Intelligent explanation method for inspection result of mart based on AI algorithm

Abstract

The invention relates to the technical field of clinical laboratory digital intelligence, and particularly discloses an intelligent explanation method for an inspection result based on an AI algorithm mart, and the process comprises the steps: constructing N disease data sets, and training N AI disease prediction models based on the N disease data sets to obtain N AI disease prediction models. A plurality of disease prediction models form an AI market and are applied to a real scene, a plurality of disease prediction models meeting conditions are triggered based on a target patient inspection result to generate a prediction result, a decision-making path is automatically adjusted by using a decision-making tree algorithm, the prediction results of the plurality of models are fused to obtain an optimal explanation, and the model performance is continuously monitored through a sliding window. And when the performance of the model is reduced, re-training or parameter adjustment is triggered in time so as to ensure the accuracy and stability of a prediction result. According to the method, through multi-model collaboration, dynamic decision and full-process AI algorithm digital closed-loop management, the comprehensive interpretation capability of clinical examination results can be improved, and the examination value is further improved.

Description

Technical Field

[0001] The present application relates to the field of digital technology for clinical laboratories, and more specifically, to a method for intelligent interpretation of test results based on an AI algorithm marketplace. Background Art

[0002] Clinical laboratories are crucial support departments for healthcare institutions, with approximately 70% of clinical decision-making information derived from laboratory tests. With the continuous advancement of medical technology, the number of test items is increasing, the spectrum of diseases is becoming increasingly complex, and individual patient differences are significant, making the interpretation of test results a major challenge in clinical work. Traditional interpretation of test results relies heavily on the clinician's extensive professional knowledge and clinical experience. This is especially true when interpreting interdisciplinary, multi-dimensional, or abnormal combination test data, requiring the physician to possess a deep background in specialized or even sub-specialized fields. This results in inefficient interpretation and is prone to subjective bias or information omissions when dealing with rare diseases, early-stage diseases, or atypical presentations, making it difficult to provide a comprehensive, accurate, and standardized interpretation.

[0003] Although new-generation information technologies such as AI, big data, and big models provide an intelligent foundation for the laboratory information ecosystem, it is still difficult to systematically implement digital closed-loop management of the entire process of result interpretation based on AI algorithms, including building high-quality disease datasets from massive test big data, conducting effective feature engineering, training, validating, evaluating, and optimizing models, and ultimately implementing their application, deployment, and continuous supervision in real clinical scenarios. Currently, even if some AI-driven test result interpretation models can be built, they are often limited to single models for specific diseases or specific test items. However, the manifestation of a disease is often the result of the combined effects of multiple factors and multiple indicators. Due to its inherent limitations or biased training data, a single AI model or algorithm may have biased interpretation results, lack stability, or even be completely unable to provide effective predictions or explanations when faced with patients with different types of diseases, different age groups, or different clinical characteristics, making it difficult to meet complex and changing clinical needs.

[0004] Therefore, an optimized method for intelligent interpretation of test results based on the AI ​​algorithm marketplace is expected. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an intelligent interpretation method for test results based on the AI ​​algorithm mart, which extracts and classifies test big data, constructs N disease data sets, trains the data sets to obtain N AI disease prediction models, and applies them to real scenarios after model interpretation and verification. Based on the test results of target patients, multiple disease prediction models that meet the conditions are triggered to generate prediction results. At the same time, the decision path is automatically adjusted by using a fusion device based on a decision tree algorithm, and the prediction results of multiple models are fused to obtain the best explanation. The model performance is monitored through a sliding window, and retraining or parameter adjustment is triggered in time when the model performance declines to ensure the accuracy and stability of the prediction results. This method can improve the comprehensive interpretation ability of clinical test results and give full play to its clinical value through multi-model collaboration and dynamic decision-making.

[0006] Accordingly, according to one aspect of the present application, a method for intelligent interpretation of test results based on an AI algorithm mart is provided, comprising: Construct N disease datasets; Based on the N disease data sets, N AI disease prediction models are trained to obtain N AI disease prediction models Obtain test results for target patients; Inputting the test results of the target patient into the N disease prediction models respectively to obtain multiple AI model prediction results; An algorithm fusion device based on a decision tree algorithm is used to obtain a decision path by combining the prediction results of multiple AI models of the target patient; The prediction results of the multiple AI models are processed based on the decision path to obtain the best explanation result.

[0007] In the above-mentioned AI algorithm mart-based test result intelligent interpretation method, N disease datasets are constructed, including: Medical records were collected based on the inclusion and exclusion criteria to obtain a data set including the disease group, healthy control group, and disease control group; Based on the medical record number in the data set, query the corresponding patient's medical records, test items and results to obtain the parameters of the disease data set; performing dataset cleaning on the parameters of the disease dataset to obtain a cleaned disease dataset; The cleaned disease dataset is subjected to dataset quantity evaluation, feature parameter selection, model dataset determination, and data desensitization processing to obtain the disease dataset.

[0008] In the above-mentioned test result intelligent interpretation method based on the AI ​​algorithm mart, based on the N disease data sets, N AI disease prediction models are trained to obtain N AI disease prediction models, including: Identifying and selecting a subset of input variables that are most correlated with the target variable from the training samples of the N disease datasets; Dividing the N disease datasets into a training set and a validation set according to a predetermined ratio; Training N AI disease prediction models based on the training set and performing model verification and evaluation on the N AI disease prediction models based on the verification set; Add additional explanations for each of the N AI disease prediction models.

[0009] In the above-mentioned AI algorithm mart-based test result intelligent interpretation method, identifying and selecting a subset of input variables most relevant to the target variable from the training samples of the N disease datasets includes: The preliminary input variable set was determined by rank sum test, with the patient's diagnosis type as the dependent variable and the test items and results as the independent variables; Using the preliminary input variable set as an independent variable and the patient diagnosis type as a dependent variable, performing feature parameter selection based on the XGBoost model to obtain a screened input variable set; The filtered input variable set is pruned and combined to obtain the most relevant input variable subset.

[0010] In the above-mentioned intelligent interpretation method of test results based on the AI ​​algorithm mart, the preset ratio between the training set and the validation set is 7:3 or 8:2.

[0011] In the above-mentioned intelligent interpretation method of test results based on the AI ​​algorithm mart, N AI disease prediction models are trained based on the training set and N AI disease prediction models are verified and evaluated based on the verification set. Each model evaluation indicator includes AUC, accuracy, precision, recall rate, F1 score, sensitivity, specificity, PPV, NPV, and AUC_CI.

[0012] In the above-mentioned test result intelligent interpretation method based on the AI ​​algorithm mart, the test results of the target patient are respectively input into the N AI disease prediction models to obtain N AI model prediction results, including: In response to the test results of the target patient meeting the model parameter triggering condition, the test results of the target patient are input into the corresponding AI disease prediction model to obtain the AI ​​model prediction result.

[0013] In the above-mentioned intelligent interpretation method of test results based on the AI ​​algorithm mart, the test results of the target patient are input into an algorithm fusion device based on a decision tree algorithm to obtain a decision path, including: the algorithm fusion device based on the decision tree algorithm automatically adjusts the decision path based on the patient's gender, age, symptoms, model performance, and matching degree with the patient's test items.

[0014] In the above-mentioned method for intelligent interpretation of test results based on the AI ​​algorithm mart, the method further includes: Continuously monitoring performance data of each of the N AI disease prediction models in real-world big data; Setting key indicator thresholds based on historical performance data of each of the AI ​​disease prediction models; Using a sliding window method to perform segmented evaluation on the performance data of each AI disease prediction model in real-world big data to obtain evaluation results; In response to the evaluation result that the average performance within the window is less than the key indicator threshold, a prompt signal for model retraining or parameter adjustment is triggered.

[0015] Compared with the existing technology, the intelligent interpretation method of test results based on the AI ​​algorithm mart provided in this application extracts and classifies test big data, constructs N disease data sets, trains the data sets to obtain N AI disease prediction models, and applies them to real scenarios after model interpretation and verification. Based on the test results of target patients, multiple disease prediction models that meet the conditions are triggered to generate prediction results. At the same time, a fusion device based on a decision tree algorithm is used to automatically adjust the decision path, fuse the prediction results of multiple models to obtain the best explanation, and monitor the model performance through a sliding window. When the model performance declines, retraining or parameter adjustment is triggered in time to ensure the accuracy and stability of the prediction results. This method can improve the comprehensive interpretation ability of clinical test results through multi-model collaboration and dynamic decision-making, and digital closed-loop management of AI algorithms throughout the entire process, further promote the digital application of clinical testing, and give full play to its clinical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 This is a flowchart of a method for intelligent interpretation of inspection results based on an AI algorithm mart according to an embodiment of the present application.

[0018] Figure 2 This is a flowchart of step S1 in the method for intelligent interpretation of inspection results based on the AI ​​algorithm mart according to an embodiment of the present application.

[0019] Figure 3 This is a flowchart of step S2 in the method for intelligent interpretation of inspection results based on the AI ​​algorithm mart according to an embodiment of the present application.

[0020] Figure 4 Schematic diagram of model evaluation indicators according to an embodiment of the present invention.

[0021] Figure 5 Schematic diagram of unified management of model parameter platform according to an embodiment of the present application.

[0022] Figure 6 Schematic diagram of the correspondence between model parameters and inspection item codes according to an embodiment of the present application.

[0023] Figure 7 Schematic diagram of the predicted result probability return according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0025] Figure 1 Flowchart of the method for intelligent interpretation of test results based on AI algorithm mart according to an embodiment of the present application. Figure 1 As shown, according to the embodiment of the present application, the intelligent interpretation method of test results based on the AI ​​algorithm mart includes the following steps: S1, constructing N disease data sets; S2, based on the N disease data sets, training N AI disease prediction models to obtain N AI disease prediction models; S3, obtaining the test results of the target patients; S4, inputting the test results of the target patients into the N disease prediction models respectively to obtain multiple AI model prediction results; S5, the algorithm fusion of the multiple AI model prediction results of the target patients is based on the decision tree algorithm to obtain a decision path; S6, processing the multiple AI model prediction results based on the decision path to obtain the best interpretation result.

[0026] In the above-mentioned intelligent interpretation method of test results based on the AI ​​algorithm mart, the step S1 is to construct N disease data sets. It should be understood that since the training of AI models is highly dependent on large amounts of high-quality and representative data, and the existing test big data, although huge in volume, is directly used for model training, it often faces problems such as data dispersion, non-uniform format, a large amount of noise and missing values. Therefore, in order to establish a high-quality disease data set for AI model training, this application extracts case data of disease groups, healthy controls, and different disease control groups from massive clinical test big data by setting inclusion and exclusion criteria and applying intelligent data extraction and classification mechanisms to establish N disease data sets.

[0027] Figure 2 FIG. 1 is a flow chart of step S1 in the method for intelligent interpretation of test results based on the AI ​​algorithm mart according to an embodiment of the present application. Figure 2 As shown, step S1 includes: S11, collecting disease medical records based on inclusion and exclusion criteria to obtain a data set including N disease groups, a healthy control group and a disease control group; S12, querying the medical records and test items and results of the corresponding patients based on the medical record number in the data set to obtain the parameters of the disease data set; S13, performing data set cleaning on the parameters of the disease data set to obtain the cleaned N disease data sets; S14, performing data set quantity evaluation, feature parameter selection, model data set determination and data desensitization processing on the cleaned disease data set to obtain the N disease data sets.

[0028] Specifically, step S11 involves collecting medical records based on inclusion and exclusion criteria to obtain a dataset containing N disease groups, a healthy control group, and a disease control group. The principles for selecting disease groups include: common clinical diseases with test indicators that have clear clinical value in the disease; a reasonable prediction time window for the disease that satisfies the causal relationship; and sufficient and representative clinical case resources are available. In other words, case screening is performed by setting a date range and diagnostic criteria. In the specific implementation process, cases are first labeled as the target disease group (labeled 1) and other disease control groups (labeled 2-5) based on the selected date range and disease diagnosis information in the medical records. At the same time, test information on healthy people without disease diagnoses is selected from physical examination or outpatient data to serve as the healthy control group (labeled 0). For example, in constructing a bile duct cancer dataset, patients newly diagnosed with bile duct cancer between a certain date and a certain year were included as the disease group, excluding those with other malignancies. Healthy individuals undergoing physical examinations during the same period served as the healthy control group, excluding samples with carbohydrate antigen 19-9 levels greater than 37 U / mL. Patients with other diseases, such as gallstones, acute cholecystitis, and primary biliary cirrhosis, were also included as the disease control group. This hierarchical labeling approach ensures clear distinction between positive and negative samples and interference categories in the dataset, providing a reliable labeling system for subsequent model learning of disease-specific features and avoiding model misjudgments due to case confounding.

[0029] Specifically, step S12 queries the corresponding patient's medical records, test items, and results based on the medical record number in the dataset to obtain the parameters of the disease dataset. To extract numerical features from the cases that the AI ​​model can understand and process, this application integrates multi-source data using the medical record number as the primary key based on the principles of data association analysis. Specifically, for each case's medical record number in the dataset, the corresponding patient's basic information, such as age, gender, disease name, diagnosis time, race, and education level, lifestyle and environmental information such as dietary habits, smoking status, exercise frequency, and living environment, test results such as blood routine, biochemical, and immune indicators, as well as imaging findings such as CT and MRI features and ultrasound imaging indicators, medical history indicators such as symptoms, signs, and medication history, and genetic characteristics such as family history, are queried. For multiple test results for the same patient, the earliest or most recent value is selected for the numerical result. The average or mode of multiple test results can also be taken. Averaging is not currently supported for text test results. After the value extraction is completed, the test item ID is used as a field and the result as a value to be added to the patient's result data. Next, the patient's test items are sorted in descending order of frequency. If the number of test items exceeds 500, the first 500 are selected (for example, in the bile duct cancer dataset, high-frequency indicators such as carbohydrate antigen 19-9, carcinoembryonic antigen, alkaline phosphatase, glutamyl transpeptidase, albumin, lymphocytes (%), and direct bilirubin are retained). Finally, the Chinese names of the test items are retrieved based on the test item ID and added to the Excel header array, ensuring that they correspond to the results of each test item mentioned above. When exporting, these are filled into the corresponding cells, forming a dataset containing the test results. This method helps provide structured input for subsequent feature engineering and model training.

[0030] Specifically, in step S13, the parameters of the disease data set are cleaned to obtain a cleaned disease data set. It should be understood that since there are often outliers, missing values ​​and duplicate data in the original test data, direct use will lead to model training deviation or overfitting, affecting the prediction accuracy. Therefore, in order to improve data quality, this application uses statistical methods and rule engines to identify and process the problem data in the disease data set and perform manual verification. Specifically, first, the detection methods and units are unified, the qualitative / quantitative mixed data are classified according to clinical thresholds, and the outliers outside the detection limit are standardized to eliminate the interference of technical differences on the model. Secondly, outliers are detected using the IQR (interquartile range) or Z-score method. For example, samples with white blood cell counts exceeding ±3 standard deviations of the mean are labeled as abnormal. Based on clinical justification, the data can be deleted, corrected (e.g., by referencing historical results), or retained (e.g., to confirm true outliers). Thirdly, for missing values, if the missing rate for a test item is >30%, the item is deleted directly. Otherwise, model-driven methods such as the median, mean, or random forest are used to fill in missing values ​​(e.g., missing albumin values ​​are filled in with the mean of healthy individuals of the same gender). Finally, parameters may need to be properly transformed or transformed to address class imbalance. Common methods include logarithmic transformation, polynomial incorporation, spline incorporation, standardization, principal component extraction, or variable combination. Through data cleaning and exploration, data noise can be effectively reduced to ensure the quality, accuracy, and consistency of the dataset.

[0031] Specifically, in step S14, the cleaned disease dataset is subjected to dataset quantity assessment, feature parameter selection, model dataset determination, and data desensitization processing to obtain the disease dataset. It should be understood that the size and feature quality of the dataset directly affect the generalization ability of the model, and clinical data involve patient privacy. Therefore, in order to balance model performance and data compliance, this application is based on the principles of sample size estimation and feature importance analysis, combined with ethical requirements for multi-dimensional processing. Specifically, (1) Dataset quantity assessment: The sample size is estimated based on the disease incidence, the number of feature variables, and the complexity of the model. It is usually required that the number of disease cases is 10-20 times the number of features (e.g., 60 features require at least 600 cases), or learning curves generated by training sets of different sizes are used to assess whether the amount of training data is sufficient. The higher the model complexity, the greater the sample requirement to reduce the risk of overfitting. In addition, the sample size of the training set must follow the principle of balance, with a moderate ratio of positive and negative classes to avoid the model tending to predict large classes. (2) Feature parameter screening: Recent data should be selected first to ensure synchronization with the current diagnosis and treatment environment; at the same time, considering feature accessibility and standardization, select projects that are routinely carried out by each institution and parameters that are homogeneous across different institutions. Feature parameter screening converts data into features that can better represent potential problems, thereby improving the performance of the model, including feature enhancement, feature construction, feature selection, feature conversion, etc. Feature parameter screening is based on empirical consensus, such as literature review and expert advice, or it can be data-driven, that is, screening using various statistical or machine learning algorithms. (3) Model data set determination: The automatically collected data set is compared with the original medical records to verify data accuracy (such as whether the test results are logically consistent with the diagnosis), parameter integrity (such as whether key indicators are omitted) and disease knowledge compliance (such as whether liver cancer patients include alpha-fetoprotein indicators) to ensure that the data conforms to clinical logic. (4) Data desensitization: Sensitive information such as patient names and ID numbers are anonymized, such as replacing real names with hash values, to ensure data fairness, impartiality and non-discrimination.

[0032] In the above-mentioned method for intelligent interpretation of test results based on the AI ​​algorithm mart, the step S2, based on the disease data set, trains N disease prediction models to obtain N AI disease prediction models. It should be understood that since a single AI model or algorithm often has inherent limitations, it may be unstable or insufficiently explained when faced with different disease types, patient characteristics or data patterns, and it is difficult to provide a comprehensive and reliable diagnosis or prediction. Therefore, in order to build an AI algorithm mart with a wider range of applications, this application is based on the principles of ensemble learning and multi-model strategies, and utilizes the disease data set to select and train multiple AI disease prediction models to build an AI algorithm mart covering multiple disease types, different age groups and clinical characteristics, to meet the multidimensional heterogeneity of disease states and the dynamic complexity of test results.

[0033] Figure 3 FIG. 1 is a flow chart of step S2 in the method for intelligent interpretation of test results based on the AI ​​algorithm mart according to an embodiment of the present application. Figure 3 As shown, the step S2 includes: S21, identifying and selecting a subset of input variables that are most relevant to the target variable from the training samples of the certain disease data set; S22, dividing the certain disease data set into a training set and a validation set according to a predetermined ratio; S23, constructing a certain disease prediction model based on the training set using machine learning, and verifying and evaluating the certain disease prediction model based on the validation set; S24, adding additional explanations to the certain AI disease prediction model; and repeating steps S21-S24 to eventually form N disease prediction models.

[0034] Specifically, step S21 identifies and selects the subset of input variables most relevant to the target variable from the training samples of the disease dataset. It should be understood that since the original disease dataset often contains a large number of redundant or irrelevant features, directly using them for model training will lead to increased computational complexity, increased risk of overfitting, and difficulty in highlighting key disease indicators. Therefore, in order to reduce the dimension of the feature space and improve model training efficiency and interpretability, this application further screens the subset of input variables most relevant to the target variable based on statistical methods and machine learning algorithms.

[0035] In a specific example of the present application, step S21 includes: using a certain diagnosis type as the dependent variable and the test items and results as independent variables, determining a preliminary input variable set through hypothesis testing or the contribution of model prediction performance; using the preliminary input variable set as the independent variable and a certain diagnosis type as the dependent variable, based on empirical consensus or data-driven, selecting feature parameters to obtain a screened input variable set; deleting and combining the screened input variable set to obtain the most relevant input variable subset.

[0036] Specifically, the method first uses a specific diagnostic type (such as cholangiocarcinoma, health checkup, gallstones, acute cholecystitis, and primary biliary cirrhosis) as the dependent variable and all numerical test items as independent variables. Hypothesis tests (such as between-group mean differences and distribution independence) such as t-tests, Wilcoxon rank sum tests, ANOVA, and chi-square tests are used to screen for features that are significantly associated with the target variable. Alternatively, machine learning-based strategies (such as information gain and random forest importance) are used to quantify the contribution of features to model predictive performance. Alternatively, a combination of hypothesis testing and machine learning is used, with this step repeated through multiple deletions and combinations to identify the subset of input variables most relevant to the target variable. This screening mechanism effectively eliminates irrelevant features while retaining core indicators such as carbohydrate antigen 199, carcinoembryonic antigen, and alkaline phosphatase (a key marker for cholangiocarcinoma). This allows the model to focus on features that contribute most to disease diagnosis and enhances the causal relationship between features and disease.

[0037] Secondly, due to the significant differences in the adaptability of different AI models to the number of features and sample size (such as linear models are suitable for low-dimensional small samples, and deep learning models are suitable for high-dimensional large samples), a single model cannot take into account all data scenarios, which may lead to model bias or underfitting. Therefore, in order to ensure that the model matches the data features and improve prediction accuracy, this application establishes a data feature-model selection dynamic matching mechanism based on the model adaptation principle. In a specific example of this application, step S21 includes: when the number of features in the input variable subset or the training sample size of the disease dataset is less than or equal to a preset threshold, selecting a linear regression model, a logistic regression model, and a COX regression clinical prediction model as the AI ​​disease prediction model; when the number of features in the input variable subset or the training sample size of the disease dataset is greater than the preset threshold, selecting SVM, Linear SVC, Decision Tree, Random Forest, Ada Boost, GradientBoosting, XGBoost, Bernoulli NB, Gaussian NB, K Nearest Neighbors, Linear Discriminant Analysis, SGD, and Multilayer Perceptron as AI disease prediction models.

[0038] Specifically, a threshold for the number of features or training sample size is set (e.g., 50 features, 1000 training samples). If the number of features in the input variable subset is ≤ the threshold and the training sample size is small (e.g., <1000 cases), linear regression, logistic regression, and Cox regression clinical prediction models with simple logic and strong interpretability are selected as AI disease prediction models. If the number of features is greater than the threshold or the training sample size is sufficient (e.g., >3000 cases), support vector machines (SVMs), linear SVCs, decision trees, random forests, Ada Boost, gradient boosting, XGBoost, Bernoulli NB, Gaussian NB, K Nearest Neighbors, linear discriminant analysis, SGD, and multilayer perceptrons are selected as AI disease prediction models (e.g., high-dimensional test metrics for cholangiocarcinoma datasets). This on-demand model selection strategy avoids overfitting caused by large models with small data or information loss caused by small models with large data, helping to ensure that each model performs optimally in its adapted data scenario and forming a comprehensive model portfolio covering diverse data characteristics.

[0039] In another preferred embodiment of the present application, step S21 includes: taking a certain diagnosis type as the dependent variable and the test items and results as independent variables, determining a preliminary input variable set through hypothesis testing or the contribution of model prediction performance; and performing variable cyclic distillation based on distribution redundancy constraints on the initial input variable set to obtain the most relevant input variable subset.

[0040] Specifically, in this preferred embodiment, the initial set of input variables is subjected to variable cyclic distillation based on distributional redundancy constraints to obtain the most relevant subset of input variables, including the following process: First, the initial set of input variables is arranged into an initial input variable vector. It should be understood that in the initial stages of building an AI disease prediction model, in order to accurately extract the core indicators that are truly decisive for the diagnosis of a specific disease from the massive amount of raw test data, which is often mixed with redundancy and noise, this preferred embodiment proposes a variable cyclic distillation method based on distributional redundancy constraints to identify and select the subset of input variables most relevant to the target variable. This process is not a one-time rigid screening process, but rather draws on the core concept of adaptive computational depth in feature cyclic refinement networks, elevating variable selection into a dynamic, intelligent polishing process driven by information entropy convergence. Its purpose is to eliminate wasteful computational resources while maximizing the discovery of deep connections hidden between data variables, ensuring that the raw materials fed to downstream AI models are the most efficient and informative.

[0041] The process begins by structuring the initial set of input variables selected for building a disease model—for example, a complex set containing patient demographic information and dozens or even hundreds of clinical test results—into a unified, machine-readable vector of initial input variables. This transformation is performed because the disordered and heterogeneous nature of the original data set precludes direct processing by the subsequent refinement modules. Its purpose is to create a standardized, vectorizable starting state for the entire cyclic distillation process, laying the foundation for all subsequent iterative calculations.

[0042] Next, the initial input variable vector undergoes variable refinement to produce a refined input variable vector. This initial input variable vector is then fed into a core variable refinement module. This is done because the initial vector contains a large number of noise variables that are weakly or even irrelevant to disease diagnosis and fails to explicitly represent the nonlinear synergistic effects between variables. Directly using this vector would significantly increase the model's computational burden, prolong the training cycle, and easily lead to overfitting, thereby compromising the model's generalization and ultimate predictive accuracy. Therefore, the purpose of this variable refinement step is to perform a nonlinear remapping of the entire variable space using a deep transformation composed of learnable parameters (such as an attention mechanism or a multi-layer perceptron). This process is equivalent to a collective deliberation and reweighting of the importance of all variables, aiming to simulate the thought process of human experts conducting comprehensive analysis when interpreting complex cases. The result is a refined input variable vector in which the representation of key indicators strongly related to the disease (such as specific tumor markers or biochemical indicators) is significantly enhanced, while the representation of irrelevant or redundant variables is effectively suppressed, resulting in a clearer differentiation of the diagnostic value between variables.

[0043] In a specific example, the process can be expressed as follows: in, is the learnable weight matrix, is the learnable bias vector, is the random disturbance coefficient, Introducing random noise to improve robustness, is the initial input variable vector.

[0044] Next, the variable distribution redundancy of the refined input variable vector relative to the initial input variable vector is calculated. However, the essence of this methodology lies in its unique feedback control loop, which immediately self-evaluates information gain after each refinement operation. Specifically, the system calculates the variable distribution redundancy of the newly generated refined input variable vector relative to its predecessor—the variable vector before entering the refinement module. The rationale behind this calculation stems from the need for an objective criterion to judge the effectiveness of a single refinement operation: whether it has introduced new and valuable insights into the variable set. Its purpose is to quantify this change, shifting the decision to continue the refinement process from relying on a fixed, empirical number of iterations to one driven by the actual changes in the data's representation. The result is a feedback signal: a low redundancy value indicates a significant and beneficial reconstruction of the diagnostic information distribution of the variable vector, indicating that further refinement is necessary. Conversely, an extremely high redundancy approaching a threshold indicates that the variable representation has reached a fixed point, and subsequent refinement operations will no longer provide substantial information gain, indicating that the system has reached representation equilibrium within the current model's capabilities.

[0045] In a specific example, the variable distribution redundancy is calculated as follows: in, is the mutual information of the refined input variable vector relative to the initial input variable vector, is the information entropy of the initial input variable vector; In response to the variable distribution redundancy satisfying a preset condition, the refined input variable vector is used as the most relevant input variable subset. Ultimately, this feedback signal based on variable distribution redundancy constitutes the core strategy of the loop iteration. This solution presets a high redundancy threshold as the criterion for loop termination. In response to the calculated variable distribution redundancy not meeting this preset condition (i.e., the redundancy is below the threshold, indicating that there is still room for refinement), the loop continues, using the newly generated refined vector as the input for the next round of further refinement. This design is implemented to achieve on-demand allocation of computing resources, avoiding excessive computation on simple, easily feature-identifiable disease datasets while ensuring sufficient computing cycles for complex disease datasets that require in-depth exploration of multivariate associations. The purpose of this implementation is to dynamically find the optimal variable representation fixed point for each disease model through this self-supervised iteration. Once the variable distribution redundancy satisfies the preset condition (i.e., exceeds the threshold), the variable set has been fully distilled and its inherent, core diagnostic information patterns have been stably captured, and the loop process terminates. At this point, the last refined input variable vector generated, which has been proven to be the most discriminative within the current model framework and cannot be significantly optimized further, is officially designated as the most relevant input variable subset. The ultimate result of this move is that we not only efficiently complete feature selection, but more importantly, the resulting subset is not the result of a simple screening, but an optimized representation refined through deep learning and intelligent polishing, achieving an elegant balance between efficiency and performance, laying a solid foundation for the subsequent construction of high-precision and highly robust AI disease prediction models.

[0046] In a specific example, the process can be expressed as follows: in, is the target redundancy, is the tolerance.

[0047] Specifically, in step S22, the disease dataset is divided into a training set and a validation set according to a predetermined ratio. Specifically, in order to scientifically evaluate the generalization ability of the model, the present application divides the disease dataset into a subset for training and a subset for validation based on the cross-validation principle in statistics. In a specific example of the present application, the preset ratio between the training set and the validation set is 7:3 or 8:2. In addition, when the data volume of the disease dataset is less than 1000 cases, K-fold cross-validation is used by default (dividing the dataset into K parts, using K-1 parts for training and 1 part for validation each time, and taking the average of the K cycles), the K value is 5 or 10, and the user is allowed to customize the split ratio. For example, in a small sample diabetes dataset (n=800), 10-fold cross-validation is used, using 720 cases for training and 80 cases for validation each time to avoid deviations caused by random division.

[0048] Specifically, in step S23, N AI disease prediction models are trained based on the training set and N AI disease prediction models are verified and evaluated based on the verification set. In a specific example of the present application, Figure 4 As shown, model evaluation metrics include AUC, accuracy, precision, recall, F1 score, sensitivity, specificity, PPV, NPV, and AUC_CI. To ensure the model's disease prediction capabilities and clinical applicability, this application optimizes model parameters through iterative training and multi-dimensional evaluation based on supervised learning principles. In the specific implementation process, each model's parameters (such as the regression coefficient for logistic regression and the tree depth and learning rate for XGBoost) are first learned using the training set. Hyperparameters are then tuned using grid search or random search (for example, adjusting the number of decision trees in a random forest from 50 to 200). The trained model is then fed into the validation set, and model performance is evaluated using metrics such as AUC, accuracy, precision, recall, F1 score, sensitivity, specificity, PPV, NPV, and AUC_CI. Five-fold or ten-fold cross-validation is performed to validate model selection and hyperparameter tuning. Furthermore, receiver operating characteristic (ROC) curves and confusion matrices of different models are compared to analyze differences in model performance in distinguishing disease from health and between different disease types. For models with poor performance (e.g., AUC < 0.7), we re-adjust the feature set or change the model type (e.g., switching from SVM to Gradient Boosting) until they meet clinical application standards. This closed-loop process of training, validation, and tuning ensures that each model has good fit and generalization capabilities on the test data.

[0049] Specifically, step S24 adds additional explanations to each of the N AI disease prediction models. It should be understood that due to the complexity of machine learning models and the characteristics of black-box algorithms, their prediction results are difficult for clinicians to understand and trust, which may affect the model's practical acceptance. Therefore, to enhance the model's interpretability and clinical credibility, this application, based on Explainable Artificial Intelligence (XAI) technology, provides transparent explanations for each model. In specific implementation, the degree of causal relationship between the machine learning model's input features and output predictions needs to be explained. This helps doctors and patients understand the logic behind the model's decisions, builds trust, and helps verify the model's reliability. For example, SHAP values ​​or the LIME algorithm can be used to explain how elevated levels of carbohydrate antigen 19-9 affect the risk of cholangiocarcinoma. Furthermore, a bar chart is used to display the ranking of the contribution of each feature variable to disease prediction (for example, in the cholangiocarcinoma model, carbohydrate antigen 19-9 accounts for 25%, significantly higher than other indicators). Furthermore, the dose-response relationship is demonstrated by using restricted cubic spline functions to fit the nonlinear relationship between characteristic variables and outcome risks, for example, showing a trend of 15% increase in the risk of bile duct cancer for every 10 years of age increase. Finally, xlsx files can be uploaded to the system to generate model files such as PMML, and the above explanatory information can be embedded in the model file or visualization interface for doctors to refer to when calling the model. In this way, the prediction logic of the model is converted into a medical language that can be understood by the clinician. For example, doctors can identify key diagnostic indicators through feature importance and judge the clinical significance of test results through dose-response relationships, which helps to enhance trust in the model's prediction results and promote the practical application of AI models in clinical decision-making.

[0050] In addition, this application also supports unified management of AI market parameters, such as Figure 5 As shown, it supports platform-based unified addition, modification, and deletion of AI algorithms or models and parameters, and can customize names, detailed descriptions, parameter configurations, operating environments, model files (pmml), WebService configurations, etc. In addition, since different AI models require different parameters, each model needs to set its own parameters. By entering the corresponding test item code after each parameter, the model parameters and LIS test items can be bound. This helps doctors directly obtain test data from the LIS system when calling the model, and automatically fill it into the model input parameters, thereby improving the convenience and efficiency of model use. Figure 6 As shown in the figure, after the parameter settings are completed, the test results can be input to test the model.

[0051] In the aforementioned AI algorithm-based intelligent interpretation method for test results, step S3 involves obtaining the test results of the target patient. It should be understood that clinical decisions require real-time individual patient test data. Therefore, to enable accurate predictions for specific patients using the AI ​​model, this application obtains the patient's test results through a laboratory information system (LIS). In a specific embodiment of this application, the test results of the target patient include basic information such as gender, age, and symptoms, as well as test item results (e.g., carbohydrate antigen 19-9, carcinoembryonic antigen, alkaline phosphatase, glutamyl transpeptidase, albumin, direct bilirubin, etc.).

[0052] In the above-mentioned method for intelligent interpretation of test results based on the AI ​​algorithm mart, the step S4 inputs the test results of the target patient into the N AI disease prediction models respectively to obtain multiple AI model prediction results. In a specific example of the present application, the step S4 includes: in response to the test results of the target patient meeting the model parameter triggering condition, the test results of the target patient are input into the corresponding AI disease prediction model to obtain the AI ​​model prediction result. Specifically, since the N AI disease prediction models may use different parameters (such as gender, age, symptoms, test results and other data) to calculate the possible probability of a certain disease or risk, when the test results of the target patient are passed to the AI ​​algorithm mart, it is first determined whether the test results of the target patient meet the triggering conditions of each model in the AI ​​algorithm mart (for example, a certain model may only be triggered for specific test items or patients of a specific age group). Figure 7 As shown, once the conditions are met, the patient's test results are used as input to invoke the corresponding AI disease prediction model for calculation. Each model independently returns a prediction result (probability of disease or risk score) for a specific disease or risk based on its training logic and patient data. By running multiple AI disease prediction models in parallel, it is possible to cover multiple dimensions of disease characteristics, avoiding the limitations of a single model and generating a multi-faceted set of prediction results.

[0053] In the above-mentioned method for intelligent interpretation of test results based on the AI ​​algorithm mart, in step S5, the test results of the target patient are input into an algorithm fusion device based on a decision tree algorithm to obtain a decision path. In a specific example of the present application, the algorithm fusion device based on the decision tree algorithm automatically adjusts the decision path based on the patient's gender, age, symptoms, model performance, and matching degree with the patient's test items. That is, given that different AI models may be good at processing different types of data, their prediction capabilities, scope of application, and matching degree with current patient characteristics are different, and simply averaging or voting the prediction results of multiple models cannot effectively utilize the patient's individualized clinical information to optimize the fusion process. Therefore, in order to dynamically select or combine the prediction results of multiple models based on the specific characteristics of the patient and the applicability and performance of each AI model to generate the best explanation with more clinical guidance significance, the present application is based on the logical judgment and rule reasoning principles of the decision tree, and constructs an algorithm fusion device based on the decision tree algorithm to perform branch judgment on patient parameters, thereby determining the dynamic decision structure of the subsequent processing flow.

[0054] During the specific implementation process, the algorithm fusion device based on the decision tree algorithm gives priority to AI models that perform better in the corresponding population based on the patient's gender, age, and symptoms. For example, a certain AI disease prediction model has been verified to be highly effective in diagnosing lung diseases in elderly men. When encountering an elderly male patient with related symptoms, the decision tree will give the model a higher weight. At the same time, the performance of the model (that is, the verification results of the AI ​​model in the training phase or the continuous monitoring performance data in real-world applications) serves as an important reference for selecting the model. The decision tree will give priority to calling models that perform well in predicting specific diseases. In addition, different AI models may rely on different test items as input features. When obtaining the test results of the target patient, the decision tree determines the fusion weight of the model by judging whether the core input features required by a model are complete in the current patient data, and may even completely exclude its prediction results and rely on models with more complete input data.

[0055] By making judgments based on these multiple factors, the algorithm amalgamator can traverse the optimal decision path, revealing the optimal result processing strategy for the current patient and the current model state. For example, the decision path might indicate: "For patients with this gender / age / symptom combination, and models A, B, and C have all been triggered and returned results, given that model A has the best validation performance in this population and all its required input items are present in the patient data, please prioritize the prediction results of model A." Or "Given that the prediction results of models B and C are similar, and the key test items they rely on are all well matched with the patient data, please average the prediction results of models B and C as the final prediction conclusion."

[0056] In the above-mentioned intelligent interpretation method of test results based on the AI ​​algorithm mart, step S6 processes the N AI model prediction results based on the decision path to obtain the best interpretation result. That is, based on the guidance of the decision path, corresponding operations are performed on the N AI model prediction results, which may include but are not limited to: weighted averaging of multiple model results according to the weights specified by the path, directly selecting the best model output indicated by the path, or combining the outputs of multiple models and applying specific thresholds or rules for judgment, and combining the final prediction result obtained after processing (such as the risk probability of a certain disease) with the additional explanatory information added during model training to form a complete intelligent interpretation of the patient's test results. In this way, the clinical applicability of the interpretation results can be improved, assisting doctors to quickly locate high-risk diseases, reducing misdiagnosis and missed diagnosis, and at the same time enhancing the credibility of the interpretation through decision path tracing.

[0057] In one embodiment of the present application, the intelligent interpretation method of test results based on the AI ​​algorithm mart also includes: continuously monitoring the performance data of each AI disease prediction model in the N AI disease prediction models in real-world big data; setting key indicator thresholds based on the historical performance data of each AI disease prediction model; using a sliding window method to perform segmented evaluation of the performance data of each AI disease prediction model in real-world big data to obtain an evaluation result; in response to the evaluation result that the average performance within the window is less than the key indicator threshold, triggering a prompt signal for model retraining or parameter adjustment.

[0058] It should be understood that since the distribution of real-world data may change over time, model performance may decline due to data drift. Therefore, in order to ensure the stability of the model in long-term applications, this application dynamically evaluates the model performance through a sliding window based on the principles of real-time monitoring and adaptive optimization. Specifically, first, continuously monitor the operating performance and prediction results of each AI disease prediction model in real-world big data, collect its performance data, and the historical performance data of the AI ​​model before it is put into use or during the stable operation period, and set key performance indicator thresholds (such as a decrease in accuracy AUC of >5%) as a standard for judging whether the model needs intervention. Then, use the sliding window method to perform segmented evaluation on the performance data collected in real time, calculate the average performance of the model over a recent period of time (within the window), and if the average performance degradation within the window reaches the preset key indicator threshold, it triggers model retraining or parameter adjustment. Among them, the accuracy evaluation calculation formula is as follows: in, It is the present moment, is the window size. If accuracy If the average value of the previous window drops by more than the threshold, a performance drift alert is triggered.

[0059] In addition, when an expert scores an AI model 0, the model is marked as "temporarily disabled". When the mean expert score is lower than the threshold, the model is re-evaluated, and the parameters are retrained and optimized if necessary. The model can be manually disabled through the unified management of the AI ​​algorithm market. At the same time, the AI ​​algorithm market unified management model described in this application can record and query the latest performance data of the model, including activation time, update time, update reason, patient usage, etc., which helps administrators quickly understand the usage status and performance of the model for subsequent maintenance and optimization.

[0060] The aforementioned AI algorithm marketplace-based intelligent interpretation method for test results also implements digital closed-loop management of AI algorithms throughout the entire process, from identifying N diseases, collecting disease datasets, building AI disease prediction models, evaluating and validating algorithm performance, to clinical application and oversight. Clinical laboratories possess multidimensional and multimodal test data, but this data, which involves patient privacy, strictly complies with regulations such as the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law, making medical data sharing difficult. Within medical institutions, disease categories and prediction or diagnostic needs are determined based on patient populations, specialty characteristics, and treatment plans. Disease databases are constructed based on real-world data to meet the personalized needs of medical institutions. A management framework for AI predictive models in the testing field should be established, encompassing data exploration, feature selection, algorithm implementation, model evaluation, model validation and interpretation, model application, and oversight. A multidimensional evaluation system encompassing clinical outcomes, health economics, and user experience should be established, and the clinical value of the models should be validated through real-world studies. Digital closed-loop management of AI algorithms can solve problems such as data security protection requirements, personalized AI algorithm training, data drift and concept drift, insufficient number of prediction models, and few clinical applications, and enhance clinical trust and acceptance of AI algorithm decision-making. Laboratory medicine is ushering in a period of transformation driven by data intelligence, which is expected to fully unleash the medical value of data elements and help precision medicine enter a more efficient and intelligent development stage.

[0061] In summary, according to the embodiment of the present application, the intelligent interpretation method of the test results based on the AI ​​algorithm mart is explained, which constructs a data set of N diseases by extracting and classifying the test big data, and. Based on a certain disease data set, after feature selection and extraction, model training and verification, model evaluation and interpretation, a certain disease prediction model is established. Numerous disease prediction models form an AI mart and are applied to real scenarios. Based on the test results of target patients, the AI ​​disease prediction model that meets the conditions is triggered to generate prediction results. At the same time, the decision path is automatically adjusted by the fusion device based on the decision tree algorithm, and the prediction results of multiple models are integrated to obtain the best explanation. The model performance is monitored through a sliding window. When the model performance declines, retraining or parameter adjustment is triggered in time to ensure the accuracy and stability of the prediction results. This method can improve the comprehensive interpretation ability of clinical test results through multi-model collaboration and dynamic decision-making, and digital closed-loop management of AI algorithms throughout the entire process, further promote the digital application of clinical tests, and give play to the clinical value of clinical tests.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent interpretation of test results based on AI algorithm marketplace, characterized in that: include: Construct N disease datasets; Based on the N disease data sets, training N AI disease prediction models to obtain N AI disease prediction models; Obtain test results for target patients; Inputting the test results of the target patient into the N disease prediction models respectively to obtain multiple AI model prediction results; An algorithm fusion device based on a decision tree algorithm is used to obtain a decision path by combining the prediction results of multiple AI models of the target patient; The prediction results of the multiple AI models are processed based on the decision path to obtain the best explanation result.

2. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 1, characterized in that: Construct N disease datasets, including: Medical records were collected based on the inclusion and exclusion criteria to obtain a data set including the disease group, healthy control group, and disease control group; Based on the medical record number in the data set, query the corresponding patient's medical records, test items and results to obtain the parameters of the disease data set; performing dataset cleaning on the parameters of the disease dataset to obtain a cleaned disease dataset; The cleaned disease dataset is subjected to dataset quantity evaluation, feature parameter selection, model dataset determination, and data desensitization processing to obtain the disease dataset.

3. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 1, characterized in that: Based on the N disease data sets, N AI disease prediction models are trained to obtain N AI disease prediction models, including: Identifying and selecting a subset of input variables that are most correlated with the target variable from the training samples of the N disease datasets; Dividing the N disease datasets into a training set and a validation set according to a predetermined ratio; Training N AI disease prediction models based on the training set and performing model verification and evaluation on the N AI disease prediction models based on the verification set; Add additional explanations for each of the N AI disease prediction models.

4. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 3, characterized in that: Identify and select a subset of input variables that are most correlated with the target variable from the training samples of the N disease datasets, including: The preliminary input variable set was determined by rank sum test, with the patient's diagnosis type as the dependent variable and the test items and results as the independent variables; Using the preliminary input variable set as an independent variable and the patient diagnosis type as a dependent variable, performing feature parameter selection based on the XG Boost model to obtain a screened input variable set; The filtered input variable set is pruned and combined to obtain the most relevant input variable subset.

5. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 3, characterized in that: The preset ratio between the training set and the validation set is 7:3 or 8:

2.

6. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 3, characterized in that: When N AI disease prediction models are trained based on the training set and N AI disease prediction models are verified and evaluated based on the validation set, each model evaluation indicator includes AUC, accuracy, precision, recall rate, F1 score, sensitivity, specificity, PPV, NPV, and AUC_CI.

7. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 1, characterized in that: Inputting the test results of the target patient into the N AI disease prediction models respectively to obtain N AI model prediction results, including: In response to the test results of the target patient meeting the model parameter triggering condition, the test results of the target patient are input into the corresponding AI disease prediction model to obtain the AI ​​model prediction result.

8. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 7, characterized in that: The test results of the target patient are input into an algorithm fusion device based on a decision tree algorithm to obtain a decision path, including: the algorithm fusion device based on the decision tree algorithm automatically adjusts the decision path based on the patient's gender, age, symptoms, model performance, and matching degree with the patient's test items.

9. The method for intelligent interpretation of test results based on the AI ​​algorithm marketplace according to claim 1, characterized in that: Also includes: Continuously monitoring performance data of each of the N AI disease prediction models in real-world big data; Setting key indicator thresholds based on historical performance data of each of the AI ​​disease prediction models; Using a sliding window method to perform segmented evaluation on the performance data of each AI disease prediction model in real-world big data to obtain evaluation results; In response to the evaluation result that the average performance within the window is less than the key indicator threshold, a prompt signal for model retraining or parameter adjustment is triggered.

Images