TWI845274B - A cell quality prediction system and method, as well as an expert knowledge parameterization method. - Google Patents

Abstract

The present invention discloses a cell quality prediction system and method, as well as an automatic anomaly data detection method. A certain amount of retrospective data is preprocessed, expertly labeled to produce a training data set and structured data. The training data set is subjected to big data analysis or machine learning algorithms to obtain a first probability distribution model. The structured data contains expert knowledge and is used to tune the first probability distribution model to obtain a second probability distribution model. The second probability distribution model is used to predict the cell quality. The present invention is characterized in that the first probability distribution model can be indirectly modulated by a user's expert knowledge. During the prediction stage, the user can select parameters they wish to include or exclude the structured data.

Description

Cell quality prediction system, method and abnormal data automatic detection method

The present invention discloses a cell quality prediction system, method and abnormal data automatic detection method thereof. The abnormal data automatic detection method detects a certain amount of retrospective data and confirms and modifies it with the user to ensure the data quality. Then, experts annotate it to obtain the state of cell production under a specific parameter set. Then, the annotated data is calculated by big data analysis or learning algorithm to obtain a probability distribution model to predict the probability of future events. It can be applied to the fields of breeding and artificial reproduction.

According to World Population Review statistics, Taiwan's fertility rate has dropped from 3.87% to 0.18% between 1955 and 2022. The country is facing a national security crisis with a decrease in the number of newborns, an aging population structure, and a decline in national competitiveness in the future. According to statistics, the prevalence of infertility in Taiwan is over 15%, which means that one in eight couples is troubled by infertility. Fortunately, Taiwan's reproductive medicine technology is now a world leader, with a success rate that is second in the world and first in Asia. Currently, the main treatment for infertility is artificial insemination (IVF), but in traditional treatments, embryologists rely on the naked eye through a microscope to judge the quality of the blastocyst, which may cause human judgment errors due to vision, brightness, angle, personal experience and current physical condition. If AI technology can be added to learn the experience of different embryologists, their evaluation standards can be integrated to produce a scientific and objective evaluation model to reduce subjective errors.

In terms of blastocyst quality assessment, as early as 1999, there was an embryo grading system (Gardner system) that is still in use today. It grades embryos based on the degree of blastocyst expansion and hatching status, as well as the texture of its inner cell mass (ICM) and the density of trophoblast cells (TE). Although many domestic and foreign methods use AI technology to assist in grading blastocyst quality, most of them still only use a single image and do not deviate from the morphokinetic method.

Patent Publication No. I781408 of the Republic of China discloses an artificial intelligence cell detection method and system using hyperspectral data analysis technology. Its purpose is to reduce the problem that during the IVF process, doctors judge the quality of embryos subjectively, which makes it difficult to improve the success rate of infertility treatment and is easily misjudged by the subjective opinions of different doctors, resulting in a lower success rate. It uses a hyperspectrometer to continuously shoot during embryo growth to obtain several images, and uses artificial intelligence methods to analyze the changes in various chemical components produced by the embryo at two specific time points, and then detect cell quality or identify cells. Although its application scenarios and purposes are similar to those of the present invention, it only uses hyperspectrometer images for analysis, and the patent scope of the case does not teach that the system can predict cell implantation rate, while the patent scope of the present invention has disclosed a method for combining medical records, expert knowledge and structured data to predict cell implantation rate and cell quality.

Chinese Patent Publication No. CN109544512A discloses a multi-modal pregnancy prediction device, which aims to improve the success rate of test tube babies. It uses artificial methods to focus and shoot the image of the developing embryo to obtain a total of three images, namely, an image of the blastocyst, an image of the inner cell mass, and an image of the trophoblast cell. The pregnancy rate result is then used as a label to train a model that can predict the pregnancy rate result using a neural network. The method requires manual zooming of the embryo image and taking three photos with different focal lengths to predict the pregnancy rate. This method obviously has the problem of subjective recognition, which causes errors when each user takes photos with different focal lengths, making the results difficult to verify. The method of the present invention automatically detects abnormal data in medical records containing images, including data cleaning, data standardization and normalization, and automatically predicts. Users no longer need to manually adjust the acquired data during data prediction, and there is no longer the problem of human subjective judgment.

The above previous cases are all methods for improving the success rate of IVF, but there are several major differences with the present invention: both cases only disclose the prediction of blastocyst quality or pregnancy rate by imaging methods, and do not disclose the prediction method using medical records combined with expert knowledge; both cases do not disclose the method for users to select the required parameters for analysis; both cases do not disclose the method of automatically fitting the medical records added later; both cases do not disclose the method for users to set up multiple parameter intervals for multiple parameters in the data set, and then tune the probability distribution model.

The retrospective data used in this invention is based on the regulations of each institution and the habits of the recorder, resulting in different data recording methods and contents. In order to facilitate the application to various domestic and foreign institutions, this invention uses an abnormal data automatic detection method to automatically organize the retrospective data according to the user's settings, and select abnormal or outlier data for the user to confirm or modify, and for the user to confirm again.

The retrospective data used in this invention are mostly sensitive personal data involving commercial secrets or personal privacy. In order to reduce the risk of data leakage, the frequency of data transmission between users and software suppliers must be reduced as much as possible. Therefore, the present invention is characterized in that after the big data analysis or learning algorithm training is completed, an expert knowledge parameterization method and an abnormal data automatic detection method can be used to ensure data quality and adjust the probability distribution model generated after the big data analysis or learning algorithm training is completed, so that it can fit the newly added data. There is no need to regularly hand over the retrospective data to the software supplier for processing, thereby reducing the user's workload and the risk of data exchange.

In addition to automatically evaluating the cell implantation rate and the cell quality based on the medical record data and the probability distribution model, the method of the present invention can also dynamically select several parameters to be included in the analysis or excluded based on the user's experience. By using big data analysis methods, the influence of various input parameters on the prediction results can be displayed in the prediction results, and the prediction results can be interpreted by post-hoc explanations to achieve the interpretability of the method.

The effects of the present invention are: to disclose a cell quality prediction system and method thereof, which can be applied to the fields of breeding and biotechnology and medicine based on the characteristics of the training data set and the annotation method, and to use retrospective data to predict the probability of future events; to disclose an expert knowledge parameterization method, and big data analysis or learning algorithms need to disclose the data annotation quality that is highly dependent on the training data set. However, the application fields of the present invention are mostly based on experience or naked eye judgment as the event evaluation standard (expert knowledge), but experience or naked eye judgment cannot be input into the algorithm for calculation. , so a way to convert it into text is needed to use it as training data to improve the quality of labeled data; the probability distribution model can be adjusted manually or automatically to automatically fit the newly added data. In addition to reducing the risk of data leakage caused by frequently transmitting data to software vendors to retrain the probability distribution model, the most notable feature is that the probability distribution model can be customized according to users, countries, regions and usage scenarios; an abnormal data automatic detection method is disclosed to automatically filter out abnormal or outlier data for users to confirm and modify, thereby improving data quality.

The training data set and its annotation method of the present invention are exemplified as follows: if the retrospective data is maternal medical history parameters or blastocyst parameters in reproductive medicine, the annotation method is: whether the data is pregnant; if the retrospective data is data obtained from chromosome testing, the annotation method is: whether the data may have a special disease and the name of the special disease.

The method of using expert knowledge to adjust the structured data of the present invention is that, given at least one probability distribution model, the user sets a plurality of parameter intervals for a plurality of parameters in the training data set according to his expert knowledge, or excludes specific parameters in the training data set to adjust the probability distribution model.

In a feasible embodiment (please refer to FIG. 1), the present invention proposes a cell quality prediction system, method and abnormal data automatic detection method (a), which includes a cell carrier (b), a lens module (c), medical record data (d), a storage unit (e), a computing unit (f) and a user interface (g).

In the present invention, the cell carrier (b) can be a test tube or a culture dish.

The sensing module (c) in the present invention refers to a lens module capable of autofocusing, and its focusing method can be optical, digital or hybrid zoom, such as an electronic microscope.

The medical record data (d) in the present invention refers to the parameters obtained during the IVF treatment process and the images obtained by the sensing module (c) during the embryo growth process. The parameters here include but are not limited to clinical data, age, BMI, female hormone (E2), preimplantation chromosome screening (PGS/PGT-A) data, preimplantation genetic diagnosis (PGD/PGT-M) data, chromosome karyotype analysis data, amniotic fluid chip (Array Comparative Genomic Hybridization, aCGH) data, non-invasive fetal chromosome gene detection data (NIFTY/NIPT/NIPS/niPGS/niPGT-A), chromosome number, luteinizing hormone (LH), follicle stimulating hormone (FSH), sperm count, sperm motility, sperm morphology, active sperm density, total active sperm, semen liquefaction time, pH value, anti-Mullerian hormone (AMH), endometrial thickness, number of follicles, progesterone (P4), implantation method, blastocyst quality rating (Gardner Grading System), beta-chorionic gonadotropin, single picture and TLI time-lapse image (Time Lapse Incubator).

The above parameters can be dynamically added and deleted according to user needs, and automatically perform data cleaning, data standardization and normalization.

Data cleaning methods at least include: screening out duplicate data, irrelevant data, extreme values, noise, missing data or structurally incorrect data, and deleting, filling in values, taking the average of previous and subsequent values, and filling in with approximate values.

Data standardization methods include at least: Minimization and maximization algorithm (Max-Min), standard score (Z-Score), maximum absolute value standardization (MaxAbs), Robust Scaler, mean (Means), standard deviation (Standard Deviation) and algorithms that make the data decimal between 0 and 1

Data normalization methods at least include Label Encoding, One-hot Encoding, and algorithms that make the mean value of the data 0 and the standard deviation 1, which are used by information engineers or data scientists to perform big data analysis or training learning algorithms to obtain optimized probability distribution models.

The storage unit (e) in the present invention includes at least a hard disk (HDD), a solid state hard disk (SSD) and a memory (RAM).

The computing unit (f) in the present invention includes at least a central processing unit (CPU), a graphics processing unit (GPU) and an accelerated processing unit (APU).

The storage unit € and computing unit (f) in the present invention can be mounted on one or a combination of a personal computer, an industrial computer, a computer cluster, cloud computing, a laptop computer, a mobile phone or an edge computing device.

The big data analysis and learning algorithms mentioned in this invention include: Backpropagation, supervised learning, semi-supervised learning, ensemble learning, active learning, reinforcement learning, generative model, discriminative model, long short-term memory, object detection, instance segmentation and diffusion model.

The generative models mentioned above include at least Gaussian Mixture Model, Maximum Likelihood Estimation, Hidden Markov Model, and Naive Bayes classifier.

The discriminant model mentioned above includes at least logistic regression, linear regression, support vector machine, decision tree, and eXtreme Gradient Boosting.

The abnormal data automatic detection method of the present invention utilizes the aforementioned data cleaning method and data normalization method to automatically screen abnormal or outlier data, and cooperates with the following methods to achieve the functions of automatically fitting the original data, reducing the data dimension, accelerating the computing performance, reducing the computing equipment requirements and improving the system interpretability, which at least includes genetic algorithm (Genetic Algorithm), data clustering (Clustering), image processing, edge detection (edge detection), Chi-square test (Chi-square test), EM algorithm (Expectation-Maximization Algorithm).

The present invention provides a cell quality prediction system, method and abnormal data automatic detection method, which at least includes the following steps:

(I) Retrospective data is provided by users, who can be practitioners in related fields or persons with relevant licenses.

(Ⅱ) Information engineers or data scientists formulate automatic detection methods for abnormal data based on the characteristics of retrospective data.

(III) Automatically detect abnormal data in the retrospective data, and use experts in related fields to confirm, modify and annotate the data to form a training data set for structured data and big data analysis or learning algorithm training.

(III) Input the training data set into the aforementioned big data analysis or learning algorithm to perform learning training, obtain a first probability distribution model and a first prediction model, and use the first prediction model and the first probability distribution model to calculate a first prediction result.

(IV) Using the structured data generated by the anomaly data automatic detection method in step (III), or the user setting multiple parameter intervals for multiple parameters in the training data set according to his expert knowledge, or excluding specific parameters in the training data set, the probability distribution model is adjusted within the parameter interval range set by the user, so that the abstract expert knowledge is incorporated into the algorithm as a parameter to adjust the first probability distribution model in step (III) to generate a second probability distribution model and a second prediction model.

(V) Inputting the retrospective data in step (I) into the second prediction model and the second probability distribution model in step (IV) to obtain a second prediction result, which at least includes the cell implantation rate, the cell quality and the influence of various input parameters on the prediction result, wherein the parameters input into the second prediction model and the second probability distribution model can be selected by the user in step (IV).

The structured data in step (IV) includes text data, digital data, image data, time series data, embryo images, data grouping results, and multiple parameter intervals of multiple parameters automatically defined by the system or by the user.

The prediction results in step (V) have multiple prediction output values, including: cell implantation rate prediction, cell growth time point, object detection, instance segmentation, object counting, blastocyst zona pellucida thickness, trophoblast cell number, trophoblast cell total area ratio, inner cell mass total area ratio, cell size, cell division uniformity and cell fragmentation degree, and cell quality assessment, where the cell quality assessment method adopts the Gardner blastocyst grading system.

The multiple predicted output values can be obtained by concatenating the predicted output values of multiple single images into a time series data.

The cell implantation rate of the present invention is obtained by analyzing the cell growth time point, object detection, instance segmentation, object counting, blastocyst zona pellucida thickness, trophoblast cell number, trophoblast cell total area ratio, inner cell mass total area ratio, cell size, cell division uniformity and cell fragmentation degree obtained by big data analysis or learning algorithms.

The method of automatically fitting the newly added data in the present invention is: when new data is generated, the data is automatically detected as abnormal data in step (III), and then handed over to the user for confirmation, modification and annotation, and then the structured data in the modification step (IV) is used to adjust the second probability distribution model.

Examples of multiple parameter ranges for multiple parameters defined automatically by the system or by the user are: if the training data set is related to aquaculture, the parameter range can be an appropriate range of water temperature, pH value or dissolved oxygen; if the training data set is related to reproductive medicine, the parameter range can be an appropriate range of age, BMI or AMH, where multiple ranges are allowed, for example, age is defined as [18-24] and [30-40].

An example of modifying the first probability distribution model and the first prediction model into the second probability distribution model and the second prediction model is: if the probability distribution model is a hidden Markov model, the probability value of each state in the Markov chain is modified by the parameter interval; if the probability distribution model is a Bayesian classifier, the maximum likelihood estimate is modified by the parameter interval.

This invention will generate multiple probability distribution models based on the aforementioned big data analysis and learning algorithms, and then present the prediction results in an independent display or integrated learning manner.

The ensemble learning method at least includes averaging, taking the maximum value, taking the minimum value, or giving a valve value to binarize the results generated by multiple algorithms, and performing one or a combination of voting methods for the final result.

The cell quality prediction system of the present invention has a user interface, which is used for users to operate and present the system execution results in the user interface. It can load and calculate multiple learning algorithms, and can automatically adjust the size of the probability distribution model running in the system according to the specifications of the learning algorithm or the hardware device to optimize the system computing performance.

The user interface can be a text-only interface or a graphical interface.

The operations provided by the user interface include: switching between the foreground and background of the user interface, user login, user logout, modifying user information, inputting text, selecting files to upload, selecting or excluding parameters to be input into the learning algorithm, and using expert knowledge to set multiple parameter ranges for multiple parameters in the training data set.

The system execution results of the user interface include: training data sets, prediction results of learning algorithms, visual data, sound data, descriptive text or digital data, or a combination thereof.

a: Cell quality prediction system, method and automatic detection method of abnormal data

b: Cell carrier

c:Sensor module

d: Original data

e: Storage unit

f: Operational unit

g: User interface

g1: Check the combination of block components

g2: Combination of text label component and input field component

g3: Image display area

g4: Comparison of results of multiple prediction methods

g5: Information display area

g6: Detailed information of multiple prediction results

g7: button component

g8: button component

g9: Text label component

g10: Slider component

h1: Data points in the training data set

h2: Coverage of the original probability distribution model

h3: Newly added data points

h4: Modulated probability distribution model

i1:A The user’s expert knowledge considers the appropriate numerical range

i2: PDF before modulation

i3:B The user’s expert knowledge considers the appropriate numerical range

i4: Modulated PDF

The first figure is a system block diagram of the cell quality prediction system, method and abnormal data automatic detection method of the present invention

The second figure is a schematic diagram of the front-end interface of the cell quality prediction system, method and abnormal data automatic detection method of the present invention.

The third figure is a schematic diagram of the enlarged view of the information interface of the cell quality prediction system, method and abnormal data automatic detection method of the present invention.

The fourth figure is an enlarged view of the prediction result interface of the cell quality prediction system, method and abnormal data automatic detection method of the present invention

The fourth figure is a schematic diagram of the background interface of the cell quality prediction system, method and abnormal data automatic detection method of the present invention

Cell quality prediction system, method and abnormal data automatic detection method

Please refer to the first figure. The user uses the sensing module (c) to manually take pictures with an electron microscope, or uses the embryo time-lapse monitoring incubator to take pictures at fixed intervals to take time-lapse pictures of the embryos cultured in the cell carrier (b).

After the embryo develops to the blastocyst stage (usually 120 hours), the user operates on the user interface (g), and the system automatically combines the aforementioned images and medical records into a retrospective data (original data). After the abnormal data is automatically detected, the detection result is provided for the user to confirm or modify, and is stored in a storage unit (e). Then, the calculation unit (f) performs calculations to obtain the predicted result.

If the image is a single image, at least the embryo information in the single image can be obtained, including the number and area ratio of the trophoblast cells (TE) and inner cell mass (ICM) in the embryo and the thickness of the embryonic zona pellucida.

If the image taken is a continuous image (video), such as a time-lapse image, in addition to the aforementioned information, at least the time points of changes in the embryonic growth process can be obtained, including but not limited to: the time point from cleavage to two to eight cells (t2~t8), the time point from fertilization to morula (tM), the time point from fertilization to early blastocyst, blastocyst and expanded blastocyst (tSB, tB, tEB).

The above results can be presented in the user interface (g) for user reference, and users can make changes by text or voice input. If the user confirms that the information is correct, it will be automatically saved to the storage unit (e) together with the original information and the information changed by the user for future reference.

The dotted lines in the second to fifth figures are the parts that are not designed, and the rest are the parts that are designed, but the position, size and distribution relationship between the appearance and the environment are not included.

Please refer to the second figure, where number (g) represents the user interface, number (g1) is a combination of checkboxes in a form component; number (g2) is a combination of a text label and an input field in a form component; number (g3) is an image display area; number (g4) is a comparison of the results of multiple prediction methods; number (g5) is an information display area; number (g6) is detailed information of multiple prediction results; and numbers (g7) and (g8) are button components.

Please refer to the second figure, where the component numbered (g2) is not a design advocated part. It is only used for illustration of this embodiment and can be replaced by any component, but it must at least include text display and text input functions.

Please refer to the second and third figures, where the figure numbered (g4) is not a part of the design advocated and is only used to illustrate the visual display of multiple different information in this embodiment. It can be transformed into but not limited to: a bar chart, a pie chart, a line chart and an area chart.

Please refer to the second figure, where the component numbered (g7) is not a design part, it is only used for illustration of this embodiment and can be replaced by any component with a click event function.

Please refer to the second to fifth figures, where the component numbered (g8) is not a design part, which is only used for illustration of this embodiment and can be replaced by any component with a click event function.

Please refer to the second figure for the operation of the user interface (g) of the cell quality prediction system and method and its abnormal data automatic detection method in this case. The user can enter the medical record number in the input field (g2). If the medical record number data entered has been archived, the system will automatically display other parameters in the fields of (g2). Then the user can select the parameters to be included or excluded in the calculation in the check box (g1), and then click the button (g7) to perform the calculation.

If the entered medical record number has not been created, the user can manually enter other information into other fields in the input field (g2) and click the button (g7) to perform the calculation.

After clicking the button (g7), the calculation results will be displayed in the image display area (g3), the result comparison of multiple prediction methods (g4), the information display area (g5) and the detailed information of multiple prediction results (g6).

Detailed information (g6) of multiple prediction results includes at least the prediction results (g4) of multiple probability distribution models and the range of values of each parameter in the input field (g2), as well as possible reasons for the prediction results.

Please refer to the third picture, which is a design in which the image display block (g3) and the information display block (g5) in the second picture are enlarged independently by clicking on the image display block (g3) in the second picture, so that the appearance of the user interface (g) changes. This function allows users to view multimedia information more clearly.

Please refer to the fourth figure, which is a design that clicks on the result comparison of multiple prediction methods (g4) in the second figure, so that the result comparison of multiple prediction methods (g4) and the detailed information of multiple prediction results (g6) in the second figure are enlarged independently, causing the appearance of the user interface (g) to change. This function allows users to more clearly view the prediction results, the distribution range of input parameters and the possible reasons for the prediction results.

Please refer to the fifth figure, which shows a design in which the appearance of the user interface (g) changes between the second figure and the fifth figure by clicking the button component (g8) in the user interface (g). This function allows the user to switch to the area where various parameters are set in the background.

Please refer to the fifth figure, in which number (g) is the user interface; number (g8) is a button component; number (g9) is the text label (label) in the form component; number (g10) is the slider component in the form.

Please refer to the fifth figure, where the component numbered (g9) is not a design part, which is only used for illustration of this embodiment and can be replaced by any component with text display function.

Please refer to the fifth figure, where the component numbered (g10) is not a design part, it is only used for illustration of this embodiment and can be replaced by any component that can input or set a value.

Please refer to Figure 5 for the cell quality prediction system, method and abnormal data automatic detection method of this case, its user interface (g) and the operation method of its background page. The text label component (g9) will display the name of each parameter. The user can set the parameter value to multiple ranges in the slider component (g10). Each range can be defined by the user. For example, a parameter in a certain range represents a good, moderate or poor state. After the user completes the setting, he can click the button component (g8) to switch to the front page, and then click the button component (g7) to execute the operation. The system will respond to the final prediction result according to the range adjusted by the user and present it on the user interface (g).

Example of how to adjust the probability distribution model

Please refer to Figure 6, which shows the method of adjusting the probability distribution model. In the figure, (h1) is the data point in the training data set, (h2) is the coverage range of the original probability distribution model, (h3) is the newly added data point, and (h4) is the adjusted probability distribution model. The implementation method is to change the coverage range of the probability distribution model to cover the newly added data points.

Please refer to Figure 7, which uses the Probability Distribution Function (PDF) to illustrate a simple example of a single parameter, where the horizontal axis x is the value of the parameter, the vertical axis y is the probability of the event, (i1) is the range of values considered appropriate by user A’s expert knowledge, (i2) is the PDF before the adjustment, (i3) is the range of values considered appropriate by user B’s expert knowledge, and (i4) is the PDF after the adjustment. If user A’s expert knowledge considers the parameter range appropriate, The interval is [1,10], and the trained PDF is (i2). Another user B, an expert, believes that the appropriate parameter range is [6,20]. At this time, it is only necessary to change the parameter range in the system to [6,20], and the system will adjust the mean and standard deviation for fitting, and automatically adjust the PDF (i2) before adjustment to (i4), so as to achieve the effect of customizing the probability distribution model according to different users.

The protection scope of this creation should not be limited to what is disclosed in the embodiments, but should include various substitutions and modifications that do not deviate from this creation, and are covered by the following patent application scope.

a: Cell quality prediction system, method and automatic detection method of abnormal data

b: Cell carrier

c:Sensor module

d: Original data

e: Storage unit

f: Operational unit

g: User interface

Claims (26)

A method for predicting cell quality is provided, wherein abnormal data of at least one patient's medical history data is automatically detected to generate a training data set and a structured data, and then the training data set is operated by a learning algorithm to generate a first probability distribution model and a first prediction model, and the first prediction model and the first probability distribution model are used to calculate a first prediction result, and the structured data with expert knowledge is modulated to obtain a second probability distribution model and a second prediction model, and the second prediction model and the second probability distribution model are used to calculate a first prediction result. The invention is characterized in that a user with expert knowledge can select parameters to be included or excluded from the structured data, and adjust the structured data to indirectly adjust the first probability distribution model and the first prediction model; wherein the structured data is adjusted by the user to set a plurality of parameter intervals for a plurality of parameters in the training data set or to exclude specific parameters in the training data set according to his expert knowledge, wherein each interval can be defined and adjusted by the user. The method for predicting cell quality as described in claim 1, wherein the medical history data includes at least one or a combination of age, BMI, female hormone, luteinizing hormone, follicle stimulating hormone, semen liquefaction time, pH value, anti-Mullerian hormone, endometrial thickness, number of follicles, progesterone, implantation method, blastocyst quality rating, beta-chorionic gonadotropin, sperm motility, total amount of active sperm and semen liquefaction time. The cell quality prediction method as described in claim 1, wherein the learning algorithm comprises at least one of back propagation algorithm, big data analysis, image classification, object detection, generative model, discriminative model, long short-term memory, reinforcement learning, generative pre-trained transformation model and ensemble learning or a combination thereof. The cell quality prediction method as described in claim 3, wherein the generative model includes at least one of a Gaussian mixture model, a maximum likelihood estimation, a hidden Markov model, a simple Bayesian classifier, and an algorithm that can use retrospective data to calculate the probability of future events, or a combination thereof. The cell quality prediction method as described in claim 3, wherein the discriminant model comprises at least one of logical regression, linear regression, support vector machine, artificial neural network, decision tree, and extreme gradient boosting, or a combination thereof. The cell quality prediction method as described in claim 3, wherein the ensemble learning at least includes averaging, taking the maximum value, taking the minimum value, or giving a threshold value to binarize the results generated by multiple algorithms, and performing one or a combination of the voting methods for the final result. The cell quality prediction method as described in claim 1, wherein the user includes at least one of practitioners in the relevant field and persons with relevant licenses, or a combination thereof. The cell quality prediction method as described in claim 1, wherein the structured data at least includes one or a combination of text data, digital data, image data, time series data, embryo images, and data grouping results with corresponding relationships. The cell quality prediction method as described in claim 1, wherein the second prediction result includes at least one of the cell implantation rate, the cell quality, and the influence of various input parameters on the prediction result, or a combination thereof. The cell quality prediction method as described in claim item 9, wherein the cell implantation rate is obtained by using the learning algorithm to predict the medical record data, obtaining a plurality of predicted output values and analyzing them. The cell quality prediction method as described in claim 10, wherein the prediction output value includes at least one of the results of cell growth time point, object detection, instance segmentation, object counting, blastocyst zona pellucida thickness, trophoblast cell number, trophoblast cell total area ratio, inner cell mass total area ratio, cell size, cell division uniformity and cell fragmentation degree, or a combination thereof. As described in claim 11, the cell quality prediction method, wherein the predicted output value can be obtained by concatenating the predicted output values of multiple single images into a time series data. As described in claim 11, the cell quality prediction method, wherein the cell growth time point is obtained by analyzing the medical record data using image processing or machine learning, and includes at least one of the time points of cell cleavage into two cells, three cells, four cells, five cells, six cells, seven cells, eight cells, fertilization to morula, fertilization to early blastocyst, fertilization to blastocyst, and fertilization to expanded blastocyst, or a combination thereof. A method for parameterizing expert knowledge, the method comprises at least one training data set, at least one expert and at least one given probability distribution model. The method is characterized in that the expert sets a plurality of parameter intervals for a plurality of parameters in the training data set according to his expert knowledge, or excludes specific parameters in the training data set to generate a structured data, and uses the structured data to adjust the probability distribution model so that the abstract expert knowledge is incorporated into the algorithm as a parameter; wherein the expert includes at least one of an information engineer, a data scientist, a practitioner in a field related to the data set, a professional with a relevant license, or a combination thereof. A method for automatically detecting abnormal data, which requires at least one training data set and at least one expert in a field related to the training data set. The method is characterized by automatically screening out abnormal or outlier data in the training data set and aggregating them through data cleaning, data standardization and data normalization methods for the expert to confirm or modify. The automatic screening method is that the user sets multiple parameter intervals for multiple parameters in the training data set according to his expert knowledge. The method for automatically detecting abnormal data as described in claim 15, wherein the pre-processing method includes one or a combination of data cleaning, data standardization, data normalization, genetic algorithm, data clustering, image processing, edge detection, chi-square test, and EM algorithm. The method for automatically detecting abnormal data as described in claim item 15, wherein the data cleaning method at least includes screening out duplicate data, irrelevant data, extreme values, noise, missing data or structurally incorrect data, and deleting, filling in values, averaging previous and subsequent values, filling in with approximate values, or a combination thereof. The method for automatically detecting abnormal data as described in claim 15, wherein the data normalization method includes at least one of the mini-max algorithm, standard score, maximum absolute value normalization, Robust Scaler, mean value, standard deviation, and an algorithm that causes the data to be a decimal between 0 and 1, or a combination thereof. The method for automatically detecting abnormal data as described in claim 15, wherein the data normalization method is at least one of Label Encoding, One-hot Encoding, and an algorithm that makes the mean value of the data equal to 0 and the standard deviation equal to 1, or a combination thereof. , a cell quality prediction system, the system at least includes: a hardware device for operation; a user interface for users to operate and present the system execution results in the user interface; a patient's medical history data; an algorithm; a training data set; and a structured data, which is characterized by: the training data set is operated in the hardware device with a learning algorithm to generate a first probability distribution model and a first prediction model, and the first prediction model is used to generate a first probability distribution model and a first prediction model. The first probability distribution model calculates the first prediction result. The user sets a plurality of parameter intervals for the plurality of parameters in the training data set according to his expert knowledge and selects the parameters to be included or excluded from the structured data to adjust the structured data. The first probability model is then adjusted by the structured data to obtain a second probability distribution model and a second prediction model. The second prediction model and the second probability distribution model are used to calculate the second prediction result. A cell quality prediction system as described in claim 20, wherein the hardware device comprises at least one or a combination of a storage unit, a computing unit, a personal computer, an industrial computer, a computer cluster, cloud computing, a laptop computer, a mobile phone or an edge computing device. A cell quality prediction system as described in claim 20, wherein the storage unit comprises one or a combination of a hard drive, a solid state drive, and a memory. The cell quality prediction system as described in claim 20, wherein the computing unit includes one or a combination of a central processing unit, a graphics processing unit, and an acceleration processing unit. A cell quality prediction system as described in claim 20, wherein the user interface comprises at least one of a pure text interface and a graphical interface or a combination thereof. The cell quality prediction system as described in claim 20, wherein the operation at least includes the foreground and background switching of the user interface, user login, user logout, user information modification, text input, file selection upload, selection or exclusion of parameters input to the machine learning algorithm, and using expert knowledge to set multiple parameter intervals for multiple parameters in the training data set, one of them or a combination thereof. A cell quality prediction system as described in claim 20, wherein the system execution result includes at least the original data for the learning algorithm to operate, the prediction result of the learning algorithm, visualized data, sound data, descriptive text or digital data, or one or a combination thereof.

Images