System for predicting outcome of ICU patient based on bronchoalveolar lavage fluid mNGS flora
By performing mNGS microbial community analysis and machine learning on bronchoalveolar lavage fluid, differential microbial community markers were screened out, and a classifier model was constructed. This solved the problems of lag and insufficient specificity in predicting outcomes of ICU patients, and achieved more accurate prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DINFECTOME
- Filing Date
- 2025-12-05
- Publication Date
- 2026-05-05
AI Technical Summary
Existing technologies suffer from lag and insufficient specificity in predicting outcomes for ICU patients. Traditional clinical indicators are difficult to reflect changes in the condition in the early stages, and the lack of specificity of single or partial indicators limits the accuracy of predictions.
Metagenomic sequencing (mNGS) of bronchoalveolar lavage fluid was performed to screen for differentially expressed microbial biomarkers. A classifier was then constructed using machine learning algorithms to predict microbial community characteristics and clinical features in bronchoalveolar lavage fluid.
It enables more accurate and timely prediction of outcomes for ICU patients, provides a valuable window of opportunity for intervention, and improves the accuracy of prediction and its clinical application value.
Smart Images
Figure CN121983125A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of molecular biomedical technology, and in particular to a method for accurately predicting ICU patient outcomes based on mNGS microbiota and machine learning. Background Technology
[0002] Intensive Care Units (ICUs) house the most critically ill patients in hospitals, whose conditions progress rapidly and are highly complex and variable. Patient survival outcomes are influenced by a combination of factors, including the severity of underlying diseases, organ function, infection status, and the effectiveness of treatment interventions. For a long time, clinical prediction of ICU patient outcomes (survival or death) has primarily relied on traditional clinical indicators, such as vital signs (heart rate, blood pressure, respiratory rate, etc.), laboratory test indicators (complete blood count, liver and kidney function, inflammatory markers, etc.), and organ function scores (such as APACHE II score, SOFA score, etc.). However, these traditional methods have significant limitations: firstly, many clinical indicators often exhibit lag effects, making it difficult to reflect prognostic trends in the early stages of critical turning points in a patient's condition; secondly, the specificity of single or partial indicators is insufficient, as patients with different etiologies and at different stages may exhibit similar indicator changes, thus limiting the accuracy of predictions.
[0003] With the development of molecular biology techniques, metagenomic sequencing (mNGS) technology has provided a powerful tool for the study of microbial communities. It can sequence the nucleic acids of all microorganisms in a sample directly without culturing, comprehensively and rapidly analyzing the composition of the microbial community and information on pathogenic microorganisms, demonstrating unique advantages in the pathogen diagnosis of infectious diseases. Meanwhile, machine learning, as an important branch of artificial intelligence, has been increasingly widely used in disease prediction and diagnosis in the medical field due to its powerful data analysis and pattern recognition capabilities, enabling it to uncover potential patterns and correlations from massive amounts of clinical and biological information.
[0004] While mNGS technology has made progress in microbial analysis and machine learning in medical prediction, there is still a lack of techniques that combine mNGS microbial analysis of bronchoalveolar lavage fluid (BALF) with machine learning for predicting outcomes in ICU patients. The microbial state of the lungs of ICU patients is closely related to lung infection and immune response, and may thus affect patient prognosis. Analyzing the microbial characteristics of BALF using mNGS technology and then combining it with machine learning to build predictive models holds promise for providing a new and more accurate method for predicting outcomes in ICU patients, thus overcoming the shortcomings of traditional prediction methods. Summary of the Invention
[0005] This application addresses the problems mentioned in the background art by analyzing the microbial community in bronchoalveolar lavage fluid, screening out different species from different populations, and using machine learning algorithms to incorporate clinical features to construct a classifier that accurately predicts the outcomes of ICU patients.
[0006] A diagnostic system for predicting outcomes in ICU patients, comprising:
[0007] The data processing module is configured to receive metagenomic sequencing (mNGS) data from bronchoalveolar lavage fluid samples of target ICU patients, and quantify the abundance information of at least one species marker selected from the following species marker groups from the mNGS data to generate feature data.
[0008] The group of bacterial biomarkers includes: *Streptococcus oralis*, *Pasteurellales*, *Pasteurellaceae*, *Actinomyces*, *Actinomycetaceae*, *Actinomycetales*, *Eubacterium saphenum*, *Eubacterium*, *Mogibacteriaceae*, *Peptostreptococcaceae*, *Pseudomonas wittichii*, *Lactococcus lactis*, *Lactococcus*, *Ralstonia pickettii*, *Acetobacteraceae*, *Rhodospirillales*, and *Roseomonas*. Mucosa, Roseomonas, and Corynebacterium tuberculostearicum.
[0009] The model prediction module is connected to the data processing module and is configured to input the feature data into a pre-built machine learning classification model, and the model calculates the predicted outcome of the target ICU patient.
[0010] The result output module is connected to the model prediction module and is configured to output the predicted outcome, which is survival or death.
[0011] The data processing module is also configured to receive clinical feature data associated with the target ICU patient; and the model prediction module is configured to combine the abundance information of the bacterial marker with the clinical feature data as a combined feature and input it into the machine learning classification model.
[0012] The clinical characteristic data include at least one of the following: length of hospital stay, sex, age, HLA-DR, total lymphocyte count, CD3+ cells, CD3+CD4+ cells, CD3+CD8+ cells, B cells, NK cells, and CD4+ / CD8+ ratio.
[0013] The machine learning classification model is a random forest model, a gradient boosting machine model, a generalized linear model, a deep learning model, or a stacked model that integrates at least two of the aforementioned models.
[0014] A computer-implemented method for predicting outcomes in ICU patients, comprising the following steps:
[0015] a) Obtain metagenomic sequencing (mNGS) data from bronchoalveolar lavage fluid samples from the target ICU patients;
[0016] b) From the mNGS data, quantify the abundance information of at least one species marker selected from the group of species markers defined in claim 1 to obtain feature data;
[0017] c) Input the feature data into a pre-built machine learning classification model to calculate the predicted outcome of the target ICU patient;
[0018] d) Output the predicted outcome, which is survival or death.
[0019] After step b) and before step c), it also includes:
[0020] The clinical characteristic data of the target ICU patient are obtained, and the clinical characteristic data is combined with the abundance information of the bacterial species markers, and used together as feature data input to the machine learning classification model.
[0021] The clinical characteristic data include at least one of the following: length of hospital stay, sex, age, HLA-DR, total lymphocyte count, CD3+ cells, CD3+CD4+ cells, CD3+CD8+ cells, B cells, NK cells, and CD4+ / CD8+ ratio.
[0022] The machine learning classification model is trained using a method that includes the following steps:
[0023] i) Collect bronchoalveolar lavage fluid samples from patients with a history of ICU care and divide them into survival and death groups based on their final outcomes;
[0024] ii) mNGS sequencing was performed on the historical samples, and the bacterial biomarker group was screened out by differential abundance analysis;
[0025] iii) Use the abundance information of the bacterial species markers as training features to train the machine learning classification model.
[0026] The use of a group of bacterial biomarkers in the preparation of systems or kits for predicting outcomes in ICU patients, said group of bacterial biomarkers includes: oral streptococci, Pasteurales, Pasteuraceae, Actinomycetes, Actinomycetes, Erwinia gingivalis, Oulsaceae, Peptostreptococci, Pseudomonas aeruginosa, Lactococcus lactis, Lactococcus spp., Roldstoneella pylori, Acetobacteraceae, Rhodospirillum, Rosomonas mucosa, Rosomonas spp., and Corynebacterium tuberculosis.
[0027] The purpose is to assess the risk of survival or death of an ICU patient by detecting the abundance of at least one of the group of bacterial markers in a bronchoalveolar lavage fluid sample.
[0028] The bacterial biomarker is selected from at least one of the following: oral streptococci, Pasteurales, Pasteuraceae, Actinomycetes, Actinomycetes, Actinomycetes, Erwinia gingivalis, Oulsaceae, Miscellanea, and Peptostreptococci; and an increase in the abundance of the biomarker indicates an increased risk of survival for the patient.
[0029] The bacterial biomarker is selected from at least one of the following: *Pseudomonas aeruginosa*, *Lactococcus lactis*, *Lactococcus* spp., *Rolstonia pinnili*, *Acetobacter* family, *Rhodospirillum* order, *Rhodospirillum mucosae*, *Rhodospirillum* spp., and *Corynebacterium tuberculosis*; and an increase in the abundance of the biomarker indicates an increased risk of death in the patient.
[0030] The beneficial effects of this invention are:
[0031] This invention innovatively combines mNGS microbiota data from bronchoalveolar lavage fluid (BALF) with machine learning, overcoming the shortcomings of traditional clinical indicators, such as lag and lack of specificity. By deeply mining pulmonary microecological imbalance—an early signal of disease progression—this method can achieve more accurate and timely prediction of ICU patient outcomes (survival / death), providing clinicians with a valuable window of opportunity for intervention.
[0032] This invention uses bioinformatics analysis to screen and validate a series of differentially expressed microbiota that are highly correlated with the outcomes of ICU patients. For example, oral streptococci and peptostreptococci enriched in the survival group, and Pseudomonas aeruginosa and lactococci enriched in the death group, together constitute a novel and objective set of biomarkers, providing a reliable and quantifiable biological basis for the construction of predictive models.
[0033] The machine learning model built based on this biomarker combination has demonstrated good predictive efficacy (AUC value of 0.722) through practical examples, and can output clear and intuitive risk warnings. This helps doctors quickly stratify patients by risk and develop more targeted and personalized intervention strategies for patients predicted to be at high risk, thereby achieving precise management of critically ill patients and demonstrating extremely high clinical translational value and application prospects. Attached Figure Description
[0034] Figure 1 The performance results of the classifier were constructed;
[0035] Figure 2 mNGS microbial community analysis LEfSe results; LEfSe analysis identified 19 discriminative biomarkers between the two groups. The surviving group contained 10 species, namely Oral Streptococcus, Pasteurales, Pasteuraceae, Actinomycetes, Actinomycetes, Actinomycetes, Erwinia gingivalis, Oltrusella, Miscellanea, and Peptostreptococci. The dead group contained 9 species, including Pseudomonas aeruginosa, Lactococcus lactis, Lactococcus, Ralstonia pinnatifida, Acetobacteraceae, Rhodospirillum, Rosomonas mucosa, Rosomonas spp., and Corynebacterium tuberculosis.
[0036] Figure 3 shows the performance of classifiers built using four algorithms: RF, GLM, GBM, and DP.
[0037] Figure 4 The classification performance of a classifier model using clinical features is evaluated. Detailed Implementation
[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0039] The specific technical solution is as follows:
[0040] A method for predicting ICU patient outcomes based on bronchoalveolar lavage fluid mNGS microbiota combined with machine learning includes the following steps:
[0041] S1. Sample collection and grouping: Collect bronchoalveolar lavage fluid samples from ICU patients and divide them into survival group and death group according to the final outcome of the patients;
[0042] S2. mNGS microbial community detection and analysis: Nucleic acid was extracted from the bronchoalveolar lavage fluid samples. After quality control, sequencing adapters were added to construct a library, and mNGS sequencing was performed. After the data was processed, the pathogenic microorganisms in the samples and the differential biomarkers between the two groups were analyzed.
[0043] S3. Machine Learning Model Construction and Prediction: Based on the differential biomarkers, construct machine learning models, evaluate model performance using receiver operating characteristic (ROC) curves and area under the curve (AUC), and select the best-performing model for predicting outcomes in ICU patients.
[0044] The mNGS sequencing data volume is 10-30M reads.
[0045] The differentially expressed species in the surviving patient population are oral streptococci, Pasteurella, Actinomycetes, Erwinia gingivalis, Mycobacteriaceae, and Peptostreptococci.
[0046] The differentially expressed species in the deceased patient population were *Pseudomonas aeruginosa*, *Lactococcus lactis*, *Rolstonia pinnatifida*, *Acetobacter*, *Rhodospirillum*, *Rhodospirillum mucosae*, and *Corynebacterium tuberculosis*.
[0047] The classifier model is a stacking model.
[0048] The stacked model combines Deep Learning (DL), Gradient Boosting Machine (GBM), Generalized Linear Model (GLM), and Random Forest (RF).
[0049] Example 1
[0050] 1. Collection and analysis of clinical samples
[0051] A total of 260 bronchoalveolar lavage fluid (BALF) specimens from ICU patients were included, divided into a survival group (n=198) and a death group (n=62) based on patient outcomes. Inclusion criteria were as follows: critically ill patients admitted to the ICU of a tertiary hospital between August 2021 and January 2024 were selected, meeting the following criteria: 1) Patients admitted to the ICU for ≥48 hours; 2) Patients exhibiting respiratory symptoms (e.g., cough, sputum production, dyspnea) requiring clinical BALF collection for lung assessment; 3) Patients with sufficient quality lower respiratory tract samples for mNGS analysis; 4) Complete clinical data allowing for 30-day survival outcome tracking (survival / death). Exclusion criteria were as follows: 1) Patients with an ICU stay of less than 24 hours; 2) Patients lacking complete clinical data (including microbiological test results and treatment records) required for analysis; 3) Patients and their families refusing to participate in this study.
[0052] Metagenomic sequencing was performed on bronchoalveolar lavage fluid from all patients. High-quality sequences obtained after filtering were compared with the human genome sequence (hs37d5) using bowtie2 software. Sequences that did not match hs37d5 were retained and compared with pathogen databases. The Kruskal-Wallis rank-sum test was used to identify differences in genus-level microbial abundance between the two groups.
[0053] 2. Machine Learning Model Construction and Outcome Prediction
[0054] The relative abundance of 19 differentially expressed species selected by LEfSe analysis was used as the feature value (the abundance information of the species was obtained by calculating the proportion of sequences aligned to a specific species to the total number of microbial sequences). A Random Forest (RF) algorithm was applied in conjunction with patient clinical characteristics (ICU length of stay, gender, age, HLA-DR (%), TLC (×10⁻⁶)). 9 / L), CD3+ cells (×10) 9 / L), CD3+CD4+ cells (×10 9 / L), CD3+CD8+ cells (×10 9 / L), B cells (×10 9 / L), NK cells (×10 9 A classifier model was constructed using CD4+ / CD8+ ratios ( / L) to further assess patient outcomes, and LOOCV was used to evaluate predictive ability. Model parameters were determined using grid search or similar optimization methods. In the random forest model, the number of trees was set to 50, and the number of variables to attempt in each split was determined to be 10 through cross-validation.
[0055] When modeling using the `h2o.randomForest` function in the R package `H2O`, leave-one-out cross-validation (LOOCV) was employed, and model parameters were determined by averaging the calculated results. The AUC value, sensitivity, and specificity were calculated using the `roc` function in the `pROC` package, and finally, an ROC curve was plotted. In this embodiment, when the model output probability of death is greater than 0.5, the predicted outcome is determined to be death; otherwise, it is determined to be survival (in actual implementation, this threshold can also be determined based on the maximum Youden exponent on the ROC curve to balance sensitivity and specificity). The results show that modeling using the random forest algorithm has a sensitivity of 64.6% and a specificity of 75%, with an area under the receiver operating characteristic (ROC) curve (AUC) of 0.722. Figure 3A ).
[0056] In addition, during the modeling process, the input variables of the aforementioned classifier model were modified, incorporating only the same clinical features that did not include the 19 differentially expressed clinical features, and training was performed using the same random forest model. The test results obtained are as follows. Figure 4 As shown, the area under the curve (AUC) is 0.689, indicating that the model obtained by using different species has better prediction accuracy.
[0057] In addition, we also used Generalized Linear Model (GLM), Gradient Boosting Machine (GBM), and Deep Learning (DP) algorithms combined with patient clinical features. We constructed the same classifier model using the h2o.glm, h2o.gbm, and h2o.deeplearning functions from the H2O package in R, under default parameter conditions, employing leave-one-out cross-validation. We then analyzed the performance of the three models. The GLM model achieved 73.8% sensitivity, 66.7% specificity, and an ACU value of 0.686. Figure 3B The GBM model has a sensitivity of 80%, a specificity of 63.9%, and an ACU value of 0.672. Figure 3C The DP model had a sensitivity of 86.2%, a specificity of 52.8%, and an ACU value of 0.667. Figure 3D The RF model used in this patent exhibits better classification characteristics.
[0058] 3. Validation of the prediction model
[0059] Ten ICU patients admitted in February 2024 who met the inclusion criteria were selected. Bronchoalveolar lavage fluid samples were collected from the patients according to the steps of this implementation method, mNGS microbiota analysis was performed, differential species characteristic variables were extracted, and the results were input into the optimized RF model for outcome prediction. The actual outcomes of the patients were tracked for 30 days. The results showed that the predicted outcomes of 7 out of the 10 patients were consistent with the actual outcomes (consistency rate of 70%), which further verified the practicality of this method.
Claims
1. A diagnostic system for predicting outcomes in ICU patients, characterized in that, include: The data processing module is configured to receive metagenomic sequencing (mNGS) data from bronchoalveolar lavage fluid samples of target ICU patients, and quantify the abundance information of at least one species marker selected from the following species marker groups from the mNGS data to generate feature data. The group of bacterial biomarkers includes: *Streptococcus oralis*, *Pasteurellales*, *Pasteurellaceae*, *Actinomyces*, *Actinomycetaceae*, *Actinomycetales*, *Eubacterium saphenum*, *Eubacterium*, *Mogibacteriaceae*, *Peptostreptococcaceae*, *Pseudomonas wittichii*, *Lactococcus lactis*, *Lactococcus*, *Ralstonia pickettii*, *Acetobacteraceae*, *Rhodospirillales*, and *Roseomonas*. Mucosa, Roseomonas, and Corynebacterium tuberculostearicum. The model prediction module is connected to the data processing module and is configured to input the feature data into a pre-built machine learning classification model, and the model calculates the predicted outcome of the target ICU patient. The result output module is connected to the model prediction module and is configured to output the predicted outcome, which is survival or death.
2. The system according to claim 1, characterized in that, The data processing module is also configured to receive clinical feature data associated with the target ICU patient; and the model prediction module is configured to combine the abundance information of the bacterial marker with the clinical feature data as a combined feature and input it into the machine learning classification model.
3. The system according to claim 2, characterized in that, The clinical characteristic data include at least one of the following: length of hospital stay, sex, age, HLA-DR, total lymphocyte count, CD3+ cells, CD3+CD4+ cells, CD3+CD8+ cells, B cells, NK cells, and CD4+ / CD8+ ratio.
4. The system according to claim 1, characterized in that, The machine learning classification model is a random forest model, a gradient boosting machine model, a generalized linear model, a deep learning model, or a stacked model that integrates at least two of the aforementioned models.
5. A computer-implemented method for predicting outcomes in ICU patients, characterized in that, Includes the following steps: a) Obtain metagenomic sequencing (mNGS) data from bronchoalveolar lavage fluid samples from the target ICU patients; b) From the mNGS data, quantify the abundance information of at least one species marker selected from the group of species markers defined in claim 1 to obtain feature data; c) Input the feature data into a pre-built machine learning classification model to calculate the predicted outcome of the target ICU patient; d) Output the predicted outcome, which is survival or death.
6. The method according to claim 5, characterized in that, After step b) and before step c), it also includes: The clinical characteristic data of the target ICU patient are obtained, and the clinical characteristic data is combined with the abundance information of the bacterial species markers, and used together as feature data input to the machine learning classification model.
7. The method according to claim 6, characterized in that, The clinical characteristic data include at least one of the following: length of hospital stay, sex, age, HLA-DR, total lymphocyte count, CD3+ cells, CD3+CD4+ cells, CD3+CD8+ cells, B cells, NK cells, and CD4+ / CD8+ ratio.
8. The method according to claim 5, characterized in that, The machine learning classification model is trained using a method that includes the following steps: i) Collect bronchoalveolar lavage fluid samples from patients with a history of ICU care and divide them into survival and death groups based on their final outcomes; ii) mNGS sequencing was performed on the historical samples, and the bacterial biomarker group was screened out by differential abundance analysis; iii) Use the abundance information of the bacterial species markers as training features to train the machine learning classification model.
9. The use of a group of bacterial biomarkers in the preparation of a system or kit for predicting outcomes in ICU patients, characterized in that, The group of bacterial markers includes: oral streptococci, Pasteurales, Pasteuraceae, Actinomycetes, Actinomycetes, Erwinia gingivalis, Euglena, Miscellanea, Peptostreptococci, Pseudomonas aeruginosa, Lactococcus lactis, Lactococcus, Ralstonia pinnatifida, Acetobacteraceae, Rhodospirillum, Rosomonas mucosa, Rosomonas spp., and Corynebacterium tuberculosis. The purpose is to assess the risk of survival or death of an ICU patient by detecting the abundance of at least one of the group of bacterial markers in a bronchoalveolar lavage fluid sample.
10. The use according to claim 9, characterized in that, The bacterial biomarker is selected from at least one of the following: oral streptococci, Pasteurales, Pasteuraceae, Actinomycetes, Actinomycetes, Actinomycetes, Erwinia gingivalis, Oulsaceae, Miscellanea, and Peptostreptococci; and an increase in the abundance of the biomarker indicates an increased risk of survival for the patient. The bacterial biomarker is selected from at least one of the following: *Pseudomonas aeruginosa*, *Lactococcus lactis*, *Lactococcus* spp., *Rolstonia pinnili*, *Acetobacter* family, *Rhodospirillum* order, *Rhodospirillum mucosae*, *Rhodospirillum* spp., and *Corynebacterium tuberculosis*; and an increase in the abundance of the biomarker indicates an increased risk of death in the patient.