Potential serum marker and kit for diagnosing inhalation pneumonia

Abstract

The invention relates to the technical field of medical kits, and provides a potential serum marker for diagnosing inhalation pneumonia, the biomarker is selected from six kinds of materials including quinolinic acid, phosphatidylcholine (18: 0 / 20: 4 (5Z, 8Z, 11Z, 14Z)), N6, N6, N6-trimethyl-L-lysine, lysophosphatidylcholine (17: 0 / 0: 0), DL-piperidine acid and galactitol. The kit comprises a calibration product, a low-value quality control product, a high-value quality control product, an internal standard product, a protein precipitant, a diluent, an electrolyte concentrated solution I and an electrolyte concentrated solution II, the six biomarkers can accurately identify the inhaled pneumonia and the non-inhaled pneumonia, and the kit is developed around the six biomarkers. The method can well identify inhalation pneumonia and non-inhalation pneumonia, and solves the technical problem of missed diagnosis of patients with early-stage inhalation pneumonia.

Description

Technical Field

[0001] This invention relates to the field of medical reagent kit technology, and more particularly to a potential serum biomarker and kit for diagnosing aspiration pneumonia. Background Technology

[0002] Aspiration pneumonia (AP) is an inflammation of the lung parenchyma caused by inhaling food, oropharyngeal secretions, or gastric contents. With the increasing aging of the population, this condition is receiving growing clinical attention, accounting for 5% to 15% of community-acquired pneumonia cases. Risk factors for aspiration pneumonia commonly include altered mental status, dysphagia, impaired cough reflex, and increased gastric reflux. Clinical features of aspiration pneumonia range from asymptomatic to respiratory failure. Based on the presence or absence of a clear history of aspiration, aspiration pneumonia can be further divided into overt aspiration pneumonia and covert aspiration pneumonia. The former presents with immediate symptoms such as irritating cough, shortness of breath, or respiratory failure after inhalation; the latter does not present with acute symptoms such as irritating cough and often manifests as a long-term chronic inflammatory state, weight loss, and malnutrition.

[0003] Currently, the internationally accepted diagnosis of aspiration pneumonia mainly relies on three aspects: a visible history of aspiration, risk factors, and imaging findings. A visible history of aspiration can be determined using several methods: overt aspiration history or a positive Kubota water swallowing test, fruit green test, video-assisted fluoroscopy (VFSS), or fiberoptic nasopharyngoscopy (FEES). However, this may present several problems: First, patients with latent aspiration pneumonia may experience intermittent aspiration due to poor tolerance, and those with milder cases may not have aspirated at the time of examination. Therefore, the commonly used Kubota water swallowing test may not accurately reflect the diagnosis of latent aspiration pneumonia. While FEES or VFSS can relatively accurately assess aspiration, these invasive examinations are also prone to false negatives, and patient compliance is often poor. Second, inferring aspiration based solely on past medical history (risk factors) is highly subjective. Finally, the imaging findings in the diagnostic criteria for aspiration pneumonia are often less reliable than those in other types of pneumonia, making it difficult to determine the presence of an aspiration cause.

[0004] Therefore, the aforementioned diagnostic criteria often lead doctors to only diagnose patients with obvious clinical symptoms of aspiration pneumonia, while a large number of patients with latent aspiration pneumonia are missed. According to relevant research reports, after autopsy pathological examination of 59 cases of aspiration pneumonia, only 4 patients were definitively diagnosed with aspiration pneumonia before death, accounting for only 6.8%. This means that in this study sample, as many as 93.2% of aspiration pneumonia cases were not diagnosed in a timely manner before death, indicating a serious problem of missed diagnosis. A survey focusing on the current status of aspiration pneumonia diagnosis in routine clinical practice found that doctors exhibit a specific tendency when determining aspiration pneumonia. When faced with patients who are physically weak, suffer from multiple comorbidities, and whose clinical manifestations indicate a possibility of aspiration, even in the absence of conclusive direct evidence of aspiration, doctors are often more likely to diagnose aspiration pneumonia. This significantly increases the false positive rate and exacerbates the risk of misdiagnosis. Further investigation revealed that, in terms of diagnostic procedure implementation, the actual proportion of patients with aspiration pneumonia receiving FEES and VFSS was quite low, at only 4%. This highlights a gap between current diagnostic practices for aspiration pneumonia and the ideal standardized procedure. Therefore, there is an urgent need for detection methods with high sensitivity and specificity. Summary of the Invention

[0005] Therefore, to address the aforementioned problems, this invention proposes a potential serum biomarker and kit for diagnosing aspiration pneumonia. This solves the technical problem of missed diagnoses of early-stage aspiration pneumonia.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A potential serum biomarker for the diagnosis of aspiration pneumonia, the biomarker being selected from six biomarkers: quinolinic acid, phosphatidylcholine (18:0 / 20:4(5Z,8Z,11Z,14Z)), N6,N6,N6-trimethyl-L-lysine, lysophosphatidylcholine (17:0 / 0:0), DL-piperidinic acid, and galactitol.

[0008] The biomarkers were derived from serum.

[0009] The present invention also provides a kit for diagnosing potential serum biomarkers for aspiration pneumonia, the kit comprising: calibrators, low-value control samples, high-value control samples, internal standards, protein precipitants, diluents, electrolyte concentrate I, and electrolyte concentrate II.

[0010] The diluent for the high-value quality control product includes the biological preservative Proclin 300, BSA solution, and water. The internal standard includes acetonitrile and isopropanol. The protein precipitant includes isopropanol solution. The diluent includes the biological preservative Proclin 300, BSA solution, and water. Electrolyte concentrate I includes formic acid, ammonium formate, and an aqueous solution. Electrolyte concentrate II includes formic acid, ammonium formate, and an aqueous solution. The calibrator, low-value quality control product, high-value quality control product, and diluent are lyophilized products, while the internal standard, electrolyte concentrate I, and electrolyte concentrate II are liquids.

[0011] The diluent for calibrators, low-value quality control products, and high-value quality control products consists of the biological preservative Proclin 300, BSA solution, and water. The ratio of biological preservative Proclin 300: 5% BSA solution: water is 0.1: 20: 79.9 (by volume). After mixing, the target substance is prepared using this diluent and then lyophilized to form a lyophilized product.

[0012] The internal standard includes acetonitrile and isopropanol, with an acetonitrile:isopropanol ratio of 4:3, by volume.

[0013] The protein precipitant includes an isopropanol solution.

[0014] The diluent comprises the biological preservative Proclin 300, BSA solution, and water, with the ratio of biological preservative Proclin 300: 5% BSA solution: water being 1:20:979, and is prepared by mixing them thoroughly by volume.

[0015] The electrolyte concentrate I comprises 50% formic acid, 0.5M ammonium formate, and an aqueous solution.

[0016] The electrolyte concentrate II comprises 50% formic acid, 0.1M ammonium formate, and an aqueous solution.

[0017] By adopting the aforementioned technical solution, the beneficial effects of the present invention are:

[0018] This study employed targeted metabolomics, utilizing selective reaction / multiple reaction monitoring (MRM) technology with triple quadrupole tandem mass spectrometry, and using standards as a reference, to identify six biomarkers that can accurately distinguish between aspiration pneumonia and non-aspiration pneumonia. Therefore, targeted and specific detection and analysis were conducted on these six potential AP biomarkers. Simultaneously, machine learning was combined to construct a predictive model for AP and non-aspiration pneumonia (Non-AP), aiming to identify and compare metabolic characteristics in the serum of AP patients and develop a clinically applicable diagnostic kit, thus filling the aforementioned research gaps. These findings are expected to advance the development of biomarkers for early AP diagnosis, providing strong support for overcoming the current challenges in AP diagnosis. The kit formulation was developed around these six biomarkers and can effectively distinguish between aspiration pneumonia and non-aspiration pneumonia. Attached Figure Description

[0019] Figure 1 These are the ROC and PRC curves for the Model_0001 model, representing its performance evaluation.

[0020] Figure 2 This is a graph showing the analysis results of the ROC and PRC curves for the Model_0001 model.

[0021] Figure 3 This is the confusion matrix diagram of the Model_0001 model.

[0022] Figure 4 This is a graph showing the analysis results of the ROC and PRC curves for the Model_0002 model.

[0023] Figure 5 This is the confusion matrix diagram of the Model_0002 model.

[0024] Figure 6 This is a graph showing the ROC and PRC curve analysis results of Model_0001 model in predicting acute and chronic aspiration pneumonia.

[0025] Figure 7 This is a diagram showing the confusion matrix analysis results of Model_0001's predictions of acute and chronic aspiration pneumonia.

[0026] Figure 8 This is a graph showing the ROC and PRC curve analysis results of Model_0002's predictions of acute and chronic aspiration pneumonia.

[0027] Figure 9 This is a diagram showing the confusion matrix analysis results of Model_0002's predictions of acute and chronic aspiration pneumonia.

[0028] Figure 10This is a bar chart ranking the feature importance of Model_0001 based on SHAP analysis.

[0029] Figure 11 This is a two-dimensional scatter plot of the feature SHAP values ​​of the Model_0001 model.

[0030] Figure 12 This is a bar chart ranking the feature importance of Model_0002 based on SHAP analysis.

[0031] Figure 13 This is a two-dimensional scatter plot of the feature SHAP values ​​of the Model_0002 model. Detailed Implementation

[0032] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0033] refer to Figures 1 to 13 This embodiment provides a potential serum biomarker for diagnosing aspiration pneumonia. The biomarker is selected from six biomarkers: quinolinic acid, phosphatidylcholine (18:0 / 20:4(5Z,8Z,11Z,14Z)), N6,N6,N6-trimethyl-L-lysine, lysophosphatidylcholine (17:0 / 0:0), DL-piperidine acid, and galactitol. The biomarkers are derived from serum.

[0034] I. Sample Collection: A phased inclusion strategy was adopted. From January 2022 to December 2023, confirmed cases of aspiration pneumonia and non-aspiration pneumonia were directly included from hospitals to construct a training cohort. From October 2024 to May 2025, newly confirmed cases of aspiration pneumonia and non-aspiration pneumonia were included from hospitalized patients using the same diagnostic criteria to construct an independent testing cohort, which was divided into AP group and Non-AP group.

[0035] The diagnostic criteria are as follows:

[0036] a. Diagnostic criteria for community-acquired pneumonia: Meeting the criteria for community-acquired pneumonia and having any one of the following clinical manifestations: fever, cough, expectoration or worsening of pre-existing respiratory symptoms, with or without purulent sputum, chest pain, dyspnea and hemoptysis; signs of pulmonary consolidation and / or moist rales; peripheral blood leukocyte count >10×10^9 / L or <4×10^9 / L, with or without left shift of the nuclei; and chest imaging showing newly appearing patchy infiltrates, lobar or segmental consolidation, ground-glass opacities or interstitial changes, with or without pleural effusion.

[0037] b. Diagnostic criteria for hospital-acquired pneumonia: The patient was not in the incubation period of the pathogen infection at the time of admission, and the pneumonia occurred in the hospital ≥48 hours after admission.

[0038] c. Diagnostic criteria for aspiration pneumonia: Meets the diagnostic criteria for acquired pneumonia or hospital-acquired pneumonia, and meets one of the following conditions: an aspiration event is directly observed during bronchoscopy or a clear aspiration event is clinically confirmed during hospitalization.

[0039] d. Diagnostic criteria for non-aspiration pneumonia: The diagnosis must first meet the criteria for community-acquired pneumonia or hospital-acquired pneumonia, and simultaneously satisfy the following conditions: ① No definite aspiration events were observed during bronchoscopy and hospitalization; ② The following factors are absent: esophageal disease (dysphagia, esophageal cancer, esophageal stricture); degenerative neurological diseases (dementia, multiple sclerosis, Parkinson's disease); history of cerebral infarction or cerebral hemorrhage; post-extubation status after invasive mechanical ventilation; history of cardiac arrest; drug overdose or alcohol poisoning; use of general anesthetics; history of epileptic seizures; gastroesophageal reflux disease; prolonged tube feeding (duration > 4 weeks); prolonged endotracheal intubation (duration > 14 days); tracheostomy; swallowing dysfunction due to other causes; impaired cough reflex due to other causes; altered consciousness due to other causes.

[0040] The grouping criteria for aspiration pneumonia are as follows: a. Acute aspiration pneumonia: meets the diagnostic criteria for aspiration pneumonia and the onset time is less than 1 month; b. Chronic aspiration pneumonia: meets the diagnostic criteria for aspiration pneumonia, the onset time is more than 1 month or the family reports that the patient has experienced repeated aspiration events over a long period of time.

[0041] Exclusion criteria are as follows: a. Patients with evidence of airway bleeding, pulmonary tuberculosis, malignancy, and pulmonary embolism (for training cohort patients only); b. Patients who cannot receive bronchoalveolar lavage for other reasons.

[0042] II. Sample processing:

[0043] Take an appropriate amount of the mixed stock solution of 6 compound standards and dilute it with 5% rat plasma to prepare a mixed standard solution with a concentration ranging from 1 ng / mL to 5000 ng / mL. After thawing the serum sample at 4°C, accurately transfer 50 μL of the sample and the standard curve solution into a 1.5 mL centrifuge tube, add 200 μL of methanol (containing the mixed isotope internal standard solution), and vortex at 2000 rpm for 5 min to mix. After mixing, place the 1.5 mL centrifuge tube in a centrifuge and centrifuge at 14000 rpm at 4°C for 10 min. After centrifugation, take 160 μL of the supernatant and freeze-dry it. When performing the detection, take the freeze-dried sample and add 50 μL of 0.1% formic acid solution to reconstitute it. Analyze all samples using mass spectrometry, with 10 μL of each sample being analyzed.

[0044] III. Mass Spectrometry Procedures:

[0045] 1. Chromatographic Detection Conditions: Metabolites in the sample were separated using a reversed-phase column and a hydrophilic column on an Ultimate™ 3000 ultra-high performance liquid chromatograph coupled with a TSQ Endura triple quadrupole tandem mass spectrometer system, followed by high-resolution mass spectrometry detection. The mobile phase was water / methanol, with 0.1% FA and 2 mM ammonium formate buffer added to both solvents. Metabolites were separated on an Acquity™ BEH C18 column with the following elution gradient: organic phase increased from 2% to 100% within 12 min, with an additional 6 min for washing and column equilibration. The flow rate and column temperature were 0.3 ml / min and 40 °C, respectively, and detection was performed simultaneously in positive and negative ion modes.

[0046] 2. High-resolution mass spectrometry detection conditions: Metabolite detection was performed on a TSQ Endura mass spectrometer, with data acquisition in multiple reaction monitoring (MRM) mode. Mass spectrometry employed a heated electrospray ionization source, with an ionization voltage of 3.5 kV for positive ion detection mode and 3 kV for negative ion detection mode. The sheath gas flow rate was 40 arb, the auxiliary gas flow rate was 10 arb, the nebulizer temperature was 350 °C, the capillary temperature was 275 °C, the Q1 resolution was 0.7 FWHM, the Q3 resolution was 1.2 FWHM, the collision pressure was 1.5 mTorr, and ultrapure helium was used as the fragmentation gas.

[0047] 3. Key metabolic molecules:

[0048]

[0049] 4. Mass spectrometry data processing (data quality control, PCA, OPLS-DA, ROC analysis, etc.):

[0050] (1) Metabolite identification: TraceFinder quantitative analysis software was used to extract the abundance data of each target metabolite based on the characteristic retention time, parent ion and characteristic fragment ion information of the metabolite, and further data quality control and standardization were carried out.

[0051] (2) Bioinformatics analysis: TraceFinder was used to obtain the abundance data (Areaunder the cures, AUCs) of each metabolite in the sample to be tested as the quantitative information of the metabolites. The quantitative intensity of metabolites was imported into R and Python programming languages ​​for various differential analysis and enrichment analysis, including PCA, PLS-DA, cluster analysis, pathway enrichment analysis, etc.

[0052] 5. Machine Learning:

[0053] Samples collected in 2024 were used as the training set, and samples collected in 2025 were used as the validation set to evaluate the model's performance when incorporating new data. The samples were grouped into AP and NonAP categories. The AP group included acute aspiration pneumonia (AcuteAP) and chronic aspiration pneumonia (ChronicAP). All six endogenous metabolic molecules were included in the machine learning algorithm modeling process. In the machine learning algorithm, a cohort of 121 samples served as the training set, and a cohort of 54 samples served as the validation set. The goal of using the machine learning algorithm was to identify one or more combinations of metabolite biomarkers to build predictive models for AP and NonAP. These predictive models were then used to predict the performance of AcuteAP and ChronicAP.

[0054] (1) Model cohort data preprocessing: The intensity values ​​of the six metabolic molecular variables were first log-transformed to a minimum of 2 (Log2(intensity value of metabolic molecule)) to make the distribution of each variable in the dataset approximate a normal distribution. Subsequently, Z-score standardization was performed on the six variables in the training and test sets respectively. The Z-score (standard score) is used to measure the deviation of a data point from the mean of the dataset, expressed in standard deviation. Its basic formula is: Z = (X - μ) / σ. X represents the Log2(intensity value) of a variable in a certain sample, μ represents the mean of the Log2(intensity value) of a variable in all samples, and σ represents the standard deviation of the Log2(intensity value) of a variable in all samples.

[0055] (2) Screening of metabolic molecular feature models:

[0056] Models were constructed using eight common machine learning algorithms and six metabolic molecules. The eight machine learning algorithms are Logistic Regression, Linear Support Vector Machine (SVM), Radial Basis Function Support Vector Machine (RBF SVM), Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), Adaptive Boosting (AdaBoost), and Extreme Gradient Boosting (XGBoost).

[0057] The XGBoost algorithm package used was XGBoost (v2.1.4), while the other algorithms used the scikit-learn package (v1.6.1). All programs were executed in a Python programming environment (v3.11). Feature selection was performed using the `feature_selection.SequentialFeatureSelector` function from the scikit-learn package. The input data consisted of the Z-scores of six molecules standardized from the training and test sets, which had been preprocessed to meet the model's input requirements. The `estimator` parameter of the feature selection function was sequentially input into the execution functions of the eight machine learning algorithms mentioned above to comprehensively evaluate the feature selection performance of different models.

[0058] Each algorithm performs the following process: First, the direction parameter is set to forward for forward feature selection, and the n_features_to_select parameter is set to an integer range from 1 to 17 with a step size of 1. A grid search method combined with a 5-fold cross-validation strategy is used to optimize the model hyperparameters in each iteration and calculate the AUROC performance index for the training and test sets respectively. Subsequently, the optimal feature set is selected based on the AUROC. After completing the forward selection, keeping other parameters unchanged, the direction parameter is switched to backward for reverse feature elimination, and the model performance index AUROC for each feature subset size is recorded. Random seeds are set for all the above experimental processes to ensure the repeatability of the results, and the n_jobs parameter is set to 20 for parallel computation. Finally, the AUROC values ​​of all models are summarized, sorted from high to low, and the top model with the highest AUROC value on the validation set is selected as the final feature selection scheme. The corresponding optimal feature subset and machine learning algorithm are also recorded.

[0059] After the above steps, each algorithm screens out multiple biomarker combinations of metabolites.

[0060] (3) Model performance evaluation:

[0061] This invention evaluates model performance using training and testing sets. The evaluation metrics include: ROC curve (Receiver Operating Characteristic Curve) and its area under the curve (AUC), Precision-Recall curve, Accuracy, Balanced Accuracy, Sensitivity, Specificity, Precision, Negative Predictive Value (NPV), and F1 Score.

[0062] The ROC curve is a key indicator for evaluating the performance of binary classification models. It uses the false positive rate (FPR) as the horizontal axis and the true positive rate (TPR / sensitivity) as the vertical axis. The area under the curve (AUC) quantifies the model's discrimination ability. The closer the AUC is to 1, the better the performance.

[0063] The Precision-Recall curve plots Recall on the x-axis and Precision on the y-axis. The Average Precision value of the curve reflects the model's precision; a value closer to 1 indicates better performance.

[0064]

[0065] Parameter description: k represents a threshold (ranging from 0 to 1, with a total of n). P k R represents the precision value corresponding to k thresholds. k This represents the Recall value corresponding to k thresholds.

[0066]

[0067] Furthermore, the Bootstrap resampling method (1000 repeated samplings) was used to calculate the 95% confidence interval (95% CI) for each performance index. By performing random sampling with replacement on the predicted sample probability values ​​and the true group labels and repeatedly modeling and evaluating, the 2nd and 97th percentiles of the index values ​​were taken as the upper and lower limits of the confidence interval after obtaining the data distribution. The results show that at a 95% confidence level, the probability that the model's true performance index falls within the interval is 95%.

[0068] Based on the evaluation of eight commonly used machine learning algorithms, the biomarker combination with the best performance on the validation set (i.e., the model with the highest AUROC value) was selected, and ROC and PRC curves were plotted for comprehensive performance comparison and analysis. Experimental data shows that the biomarker combination of the XGBoost algorithm exhibits the best discriminative performance: the top-1 model (model_0001) has an area under the curve (AUROC) of 0.984 on the training set and 0.905 on the validation set; the top-2 model (model_0002) has an area under the curve (AUROC) of 0.909 on the training set and 0.900 on the validation set.

[0069] This invention uses the XGBoost algorithm's Model_0001 and Model_0002 models as representative embodiments for detailed description. The biomarker combinations and performance parameters corresponding to these models have been fully recorded. It should be noted that the scope of protection of this invention is not limited to this specific biomarker combination; complete model performance data and corresponding biomarker combinations for seven other machine learning algorithms have also been systematically recorded.

[0070] refer to Figure 1 , Figure 1 The chart shows the performance evaluation results for the ROC curve (first row) and PRC curve (second row), where the first column represents the test set data and the second column represents the independent validation set data. Each graph includes the top-1 models of eight machine learning algorithms, and each curve represents the prediction result of one algorithm. In the ROC curve, the horizontal axis (X-axis) represents the false positive rate, and the vertical axis (Y-axis) represents the true positive rate; in the PRC curve, the horizontal axis (X-axis) represents the recall rate, and the vertical axis (Y-axis) represents the precision rate.

[0071] The molecular combination of the XGBoost algorithm's Model_0001 model is: ['Galactitol','LysoP C(17:0 / 0:0)','N6,N6,N6-Trimethyl-L-lysine','PC(18:0 / 20:4(5Z,8Z,11Z,14Z))','Pipecolicacid']. The AUROC of the training set is 0.984 (95% CI: 0.961–1.000), and the AUROC of the validation set is 0.905 (95% CI: 0.826–0.983). Based on the probability values ​​of the validation set samples, the optimal cutoff value of the model is calculated to be 0.616 using the Youden's index method. The Youden index can be expressed as J = Sensitivity + Specificity - 1, with a value ranging from -1 to 1. A value closer to 1 indicates better overall model performance in terms of both sensitivity and specificity. Based on a cutoff value of 0.616, if the model predicts a sample with a probability greater than or equal to 0.616, then the sample is considered AP (Average Probability); otherwise, it is considered Non-AP (Non-AP). In the training set, 77 out of 79 AP samples were correctly predicted as AP samples, and 36 out of 42 Non-AP samples were correctly predicted as Non-AP samples. In the validation set, 34 out of 38 AP samples were correctly predicted as AP samples, and 12 out of 16 Non-AP samples were correctly predicted as Non-AP samples.

[0072] refer to Figure 2The training set accuracy was 0.934 (95% CI: 0.884–0.975), and the validation set accuracy was 0.852 (95% CI: 0.759–0.944). The balanced training set accuracy was 0.916 (95% CI: 0.851–0.969), and the validation set accuracy was 0.822 (95% CI: 0.697–0.928). The training set mean precision was 0.992 (95% CI: 0.976–1.000), and the validation set precision was 0.963 (95% CI: 0.913–1.000). The training set negative predictive value was 0.947 (95% CI: 0.857–1.000), and the validation set negative predictive value was 0.750 (95% CI: 0.529–0.938). The F1 score reached 0.951 (95% CI: 0.912–0.982) on the training set and 0.895 (95% CI: 0.817–0.961) on the validation set. Sensitivity metrics showed 0.975 (95% CI: 0.934–1.000) on the training set and 0.895 (95% CI: 0.795–0.975) on the validation set; specificity was 0.857 (95% CI: 0.744–0.954) on the training set and 0.750 (95% CI: 0.526–0.938) on the validation set. Precision was 0.928 (95% CI: 0.867–0.976) on the training set and 0.895 (95% CI: 0.794–0.975) on the validation set. These results demonstrate that the combination of biomarkers has good discriminative ability and generalized predictive performance.

[0073] refer to Figure 3 , Figure 3 The image shows the confusion matrix of the XGBoost algorithm Model_0001 model. The left image represents the training set, and the right image represents the validation set.

[0074] The molecular combination of the XGBoost algorithm's Model_0002 model is: ['Galactitol','N6,N6,N6-Trimethyl-L-lysine','PC(18:0 / 20:4(5Z,8Z,11Z,14Z))','Quinolinicacid']. The AUROC of the training set is 0.909 (95% CI: 0.857-0.960), and the AUROC of the validation set is 0.900 (95% CI: 0.820-0.980). Based on the probability values ​​of the validation set samples, the optimal cutoff value of the model is calculated to be 0.613 using the Youden's index method. The formula for the Youden's index can be expressed as J = Sensitivity + Specificity - 1, and the value of the Youden's index J ranges from [-1,1]. The closer the value is to 1, the better the overall performance of the model in terms of sensitivity and specificity. Based on a cutoff value of 0.613, if the model predicts a sample with a probability greater than or equal to 0.613, then the sample is considered AP; otherwise, it is considered NonAP. In the training set, 71 out of 79 AP samples were correctly predicted as AP samples, and 31 out of 42 NonAP samples were correctly predicted as NonAP samples. In the validation set, 34 out of 38 AP samples were correctly predicted as AP samples, and 13 out of 16 NonAP samples were correctly predicted as NonAP samples.

[0075] refer to Figure 4 and Figure 5The training set accuracy was 0.843 (95% CI: 0.777–0.909), and the validation set accuracy was 0.870 (95% CI: 0.777–0.944). The balanced training set accuracy was 0.818 (95% CI: 0.742–0.893), and the validation set accuracy was 0.854 (95% CI: 0.724–0.958). The mean training set precision was 0.950 (95% CI: 0.912–0.989), and the validation set precision was 0.961 (95% CI: 0.910–1.000). The negative predictive value on the training set was 0.795 (95% CI: 0.666–0.925), and the validation set precision was 0.765 (95% CI: 0.538–0.944). The F1 score reached 0.882 (95% CI: 0.827–0.932) on the training set and 0.907 (95% CI: 0.827–0.966) on the validation set. Sensitivity metrics were 0.899 (95% CI: 0.831–0.962) on the training set and 0.895 (95% CI: 0.786–0.976) on the validation set; specificity was 0.738 (95% CI: 0.605–0.861) on the training set and 0.812 (95% CI: 0.579–1.000) on the validation set. Precision was 0.866 (95% CI: 0.788–0.937) on the training set and 0.919 (95% CI: 0.816–1.000) on the validation set. These results demonstrate that the Model_0002 group of biomarker combinations also has good discriminative ability and generalization prediction performance.

[0076] refer to Figure 6Performance analysis of Model_0001 model on chronic aspiration pneumonia (Chronic AP) and the control group (NonAP) showed the following: AUROC for the training set was 0.976 (95% CI: 0.944–1.000), and for the validation set it was 0.931 (95% CI: 0.847–1.000); accuracy for the training set was 0.907 (95% CI: 0.837–0.965), and for the validation set it was 0.857 (95% CI: 0.742–0.971); and the balanced accuracy for the training set was 0.906 (95% CI: 0.839–0.965), and for the validation set it was 0.849 (95% CI: 0.717–0.955). Regarding precision metrics, the mean precision on the training set was 0.981 (95% CI: 0.952–1.000), and on the validation set it was 0.944 (95% CI: 0.868–1.000). The negative predictive value on the training set was 0.947 (95% CI: 0.865–1.000), and on the validation set it was 0.923 (95% CI: 0.733–1.000). The F1 score reached 0.913 (95% CI: 0.846–0.968) on the training set and 0.878 (95% CI: 0.755–0.971) on the validation set. Sensitivity was 0.955 (95% CI: 0.886–1.000) on the training set and 0.947 (95% CI: 0.812–1.000) on the validation set. Specificity was 0.857 (95% CI: 0.745–0.953) on the training set and 0.750 (95% CI: 0.500–0.941) on the validation set. Precision was 0.875 (95% CI: 0.775–0.960) on the training set and 0.818 (95% CI: 0.652–0.957) on the validation set.

[0077] Performance analysis of Model_0001 in predicting acute aspiration pneumonia (Acute AP) and the control group (NonAP) showed the following: AUROC for the training set was 0.993 (95% CI: 0.975–1.000), and for the validation set it was 0.878 (95% CI: 0.770–0.987); the accuracy for the training set was 0.922 (95% CI: 0.857–0.974), and for the validation set it was 0.800 (95% CI: 0.657–0.914); the balanced accuracy for the training set was 0.929 (95% CI: 0.875–0.976), and for the validation set it was 0.796 (95% CI: 0.652–0.917). In terms of precision metrics, the mean precision on the training set was 0.992 (95% CI: 0.972–1.000), and on the validation set it was 0.928 (95% CI: 0.842–1.000). The negative predictive value on the training set was 1.000 (95% CI: 1.000–1.000), and on the validation set it was 0.800 (95% CI: 0.571–1.000). The F1 score reached 0.921 (95% CI: 0.841–0.976) on the training set and 0.821 (95% CI: 0.667–0.933) on the validation set. Sensitivity metrics showed a value of 1.000 (95% CI: 1.000–1.000) for the training set and 0.842 (95% CI: 0.667–1.000) for the validation set. Specificity was 0.857 (95% CI: 0.750–0.952) for the training set and 0.750 (95% CI: 0.500–0.941) for the validation set. Precision was 0.854 (95% CI: 0.725–0.953) for the training set and 0.800 (95% CI: 0.619–0.952) for the validation set.

[0078] refer to Figure 7 The confusion matrix analysis results of the XGBoost algorithm Model_0001 model (cutoff value 0.616) predicting acute aspiration pneumonia (Acute AP) and chronic aspiration pneumonia (Chronic AP) are shown. The left figure shows the confusion matrix of the training set data, and the right figure shows the confusion matrix of the independent validation set data. The horizontal axis (x-axis) of each sub-figure represents the model prediction group (Predicted Value), including the NonAP control group and the AP group; the vertical axis (y-axis) represents the true clinical diagnosis group (True Value), including the NonAP control group, the Acute AP group, and the Chronic AP group.

[0079] Performance analysis of Model_0002 in predicting chronic aspiration pneumonia (Chronic AP) and the control group (NonAP) showed the following: AUROC for the training set was 0.969 (95% CI: 0.930–1.000), and for the validation set it was 0.905 (95% CI: 0.807–1.000); the accuracy for the training set was 0.844 (95% CI: 0.766–0.922), and for the validation set it was 0.857 (95% CI: 0.743–0.971); the balanced accuracy for the training set was 0.855 (95% CI: 0.782–0.925), and for the validation set it was 0.854 (95% CI: 0.725–0.967). In terms of precision metrics, the mean precision on the training set was 0.968 (95% CI: 0.929–1.000), and on the validation set it was 0.940 (95% CI: 0.861–1.000); the negative predictive value on the training set was 0.969 (95% CI: 0.900–1.000), and on the validation set it was 0.867 (95% CI: 0.667–1.000). The F1 score reached 0.850 (95% CI: 0.747–0.923) on the training set and 0.872 (95% CI: 0.741–0.971) on the validation set. Sensitivity was 0.971 (95% CI: 0.903–1.000) on the training set and 0.895 (95% CI: 0.750–1.000) on the validation set; specificity was 0.738 (95% CI: 0.595–0.860) on the training set and 0.812 (95% CI: 0.600–1.000) on the validation set. Precision was 0.756 (95% CI: 0.614–0.872) on the training set and 0.850 (95% CI: 0.681–1.000) on the validation set.

[0080] refer to Figure 8Performance analysis of Model_0002 for predicting acute aspiration pneumonia (Acute AP) and the control group (NonAP) showed the following: AUROC for the training set was 0.861 (95% CI: 0.788–0.934), and for the validation set it was 0.895 (95% CI: 0.793–0.996); accuracy for the training set was 0.791 (95% CI: 0.709–0.872), and for the validation set it was 0.857 (95% CI: 0.743–0.971); and the balanced accuracy for the training set was 0.790 (95% CI: 0.702–0.872), and for the validation set it was 0.854 (95% CI: 0.717–0.957). In terms of precision metrics, the mean precision on the training set was 0.864 (95% CI: 0.791–0.936), and on the validation set it was 0.926 (95% CI: 0.840–1.000); the negative predictive value on the training set was 0.816 (95% CI: 0.692–0.930), and on the validation set it was 0.867 (95% CI: 0.667–1.000). The F1 score reached 0.804 (95% CI: 0.707–0.884) on the training set and 0.872 (95% CI: 0.737–0.968) on the validation set. Sensitivity was 0.841 (95% CI: 0.718–0.943) on the training set and 0.895 (95% CI: 0.733–1.000) on the validation set; specificity was 0.738 (95% CI: 0.600–0.865) on the training set and 0.812 (95% CI: 0.600–1.000) on the validation set. Precision was 0.771 (95% CI: 0.651–0.880) on the training set and 0.850 (95% CI: 0.684–1.000) on the validation set.

[0081] Figure 9 The confusion matrix analysis results for the Boost algorithm Model_0002 model (cutoff value 0.613) predicting acute aspiration pneumonia (Acute AP) and chronic aspiration pneumonia (Chronic AP) are presented. The left figure shows the confusion matrix of the training set data, and the right figure shows the confusion matrix of the independent validation set data. The horizontal axis (x-axis) of each subplot represents the model prediction group (Predicted Value), including the NonAP control group and the AP group; the vertical axis (y-axis) represents the true clinical diagnosis group (True Value), including the NonAP control group, the Acute AP group, and the Chronic AP group.

[0082] Model Feature Importance Evaluation Method: To analyze the contribution of each feature to the prediction model, this invention employs the SHAP (SHapley Additive exPlanations) algorithm for system evaluation. This algorithm is a commonly used feature importance interpretation method in the field of machine learning and can be implemented using the Python SHAP (v0.41.2) package. The core principle of the SHAP algorithm is based on Shapley values ​​in game theory. It quantifies the importance of each feature by calculating the mean marginal contribution of each feature in multiple feature combinations. The algorithm assigns a SHAP value to each feature, which represents the contribution of that feature relative to the baseline prediction value (positive values ​​indicate promotion of target classification, negative values ​​indicate inhibition). The feature contribution of complex models such as XGBoost is efficiently calculated using the TreeSHAP algorithm.

[0083] The `Explainer` function, implemented in Python's `SHAP` package, requires two key parameters: a trained prediction model (such as an XGBoost classifier) ​​and a training set as background data (used to construct a reference baseline for feature distribution). After the explainer is built, the validation set data of the biomarker combination is input into the `shap_values()` function, which calculates the `SHAP` value of each biomarker feature in each sample of the validation set. This value quantifies the contribution of a specific feature to the model's individual prediction results (positive values ​​indicate promotion of the target classification, negative values ​​indicate inhibition).

[0084] This invention uses the mean absolute value of SHAP values ​​as a global feature importance index, and plots bar charts and two-dimensional scatter plots (also called Beeswarm plots) of the mean absolute value of SHAP values ​​to show the contribution of each feature to the model output, i.e., the SHAP value. SHAP calculation results show that the contribution of the five metabolic molecules in the Model_0001 model to the model prediction from high to low is Galactitol, N6,N6,N6-Trimethyl-L-lysine, Pipecolic acid, PC (18:0 / 20:4(5Z,8Z,11Z,14Z)) and LysoPC (17:0 / 0:0). SHAP calculation results show that the four metabolic molecules in the Model_0002 model contribute to the model prediction in the following order from highest to lowest: Galacitol, Quinolinic acid, N6,N6,N6-Trimethyl-L-lysine, and PC (18:0 / 20:4(5Z,8Z,11Z,14Z)).

[0085] Figure 10This is a bar chart ranking the feature importance of Model_0001 based on SHAP analysis. The vertical axis (Y-axis) arranges feature names in descending order of importance. The horizontal axis (X-axis) represents the mean absolute value of the corresponding feature's SHAP score, which directly reflects the feature's contribution to the model's prediction results—a higher mean value indicates greater importance of the feature in the overall model decision.

[0086] Figure 11 This is a two-dimensional scatter plot of the feature SHAP values ​​for the Model_0001 model. The X-axis represents the Shap values, divided into positive values ​​(indicating a positive impact on prediction) and negative values ​​(indicating a negative impact on prediction). The larger the absolute value of the Shap value, the greater the impact on prediction. The Y-axis represents the features, usually listed in descending order of the distribution width of the feature Shap values. Each point represents a sample, typically with red indicating high feature values ​​and blue indicating low feature values.

[0087] Figure 12 This is a bar chart ranking the feature importance of Model_0002 based on SHAP analysis. The vertical axis (Y-axis) arranges feature names in descending order of importance. The horizontal axis (X-axis) represents the mean absolute value of the corresponding feature's SHAP score, which directly reflects the feature's contribution to the model's prediction results—a higher mean value indicates greater importance of the feature in the overall model decision.

[0088] Figure 13 This is a two-dimensional scatter plot of the feature SHAP values ​​for the Model_0002 model. The X-axis represents the Shap values, divided into positive values ​​(indicating a positive impact on prediction) and negative values ​​(indicating a negative impact on prediction). The larger the absolute value of the Shap value, the greater the impact on prediction. The Y-axis represents the features, usually listed in descending order of the distribution width of the feature Shap values. Each point represents a sample, typically with red indicating high feature values ​​and blue indicating low feature values.

[0089] A mass spectrometry detection method for six metabolic biomarkers was developed into a kit for convenient clinical testing. After protein precipitation, plasma or serum samples, using isotopes as internal standards, undergo a series of pretreatments, followed by separation by liquid chromatography. Mass spectrometry in multiple reaction monitoring (MRM) mode acquires the specific precursor / daughter ion mass spectrometry signals, recording the peak area of ​​the target analyte and the peak area of ​​the internal standard in each sample. The ratio of the peak area of ​​the target analyte to the corresponding internal standard in each sample is calculated. A standard curve is plotted and fitted using the ratio of the labeled concentration (x) of the target analyte in the calibrator to the corresponding peak area (y). Then, by substituting the ratios of the peak area of ​​the target analyte to the corresponding internal standard in other samples into the fitted standard curve equation, the concentration of the target analyte in different samples can be calculated.

[0090] A kit for diagnosing potential serum biomarkers for aspiration pneumonia, comprising: calibrators, low-value controls, high-value controls, internal standards, protein precipitants, diluents, electrolyte concentrate I, and electrolyte concentrate II.

[0091] The diluent for calibrators, low-value quality control products, and high-value quality control products consists of the biological preservative Proclin 300, BSA solution, and water. The ratio of biological preservative Proclin 300: 5% BSA solution: water is 0.1: 20: 79.9 (by volume). After mixing, the target substance is prepared using this diluent and then lyophilized to form a lyophilized product.

[0092] The internal standard concentrations of the target analyte are shown in the table below:

[0093]

[0094] The internal standard includes acetonitrile and isopropanol, with an acetonitrile:isopropanol ratio of 4:3. This solution is used to prepare an internal standard of a certain concentration of the target analyte by volume.

[0095] The internal standard concentrations of the target analyte are shown in the table below:

[0096]

[0097] The protein precipitant includes an isopropanol solution. The diluent includes the biological preservative Proclin 300, BSA solution, and water, in a ratio of Proclin 300:5% BSA solution:water = 1:20:979 (volume ratio). Electrolyte concentrate I consists of 50% formic acid, 0.5M ammonium formate, and an aqueous solution. Electrolyte concentrate II consists of 50% formic acid, 0.1M ammonium formate, and an aqueous solution. The liquid chromatography-tandem mass spectrometry kit mainly includes: calibrators (lyophilized), low-value quality control samples (lyophilized), high-value quality control samples (lyophilized), internal standard (liquid), protein precipitant, diluent (lyophilized), electrolyte concentrate I (liquid), and electrolyte concentrate II (liquid).

[0098] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A potential serum marker for diagnosing aspiration pneumonia, characterized by: The biomarkers are selected from the group consisting of quinolinic acid, phosphatidylcholine (18:0 / 20:4 (5Z,8Z,11Z,14Z)), N6,N6,N6-trimethyl-L-lysine, lysophosphatidylcholine (17:0 / 0:0), DL-piperidinic acid and galactitol.

2. A potential serum marker required for diagnosis of aspiration pneumonia according to claim 1, wherein: The biomarkers are from serum.

3. A kit for diagnosing the potential serum markers required for aspiration pneumonia, characterized by: The kit comprises: a calibrator, a low-value quality control, a high-value quality control, an internal standard, a protein precipitant, a diluent, electrolyte concentrate I and electrolyte concentrate II.

4. A kit for diagnosing the required potential serum markers for aspiration pneumonia according to claim 3, characterized by: The diluent configuration of the high-value quality control comprises biological preservative Proclin 300, BSA solution and water, the internal standard comprises acetonitrile and isopropanol, the protein precipitant comprises isopropanol solution, the diluent comprises biological preservative Proclin 300, BSA solution and water, electrolyte concentrate I comprises formic acid, ammonium formate and aqueous solution, and electrolyte concentrate II comprises formic acid, ammonium formate and aqueous solution, wherein the calibrator, the low-value quality control, the high-value quality control and the diluent are lyophilized products, and the internal standard, electrolyte concentrate I and electrolyte concentrate II are liquids.

5. A kit for diagnosing the required potential serum markers for aspiration pneumonia according to claim 3, characterized by: The diluent configuration for the calibrator, the low-value quality control and the high-value quality control comprises biological preservative Proclin 300, BSA solution and water, and the biological preservative Proclin 300:5% BSA solution:water = 0.1:20:79.9 in terms of volume ratio, which is mixed and then used to prepare the target lyophilized product.

6. A kit for diagnosing the required potential serum markers for aspiration pneumonia according to claim 3, characterized by: The internal standard comprises acetonitrile and isopropanol, and acetonitrile:isopropanol = 4:3 in terms of volume ratio.

7. A kit for diagnosing the required potential serum markers for aspiration pneumonia according to claim 3, characterized by: The protein precipitant comprises isopropanol solution.

8. A kit for diagnosing the required potential serum markers for aspiration pneumonia according to claim 3, characterized by: The diluent comprises biological preservative Proclin 300, BSA solution and water, and the biological preservative Proclin 300:5% BSA solution:water = 1:20:979 in terms of volume ratio, which is mixed to prepare the diluent.

9. A kit for diagnosing the required potential serum markers for aspiration pneumonia according to claim 3, characterized by: Electrolyte concentrate I comprises 50% formic acid, 0.5M ammonium formate and aqueous solution.

10. A kit for diagnosing the required potential serum markers for aspiration pneumonia according to claim 3, characterized by: Electrolyte concentrate II comprises 50% formic acid, 0.1M ammonium formate and aqueous solution.

Images