point, addressing the core requirements of fresh products for timeliness and accuracy. Simultaneously, an adaptive model optimization mechanism is introduced, enabling the system to continuously learn as the process evolves, ensuring long-term reliability. Furthermore, through interpretable causal traceability and blockchain evidence storage, the entire decision-making process achieves transparency and immutability, providing an auditable and trustworthy intelligent solution for quality control. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 The main technology roadmap for the intelligent quality release method for fresh stem cell products provided in this embodiment; Figure 2 A flowchart of a GNN-based spatiotemporal graph prediction model provided for an example. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] Example 1: like Figure 1 As shown, this application provides a smart quality release method for fresh stem cell products, including: S1: Collect multi-source data during the stem cell preparation process, align and fuse them in the time dimension to generate a structured process dataset; S2: Input the process dataset into the prediction model. The prediction model analyzes the dynamic correlation between multi-source data and outputs the predicted values ​​of key quality attributes for the planned feedback time point. Based on the predicted values ​​of key quality attributes, a product release recommendation is generated. At the same time, a trend evaluation conclusion on the predicted values ​​of key quality attributes is output. S3: Obtain the actual quality observation data of the stem cell product at the planned reinfusion time point or a nearby time point, compare the actual quality observation data with the predicted value of the key quality attribute, and when the comparison deviation exceeds a preset threshold or the system detects a preset process change event, use new batch data containing the actual quality observation data to fine-tune the parameters of the prediction model to obtain an updated prediction model.

[0019] S4: Record and fix the causal relationship graph on which the release suggestion is based; record and fix the process record that triggers the optimization; jointly store the causal relationship graph, the optimization process record, the core feature vector hash value of the process dataset and the release suggestion on the blockchain to form an auditable full-chain traceability record of decision-making and optimization.

[0020] Specifically, the multi-source data in step S1 includes: time-series data of biochemical indicators of the culture environment obtained through continuous monitoring by embedded biosensors; time-series data of cell dynamic morphological parameters obtained through image analysis of periodically captured microscopic images; structured operational event data recorded during the preparation process; and cell viability, purity, and microbial detection data performed at preparation nodes or endpoints. The biochemical indicators of the culture environment include pH, dissolved oxygen, glucose concentration, and lactate concentration; the cell dynamic morphological parameters include cell spreading area, aggregation degree, and pseudopodia activity.

[0021] For example, step S1 specifically involves multi-source data acquisition, time alignment, and structured modeling; in generating a unified feature vector... Previously, for missing multi-source data, linear interpolation was used to fill in the gaps; for outliers, according to... The principle is to perform detection and replace it with the mean of nearby time points. For data with different sampling frequencies, they are aligned to a unified time axis by upsampling or downsampling. .

[0022] Preparation cycle and time definition: Let the starting time of the single-batch stem cell preparation process be... The end time is The time interval corresponding to the entire preparation process Defined as:

[0023] in: Preparation start time; The last observation time before preparation is completed or before planned retransmission.

[0024] To facilitate the fusion of multi-source data, the time interval is discretized into a set of unified time points. :

[0025] in: : No. Each discrete observation time point; : The total number of discrete time points within the preparation cycle.

[0026] Biochemical indicators of the culture environment: At each point in time Biochemical indicators of the culture environment are collected using embedded biosensors and defined as environmental feature vectors. :

[0027] in: At a certain point in time pH value of the culture medium; At a certain point in time Dissolved oxygen concentration; At a certain point in time glucose concentration; At a certain point in time The lactic acid concentration.

[0028] Cell dynamic morphological parameter data: By analyzing the microscopic images acquired at regular intervals, cell morphological features are extracted and defined as morphological feature vectors. :

[0029] in: Cells at time points The average spreading area; Cells at time points Aggregation index; Cells at time points The pseudofoot activity.

[0030] Preparation process operation event data: Transform the operational events recorded during the preparation process into structured features. :

[0031] in: : indicates the first Class operation events at time points Whether it occurs or its intensity; : The number of predefined operation event types.

[0032] Quality inspection data: Quality detection results are collected at the preparation node or endpoint and defined as a quality feature vector. :

[0033] in: Cell viability test value; Cell purity test value; Microbial test results.

[0034] Structured process dataset: At each point in time The above multi-source data are fused to form a unified feature vector. :

[0035] The process dataset consists of feature vectors from all time points.

[0036]

[0037] in: A structured process dataset for a single batch of stem cell preparation. The discrete time points N for each batch of stem cell preparation are recommended to be 50-100. Each time point includes 4 dimensions of environmental indicators, 3 dimensions of morphological features, 5-10 dimensions of operational events, and 3 dimensions of quality detection. It is recommended to use 10-20 historical batches of data for the initial model training.

[0038] Specifically, the prediction model is a spatiotemporal graph model based on a graph neural network (GNN); the specific processing steps include: In a preferred embodiment of this application, the specific architecture of the spatiotemporal prediction model based on graph neural networks is as follows: the input layer contains input units with the same dimension as the node features; the graph convolutional layer is set to 3 layers, each with 64 hidden units, and uses the ReLU activation function; time series modeling uses a single-layer GRU unit with a hidden state dimension of 64; the output layer maps to the predicted values ​​of key quality attributes through a linear transformation. Initial learning rate. Set to 0.001, elastic constraint strength Set to 0.1, historical parameter importance weight Calculated based on the Fisher information matrix.

[0039] Spatiotemporal graph construction: Each key observation moment in the structured process dataset is defined as a graph structure; wherein, data features from different sources, including operational events, environmental parameters, metabolic signals and morphological parameters, are modeled as graph nodes, and the connection edges between nodes are defined according to the physical source of the data, temporal adjacency and process logic relationship; Spatiotemporal correlation learning; the spatiotemporal graph model aggregates the feature information of the neighboring nodes of each node at the same time through graph convolutional layers to learn the spatial correlation between multi-source data; and captures the dynamic evolution pattern of node features along the time dimension through embedded temporal attention mechanism or recurrent connection to learn the temporal correlation. Dynamic prediction generation; the spatiotemporal graph model, based on the spatiotemporal correlation, infers the state of key quality attributes at the planned feedback time point through the readout layer, and outputs the predicted value of the key quality attributes.

[0040] like Figure 2 As shown, the construction of a spatiotemporal prediction model based on graph neural networks includes the following process: Definition of spacetime graph structure; At every point in time Construct a graph structure :

[0041] in: Time point The corresponding set of nodes; Time point The corresponding set of edges.

[0042] Node definition and characteristics; Node set Defined as:

[0043] in: : Represents a node in the culture environment; : Represents a cell morphology node; : Indicates an operation event node; : Represents a metabolic state node.

[0044] The features of each node are defined as follows: Cultivation environment node characteristics :

[0045] Morphological node features :

[0046] Operation event node characteristics :

[0047] Metabolic node characteristics (Can be derived from environmental indicators):

[0048] Graph edges are associated with space; For any two nodes and If a relationship is satisfied based on technological logic, physical origin, or temporal adjacency, an edge is established between them. The rules for defining edges connecting graph nodes are as follows: Within the same time point, if nodes belong to different origins (environment, morphology, operational events, metabolic states) and have a technological logical connection, an edge is established; edges are also established between nodes of the same type at adjacent time points to reflect temporal continuity. Time dimension weight. Based on discrete time intervals calculate: , For time decay coefficient, it is recommended to take... Hour.

[0049] In the In layered graph convolution, the formula for updating node features is:

[0050] in: :node In the Layer feature representation; :node The set of neighboring nodes; :node For nodes Influence weight; : No. The trainable weight matrix of the layer; : Non-linear activation function.

[0051] Time-dimensional modeling; To model the temporal correlation of features of the same node at different time points, a temporal state vector is defined:

[0052] And update recursively:

[0053] in: Time series modeling unit; : Represents a node At a specific point in time The characteristics or state representation; The time state at the previous point in time.

[0054] Specifically, the predicted values ​​for key quality attributes include: predicted cell viability at the planned infusion time and predicted positivity rates for cell-specific surface markers, including the following processes: Prediction target definition; The design return time point is as follows: .

[0055] The key quality attributes to be predicted include: Predicted cell viability at the planned infusion time; : Predicted positivity rate of specific surface markers at the planned reinfusion time.

[0056] The prediction function is defined as:

[0057] in: Model output function; : Global time state characteristics at the time of data return.

[0058] Release recommendation generated; Set quality threshold: Minimum acceptable threshold for cell viability; : The minimum acceptable threshold for the positive rate of surface markers.

[0059] The release decision rules are as follows:

[0060] Furthermore, the prediction model is fine-tuned using parameters by employing an online learning algorithm based on gradient descent, and the following steps are performed: Loss calculation; input the new batch of data into the prediction model for forward propagation to obtain the corresponding predicted value; compare the predicted value with the actual quality observation data, and calculate the prediction loss of the model on the current new batch of data through a preset loss function; Constrained backpropagation: When calculating the update gradient of model parameters through the backpropagation algorithm, elastic constraints are applied to the core network layer parameters of the prediction model; the elastic constraints dynamically limit the update magnitude according to the importance weight of the parameters in historical batch data, and output the constrained model parameter gradient. Incremental parameter update: Using a learning rate lower than that of the original model training phase, the model parameter gradients calculated by the constrained backpropagation step are used to perform one or more small-batch iterative updates on the parameters of the current prediction model to obtain the updated prediction model, thus completing the parameter fine-tuning of the prediction model.

[0061] For example, the online fine-tuning mechanism for the prediction model specifically includes: Bias calculations obtain actual quality observations at or near the time of data transmission.

[0062] in: : Actual measured cell viability; : The actual measured positive rate of surface markers.

[0063] Prediction bias Defined as:

[0064] Loss function definition; loss function Defined as:

[0065] in: Weighting coefficients for different quality attributes.

[0066] Backpropagation of elastic constraints; for model parameters Introducing historical importance weights Construct constraint terms :

[0067] in: : Current model parameters; : Parameter values ​​before fine-tuning; : Parameter importance weight; : Constraint strength coefficient.

[0068] Parameter update; using learning rate Update parameters:

[0069] in: Partial derivatives; : indicates the updated number One model parameter; : Indicates the learning rate; Using the current model parameters Calculate the loss function using the input data. The gradient is used to obtain the gradient of each parameter. derivative .

[0070] Furthermore, the causal relationship graph is automatically generated during decision-making, displaying key data nodes, influence paths, and weights affecting the release decision in a visual graph format. The optimization process record specifically includes: the event type triggering fine-tuning, the deviation data between predicted and actual observed values, the data batch identifier used for fine-tuning, and the version hash of the updated model.

[0071] For example, causal relationship mapping and blockchain evidence storage are performed, specifically including: Causal weight generation; mapping node weights in the model to causal influence weights:

[0072] in: :node For nodes The strength of the causal influence; : Represents a node and nodes The basic connection weights between them; Time dimension weight.

[0073] Evidence hashing; perform hash calculation on the following information:

[0074] in: Hash value; Hash function; Causal relationship diagram; Core features of the process dataset; : Decision to approve.

[0075] Example 2: The following describes the system deployment and operation process of an intelligent quality release method for fresh stem cell products according to this application, with reference to specific embodiments, but this does not constitute a limitation on the scope of protection of this invention.

[0076] System hardware and operating environment deployment: In the implementation, the intelligent quality release system described in this application can be deployed on a local server, private cloud, or public cloud computing environment. The system as a whole includes a data acquisition layer, a data processing and model inference layer, an application display layer, and an evidence storage and traceability layer. The front-end interaction layer supports operation via Web or mobile terminal. Technicians can input structured data generated in the preparation process into the system by scanning or importing. The back-end computing layer integrates an intelligent analysis module for performing prediction and decision-making. The intelligent analysis module includes a prediction model for predicting key quality attributes and a logic module for trend analysis and release decision generation. In the device access layer, stem cell preparation-related equipment, including incubators, microscopic imaging systems, cell counters, biosensors, etc., are connected to the system through IoT communication interfaces or application programming interfaces (APIs) to realize automatic acquisition or periodic synchronization of biochemical indicators of the culture environment, microscopic image data, and detection data. The structured data is in the form of an operation record table or a detection result table in Excel format.

[0077] In a preferred embodiment, the prediction model deployed in the back-end computing layer is based on graph neural network (GNN) and spatiotemporal graph model, which is used to perform fusion modeling and inference on multi-source process data; in other embodiments, random forest model, neural network model or a combination thereof may also be used as the prediction model without departing from the overall concept of the present invention.

[0078] Software operation process and implementation steps; Release assessment triggered: When the stem cell preparation process is completed and the planned reinfusion time is approaching, technicians or quality personnel perform the "Initiate Release Assessment" operation on the target batch of stem cell products in the system interface. The system identifies the batch to be assessed and locks the corresponding batch identifier accordingly.

[0079] Multi-source data acquisition and structured processing: After receiving the release assessment trigger command, the system automatically retrieves multi-source data generated throughout the entire preparation process of this batch of stem cell products, including: time-series data of biochemical indicators of the culture environment continuously acquired through embedded biosensors; time-series data of cell dynamic morphological parameters obtained after image analysis of periodically acquired microscopic images; structured operation event data recorded in the preparation process; and cell viability, purity, and microbial detection data collected at preparation nodes or endpoints.

[0080] The system aligns, pads, or interpolates the aforementioned multi-source data according to a unified timeline and standardizes the format to generate a structured process dataset for model inference.

[0081] Key quality attribute prediction and release suggestion generation; the system inputs the structured process dataset into the prediction model to perform inference calculation.

[0082] In a preferred embodiment, the prediction model constructs a spatiotemporal graph structure to analyze the spatiotemporal correlations between operational events, environmental parameters, metabolic signals, and morphological parameters, outputting predicted values ​​of key quality attributes at the planned reinfusion time point. These predicted values ​​include: predicted cell viability and predicted positivity rates of specific cell surface markers. Based on these predicted values, the system further generates corresponding release recommendations and evaluates the changing trends of the key quality attributes, forming a release assessment report.

[0083] In the system interface, the release assessment report is displayed in a visual form, including: release conclusion indicator: pass, review or fail; achievement status and prediction trend chart of each key quality attribute; display of the influence weight of multi-source data features on the prediction results; when the release conclusion is "review" or "fail", the key indicators with abnormalities and the corresponding process steps are highlighted.

[0084] Manual review and result archiving: The quality manager reviews the release assessment report in the system interface and completes the confirmation operation through electronic signature. The confirmed release assessment report is automatically archived and stored with the unique identifier of the corresponding batch of stem cell products for clinical inquiry, quality audit and subsequent full-process traceability.

[0085] Prediction result verification and model adaptive update: At the planned data feedback point or a nearby point in time, the system acquires the actual quality observation data of the batch of stem cell products and compares the actual quality observation data with the predicted values ​​of the aforementioned key quality attributes. When the comparison deviation exceeds a preset threshold, or when the system detects a preset process change event, the system triggers a parameter fine-tuning process for the prediction model. Using new batch data containing the actual quality observation data, the prediction model is updated online to obtain an updated prediction model.

[0086] In the initial operation phase, when the amount of historical batch data available for model training is insufficient, the system can adopt a hybrid implementation method combining a rule engine and a simplified prediction model. For example, based on pre-set indicator weights and scoring logic, rule calculations are performed on the structured data in Excel, and release suggestions are generated in conjunction with a simple machine learning model; as the data scale gradually accumulates, the system can then gradually transition to a full-scale prediction model implementation method based on graph neural networks.

[0087] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0088] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A smart quality release method for fresh stem cell products, characterized in that, The method includes: S1: Collect multi-source data during the stem cell preparation process, align and fuse them in the time dimension to generate a structured process dataset; S2: Input the process dataset into the prediction model. The prediction model analyzes the dynamic correlation between multi-source data and outputs the predicted values ​​of key quality attributes for the planned feedback time point. Based on the predicted values ​​of key quality attributes, a product release recommendation is generated. At the same time, a trend evaluation conclusion on the predicted values ​​of key quality attributes is output. S3: Obtain the actual quality observation data of the stem cell product at the planned reinfusion time point or a nearby time point, compare the actual quality observation data with the predicted value of the key quality attribute, and when the comparison deviation exceeds a preset threshold or the system detects a preset process change event, use new batch data containing the actual quality observation data to fine-tune the parameters of the prediction model to obtain an updated prediction model.

2. The intelligent quality release method for fresh stem cell products according to claim 1, characterized in that, The method further includes: Record and fix the causal relationship graph on which the release recommendation is based; record and fix the trigger optimization process record; and jointly store the causal relationship graph, the optimization process record, the core feature vector hash value of the process dataset, and the release recommendation on the blockchain to form an auditable full-chain traceable record of decision-making and optimization.

3. The intelligent quality release method for fresh stem cell products according to claim 1, characterized in that, The multi-source data in S1 specifically includes: time-series data of biochemical indicators of the culture environment obtained through continuous monitoring by embedded biosensors; time-series data of cell dynamic morphological parameters obtained through image analysis of microscopic images captured at time intervals; structured operation event data recorded in the preparation process; and cell viability, purity, and microbial detection data performed at preparation nodes or endpoints.

4. The intelligent quality release method for fresh stem cell products according to claim 3, characterized in that, The biochemical parameters of the culture environment include pH value, dissolved oxygen, glucose concentration, and lactate concentration; the cell dynamic morphological parameters include cell spreading area, aggregation degree, and pseudopodia activity.

5. The intelligent quality release method for fresh stem cell products according to claim 1, characterized in that, The prediction model is a spatiotemporal graph model based on graph neural networks (GNNs). The specific processing steps include: Spatiotemporal graph construction: Each key observation moment in the structured process dataset is defined as a graph structure; wherein, data features from different sources, including operational events, environmental parameters, metabolic signals and morphological parameters, are modeled as graph nodes, and the connection edges between nodes are defined according to the physical source of the data, temporal adjacency and process logic relationship; Spatiotemporal correlation learning; the spatiotemporal graph model aggregates the feature information of the neighboring nodes of each node at the same time through graph convolutional layers to learn the spatial correlation between multi-source data; and captures the dynamic evolution pattern of node features along the time dimension through embedded temporal attention mechanism or recurrent connection to learn the temporal correlation. Dynamic prediction generation; the spatiotemporal graph model, based on the spatiotemporal correlation, infers the state of key quality attributes at the planned feedback time point through the readout layer, and outputs the predicted value of the key quality attributes.

6. The intelligent quality release method for fresh stem cell products according to claim 1, characterized in that, The predicted values ​​for key quality attributes include: predicted cell viability at the planned infusion time and predicted positivity rates for cell-specific surface markers.

7. The intelligent quality release method for fresh stem cell products according to claim 1, characterized in that, The prediction model is fine-tuned using an online learning algorithm based on gradient descent, and the following steps are performed: Loss calculation; input the new batch of data into the prediction model for forward propagation to obtain the corresponding predicted value; compare the predicted value with the actual quality observation data, and calculate the prediction loss of the model on the current new batch of data through a preset loss function; Constrained backpropagation: When calculating the update gradient of model parameters through the backpropagation algorithm, elastic constraints are applied to the core network layer parameters of the prediction model; the elastic constraints dynamically limit the update magnitude according to the importance weight of the parameters in historical batch data, and output the constrained model parameter gradient. Incremental parameter update: Using a learning rate lower than that of the original model training phase, the model parameter gradients calculated by the constrained backpropagation step are used to perform one or more small-batch iterative updates on the parameters of the current prediction model to obtain the updated prediction model, thus completing the parameter fine-tuning of the prediction model.

8. The intelligent quality release method for fresh stem cell products according to claim 2, characterized in that, The causal relationship graph is automatically generated during the decision-making process, and displays the key data nodes, influence paths, and weights that affect this release decision in the form of a visual graph.

9. The intelligent quality release method for fresh stem cell products according to claim 2, characterized in that, The optimization process record specifically includes: the event type that triggers fine-tuning, the deviation data between the predicted value and the actual observed value, the data batch identifier used for fine-tuning, and the version hash of the updated model.

Images