FUO etiology stratified diagnosis method, system and application based on machine learning

Abstract

This application discloses a machine learning-based method, system, and application for stratified diagnosis of FUO (Fever of Unknown Occurrence) etiology. The method includes: collecting first data on clinical immune and inflammatory markers of patients with fever of unknown origin within 24 hours of admission; preprocessing and feature selection of the first data; using an XGBoost algorithm in a first-level prediction model to distinguish between infectious and non-infectious diseases; further subdividing the etiology type using a second-level prediction model, wherein the first sub-model uses the AdaBoost algorithm to distinguish between neoplastic diseases and non-infectious inflammatory diseases, and the second sub-model uses the XGBoost algorithm to distinguish between simple infection and infection combined with non-infectious diseases; and outputting predicted probabilities and SHAP-based model interpretation information. This application addresses the problem that existing FUO prediction models struggle to fully extract high-order interaction information between multi-dimensional input variables and disease outcomes.

Description

Technical Field

[0001] This invention belongs to the fields of biomedical engineering and medical artificial intelligence technology, specifically relating to a machine learning-based FUO etiological stratification diagnostic method, system, and application. Background Technology

[0002] Fever of unknown origin (FUO) is a long-standing diagnostic challenge in clinical practice, with a broad spectrum of causes, encompassing over 200 specific diseases, including infectious diseases, non-infectious inflammatory diseases, oncological diseases, and other types of illnesses. The etiological composition of FUO has changed significantly over time and geographically: although infectious diseases remain the primary cause, their proportion is declining with advancements in diagnostic technology; simultaneously, the proportion of FUO diagnoses involving non-infectious inflammatory diseases and oncological diseases is gradually increasing.

[0003] Diagnosing FUO presents multiple challenges, including the lack of specificity in clinical presentation, limitations of existing laboratory testing methods, and interference from empirical treatment. Traditional diagnostic pathways heavily rely on the clinician's personal experience and sequential testing, often leading to prolonged treatment cycles, increased medical costs, and overuse of broad-spectrum antibiotics. Even after systematic evaluation, approximately 15% of FUO patients still cannot obtain a definitive etiological diagnosis.

[0004] Artificial intelligence, especially machine learning, has demonstrated significant advantages in medical diagnosis and risk stratification. Machine learning can not only capture linear relationships between variables but also effectively uncover complex nonlinear interactions between multidimensional predictors. Existing FUO prediction models are mostly based on traditional logistic regression, which makes it difficult to fully extract high-order interaction information between multidimensional input variables and disease outcomes. Summary of the Invention

[0005] The main purpose of this application is to provide a machine learning-based FUO etiology stratification diagnostic method, system and application, which aims to solve the problem that existing FUO prediction models are unable to fully extract high-order interaction information between multi-dimensional input variables and disease outcomes.

[0006] To achieve the above objectives, this application provides a machine learning-based FUO etiological stratification diagnostic method, comprising the following steps: S1. Collect the first data of clinical immune and inflammatory markers within 24 hours after admission of patients with fever of unknown cause. The inflammatory markers include soluble interleukin-2 receptor, C-reactive protein, β2-microglobulin, hemoglobin, platelet count and neutrophil CD64 index. S2. Preprocess the first data, including missing value imputation and class balancing. S3. Perform feature selection on the preprocessed first data to obtain the optimal feature subset required by each level of the model; S4. Input the optimal feature subset into the first-level prediction model and use machine learning algorithms to determine whether the patient belongs to the non-infectious disease group or the infectious disease group with or without non-infectious disease. S5. Based on the discrimination result, the first data is guided to the corresponding sub-model in the second-level prediction model for further classification; if it is determined to be a non-infectious disease, the first data is input into the first sub-model to further distinguish it into tumor diseases, non-infectious inflammatory diseases, and miscellaneous diseases; if it is determined to be an infectious disease, the first data is input into the second sub-model to further distinguish it into simple infection or infection combined with non-infectious disease. S6. Evaluate the confidence level of the classification results of each sub-model of the second-level prediction model, and output the probability value of each classification and model interpretation information. S7. Based on the probability values ​​and model interpretation information, generate a stratified diagnostic report and present it to clinicians through an interactive interface.

[0007] Optionally, in step S2, the missing value imputation adopts the grouped hierarchical K-nearest neighbor algorithm, and K-nearest neighbor imputation is performed independently in each etiological subgroup, with K value ranging from 5 to 10; variables with a missing value ratio exceeding 25% are removed.

[0008] Optionally, in step S2, the class balancing process uses a random downsampling method to reduce the number of majority class samples to be consistent with the number of minority class samples, thereby constructing a balanced training set.

[0009] Optionally, in step S3, the feature selection includes: using Pearson correlation analysis for initial screening, and removing features with low correlation to the outcome variable when the absolute value of the correlation coefficient between two features is greater than 0.8; further screening by combining univariate analysis and LASSO regression, and determining the optimal feature combination for each model through recursive feature elimination.

[0010] Optionally, in step S4, the first-level prediction model is constructed using the XGBoost algorithm, with input features being soluble interleukin-2 receptor, C-reactive protein, and β2-microglobulin; the model hyperparameters of the first-level prediction model are determined by grid search combined with five-fold cross-validation, with a learning rate of 0.01-0.3, a tree depth of 310, and a number of trees ranging from 50 to 500.

[0011] Optionally, in step S5, the first sub-model is constructed using the AdaBoost algorithm, with input features being hemoglobin, platelet count, and C-reactive protein, and the kernel function being the radial basis function, with the regularization parameter C being 1~10; the second sub-model is constructed using the XGBoost algorithm, with input features being soluble interleukin-2 receptor and neutrophil CD64 index.

[0012] Optionally, in step S6, SHAP, LIME, or partial dependency graph methods are used to interpret the prediction results of the first sub-model and the second sub-model, providing global feature importance analysis and individualized prediction interpretation.

[0013] Furthermore, to achieve the above objectives, this application also provides a machine learning-based stratified diagnostic system for fever of unknown origin (FUO) to implement the aforementioned FUO stratified diagnostic method, comprising: The data acquisition module is used to collect the first data of clinical immune and inflammatory markers in patients with unexplained fever within 24 hours of admission; The data preprocessing module is used to preprocess the first data, including missing value imputation and class balancing. The feature selection module is used to select features from the preprocessed first data to obtain the optimal feature subset required by each level of the model. The first-level prediction model module is used to input the optimal feature subset into the first-level prediction model and use machine learning algorithms to determine whether the patient belongs to the non-infectious disease group or the infectious disease group with or without non-infectious disease. The second-level prediction model module is used to guide the first data to the corresponding sub-model in the second-level prediction model for further classification based on the discrimination result. If the first data is discriminated as a non-infectious disease, it is input into the first sub-model to further distinguish it as a neoplastic disease, a non-infectious inflammatory disease, or a miscellaneous disease. If the first data is discriminated as an infectious disease, it is input into the second sub-model to further distinguish it as a simple infection or an infection combined with a non-infectious disease. The model evaluation and optimization module is used to evaluate the confidence of the classification results of each sub-model of the second-level prediction model and output the probability value of each classification and model interpretation information. The interactive diagnostic output module is used to generate a stratified diagnostic report based on the probability values ​​and model interpretation information and present it to clinicians through an interactive interface.

[0014] Optionally, the interactive diagnostic output module is deployed as a web application based on the R Shiny framework, including a parameter input interface, a real-time etiology classification prediction function, a visualization interpretation function, and a hierarchical diagnostic report generation function; the parameter input interface allows users to input the core feature indicators required by each model, the real-time etiology classification prediction function outputs the predicted probability of each etiology category immediately after the parameters are input, and the visualization interpretation function displays individualized prediction interpretations simultaneously.

[0015] This application also provides an application of the above-mentioned FUO etiology stratification diagnostic method or the above-mentioned fever of unknown origin etiology stratification diagnostic system in the clinical auxiliary diagnosis of patients with fever of unknown origin. The application includes: collecting peripheral venous blood samples from patients with fever of unknown origin within 24 hours of admission, detecting six indicators: soluble interleukin-2 receptor, C-reactive protein, β2-microglobulin, hemoglobin, platelet count, and neutrophil CD64 index, inputting the test results into the FUO etiology stratification diagnostic method or the fever of unknown origin etiology stratification diagnostic system to obtain etiology stratification diagnostic results, and guiding the clinical examination direction and treatment plan formulation based on the etiology stratification diagnostic results.

[0016] Compared with the prior art, this application has the following beneficial effects: This application is the first to construct a three-model hierarchical system of "Level 1 (non-infectious diseases vs. infectious diseases) → Level 2 (tumorous diseases vs. non-infectious inflammatory diseases; simple infection vs. infection combined with non-infectious diseases)", which realizes accurate diagnosis step by step from macro classification to micro subtyping, filling the technical gap in the field of FUO hierarchical diagnosis. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of the FUO etiology hierarchical diagnostic system based on machine learning in the embodiments of this application; Figure 2 This is a flowchart of the FUO etiology stratification diagnosis method based on machine learning in the embodiments of this application; Figure 3 This is the interface effect of the Web online diagnostic system in the embodiments of this application. Figure 1 ; Figure 4 This is the interface effect of the Web online diagnostic system in the embodiments of this application. Figure 2 ; Figure 5 This is the interface effect of the Web online diagnostic system in the embodiments of this application. Figure 3 ; Figure 6 This is the interface effect of the Web online diagnostic system in the embodiments of this application. Figure 4 . Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] This application provides a machine learning-based FUO etiology stratification diagnosis method, system, and application. The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] As attached Figure 1 As shown, this application provides a machine learning-based FUO etiology hierarchical diagnostic system. The overall architecture of the system includes: a data acquisition module 100, a data preprocessing module 200, a feature selection module 300, a first-level prediction model module 410, a second-level prediction model module 420, a model evaluation and optimization module 500, and an interactive diagnostic output module 600.

[0021] The output of the data acquisition module 100 is connected to the input of the data preprocessing module 200. The output of the data preprocessing module 200 is connected to the input of the feature selection module 300. The output of the feature selection module 300 is connected to the input of the first-level prediction model module 410 and the second-level prediction model module 420, respectively. The outputs of the first-level prediction model module 410 and the second-level prediction model module 420 are both connected to the input of the model evaluation and optimization module 500. The output of the model evaluation and optimization module 500 is connected to the input of the interactive diagnostic output module 600.

[0022] As attached Figure 1 As shown, the first-level prediction model module 410 is used to distinguish between the non-infectious disease (NID without infection) group and the infection with or without non-infectious disease (Infection with or without NID) group. The second-level prediction model module 420 includes two parallel sub-models: the first sub-model 421 is a model for distinguishing between neoplasm and non-infectious inflammatory diseases and miscellaneous diseases (NIID and Miscellaneous); the second sub-model 422 is a model for distinguishing between infection without NID and infection with NID.

[0023] The following is a detailed introduction to the composition and function of each module.

[0024] (1) Data acquisition module 100 The data acquisition module 100 was used to collect clinical immune and inflammatory marker data of FUO patients within 24 hours of admission. A total of 32 characteristic variables were collected, and after data preprocessing and feature screening, 26 were retained for model development. These included: white blood cell count (WBC), hemoglobin (HGB), platelet count (PLT), red blood cell count (RBC), lymphocyte count (LYM), lymphocyte percentage (LYM.pct), neutrophil count (NEU), monocyte count (MONO), monocyte percentage (MONO.pct), eosinophil percentage (EOS.pct), basophil count (BASO), and basophil saturation. Cell percentage (BASO.pct), albumin (ALB), lactate dehydrogenase (LDH), procalcitonin (PCT), C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), ferritin (FER), β2-microglobulin (β2-MG), neutrophil CD64 index (nCD64), soluble interleukin-2 receptor (sIL-2R), interleukin-6 (IL-6), fibrinogen (FIB), serum amyloid A (SAA), creatine kinase isoenzyme (CK-MB), D-dimer (DD).

[0025] (2) Data preprocessing module 200 The data preprocessing module 200 includes a missing value processing unit 201 and a class balancing processing unit 202. The missing value processing unit 201 uses a hierarchical K-Nearest Neighbors (kNN) algorithm to impute missing data. Specifically, kNN imputation is performed independently within each etiology subgroup to maintain the data distribution characteristics within each group. Variables with a missing proportion exceeding 25% are removed, including interleukin-1β (IL-1β), complement C3 (C3), complement C4 (C4), and monocyte HLA-DR (mHLA-DR). The class balancing processing unit 202 addresses the imbalance of class samples in the training set by using a random downsampling method to reduce the majority class sample size to match the minority class sample size, thus constructing a balanced training set.

[0026] (3) Feature selection module 300 The feature selection module 300 is used to select the optimal feature subset from the original 26 features. This application uses Pearson correlation analysis for feature selection: when the absolute value of the correlation coefficient between two features is greater than 0.8, features with low correlation to the outcome variable are removed. Based on this criterion, the percentage of neutrophils (NEU.pct) and eosinophil count (EOS) are removed.

[0027] For each level of model, the feature selection module 300 further filters the model by combining univariate analysis (P<0.05 or P<0.1) and LASSO regression, and determines the optimal feature combination for each model by recursive feature elimination (RFE).

[0028] Specifically: The optimal feature combination determined by the first-level prediction model module 410 after screening is: soluble interleukin-2 receptor (sIL-2R), C-reactive protein (CRP), and β2-microglobulin (β2-MG). The optimal feature combination determined by the first sub-model 421 after screening is: hemoglobin (HGB), platelet count (PLT), and C-reactive protein (CRP). The optimal feature combination determined by the screening of the second sub-model 422 is: soluble interleukin-2 receptor (sIL-2R) and neutrophil CD64 index (nCD64).

[0029] The model construction in this application adopts a two-level hierarchical modeling strategy, including a first-level prediction model module 410 and a second-level prediction model module 420.

[0030] The first-level prediction model module 410 is used to distinguish between non-infectious diseases and infections with or without non-infectious diseases. This module can be constructed using the XGBoost algorithm, with input features including sIL-2R, CRP, and β2-MG. Model hyperparameters are determined through grid search combined with manual fine-tuning, and model stability is evaluated using five-fold cross-validation.

[0031] The second-level prediction model module 420 includes two parallel sub-models: First sub-model 421: In the non-infectious disease branch, a support vector machine (AdaBoost) algorithm is used to construct the model. The input features are three items: HGB, PLT and CRP, which are used to distinguish neoplasm from NIID (neuronal intranuclear inclusion body disease) and miscellaneous diseases (unclassified diseases). The second sub-model 422: In the infectious disease branch, the XGBoost algorithm is used to construct the model. The input features are sIL-2R and nCD64, which are used to distinguish between simple infection (Infection without NID) and infection with non-infectious disease (Infection with NID).

[0032] The following seven machine learning algorithms were compared and selected for each model: Decision Tree, Random Forest, Support Vector Machine, LASSO Regression, K-Nearest Neighbors, XGBoost, and AdaBoost. Models were trained using the training set data, and the AUC value was used as the primary evaluation metric to select the optimal algorithm.

[0033] (4) Model Evaluation and Optimization Module 500 The model evaluation and optimization module 500 is used to perform multi-dimensional performance evaluation and stability verification on the trained model.

[0034] The performance metrics for the model include: Area under the receiver operating characteristic (AUC), Area under the precision-recall curve (AUC-PR), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score.

[0035] Model stability was evaluated on the training set using 5-fold cross-validation.

[0036] The SHAP (SHapley Additive exPlanations) method is used to rank the features of the XGBoost model and interpret the model, providing global feature importance analysis and individualized prediction interpretation.

[0037] External validation: Independent FUO patient cohorts from Huashan Hospital (June 2023 to May 2025), the First Affiliated Hospital of Xi'an Jiaotong University (January 2023 to May 2025), and the First Affiliated Hospital of Zhengzhou University (January 2023 to May 2025) were used as external validation sets to evaluate the model's generalization ability. External validation results showed that the AUC of the first-level model was 0.975, the AUC of the first sub-model was 0.946, and the AUC of the second sub-model was 0.896.

[0038] (5) Interactive diagnostic output module 600 The interactive diagnostic output module 600 deploys the three final optimized models as interactive web applications based on the R Shiny framework. In other implementations, besides the R Shiny framework, the web application can also be deployed using the Python Flask / Django framework, Streamlit, or a JavaScript front-end framework.

[0039] This module implements the following functions: Parameter input interface: Supports users to input the core feature indicators required for each model (the first-level prediction model module 410 requires input of sIL-2R, CRP, and β2-MG; the first sub-model 421 requires input of HGB, PLT, and CRP; the second sub-model 422 requires input of sIL-2R and nCD64). Real-time etiology classification prediction: After inputting parameters, the system immediately outputs the predicted probability of each etiology category; Visual interpretation: Simultaneously displays individualized prediction interpretations based on SHAP values ​​to help clinicians understand the basis for predictions; Stratified diagnostic report generation: Outputs final etiological classification suggestions according to a two-level stratified structure.

[0040] like Figure 2 As shown, this application provides the above-mentioned machine learning-based FUO etiology stratification diagnostic system, which includes the following steps when performing FUO etiology stratification diagnosis: Step S1: Acquire clinical immune and inflammatory marker data of FUO patients within 24 hours of admission through data acquisition module 100. After detailed medical history collection and temperature monitoring, and excluding drug fever and false fever, the collected immune and inflammatory markers include: sIL-2R, CRP, β2-MG, HGB, PLT, and nCD64.

[0041] Step S2: The data preprocessing module 200 receives the data collected in step S1. First, the missing value processing unit 201 fills in the missing values ​​using the grouped hierarchical kNN algorithm. Then, the class balancing processing unit 202 performs random downsampling on the training set and outputs the balanced feature data.

[0042] Step S3: The feature selection module 300 receives the feature data processed in step S2, performs feature screening according to the preset Pearson correlation threshold (|r|>0.8), removes features that are highly collinear and have low correlation with the outcome, and determines the optimal feature subset of each level model by combining univariate analysis and LASSO regression.

[0043] Step S4: Input the preferred feature subset (sIL-2R, CRP, β2-MG) output from step S3 into the first-level prediction model module 410, and use the XGBoost algorithm to determine whether the patient belongs to the non-infectious disease (NID without infection) group or the infection with or without non-infectious disease (Infection with or without NID) group.

[0044] Step S5: Based on the stratification results of Step S4, guide the data to the corresponding sub-model in the second-level prediction model module 420: ① If identified as a non-infectious disease, input into the first sub-model 421, using the AdaBoost algorithm, based on the three features of HGB, PLT, and CRP, to further distinguish it as a neoplastic disease (Neoplasm) or NIID and Miscellaneous diseases; if predicted as a neoplastic disease, it is recommended to screen for neoplastic diseases; if predicted as NIID and Miscellaneous diseases, it is recommended to screen for autoimmune and autoinflammatory diseases. ② If identified as an infectious disease, input into the second sub-model 422, using the XGBoost algorithm, based on the two features of sIL-2R and nCD64, to further distinguish it as a simple infection or an infection combined with a non-infectious disease. All patients in this branch should immediately start empirical anti-infective treatment and simultaneously undergo comprehensive etiological testing; for patients identified as an infection combined with a non-infectious disease, it is recommended to simultaneously screen for neoplastic diseases and autoimmune / autoinflammatory diseases.

[0045] Step S6: The model evaluation and optimization module 500 evaluates the confidence of the prediction results of each level of the model and outputs the probability values ​​of each category and the SHAP interpretation.

[0046] Step S7: The interactive diagnostic output module 600 generates a graded and hierarchical diagnostic report and presents it to clinicians through a web interface to achieve personalized diagnostic decision support.

[0047] As a preferred implementation, when performing kNN imputation independently within each etiological subgroup, the k value is preferably 5 to 10. In addition to grouped stratified kNN imputation, multiple imputation, random forest imputation, or mean / median imputation can also be used.

[0048] In a preferred embodiment, in the first-level prediction model module 410, the learning rate of the XGBoost model is preferably 0.01~0.3, the tree depth is preferably 3~10, and the number of trees is preferably 50~500.

[0049] In a preferred embodiment, in the first sub-model 421, the kernel function of the AdaBoost model is preferably a radial basis function (RBF kernel), and the regularization parameter C is preferably 1 to 10.

[0050] In a preferred embodiment, in the second sub-model 422, the learning rate of the XGBoost model is preferably 0.01~0.3, the tree depth is preferably 3~10, and the number of trees is preferably 50~500.

[0051] As a preferred implementation, the model hyperparameter optimization adopts a grid search method with five-fold cross-validation, with the optimization objective being to maximize the AUC value.

[0052] In step S2, in addition to random downsampling, methods such as SMOTE oversampling, cost-sensitive learning, or Focal Loss can also be used to handle class imbalance problems.

[0053] In step S3, in addition to Pearson correlation analysis, univariate analysis and LASSO regression combination, recursive feature elimination (RFE), mutual information, feature ranking based on SHAP value or random forest feature importance methods can also be used for feature screening, all of which can achieve the technical objectives of this invention.

[0054] In step S5, in addition to the XGBoost and AdaBoost algorithms used, other algorithms such as DecisionTree, Random Forest, LASSO regression, K-Nearest Neighbors (KNN), AdaBoost, Deep Neural Network (DNN), LightGBM, or CatBoost can also be used to build the prediction model.

[0055] In step S6, in addition to the SHAP method, other methods such as LIME (Local Interpretable Model-agnostic Explanations) or Partial Dependency Graph (PDP) can also be used for model interpretation.

[0056] The present application will be further described in detail below through specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present application.

[0057] Example 1: A unified FUO sequential auxiliary diagnostic process This embodiment illustrates the complete sequential application process of the system described in this application in patients with fever of unknown origin. For ease of clinical use, this application integrates the first-level prediction model module 410, the first sub-model 421, and the second sub-model 422 into the same interactive web application.

[0058] Clinicians can complete the FUO suitability assessment, special fever case screening, test item confirmation, primary diagnosis, secondary branch diagnosis, and comprehensive result determination by following the system guidance without having to access multiple web applications separately.

[0059] 1. Applicable Object Determination First, the clinician collects the patient's medical history, performs a physical examination, monitors body temperature, and conducts necessary routine examinations. The clinician then determines whether the patient meets the diagnostic criteria for FUO (fever lasting more than 3 weeks), with oral body temperature ≥38.3℃ at least 3 times, or body temperature fluctuations >1.2℃ at least 3 times within 24 hours, and a definitive diagnosis has not been obtained after at least one week of systematic inpatient or outpatient examinations. Simultaneously, the clinician determines whether the patient has any of the following conditions: immunodeficiency, a response to empirical anti-infective therapy before admission, or received continuous glucocorticoid therapy for ≥2 weeks within 3 months prior to admission. If any of these conditions are present, the patient is not suitable for this model.

[0060] For patients who meet the inclusion and exclusion criteria, clinicians access the unified web application mentioned above, click to confirm the inclusion and exclusion criteria, and proceed to the next step.

[0061] 2. Special Circumstances Investigation Before entering the prediction model, the system prompts clinicians to combine medical history, characteristics of body temperature changes, medication use and relevant test results to rule out conditions that require priority identification or special treatment pathways, such as masquerading fever, drug fever, and hemophagocytic lymphohistiocytosis (HLH).

[0062] For patients with the possibility of the above-mentioned special circumstances, the system will prompt them to complete the corresponding differential diagnosis and clinical treatment first, and will not directly enter the routine FUO prediction model. For patients who have not been found to have the above-mentioned special circumstances after clinical evaluation, or whose relevant circumstances have been reasonably ruled out, click to confirm the exclusion of special circumstances and proceed to the test item confirmation step.

[0063] 3. Confirmation of testing items The system prompts clinicians to confirm whether the patient has completed the testing of the indicators required by the predictive model, including: ① Soluble interleukin-2 receptor (sIL-2R, unit U / mL); ② C-reactive protein (CRP, unit mg / L); ③ β2-microglobulin (β2-MG, unit mg / L); ④ Hemoglobin (HGB, unit g / L); ⑤ Platelet count (PLT, unit ×10^9 / L); ⑥ Neutrophil CD64 index (nCD64, index).

[0064] The above indicators can be detected by collecting peripheral venous blood from patients. Among them, sIL-2R can be detected by enzyme-linked immunosorbent assay (ELISA) or chemiluminescence immunoassay; CRP and β2-MG can be detected by immunoturbidimetric assay or chemiluminescence immunoassay; HGB and PLT can be detected by blood analyzer; and nCD64 can be detected by flow cytometry.

[0065] If the relevant indicators have not yet been tested, the system will prompt you to supplement the corresponding tests. Only after all indicators required for the current diagnostic stage have been completed and valid values ​​have been entered will the system proceed to the corresponding predictive analysis process. Test results already obtained can be used directly without the need for repeat testing. After clicking to confirm the test completion status, proceed to the first level of diagnosis.

[0066] 4. First-level diagnosis After confirming the test items, enter the test values ​​for sIL-2R, CRP, and β2-MG in the unified web application and click the "Execute Level 1 Diagnostic Prediction" button. The system will automatically call the Level 1 prediction model module 410 deployed on the server for calculation.

[0067] The first-level prediction model module 410 uses the XGBoost algorithm to output the predicted probabilities of infectious and non-infectious diseases and the corresponding classification results: When the predicted probability of non-infectious diseases is ≥50%, the system will determine the patient as having a tendency for non-infectious diseases and automatically enter the second-level diagnostic branch corresponding to the first sub-model 421. When the predicted probability of an infectious disease is greater than 50%, the system will determine that the patient is prone to infectious diseases and automatically enter the second-level diagnostic branch corresponding to the second sub-model 422.

[0068] 5. Secondary diagnostic subcategory of non-infectious diseases For patients identified as having a predisposition to non-infectious diseases by the first-level prediction model module 410, the system prompts for the input of test values ​​for three indicators: HGB, PLT, and CRP. The CRP result can be directly used from the initial diagnosis and does not require repeat testing.

[0069] The system automatically calls the first sub-model 421 for calculation. The first sub-model 421 uses the AdaBoost algorithm to output the predicted probabilities and classification results for neoplastic diseases, non-infectious inflammatory diseases, and other diseases (NIID / Miscellaneous). When the predicted probability of neoplastic disease is ≥50%, the system indicates that the patient is prone to neoplastic disease and suggests further examinations such as PET-CT, tumor marker detection, bone marrow biopsy or lymph node biopsy based on the clinical situation. When the predicted probability of NIID / Miscellaneous is >50%, the system suggests that the patient is likely to have a non-infectious inflammatory disease or other diseases, and recommends further examinations such as antinuclear antibody profile, rheumatoid factor, complement, and immunoglobulins based on the clinical situation.

[0070] 6. Secondary diagnostic sub-branch of infectious diseases For patients identified as having a predisposition to infectious diseases by the first-level prediction model module 410, the system prompts for the input of test values ​​for two indicators: sIL-2R and nCD64. The sIL-2R result can be directly used from the test entered during the first-level diagnosis; only the nCD64 test result needs to be supplemented based on the actual situation.

[0071] The system automatically calls the second sub-model 422 for calculation. The second sub-model 422 uses the XGBoost algorithm to output the predicted probabilities and classification results for simple infection and infection combined with non-infectious diseases: When the predicted probability of simple infection is ≥50%, the system indicates that the patient is likely to have simple infection. It is recommended to conduct blood culture, etiological metagenomic sequencing and other tests in combination with the clinical situation, and to formulate or adjust the anti-infection regimen based on the etiological results. When the predicted probability of an infection combined with a non-infectious disease is greater than 50%, the system suggests that the patient may have both infectious and non-infectious diseases. It is recommended that while carrying out infection-related examinations and treatments, further investigations be conducted to rule out neoplastic diseases, autoimmune diseases, or autoinflammatory diseases.

[0072] 7. Final Diagnostic Recommendations After completing the second-level diagnosis, the system integrates the patient's FUO applicability assessment, special case investigation results, first-level prediction results, and second-level branch prediction results to form a unified final diagnostic recommendation.

[0073] The prediction results output by this system are intended to provide auxiliary diagnostic references for clinicians, but do not replace the final diagnosis and treatment decisions made by clinicians in combination with the patient's medical history, physical signs, laboratory tests, imaging examinations, etiological examinations, and treatment response.

[0074] Example 2 like Figure 3 , Figure 4 As shown, this embodiment uses the unified FUO sequential auxiliary diagnostic system described in this application to perform predictive analysis on a patient who was finally diagnosed with remission seronegative symmetrical synovitis with pitting edema syndrome (RS3PE syndrome).

[0075] Case information: The patient was a 73-year-old male admitted to the hospital due to "weakness in the limbs for more than a month and fever for 3 weeks." One month prior to admission, the patient developed weakness and joint pain in the limbs without any obvious cause, gradually progressing to pitting edema in the upper and lower extremities. Subsequently, he experienced chills, rigors, and fever, primarily in the evening or at night, with a highest temperature of 38.8℃. Examinations at another hospital showed hemoglobin 99 g / L, platelet count 466×10^9 / L, C-reactive protein 217.3 mg / L, ferritin 567.9 ng / mL, and negative results for autoantibodies and hepatitis markers. He had received anti-infective treatment at the other hospital with levofloxacin, amoxicillin-clavulanate potassium, and imipenem, but the patient continued to experience recurrent fever, and the weakness and joint pain in the limbs worsened, indicating that conventional anti-infective treatment was ineffective.

[0076] Physical examination upon admission revealed a body temperature of 37.8℃, clear consciousness, and several soft, mobile small lymph nodes palpable above the clavicle; muscle strength grade III in the left upper and lower extremities, and grade II in the right upper and lower extremities; pain upon passive movement of the right shoulder, elbow, and knee joints, and mild pitting edema in the right hand. A chest CT scan at another hospital showed bilateral pleural effusion and minor inflammation in both lower lobes of the lungs. Abdominal imaging revealed chronic schistosomiasis-related liver disease, fatty liver, liver cysts, and gallstones.

[0077] Clinicians access the FUO sequential auxiliary diagnostic process through a unified web application. Based on the patient's fever duration, highest body temperature, and previous medical history, the patient is confirmed to meet the system's inclusion criteria. After medical history collection, temperature monitoring, and relevant examinations, special cases such as masquerading fever, drug fever, and hemophagocytic lymphohistiocytosis are excluded before the patient enters the model prediction process.

[0078] After admission, the patient underwent relevant tests, and the results were as follows: sIL-2R was 730 U / mL, β2-MG was 1.94 mg / L, CRP was 151.81 mg / L, nCD64 index was 0.67, HGB was 101 g / L, and PLT was 578×10^9 / L.

[0079] The first step is to perform a Level 1 diagnosis. The clinician inputs the test values ​​for sIL-2R, CRP, and β2-MG into the unified web application, and the system automatically calls the Level 1 prediction model module 410 for calculation. The results show that the predicted probability of the patient developing a non-infectious disease is 91.4%, and the classification result is "non-infectious disease." Based on this, the system suggests that the patient has a high probability of developing a non-infectious disease and recommends further testing for autoimmune markers and inflammatory indicators, proceeding to the Level 2 diagnostic branch corresponding to non-infectious disease.

[0080] The second level of diagnosis is then performed. The system enters branch A, the differential diagnosis branch for "tumorous diseases, non-infectious inflammatory diseases, and others." The clinician inputs the test values ​​for HGB, PLT, and CRP, and the system automatically calls the first sub-model 421 for calculation. The results show that the predicted probability of the patient having a tumor is 26.5%, while the predicted probability of non-infectious inflammatory diseases and others is 73.5%. The model's prediction result is "non-infectious inflammatory diseases and others," suggesting that further investigations related to autoimmune or autoinflammatory diseases should be conducted, including autoantibodies, complement, and inflammatory markers.

[0081] Based on the patient's clinical presentation and subsequent examination results, the final diagnosis upon discharge was: 1. Connective tissue disease: RS3PE syndrome, i.e., remission-associated seronegative symmetrical synovitis with pitting edema; 2. Hashimoto's thyroiditis; 3. Diabetes mellitus. After independent review and comprehensive assessment by two infectious disease specialists, RS3PE syndrome was classified as "non-infectious inflammatory diseases and others" in the classification system of this application model.

[0082] Therefore, in this case, the first-level prediction model accurately identified the patient as having a non-infectious disease, and the second-level prediction model further classified it into non-infectious inflammatory diseases and other categories. The sequential prediction results of the models were consistent with the patient's final clinical diagnosis category, indicating that the system described in this application can provide auxiliary reference for the etiological classification and subsequent examination direction of FUO patients.

[0083] Example 3 like Figure 5 , Figure 6 As shown, this embodiment uses the unified FUO sequential auxiliary diagnostic system described in this application to perform predictive analysis on a patient who was ultimately diagnosed with biliary tract infection complicated with chronic lymphocytic leukemia.

[0084] Case information: The patient was a 57-year-old male admitted to the hospital due to "fever accompanied by abnormal liver function for 3 weeks". The patient experienced recurrent fever without any obvious cause, with body temperature gradually rising from 37-38℃, reaching a maximum of 40℃, accompanied by chills, rigors, and fatigue. There was no significant cough, sputum production, abdominal pain, diarrhea, urinary frequency, urgency, dysuria, rash, or joint swelling. He had received anti-infective treatment at another hospital, including ceftazidime combined with moxifloxacin and ceftriaxone combined with levofloxacin, as well as liver-protective treatment. While the peak temperature decreased somewhat, recurrent fever persisted.

[0085] The patient's examination at another hospital showed a white blood cell count of 14.12 × 10^9 / L, neutrophils 51.1%, lymphocytes 43.4%, hemoglobin 144 g / L, platelet count 197 × 10^9 / L, CRP 167.1 mg / L, ESR 42 mm / h, and PCT 0.61 ng / mL. Liver function tests showed ALT 269 U / L, AST 234 U / L, and GGT 357 U / L. Tests for influenza A and B viruses, respiratory virus antibodies, cytomegalovirus, and EBV were all negative, and blood cultures were negative. Chest CT showed ground-glass opacity in the left upper lobe, micronodules in the right upper lobe, and chronic inflammation in both lungs. Abdominal ultrasound revealed fatty liver, liver cysts, and uneven distribution of fat within the liver. After anti-infection and liver-protective treatment, ALT and AST decreased somewhat, but GGT remained significantly elevated.

[0086] The patient had a history of persistent leukocytosis and lymphocytosis. About one year prior to admission, a routine blood test showed a white blood cell count of approximately 20 × 10^9 / L and a lymphocyte percentage of approximately 70%. The patient had previously undergone bone marrow aspiration and flow cytometry, and subsequent blood tests still showed similar changes.

[0087] On admission, physical examination revealed a body temperature of 38.5℃, pulse of 102 beats / min, respiratory rate of 20 breaths / min, and blood pressure of 115 / 73 mmHg. The patient was alert and conscious. No obvious abnormalities were found in the skin and mucous membranes, and no significant enlargement of superficial lymph nodes was palpable. Breath sounds were clear in both lungs, with no dry or wet rales heard. The heart rhythm was regular. The abdomen was soft, without tenderness or rebound tenderness. There was no redness, swelling, or tenderness in the joints of the limbs, and no edema in the lower extremities.

[0088] Clinicians access the FUO sequential auxiliary diagnostic process through a unified web application. After taking a medical history, conducting a physical examination, monitoring body temperature, and performing relevant examinations, the system determines that the patient meets the inclusion criteria and excludes special cases such as masquerade fever, drug fever, and hemophagocytic lymphohistiocytosis. The patient then proceeds to the model prediction process.

[0089] After admission, the patient underwent relevant tests, and the results were as follows: CRP was 100.61 mg / L, sIL-2R was 2594 U / mL, nCD64 index was 52.01, β2-MG was 3.5 mg / L, HGB was 124 g / L, and PLT was 398×10^9 / L.

[0090] The first step is to perform a Level 1 diagnosis. The clinician inputs the test values ​​for three indicators—sIL-2R, CRP, and β2-MG—in the unified web application. The system automatically calls the Level 1 prediction model module 410 for calculation. The results show that the predicted probability of non-infectious disease is 1.2%, and the predicted probability of infectious disease with or without non-infectious disease is 98.8%. The final classification result is "infectious disease with or without non-infectious disease." The system indicates a high probability of the patient having an infectious disease and recommends further infection-related examinations and etiological testing, proceeding to the Level 2 diagnostic branch corresponding to infectious disease.

[0091] The second level of diagnosis is then performed. The system enters branch B, the differential diagnosis branch for "simple infection versus infection combined with non-infectious disease". The clinician inputs the test values ​​of sIL-2R and nCD64, and the system automatically calls the second sub-model 422 for calculation. The results show that the predicted probability of simple infection is 20.0%, and the predicted probability of infection combined with non-infectious disease is 80.0%, with the final classification result being "infection combined with non-infectious disease".

[0092] The system indicates that the patient is highly likely to have both infectious and non-infectious diseases. It is recommended that while conducting empirical anti-infective treatment and simultaneously seeking etiological evidence, further investigations be conducted to rule out neoplastic diseases, autoimmune diseases, or autoinflammatory diseases.

[0093] Based on the patient's clinical presentation, laboratory tests, and subsequent specialist examinations, the final diagnosis upon discharge was: 1. Biliary tract infection; 2. Chronic lymphocytic leukemia; 3. Sequelae of poliomyelitis. After independent review and comprehensive assessment by two infectious disease specialists, it was determined that biliary tract infection is an infectious disease, and chronic lymphocytic leukemia is a neoplastic disease; therefore, the patient falls into the category of "infectious disease combined with non-infectious disease."

[0094] Therefore, in this case, the first-level prediction model accurately identified the patient's infectious disease, and the second-level prediction model further indicated that the infection was not an infectious disease. The sequential prediction results of the model were consistent with the final clinical diagnosis category, indicating that the system described in this application can identify complex clinical situations where infectious and neoplastic diseases coexist, and provide auxiliary reference for subsequent etiological examination and non-infectious disease screening.

[0095] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope of this application is indicated by the appended claims.

[0096] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A machine learning-based stratified diagnostic method for FUO (Full Organ Occurrence) etiology, characterized in that, Includes the following steps: S1. Collect the first data of clinical immune and inflammatory markers within 24 hours after admission of patients with fever of unknown cause. The inflammatory markers include soluble interleukin-2 receptor, C-reactive protein, β2-microglobulin, hemoglobin, platelet count and neutrophil CD64 index. S2. Preprocess the first data, including missing value imputation and class balancing. S3. Perform feature selection on the preprocessed first data to obtain the optimal feature subset required by each level of the model; S4. Input the optimal feature subset into the first-level prediction model and use machine learning algorithms to determine whether the patient belongs to the non-infectious disease group or the infectious disease group with or without non-infectious disease. S5. Based on the discrimination result, the first data is guided to the corresponding sub-model in the second-level prediction model for further classification; if it is determined to be a non-infectious disease, the first data is input into the first sub-model to further distinguish it into tumor diseases, non-infectious inflammatory diseases, and miscellaneous diseases; if it is determined to be an infectious disease, the first data is input into the second sub-model to further distinguish it into simple infection or infection combined with non-infectious disease. S6. Evaluate the confidence level of the classification results of each sub-model of the second-level prediction model, and output the probability value of each classification and model interpretation information. S7. Based on the probability values ​​and model interpretation information, generate a stratified diagnostic report and present it to clinicians through an interactive interface.

2. The FUO etiological stratification diagnostic method according to claim 1, characterized in that, In step S2, the missing value imputation adopts the grouped hierarchical K-nearest neighbor algorithm, and K-nearest neighbor imputation is performed independently in each etiological subgroup, with K value ranging from 5 to 10; variables with a missing value ratio exceeding 25% are removed.

3. The FUO etiological stratification diagnostic method according to claim 1, characterized in that, In step S2, the class balancing process uses a random downsampling method to reduce the number of majority class samples to match the number of minority class samples, thereby constructing a balanced training set.

4. The FUO etiological stratification diagnostic method according to claim 1, characterized in that, In step S3, the feature selection includes: using Pearson correlation analysis for initial screening, and removing features with low correlation to the outcome variable when the absolute value of the correlation coefficient between two features is greater than 0.8; further screening by combining univariate analysis and LASSO regression, and determining the optimal feature combination for each model through recursive feature elimination.

5. The FUO etiological stratification diagnostic method according to claim 1, characterized in that, In step S4, the first-level prediction model is constructed using the XGBoost algorithm, with input features being soluble interleukin-2 receptor, C-reactive protein, and β2-microglobulin. The model hyperparameters of the first-level prediction model are determined by grid search combined with five-fold cross-validation, with a learning rate of 0.01 to 0.3, a tree depth of 310, and a number of trees ranging from 50 to 500.

6. The FUO etiological stratification diagnostic method according to any one of claims 1-5, characterized in that, In step S5, the first sub-model is constructed using the AdaBoost algorithm, with input features including hemoglobin, platelet count, and C-reactive protein, and the kernel function being the radial basis function, with the regularization parameter C ranging from 1 to 10; the second sub-model is constructed using the XGBoost algorithm, with input features including soluble interleukin-2 receptor and neutrophil CD64 index.

7. The FUO etiological stratification diagnostic method according to claim 6, characterized in that, In step S6, the prediction results of the first sub-model and the second sub-model are interpreted using SHAP, LIME, or partial dependency graph methods, providing global feature importance analysis and individualized prediction interpretation.

8. A machine learning-based stratified diagnostic system for fever of unknown origin (FUO), used to implement the FUO stratified diagnostic method as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect the first data of clinical immune and inflammatory markers in patients with unexplained fever within 24 hours of admission; The data preprocessing module is used to preprocess the first data, including missing value imputation and class balancing. The feature selection module is used to select features from the preprocessed first data to obtain the optimal feature subset required by each level of the model. The first-level prediction model module is used to input the optimal feature subset into the first-level prediction model and use machine learning algorithms to determine whether the patient belongs to the non-infectious disease group or the infectious disease group with or without non-infectious disease. The second-level prediction model module is used to guide the first data to the corresponding sub-model in the second-level prediction model for further classification based on the discrimination result. If the first data is discriminated as a non-infectious disease, it is input into the first sub-model to further distinguish it as a neoplastic disease, a non-infectious inflammatory disease, or a miscellaneous disease. If the first data is discriminated as an infectious disease, it is input into the second sub-model to further distinguish it as a simple infection or an infection combined with a non-infectious disease. The model evaluation and optimization module is used to evaluate the confidence of the classification results of each sub-model of the second-level prediction model and output the probability value of each classification and model interpretation information. The interactive diagnostic output module is used to generate a stratified diagnostic report based on the probability values ​​and model interpretation information and present it to clinicians through an interactive interface.

9. The stratified diagnostic system for unexplained fever according to claim 8, characterized in that, The interactive diagnostic output module is deployed as a web application based on the R Shiny framework, including a parameter input interface, a real-time etiology classification prediction function, a visualization interpretation function, and a hierarchical diagnostic report generation function. The parameter input interface allows users to input the core feature indicators required by each model. The real-time etiology classification prediction function outputs the predicted probability of each etiology category immediately after the parameters are input. The visualization interpretation function displays individualized prediction interpretations in a synchronized manner.

10. The application of the FUO etiology stratification diagnostic method according to any one of claims 1-7 or the unexplained fever etiology stratification diagnostic system according to any one of claims 8-9 in the clinical auxiliary diagnosis of patients with unexplained fever, characterized in that, The application includes: collecting peripheral venous blood samples from patients with unexplained fever within 24 hours of admission, detecting six indicators: soluble interleukin-2 receptor, C-reactive protein, β2-microglobulin, hemoglobin, platelet count, and neutrophil CD64 index, inputting the test results into the FUO etiological stratification diagnostic method or the unexplained fever etiological stratification diagnostic system to obtain etiological stratification diagnostic results, and guiding the direction of clinical examination and the formulation of treatment plans based on the etiological stratification diagnostic results.

Images