Multi-mode-based human umbilical cord blood stem cell detection and screening system

Abstract

The invention discloses a multimodal-based human umbilical cord blood stem cell detection and screening system, and belongs to the technical field of medical detection, and the system specifically comprises the following steps: collecting multimodal data of a human umbilical cord blood sample; performing time domain and space domain feature extraction on the cell morphology image sequence to obtain a cell morphology dynamic change track; analyzing a metabolite concentration change curve, identifying an abnormal fluctuation interval and extracting concentration fluctuation characteristics; processing a cell acoustic vibration signal, separating vibration energy distribution of different frequencies, and calculating a cell membrane vibration stability index; fusing the features to generate a multi-modal feature vector, inputting the multi-modal feature vector into a trained umbilical cord blood stem cell quality evaluation model to output a quality score, and judging whether the sample is qualified or not according to the score; when the sample is unqualified, starting traceability, and identifying a key unqualified link in combination with the multi-modal feature vector and the sample operation log; updating model parameters, and outputting an optimized model; the quality of the umbilical cord blood stem cells can be accurately evaluated, and problems can be traced.

Description

Technical Field

[0001] This invention relates to the field of medical testing technology, specifically to a multimodal human umbilical cord blood stem cell detection and screening system. Background Technology

[0002] Umbilical cord blood stem cells, rich in hematopoietic and mesenchymal stem cells, have irreplaceable value in the treatment of hematological diseases such as leukemia and aplastic anemia, and also show great promise in regenerative medicine fields such as tissue repair and immune regulation. Their quality directly determines transplant success rates and clinical efficacy. Therefore, establishing a comprehensive and accurate testing and screening system before umbilical cord blood collection, storage, and clinical application has become a core requirement for the industry to achieve standardized application and clinical translation. Currently, the industry's assessment of umbilical cord blood stem cell quality has shifted from a single indicator to a multi-dimensional comprehensive judgment, but a systematic technical integration solution has not yet been formed.

[0003] In existing technologies, morphological detection uses microscopic imaging systems to acquire static features such as cell outlines and area, providing an intuitive basis for preliminary quality assessment; metabolic analysis technology uses microfluidic chips to monitor changes in the concentration of metabolites such as lactic acid and glucose, reflecting the cell's activity status; acoustic detection can capture cell vibration signals to assess cell membrane integrity, while some quality assessment models build machine learning algorithms based on single-modal features to assist in quality grading; and operation log management systems are also used to record information such as sample collection time and transportation temperature, providing basic data support for subsequent traceability.

[0004] However, existing technologies still have certain limitations: on the one hand, single-modal detection data is one-sided. For example, morphology is difficult to reflect abnormal cell dynamic metabolism, and metabolic analysis cannot be correlated with cell morphological stability, resulting in incomplete quality assessment results that cannot fully cover core quality dimensions such as cell activity and structural stability. On the other hand, source tracing analysis relies solely on operation logs and lacks the ability to correlate with abnormal cell characteristics during the detection process. It cannot accurately match the quality problems reflected by multimodal features with specific operational steps, making it difficult to locate key quality influencing factors and to optimize sample processing procedures in a targeted manner. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal human umbilical cord blood stem cell detection and screening system, solving the following technical problems:

[0006] Existing single-modal detection data is one-sided and cannot comprehensively assess core quality dimensions such as cell viability and structural stability. Furthermore, the source tracing analysis relies solely on operation logs and lacks the ability to correlate with abnormal cell characteristics during the detection process, making it impossible to accurately locate key quality influencing factors.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A multimodal human umbilical cord blood stem cell detection and screening system includes:

[0009] A multimodal data acquisition module is used to acquire multimodal data from human umbilical cord blood samples. The multimodal data includes cell morphology image sequences, metabolite concentration change curves, and cell acoustic vibration signals.

[0010] The morphological feature extraction module is used to extract features in the time and spatial domains from cell morphology image sequences to obtain the dynamic change trajectory of cell morphology.

[0011] The metabolic feature analysis module is used to analyze the concentration change curves of metabolites, identify abnormal fluctuation ranges in the concentration of metabolites, and extract the characteristics of the concentration fluctuations of metabolites.

[0012] The acoustic signal processing module is used to process cellular acoustic vibration signals, separate the vibration energy distribution in different frequency ranges, and calculate cell membrane vibration stability indices.

[0013] The feature fusion module is used to fuse the dynamic change trajectory of cell morphology, the characteristics of metabolite concentration fluctuations, and the cell membrane vibration stability index to generate a multimodal feature vector.

[0014] The quality assessment module is used to input multimodal feature vectors into a trained umbilical cord blood stem cell quality assessment model and output an umbilical cord blood stem cell quality score; the umbilical cord blood stem cell quality score is used to determine whether the umbilical cord blood sample is qualified.

[0015] The traceability analysis module is used to initiate the traceability analysis process when umbilical cord blood samples are unqualified. It combines multimodal feature vectors and sample operation log data to identify the key operational steps that lead to the quality failure.

[0016] The model update module is used to update the parameters of the umbilical cord blood stem cell quality assessment model and output the optimized umbilical cord blood stem cell quality assessment model.

[0017] As a further aspect of the present invention: the specific process for processing the cell morphology image sequence in the morphological feature extraction module is as follows:

[0018] The morphological images of umbilical cord blood stem cells at different time points were continuously acquired using a microscopic imaging system to form a cell morphology image sequence. The cell morphology image sequence was preprocessed to enhance image contrast and eliminate background noise. The contour features of individual stem cells were extracted from the preprocessed images, and the temporal changes in cell area, perimeter, and roundness were calculated.

[0019] The time-series data is smoothed to remove abnormal fluctuations and generate dynamic trajectory of cell morphology changes. The dynamic trajectory of cell morphology changes includes cell expansion rate, contraction frequency and morphological stability indicators.

[0020] As a further aspect of the present invention: the specific process of analyzing the concentration change curve of metabolites in the metabolic characteristic analysis module is as follows:

[0021] The concentration of metabolites in the umbilical cord blood stem cell culture environment is monitored in real time using a microfluidic chip, and the concentration change curve of metabolites is collected. The concentration change curve of metabolites is segmented to identify the concentration rising phase, falling phase and stable phase.

[0022] Calculate the slope changes and durations of the concentration rise, fall, and plateau phases to extract metabolite concentration fluctuation characteristics; these characteristics include concentration peak intervals, fluctuation amplitudes, and trend consistency indices.

[0023] As a further aspect of the present invention: the specific process for processing the cell acoustic vibration signal in the acoustic signal processing module is as follows:

[0024] Vibration signals of umbilical cord blood stem cells in the culture environment are collected by acoustic sensors to obtain cell acoustic vibration signals; frequency domain transformation is performed on the cell acoustic vibration signals to separate the vibration energy distribution in different frequency ranges.

[0025] The proportion of energy in each frequency band to the total energy is calculated to generate a cell membrane vibration stability index; the cell membrane vibration stability index includes the energy concentration of the main frequency band and the energy difference coefficient between frequency bands.

[0026] As a further aspect of the present invention: the specific process for generating multimodal feature vectors in the feature fusion module is as follows:

[0027] The dynamic changes in cell morphology, the fluctuations in metabolite concentrations, and the stability of cell membrane vibrations were normalized to eliminate dimensional differences. Then, a feature-level fusion method was used to align the normalized features according to the time window and splice them into a multimodal feature vector.

[0028] Multimodal feature vectors include morphological feature dimensions, metabolic feature dimensions, and acoustic feature dimensions; dimensionality reduction is performed on multimodal feature vectors to retain the feature subset with the highest discriminative power.

[0029] As a further aspect of the present invention: the specific training process of the umbilical cord blood stem cell quality assessment model in the quality assessment module is as follows:

[0030] Collect multimodal feature vectors and corresponding quality labels from historical umbilical cord blood samples to form a training dataset; construct a deep neural network model, where the input layer of the deep neural network model receives multimodal feature vectors and the output layer generates quality scores;

[0031] An adaptive learning rate adjustment strategy is adopted during training, and the learning rate is dynamically updated according to the training error; the network structure and hyperparameters are selected through cross-validation; after training is completed, the model parameters and network structure are saved.

[0032] As a further aspect of the present invention: in the traceability analysis module, the specific process of the traceability analysis process is as follows: acquiring the multimodal feature vector of the unqualified umbilical cord blood sample and the sample operation log data; the sample operation log data includes the collection time, transportation temperature, storage conditions and detection operation records;

[0033] Analyze the degree of anomaly in each dimension of the multimodal feature vector and identify the main abnormal feature dimensions; perform time correlation analysis between the main abnormal feature dimensions and operation events in the sample operation log data; locate the key operation steps that cause feature anomalies and generate operation step anomaly reports; operation step anomaly reports include abnormal operation events, occurrence time and impact assessment.

[0034] As a further aspect of the present invention: the specific process of time correlation analysis in the source tracing analysis module is as follows:

[0035] The abnormal time points of the multimodal feature vectors are matched with the operation event time points in the sample operation logs; the time difference between the abnormal time points and the operation event time points is calculated, and if the time difference is less than a set threshold, it is determined to be a related event;

[0036] Accumulated analysis of related events is performed to statistically analyze the frequency and intensity of abnormal characteristics caused by the same operational event. Based on the frequency and intensity, key operational links are identified, including data acquisition delays, transportation temperature fluctuations, and detection operational errors.

[0037] As a further aspect of the present invention: the specific process of model updating in the model updating module is as follows:

[0038] Add the multimodal feature vectors and quality labels of newly collected qualified and unqualified umbilical cord blood samples to the training dataset; retrain the umbilical cord blood stem cell quality assessment model and adjust the model parameters and network weights.

[0039] Verify the evaluation accuracy and generalization performance of the updated model, retain the improved model version, and output the optimized umbilical cord blood stem cell quality assessment model.

[0040] The beneficial effects of this invention are:

[0041] This invention utilizes a multimodal data acquisition module to acquire cell morphology image sequences, metabolite concentration change curves, and cell acoustic vibration signals. A morphological feature extraction module performs spatiotemporal feature extraction on the image sequences to generate dynamic morphological change trajectories containing cell expansion rate, contraction frequency, and morphological stability indices. A metabolic feature analysis module segments the concentration curves to extract metabolic fluctuation features containing concentration peak intervals, fluctuation amplitudes, and trend consistency indices. An acoustic signal processing module performs frequency domain transformation on the vibration signals to calculate cell membrane vibration stability indices containing the main frequency band energy concentration and inter-band energy difference coefficients. Finally, a feature fusion module normalizes, aligns time windows, and reduces dimensions to generate multimodal feature vectors, comprehensively covering cell activity and structural stability. The module employs qualitative and other core quality dimensions to avoid the shortcomings of single-modal assessment. The source analysis module performs time correlation analysis between the abnormal time points of multimodal characteristics of non-compliant samples and sample operation logs containing data on collection time, transportation temperature, storage conditions, and testing operations. It calculates time differences to determine associated events and statistically analyzes frequency and intensity to pinpoint key factors such as collection delays, transportation temperature fluctuations, and testing errors. This addresses the problem of relying solely on logs to match abnormal features with operations, thus helping to optimize the sample processing flow. The model update module incorporates new sample multimodal features and quality labels for retraining, verifying assessment accuracy and generalization performance while retaining the best version. This ensures the model continuously adapts to needs, improves long-term accuracy, and provides reliable and efficient support for umbilical cord blood stem cell testing and screening. Attached Figure Description

[0042] The invention will now be further described with reference to the accompanying drawings.

[0043] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Please see Figure 1 As shown, the present invention is a multimodal human umbilical cord blood stem cell detection and screening system, comprising:

[0046] A multimodal data acquisition module is used to acquire multimodal data from human umbilical cord blood samples. The multimodal data includes cell morphology image sequences, metabolite concentration change curves, and cell acoustic vibration signals.

[0047] The morphological feature extraction module is used to extract features in the time and spatial domains from cell morphology image sequences to obtain the dynamic change trajectory of cell morphology.

[0048] The metabolic feature analysis module is used to analyze the concentration change curves of metabolites, identify abnormal fluctuation ranges in the concentration of metabolites, and extract the characteristics of the concentration fluctuations of metabolites.

[0049] The acoustic signal processing module is used to process cellular acoustic vibration signals, separate the vibration energy distribution in different frequency ranges, and calculate cell membrane vibration stability indices.

[0050] The feature fusion module is used to fuse the dynamic change trajectory of cell morphology, the characteristics of metabolite concentration fluctuations, and the cell membrane vibration stability index to generate a multimodal feature vector.

[0051] The quality assessment module is used to input multimodal feature vectors into a trained umbilical cord blood stem cell quality assessment model and output an umbilical cord blood stem cell quality score; the umbilical cord blood stem cell quality score is used to determine whether the umbilical cord blood sample is qualified.

[0052] The traceability analysis module is used to initiate the traceability analysis process when umbilical cord blood samples are unqualified. It combines multimodal feature vectors and sample operation log data to identify the key operational steps that lead to the quality failure.

[0053] The model update module is used to update the parameters of the umbilical cord blood stem cell quality assessment model and output the optimized umbilical cord blood stem cell quality assessment model.

[0054] In the morphological feature extraction module, the processing of cell morphology image sequences is achieved through a multi-step technical operation, specifically as follows: First, the microscopic imaging system is activated for image acquisition. Considering the transparency characteristics of umbilical cord blood stem cells and the requirements of the culture environment, an inverted phase-contrast microscope is selected as the core imaging device. This device can enhance the optical contrast between cells and the background through phase-contrast imaging technology, avoiding the impact of staining operations on cell viability. During the acquisition process, fixed time intervals and imaging parameters are set, such as acquiring images every 15 minutes. Each imaging session maintains consistent parameters such as light source intensity, objective magnification (usually 20x or 40x objective lens), and focal length. Continuous acquisition is performed for 24 to 72 hours. The single-frame cell morphology images acquired at different time points are arranged in chronological order to form a complete cell morphology image sequence, ensuring that the sequence can reflect the dynamic changes of cells during the culture period.

[0055] After image acquisition, the preprocessing stage begins, primarily addressing noise interference and insufficient contrast in the cell morphology image sequence. First, a Gaussian filtering algorithm is used to eliminate background noise. A pre-sized Gaussian kernel function is constructed and convolved onto a single frame image to smooth high-frequency interference signals caused by optical diffraction and sensor noise, while preserving key details of the cell contours. Subsequently, an adaptive histogram equalization algorithm is employed to enhance image contrast. This algorithm divides the image into multiple sub-regions and equalizes the gray-level histogram of each sub-region, effectively improving the gray-level difference between cell edges and cytoplasm regions. This avoids the loss of local details caused by global histogram equalization, ensuring the accuracy of subsequent contour extraction.

[0056] After preprocessing, the process moves to extracting individual stem cell contour features. First, the Canny edge detection algorithm is applied to identify edges in a single preprocessed image frame. By setting dual thresholds (high and low thresholds), true cell edge pixels are selected, eliminating false edge interference. Then, a contour tracking algorithm connects the edge pixels to form closed single-cell contours. Simultaneously, a region labeling algorithm distinguishes overlapping cells from independent cells in the image, ensuring each contour corresponds to a single stem cell. Based on the extracted closed contours, the temporal variation data of cell area, perimeter, and roundness are calculated: cell area is calculated by combining the number of pixels contained within the contour with pixel size calibration values ​​(derived from the objective parameters of the microscopic imaging system and the pixel size of the image sensor); cell perimeter is obtained by traversing the pixels on the contour and accumulating the Euclidean distances between adjacent pixels; cell roundness is calculated using the logic of "4π multiplied by the cell area and then divided by the square of the cell perimeter," reflecting the regularity of cell morphology. The area, perimeter, and roundness data calculated at different time points are organized chronologically to form complete temporal variation data.

[0057] After the time-series data is generated, it needs to be smoothed and outlier removed. A moving average algorithm is used to smooth the time-series data. A fixed-length sliding window is selected (e.g., 5 time points constitute one window). The average value of the data in each window is calculated as the smoothed value of the center time point of that window. By sliding the window, the smoothing of the entire time-series data is achieved, reducing the impact of short-term random fluctuations on the data trend. At the same time, the 3σ criterion is used to remove outlier points. The mean and standard deviation of the time-series data are calculated. Data that exceeds the range of "mean plus or minus 3 times the standard deviation" are identified as outliers, and the outliers are replaced with the linear interpolation result of two adjacent normal data points to ensure the continuity and reliability of the time-series data. Based on the processed time-series data, a dynamic trajectory of cell morphology change is generated. This trajectory includes three core parameters: cell expansion rate, contraction frequency, and morphological stability index. The cell expansion rate is calculated by dividing the difference in cell area between two adjacent time points by the time interval, reflecting the cell growth rate. The contraction frequency is obtained by counting the number of times the cell perimeter undergoes a "decline-rise" periodic change per unit time, reflecting the cell's activity state. The morphological stability index is obtained by calculating the standard deviation of cell roundness within a certain time range. The smaller the standard deviation, the more stable the cell morphology.

[0058] In the metabolic characteristic analysis module, the specific process of analyzing the concentration change curve of metabolites is achieved through a combination of microfluidic chip technology and data analysis algorithms. The steps are as follows: First, a microfluidic chip detection system is built, which includes a microfluidic chip body, a sample introduction device, a detection unit, and a data acquisition module. The microfluidic chip body is made of PDMS material and integrates a sample channel, a cell culture chamber, and a detection chamber. The size of the channel and chamber is designed to meet the culture requirements of umbilical cord blood stem cells (usually the chamber volume is in the microliter range), ensuring that the cells receive sufficient nutrient supply and gas exchange during culture. The sample introduction device uses a precision peristaltic pump to control the delivery rate of culture medium and detection reagents, avoiding interference from flow rate fluctuations on the cell culture environment. The detection unit selects corresponding sensing components according to the type of metabolite. For example, a glucose oxidase biosensor is used to monitor glucose concentration, and a lactate dehydrogenase biosensor is used to monitor lactate concentration. These sensors can convert metabolite concentration signals into electrical signals (such as current or voltage signals). The data acquisition module is connected to the detection unit, collects electrical signals in real time and converts them into metabolite concentration values. The acquisition time resolution is set to collect concentration data every 5 minutes, and continuous monitoring is carried out for 48 hours. The collected concentration values ​​are arranged in chronological order to form a metabolite concentration change curve.

[0059] After the concentration change curve is generated, it enters the segmented processing stage, where slope analysis is used to divide the curve into stages. First, the slope of the concentration change between two adjacent sampling time points is calculated. The slope calculation logic is: "concentration value at the later time point minus concentration value at the previous time point, then divided by the time interval." Three judgment intervals are set based on the slope value range: when the slope is greater than a preset positive threshold, the curve is determined to be in the concentration rising stage, which usually corresponds to active cell metabolism, the consumption of nutrients, and the generation of metabolites; when the slope is less than a preset negative threshold, the curve is determined to be in the concentration falling stage, which usually corresponds to the decomposition or removal of metabolites from the detection system; when the slope is between the positive and negative thresholds, the curve is determined to be in the concentration stable stage, indicating that the generation and consumption of metabolites are in a dynamic equilibrium. By traversing the slope values ​​of the entire concentration change curve, the start and end time points of each stage are determined, completing the segmented processing of the curve.

[0060] After segmented processing, the slope changes and durations of each stage were calculated, and the concentration fluctuation characteristics of metabolites were extracted. The slope change was calculated as the difference in slope between adjacent stages, such as the difference between the rising and stable stages, or the difference between the stable and falling stages. This parameter reflects the intensity of the metabolic process changes. The duration was obtained by subtracting the starting time from the end time of each stage, reflecting the length of each metabolic stage. Based on the above calculations, the concentration fluctuation characteristics of metabolites were extracted. These characteristics included three parameters: concentration peak interval, fluctuation amplitude, and trend consistency index. The concentration peak interval was obtained by identifying the local maximum value (i.e., concentration peak) in the concentration change curve and calculating the time difference between two adjacent concentration peaks. The fluctuation amplitude was obtained by calculating the difference between the concentration peak and the concentration trough (local minimum) within each fluctuation cycle. The trend consistency index was obtained by comparing the stage division results of the concentration change curves of various metabolites (such as glucose and lactic acid) and statistically analyzing the proportion of different curves in the same metabolic stage within the same time interval. A higher proportion indicates a stronger trend consistency in the metabolic process.

[0061] In the acoustic signal processing module, the specific process of processing cellular acoustic vibration signals relies on acoustic sensing technology and frequency domain analysis methods. The specific steps are as follows: First, a cellular acoustic vibration signal acquisition system is constructed, which includes an acoustic sensor, a signal amplification module, a filtering module, and a data acquisition card. A piezoelectric acoustic sensor is selected as the core component for signal acquisition. This sensor has high sensitivity and a wide frequency response range, and can capture the weak vibration signals generated by cells during metabolism and activity. The sensor should be installed close to the bottom of the umbilical cord blood stem cell culture container, and a flexible buffer layer should be set between the sensor and the culture container to avoid interference from external vibrations (such as equipment operation vibrations and ambient sound waves). A stable temperature (37℃) and CO2 concentration (5%) are maintained inside the culture container to ensure that the cells are in a normal physiological state and reduce the influence of environmental factors on the vibration signal. During the acquisition process, the frequency response range of the sensor is set to 20Hz to 20kHz, the sampling rate of the data acquisition card is 1kHz, and continuous acquisition is performed for 12 hours. The acquired time-domain electrical signal is stored as the original cellular acoustic vibration signal.

[0062] After the raw signal is acquired, it enters the frequency domain transformation stage. The Fast Fourier Transform (FFT) algorithm is used to process the time-domain cellular acoustic vibration signal, converting the time-domain signal into a power spectral density function in the frequency domain. This transformation converts the signal from a "time-amplitude" dimension to a "frequency-power" dimension, clearly presenting the energy distribution of the vibration signal within different frequency ranges. Based on the power spectral density function, the signal is divided into multiple frequency bands according to the frequency range, typically low-frequency (20Hz to 200Hz), mid-frequency (200Hz to 2000Hz), and high-frequency (2000Hz to 20000Hz). Each frequency band corresponds to vibration signals generated by different cellular activities. For example, low-frequency signals may be related to the overall deformation activity of the cell, mid-frequency signals may be related to the wave activity of the cell membrane, and high-frequency signals may be related to the movement of organelles within the cell. By calculating the integral value of the power spectral density function within each frequency band, the vibration energy of each band is obtained, thus clarifying the vibration energy distribution characteristics within different frequency ranges.

[0063] After calculating the energy of each frequency band, the proportion of energy in each band to the total energy is calculated, and a cell membrane vibration stability index is generated based on this. The proportion of energy in each frequency band to the total energy is calculated by dividing the integral value of energy in that frequency band by the sum of the integral values ​​of energy in all frequency bands. This proportion reflects the contribution of different vibration sources to the overall signal. The cell membrane vibration stability index includes two parameters: the energy concentration of the main frequency band and the energy difference coefficient between frequency bands. The energy concentration of the main frequency band is obtained by determining the frequency band with the highest energy in the power spectral density function (i.e., the main frequency band) and calculating the proportion of energy in that frequency band to the total energy. The higher the proportion, the more concentrated the energy of cell membrane vibration is in a specific frequency range, and the more stable the vibration mode. The energy difference coefficient between frequency bands is obtained by calculating the deviation between the energy proportion of each frequency band and the average energy proportion of all frequency bands, summing the squares of all deviations, dividing by the number of frequency bands, and then taking the square root of the result (the calculation logic is described in words, without involving formulas). The smaller this coefficient, the more uniform the energy distribution of each frequency band and the stronger the stability of cell membrane vibration.

[0064] In the feature fusion module, the specific process of generating multimodal feature vectors is achieved through data standardization, feature alignment, and dimensionality reduction. The specific steps are as follows: First, the three types of feature data—cell morphology dynamic change trajectory, metabolite concentration fluctuation characteristics, and cell membrane vibration stability index—are normalized. The purpose is to eliminate dimensional differences between different types of features. For example, the cell expansion rate in the cell morphology dynamic change trajectory may be measured in square micrometers per hour, the fluctuation amplitude in the metabolite concentration fluctuation characteristics may be measured in millimoles per liter, and the dominant frequency band energy concentration in the cell membrane vibration stability index is a dimensionless proportional value. Dimensional differences can lead to an imbalance in feature weights during subsequent fusion. The normalization process uses the Min-Max normalization method. The processing logic is to determine the minimum and maximum values ​​among all values ​​of each feature as the value range of that feature. Then, each feature value is subtracted from its minimum value, and then divided by the difference between its maximum and minimum values, so that all processed feature values ​​are within the range of 0 to 1, ensuring that various features are comparable in numerical terms.

[0065] After normalization, feature stitching is performed using a feature-level fusion method. The core step is to align the time windows of different feature types. First, a unified time window division standard is determined, for example, using one hour as a time window, dividing the entire detection period (e.g., 24 hours) into 24 consecutive time windows. For each time window, the following features are extracted: cell morphology dynamic change trajectory features (including cell expansion rate, contraction frequency, and morphological stability indicators), metabolite concentration fluctuation features (including concentration peak interval, fluctuation amplitude, and trend consistency indicators), and cell membrane vibration stability indicators (including main frequency band energy concentration and inter-band energy difference coefficient), ensuring a one-to-one correspondence between the three types of features in the time dimension. The feature stitching process arranges the three types of features within each time window in a preset order, for example, morphological features first, then metabolic features, and finally acoustic features, forming an initial multimodal feature vector. The dimension of the initial vector is determined by the number of features in each type; for example, when the three types of features contain 3, 3, and 2 feature parameters respectively, the initial multimodal feature vector has an 8-dimensional dimension.

[0066] After the initial multimodal feature vectors are generated, dimensionality reduction is required to remove redundant information between features, retain the feature subset with the highest discriminative power, and reduce the computational complexity of subsequent models. Dimensionality reduction employs principal component analysis (PCA). This algorithm constructs the covariance matrix of the features and calculates the eigenvalues ​​and eigenvectors of the covariance matrix. The magnitude of the eigenvalue reflects the variance contribution of the principal component represented by the corresponding feature vector. A higher variance contribution indicates that the principal component contains more original feature information and has stronger discriminative power. During dimensionality reduction, principal components are selected based on a preset variance contribution rate threshold (e.g., a cumulative variance contribution rate of 85%). Principal components corresponding to the eigenvalues ​​are selected sequentially from largest to smallest until the sum of the variance contribution rates of the selected principal components reaches the threshold. These selected principal components together constitute the final feature subset. This subset retains the most discriminative information from the original multimodal features while effectively reducing the feature dimensionality, providing high-quality input features for the subsequent umbilical cord blood stem cell quality assessment model.

[0067] In the quality assessment module, the training of the umbilical cord blood stem cell quality assessment model is achieved through four core steps: data preparation, network construction, training optimization, and validation and storage. The specific process is as follows: First, a training dataset is constructed, collecting multimodal feature vectors and quality labels corresponding to historical umbilical cord blood samples. The multimodal feature vectors originate from the output of the previous feature fusion module, covering core features in three dimensions: morphology, metabolism, and acoustics. Quality labels are determined based on clinically recognized umbilical cord blood stem cell quality standards, including key indicators such as cell viability, proliferation capacity, and differentiation potential. These standards are obtained by professional technicians using flow cytometry, cell counting, and other detection methods to acquire actual sample quality data, which is then mapped to corresponding quality labels (e.g., represented by a numerical value from 0 to 100). During the collection process, data preprocessing is required, including removing duplicate samples, correcting incorrectly labeled samples, and using data cleaning methods to handle missing values ​​in the feature vectors, ensuring the completeness and accuracy of the dataset. The final result is a training dataset containing a sufficient number of samples (typically covering sample types with different collection conditions and storage durations).

[0068] Next, the deep neural network model is constructed. The model structure design needs to adapt to the input requirements of multimodal feature vectors. The number of neurons in the input layer is consistent with the dimension of the multimodal feature vectors to ensure that the feature vectors can be completely input into the network. The hidden layers adopt a fully connected layer structure, with multiple hidden layers (commonly 3-5 layers) set according to the feature complexity. The number of neurons in each hidden layer is adjusted through empirical values ​​and subsequent validation. The ReLU function is selected as the activation function, which can effectively solve the gradient vanishing problem and improve the network training efficiency. The output layer uses a single neuron and a linear activation function. The output range corresponds to the value range of the quality score (e.g., 0-100), realizing the mapping from feature vectors to quality scores. During the network construction process, inter-layer connections and parameter initialization need to be implemented through programming frameworks (such as TensorFlow, PyTorch). Xavier initialization is used to ensure that the variance of the input and output of each layer is consistent, laying the foundation for subsequent training.

[0069] An adaptive learning rate adjustment strategy is employed during training, specifically implemented through the Adam optimizer. The optimizer calculates the gradient of the loss function in real time during training iterations. The mean squared error function (MSE) is used to measure the difference between the model's predicted quality score and the actual quality label. The adaptive learning rate adjustment logic is as follows: an initial base learning rate is set, and after each iteration, the learning rate is dynamically updated based on the trend of the loss function. When the loss function decreases more than a preset threshold over several consecutive iterations, the learning rate is appropriately increased to accelerate training; when the loss function decreases less than the preset threshold or shows an upward trend, the learning rate is decreased to avoid model oscillation and ensure stable convergence to the optimal solution. Simultaneously, cross-validation is used to select the network structure and hyperparameters. A k-fold cross-validation method is employed (common k values ​​range from 5 to 10). The training dataset is randomly divided into k subsets of similar size, and k-1 subsets are selected sequentially as the training set and one subset as the validation set, repeating the training and validation process k times. Based on the model performance of each validation (such as loss value and prediction accuracy on the validation set), the hyperparameters such as the number of hidden layers, number of neurons, batch size, and number of iterations are adjusted to finally determine the optimal network structure and hyperparameter combination.

[0070] After training, the model parameters and network structure are saved. The saved parameters include the weights and biases of neurons in each layer; these parameters are core to the model's quality assessment function. The saved network structure includes the number of input, hidden, and output layers and neurons, activation function types, optimizer types, and other configuration information. The saved format uses an industry-standard model file format (such as HDF5) to ensure the model can be loaded and reused in different programming environments. It also records the training dataset information, training time, and validation performance metrics, generating a model training report for future model calls and updates.

[0071] In the traceability analysis module, the traceability analysis process is implemented through four steps: data acquisition, anomaly identification, time correlation, and process localization. The specific process is as follows: First, two types of core data are acquired: the multimodal feature vector of the unqualified umbilical cord blood sample and the sample operation log data. The multimodal feature vector of the unqualified sample is obtained by filtering from the output results of the quality assessment module. At the same time, the time-series data of the feature vector needs to be extracted to reflect the changes of the feature over time. The sample operation log data contains multiple types of key information: the collection time record is the time node from the completion of umbilical cord blood sample collection from the mother to the transfer to the storage container, accurate to the minute; the transportation temperature record is the real-time temperature value collected at fixed time intervals during the transportation of the sample; the storage condition record is the environmental parameters such as storage temperature, relative humidity, and carbon dioxide concentration after the sample is stored in the cold storage; the testing operation record details the operation steps of the sample in each testing stage, the model of the instrument used, and the operator information. All log data is stored in a structured format to ensure that it can be called by subsequent analysis processes.

[0072] Subsequently, the degree of anomaly in each dimension of the multimodal feature vector was analyzed to identify the main anomalous feature dimensions. The analysis method employed a benchmark comparison approach. First, reference ranges were established for each dimension of the multimodal feature vector of normal umbilical cord blood samples. These reference ranges were determined by statistically analyzing the feature data of a large number of qualified samples, encompassing the mean and fluctuation range of each dimension. The feature values ​​of each dimension of unqualified samples were compared with the reference ranges, calculating the degree to which each dimension's feature value deviated from the mean of the reference range. The greater the deviation, the higher the degree of anomaly in that dimension. By quantifying and ranking the degree of anomaly in each dimension, the top 3-5 dimensions with the highest degree of anomaly were selected as the main anomalous feature dimensions. These dimensions are the key characteristics leading to unqualified sample quality. For example, if the cell expansion rate, peak interval of metabolite concentration, and energy concentration of the cell membrane main frequency band had the highest degree of anomaly, these were identified as the main anomalous feature dimensions.

[0073] Next, a time correlation analysis will be performed between the main abnormal feature dimensions and the operation events in the sample operation log data. First, the abnormal time points of the main abnormal feature dimensions will be determined. By analyzing the feature time series data, the time point when the feature value first exceeds the reference range will be located; this time point is the abnormal time point. Simultaneously, the time points of all operation events will be extracted from the sample operation log data. Operation events include time records of key steps such as completion of acquisition, start of transportation, end of transportation, storage, and initiation of detection. Abnormal time points will be matched one by one with the operation event time points, and the time difference between each abnormal time point and each operation event time point will be calculated. The time difference calculation logic is the absolute time difference between the two time points. A time difference threshold will be set, which will be determined based on the lag time of the impact of different operations on cell characteristics (for example, the lag time of the impact of acquisition operations on cell morphology is short, while the lag time of the impact of transportation operations on metabolic characteristics is long, and the threshold needs to be adjusted accordingly). If the time difference between a certain operation event time point and an abnormal time point is less than the set threshold, then the operation event and the feature anomaly are determined to be correlated.

[0074] Finally, the key operational steps leading to the abnormal characteristics are identified, and an operational step anomaly report is generated. Cumulative analysis is performed on related events to statistically analyze the frequency and intensity of abnormal characteristics caused by the same operational event: frequency is the number of times the operational event is identified as a related event in all source tracing analyses of non-compliant samples; intensity is the average degree to which the abnormal characteristic dimension associated with the operational event deviates from the reference range. Operational steps are ranked according to a comprehensive score of frequency and intensity (the scoring logic is that frequency weight and intensity weight are superimposed according to a preset ratio). The operational steps corresponding to the top-ranked operational events are the key operational steps. Key operational steps include acquisition delay, transportation temperature fluctuation, and detection operation error. Acquisition delay refers to the acquisition completion time exceeding the standard acquisition time limit; transportation temperature fluctuation refers to the temperature value exceeding the set temperature range during transportation; and detection operation error refers to the omission of detection steps or failure to calibrate the instrument according to procedures. The operational step anomaly report must include the specific type of abnormal operational event, the precise time of the event, and an assessment of the impact of the event on the abnormal sample characteristics (the impact is determined based on the correlation between the degree of deviation of the abnormal characteristics and clinical quality standards). The report is output in structured document format to provide a basis for subsequent sample processing workflow optimization.

[0075] The specific process of time correlation analysis is refined through four sub-steps: time point matching, time difference calculation, correlation determination, and cumulative sorting. The first step is time point matching, which involves compiling a time series list of abnormal time points for each major abnormal feature dimension of the multimodal feature vector. Simultaneously, the time points of all operation events are extracted from the sample operation logs and compiled into an operation time list, ensuring that the time format of both lists is consistent (e.g., both use the format "year-month-day hour:minute:second") for ease of subsequent calculation. The second step is time difference calculation. For each abnormal time point, the absolute time difference between each operation event time point in the operation time list is calculated, with the unit of time difference selected according to the operation type (e.g., minutes or hours) to ensure the accuracy of the difference calculation. The third step is correlation event determination. The calculation results are filtered based on a preset time difference threshold. If the time difference between an operation event time point and an abnormal time point is less than the threshold, the operation event is marked as a correlation event corresponding to that abnormal time point. The same operation event may be correlated with multiple abnormal time points. The fourth step involves cumulative analysis and ranking. This involves counting the total number of times each operational event is marked as an associated event (i.e., frequency), and simultaneously calculating the average deviation of all abnormal feature dimensions associated with that operational event from the reference range (i.e., intensity). Based on preset weighting ratios (e.g., frequency weighting 40% and intensity weighting 60%), a comprehensive score is calculated for each operational event. The events are then ranked from highest to lowest comprehensive score. The top-scoring operational events correspond to the key operational steps. These key operational steps specifically include data acquisition delays, transportation temperature fluctuations, and detection operation errors. Each key step must clearly define the corresponding operational event characteristics (e.g., the time range of data acquisition delays and the numerical range of transportation temperature fluctuations).

[0076] In the model update module, the model update process is implemented through four stages: data expansion, model retraining, performance verification, and optimization retention. The specific process is as follows: First, the training dataset is expanded by adding the multimodal feature vectors and quality labels of newly collected qualified and unqualified umbilical cord blood samples to the existing training dataset. The collection of new samples must follow the same detection process as historical samples to ensure that the collection conditions and processing methods for multimodal feature vectors are consistent with historical samples, avoiding model deviation due to differences in data sources. The determination of quality labels is also based on clinical quality standards, obtained by professionals through standardized detection methods to ensure label accuracy. During the addition process, the new sample data needs to be preprocessed, including aligning with the format of historical data, removing abnormal samples (such as samples with too many missing feature vector values ​​or contradictory label logic), and adjusting the dataset structure according to the sample quantity ratio. If the difference between the number of qualified and unqualified samples is too large, a data balancing method (such as oversampling the smaller sample category) is used to ensure a balanced distribution of dataset categories, providing a high-quality data foundation for model retraining.

[0077] Next, the umbilical cord blood stem cell quality assessment model is retrained. During retraining, the original model's network structure (e.g., input layer, hidden layer, and output layer configuration) is used first. If new sample data shows performance bottlenecks in the original structure (e.g., slow decrease in training loss), the number of hidden layers or neurons can be adjusted appropriately, but compatibility between the input and output layers must be maintained. The training process uses the same optimizer (e.g., Adam optimizer) and loss function (e.g., mean squared error function) as the original model. The initial learning rate is set to the learning rate at which the original model was completed, avoiding excessively large initial parameters that could cause model oscillations. The number of training iterations is adjusted according to the size of the new dataset, typically set to 50%-80% of the original training iterations. Simultaneously, the loss function is monitored in real-time. When the loss function remains stable at a low value for several consecutive iterations without a decreasing trend, training can be terminated early to avoid overfitting. Model parameters and network weights are dynamically adjusted during training. The parameter adjustment logic is to update the weight and bias values ​​based on the gradient direction of each iteration, ensuring the model gradually converges to the optimal solution adapted to the new data.

[0078] Then, the evaluation accuracy and generalization performance of the updated model are verified. The evaluation accuracy is verified using a test dataset, which must be partitioned from the newly collected samples (not included in the training dataset), and the number of samples must meet statistical validity requirements (usually no less than 20% of the total new samples). The accuracy calculation logic is the degree of matching between the model's predicted quality score and the actual quality label. This is achieved by calculating the error rate between the predicted and actual values ​​(an error rate below a preset threshold indicates a correct prediction), and then calculating the proportion of correctly predicted samples to the total number of samples in the test dataset. This proportion is the evaluation accuracy. Generalization performance is verified using a cross-scenario testing method. Samples with different collection conditions than the original training dataset (e.g., samples from different collection institutions or with different storage durations) are selected as the generalization test set. The prediction error rate of the model on the generalization test set is calculated; a lower error rate indicates stronger generalization performance. Simultaneously, the performance metrics of the model before and after the update are compared on the same test set and the generalization test set. If the evaluation accuracy of the updated model is higher than that of the original model, and the generalization test error rate is lower than that of the original model, then the model performance is considered improved.

[0079] Finally, retain the performance-improved model version and output the optimized umbilical cord blood stem cell quality assessment model. During the retention process, the improved model parameters (weights, biases) and network structure configuration information must be saved in a standard format. Simultaneously, relevant model update information (such as the number of new samples, training time, and performance improvement) should be recorded to generate a model update report. If the updated model performance does not meet the improvement criteria (e.g., accuracy does not improve or generalization error increases), this version should be discarded, and the original model version should be retained. The optimized model should be output in a reusable file format (e.g., HDF5 format), along with instructions for the model's API call to ensure that the clinical testing system can load and use it smoothly, achieving continuous optimization of the umbilical cord blood stem cell quality assessment capability.

[0080] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A multimodal human umbilical cord blood stem cell detection and screening system, characterized in that, include: A multimodal data acquisition module is used to acquire multimodal data from human umbilical cord blood samples. The multimodal data includes cell morphology image sequences, metabolite concentration change curves, and cell acoustic vibration signals. The morphological feature extraction module is used to extract features in the time and spatial domains from cell morphology image sequences to obtain the dynamic change trajectory of cell morphology. The metabolic feature analysis module is used to analyze the concentration change curves of metabolites, identify abnormal fluctuation ranges in the concentration of metabolites, and extract the characteristics of the concentration fluctuations of metabolites. The acoustic signal processing module is used to process cellular acoustic vibration signals, separate the vibration energy distribution in different frequency ranges, and calculate cell membrane vibration stability indices. The feature fusion module is used to fuse the dynamic change trajectory of cell morphology, the characteristics of metabolite concentration fluctuations, and the cell membrane vibration stability index to generate a multimodal feature vector. The quality assessment module is used to input multimodal feature vectors into a trained umbilical cord blood stem cell quality assessment model and output an umbilical cord blood stem cell quality score; the umbilical cord blood stem cell quality score is used to determine whether the umbilical cord blood sample is qualified. The traceability analysis module is used to initiate the traceability analysis process when umbilical cord blood samples are unqualified. It combines multimodal feature vectors and sample operation log data to identify the key operational steps that lead to the quality failure. The model update module is used to update the parameters of the umbilical cord blood stem cell quality assessment model and output the optimized umbilical cord blood stem cell quality assessment model.

2. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 1, characterized in that, The specific process for processing the cell morphology image sequence in the morphological feature extraction module is as follows: The morphological images of umbilical cord blood stem cells at different time points were continuously acquired using a microscopic imaging system to form a cell morphology image sequence. The cell morphology image sequence was preprocessed to enhance image contrast and eliminate background noise. The contour features of individual stem cells were extracted from the preprocessed images, and the temporal changes in cell area, perimeter, and roundness were calculated. The time-series data is smoothed to remove abnormal fluctuations and generate dynamic trajectory of cell morphology changes. The dynamic trajectory of cell morphology changes includes cell expansion rate, contraction frequency and morphological stability indicators.

3. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 1, characterized in that, The specific process of analyzing the concentration change curve of metabolites in the metabolic characteristic analysis module is as follows: The concentration of metabolites in the umbilical cord blood stem cell culture environment is monitored in real time using a microfluidic chip, and the concentration change curve of metabolites is collected. The concentration change curve of metabolites is segmented to identify the concentration rising phase, falling phase and stable phase. Calculate the slope changes and duration of the concentration rise, fall and plateau phases to extract the characteristics of metabolite concentration fluctuations; The characteristics of metabolite concentration fluctuations include the peak concentration interval, fluctuation amplitude, and trend consistency indicators.

4. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 1, characterized in that, The specific process for processing cellular acoustic vibration signals in the acoustic signal processing module is as follows: Vibration signals of umbilical cord blood stem cells in the culture environment are collected by acoustic sensors to obtain cell acoustic vibration signals; frequency domain transformation is performed on the cell acoustic vibration signals to separate the vibration energy distribution in different frequency ranges. The proportion of energy in each frequency band to the total energy is calculated to generate a cell membrane vibration stability index; the cell membrane vibration stability index includes the energy concentration of the main frequency band and the energy difference coefficient between frequency bands.

5. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 1, characterized in that, The specific process of generating multimodal feature vectors in the feature fusion module is as follows: The dynamic changes in cell morphology, the fluctuations in metabolite concentrations, and the stability of cell membrane vibrations were normalized to eliminate dimensional differences. Then, a feature-level fusion method was used to align the normalized features according to the time window and splice them into a multimodal feature vector. Multimodal feature vectors include morphological feature dimensions, metabolic feature dimensions, and acoustic feature dimensions; dimensionality reduction is performed on multimodal feature vectors to retain the feature subset with the highest discriminative power.

6. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 1, characterized in that, The specific training process for the umbilical cord blood stem cell quality assessment model in the quality assessment module is as follows: Collect multimodal feature vectors and corresponding quality labels from historical umbilical cord blood samples to form a training dataset; construct a deep neural network model, where the input layer of the deep neural network model receives multimodal feature vectors and the output layer generates quality scores; An adaptive learning rate adjustment strategy is used during training to dynamically update the learning rate based on training error; network structure and hyperparameters are selected through cross-validation. After training is complete, save the model parameters and network structure.

7. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 1, characterized in that, In the traceability analysis module, the specific process of the traceability analysis procedure is as follows: obtain the multimodal feature vector of the unqualified umbilical cord blood sample and the sample operation log data; the sample operation log data includes the collection time, transportation temperature, storage conditions and detection operation records; Analyze the degree of anomaly in each dimension of the multimodal feature vector and identify the main abnormal feature dimensions; perform time correlation analysis between the main abnormal feature dimensions and operation events in the sample operation log data; locate the key operation steps that cause feature anomalies and generate operation step anomaly reports; operation step anomaly reports include abnormal operation events, occurrence time and impact assessment.

8. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 7, characterized in that, The specific process of time correlation analysis in the source tracing analysis module is as follows: The abnormal time points of the multimodal feature vector are matched with the operation event time points in the sample operation log; the time difference between the abnormal time points and the operation event time points is calculated, and if the time difference is less than a set threshold, it is determined to be a related event. 、 Accumulated analysis of related events is performed to statistically analyze the frequency and intensity of abnormal characteristics caused by the same operational event. Based on the frequency and intensity, key operational links are identified, including data acquisition delays, transportation temperature fluctuations, and detection operational errors.

9. The multimodal human umbilical cord blood stem cell detection and screening system according to claim 1, characterized in that, The specific process of model updating in the model update module is as follows: Add the multimodal feature vectors and quality labels of newly collected qualified and unqualified umbilical cord blood samples to the training dataset; retrain the umbilical cord blood stem cell quality assessment model and adjust the model parameters and network weights. Verify the evaluation accuracy and generalization performance of the updated model, retain the improved model version, and output the optimized umbilical cord blood stem cell quality assessment model.

Images