We conducted a retrospective cohort study of 261 patients with suspected LRTIs from the ICU of Nanjing Drum Tower Hospital, China, between April 2021 and February 2024. The inclusion criteria were as follows: 1) patients admitted to the ICU for ≥24 hours; 2) patients who met the diagnostic criteria for LRTIs as defined in the Diagnostic Criteria for Hospital Infections (Trial), which included clinical symptoms such as fever, cough, abnormal sputum production, and dyspnea, as well as physical examination findings such as wet rales on lung auscultation and radiological evidence of lung abnormalities; 3) patients aged ≥18 years; 4) Lower respiratory tract samples of sufficient quality for mNGS analysis; and 5) patients with complete clinical treatment records available for review. The exclusion criteria were as follows: 1) patients had an ICU stay of less than 24 hours; 2) patients who lacked complete clinical data necessary for analysis, including microbiological testing results and treatment records; 3) patients aged <18 years. The study workflow was illustrated in Figure 1. Ethical approval was obtained from the Ethics Committee of Nanjing Drum Tower Hospital (No. 2023-061-01). Written informed consent was obtained from patients or their legally authorized representatives prior to sample collection.

A total of 261 samples were collected for concurrent conventional microbiological testing (CMT) and mNGS analysis, including 119 samples of alveolar lavage fluid (BALF), 105 samples of blood, 30 samples of cerebrospinal fluid, and 7 samples of other types. For BALF collection, patients received local anesthesia of the throat, followed by the introduction of a fiberoptic bronchoscope to irrigate the lesion site multiple times. The fluid was then adequately suctioned and collected in a sterile container. Blood samples were collected using cfDNA anticoagulant tubes under negative pressure. Pleural fluid samples were obtained by clinicians via percutaneous puncture or surgical approaches under strict aseptic conditions. The puncture site was localized by imaging or percussion, sterilized, and anesthetized, followed by extraction of the pleural fluid using a hollow-bore needle. After collection, all samples were sent to Dinfectome Inc. for mNGS testing, while conventional pathogen testing (including bacterial culture, viral serology, G test and GM test for fungi, and acid-fast staining for tuberculosis, etc.) was performed in the hospital laboratory.

Demographic data, laboratory test results, therapeutic interventions, and clinical outcomes were extracted from patient electronic medical records via the hospital’s integrated information system. Additionally, information regarding pre-admission antibiotic therapy, initial antibiotic regimens upon admission, and subsequent modifications based on mNGS results was also collected. The clinical impact of mNGS on management was adjudicated by medical record review and categorized as positive, negative, no effect, or indeterminate. Positive impact referred to clinically compatible and actionable mNGS findings that supported diagnosis and informed antimicrobial management, including escalation, de-escalation, replacement, enablement, or maintenance; negative impact referred to clinically incompatible findings that led to potentially inappropriate management changes, whereas no effect indicated unchanged management and indeterminate indicated that the contribution of mNGS could not be reliably assessed due to insufficient documentation or confounding factors.

Raw sequencing data were split by bcl2fastq2 (version 2.20), and high-quality sequencing data were generated using Trimmomatic (version 0.36) by removing low-quality reads, adapter contamination, duplicated and shot reads (length<36 bp). Human host sequences were subtracted by mapping to the human reference genome (hs37d5) using bowtie2 (version 2.2.6), and the proportion of human reads was calculated for each sample as a sequencing quality metric. Reads that could not be mapped to the human genome were retained and aligned to the microorganism genome database for microbial identification by Kraken (version 2.0.7), and for species abundance estimation by Bracken (version 2.5.0). The microorganism genome database contained genomes or scaffolds of bacteria, fungi, viruses and parasites (downloaded from GenBank release 238, ftp://ftp.ncbi.nlm.nih.gov/genomes/genbank/).

The mNGS pathogen detection pipeline was described in previous studies (Xu et al., 2023). The criteria for defining a positive detection were as follows: (1) At least one species-specific read was required for the detection of Mycobacterium, Nocardia, and Legionella pneumophila; (2) For other bacteria, fungi, viruses, and parasites, a minimum of three unique reads was required; (3) Pathogens were excluded if the ratio of microorganism reads per million of a given sample compared to the NTC was less than 10. Organisms meeting the above thresholds were subsequently interpreted for clinical relevance by integrating specimen type and clinical context. Briefly, detections were considered likely pathogens when they were concordant with the patient’s symptoms or supported by clinical microbiological test results, whereas organisms consistent with typical oral flora in the absence of supportive clinical evidence were interpreted as commensals.

Alpha diversity was estimated by the Shannon index based on the taxonomic profile of each sample, beta diversity was assessed by the Bray-Curtis measure, and compared between ICU patients by using Wilcoxon rank sum test, and was subsequently visualized by principal coordinate analysis (PCoA) plot. PERMANOVA was performed by the R package “vegan” to analyze Bray-Curtis distance in different survival and death groups. Differential relative abundance of taxonomic groups at the genus level among groups was tested by using Kruskal-Wallis rank sum test. To account for multiple comparisons across taxa, p-values from differential abundance testing were adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure, with FDR-adjusted p < 0.05 considered statistically significant. The species with mean relative abundances greater than 0.1% and penetrance greater than 10% among all samples were compared. Statistically significant differences in the relative microbial abundance among groups were assessed by the linear discriminant analysis of effect size (LEfSe) analysis.

A diagnostic prediction model for clinical outcomes in critically ill patients was developed by integrating the abundances of significantly differential microbial species with key clinical variables as predictive features. To mitigate the randomness inherent in conventional train-test splits given the limited sample size, a rigorous leave-one-out cross-validation (LOOCV) framework was adopted as the core strategy for model training and evaluation. In each LOOCV iteration, a single sample was held out as the independent test set, while all remaining samples constituted the training set, ensuring that every sample was predicted independently once. Using the H2O machine-learning platform, four distinct types of prediction models were constructed and evaluated in parallel: Gradient Boosting Machine (GBM), Generalized Linear Model with elastic-net regularization (GLM), Random Forest (RF), and Deep Learning neural network (DL). During the training of each model, an additional 10-fold cross-validation was implemented on the LOOCV training set to guide hyperparameter tuning and mitigate overfitting, thereby enhancing generalizability. Model performance was primarily assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC). The 95% confidence intervals for the AUC were calculated using DeLong’s method, and the optimal classification threshold was determined by maximizing the Youden index, from which sensitivity, specificity, and other diagnostic metrics were derived. Predicted probabilities, ROC curve coordinates, and all performance metrics for each model were systematically recorded for comprehensive comparison. The model that demonstrated the highest AUC alongside robust diagnostic performance under the LOOCV framework was selected as the final predictive tool.

Treatment outcomes were categorized as clinical improvement, deterioration, no change, death, or untraceable according to the Chinese guidelines for the diagnosis and treatment of adult community-acquired pneumonia and medical record review. Clinical improvement was defined as marked resolution of infection, evidenced by normalization of infection-related markers (e.g., C-reactive protein and white blood cell count), radiological improvement on chest CT (reduction or resolution of lesions), and substantial relief of respiratory symptoms (decreased cough frequency, sputum changing from purulent to thinner, resolution of chest tightness, and alleviation of chest pain). Deterioration was defined as infection-related clinical worsening, reflected by worsening symptoms, uncontrolled infection markers, or radiological progression. No change was defined as persistent symptoms with no meaningful improvement in infection markers or imaging findings, in the absence of clear progression. Death was defined as in-hospital mortality during the index admission. Untraceable referred to cases in which treatment response could not be reliably assessed due to transfer, discharge against medical advice, or poor adherence resulting in incomplete outcome documentation.

A 2×2 contingency table was constructed to calculate sensitivity, specificity, and accuracy. Comparative analyses were performed using McNemar’s Chi-squared test with the R statistical software package (version 3.6.2). Graphical illustrations were created using Adobe Illustrator (version 2020). A p-value of less than 0.05 was considered to indicate a statistically significant difference.

A total of 261 patients with suspected LRTIs were enrolled and provided consent for sample collection and clinical screening. The cohort comprised 146 males and 115 females, with a median age of 58 years. Baseline characteristics of the patients at the time of admission are summarized in Table 1. The most common comorbidities included hypertension (36.1%), diabetes (29.5%), and hepatopathy (28.7%). Of the total patients, 237 (90.8%) cases had underlying pulmonary disease. Additionally, 51 patients had a history of malignancy. For all patients, the average number of days of hospitalization was 39 days.

Using CMT as the reference standard, the sensitivity and specificity of mNGS across all samples were 80.1% and 35%, respectively, with an overall accuracy of 66.3% (Figure 2). Among different sample types, mNGS demonstrated the highest sensitivity in BALF at 95.7%, followed by blood at 69.8%. In contrast, specificity was highest in CSF at 70%, but only 7.4% in BALF. The highest accuracy was observed in BALF samples, reaching 75.6%.

In addition to sample type, we compared the performance of mNGS and CMT in detecting bacterial, fungal, RNA viral, and DNA viral species (Figure 3A). Compared to CMT, mNGS demonstrated superior detection rates for Gram-positive bacteria (G+, 99 vs. 17), Gram-negative bacteria (G-, 109 vs. 74) and DNA viruses (102 vs. 62). These differences were statistically significant (p < 0.001). For fungi, the positive detection rate of mNGS was only slightly higher than that of CMT (33.0% vs. 23.8%, p = 0.026). In contrast, CMT showed a slight advantage over mNGS in detecting RNA viruses, although this difference was not statistically significant (p = 0.305).

Regarding the consistency of species detection, 55.6% of the samples were positive by both mNGS and CMT. However, mNGS uniquely detected pathogens in 52 samples, while CMT identified pathogens in 36 samples that were not detected by mNGS. Additionally, 28 samples were negative for pathogens by both methods. Among the 145 samples positive by both techniques, only 8 samples showed identical pathogen detection by mNGS and CMT. Overlapping pathogen detection was observed in 79.3% of the samples, while 22 samples exhibited discordant pathogen detection between the two methods (Figure 3B).

We further analyzed the distribution of pathogenic microorganisms detected by mNGS and CMT. Two detection methods showed differences in the number of various microorganisms detected, particularly for G+ (173 vs. 17) and G- (195 vs. 89) (Figure 4A). Analysis of the proportion of different microorganisms showed that bacteria were the most frequently detected pathogens (47.88%), followed by DNA viruses (21.21%) (Figure 4B). For common bacterial pathogens such as Klebsiella pneumoniae, Pseudomonas aeruginosa, Streptococcus pneumoniae, and Stenotrophomonas maltophilia, significant differences in detection rates were observed between mNGS and CMT (p < 0.05). For fungal pathogens, the detection rates of Aspergillus fumigatus and Candida albicans were consistent between the two methods. However, Pneumocystis jirovecii was detected exclusively by mNGS, likely due to the limitations of CMT in identifying this organism. Among viral pathogens, the most frequently detected DNA and RNA viruses were Human gammaherpesvirus 4 and SFTS phlebovirus, respectively, with both methods showing largely consistent detection results. However, for Influenza A virus, CMT demonstrated a significantly higher detection rate compared to mNGS (p < 0.05) (Figure 4C).

The impact of mNGS results on subsequent clinical management and outcomes was evaluated in this study (Figure 5). Among the patients, mNGS results had a positive influence on treatment decisions in 187 cases, a negative impact in 33 cases, and no or indeterminate impact in 21 and 20 cases, respectively. In terms of diagnostic significance, mNGS facilitated the identification of pathogenic microorganisms in 158 cases and helped rule out infection in 51 patients. Among the 128 patients whose treatment regimens were modified based on mNGS results, 59.4% underwent escalated therapy, 25.8% had changes in therapeutic agents, and 1.6% initiated a new treatment regimen. Regarding clinical outcomes, 58.6% of patients who adjusted their treatment regimens based on mNGS results demonstrated clinical improvement, while 37.5% experienced disease progression, 1.6% showed no change, and 2.3% died. These findings suggest the potential of mNGS to guide clinical decision-making and improve patient outcomes, particularly in cases where traditional methods fail to provide timely or informative diagnostic information.

To investigate the relationship between microbial profiles and patient outcomes, patients were divided into survival (n = 198) and death (n = 63) groups. The pathogen spectrum was found to be richer in the survival group, particularly for bacteria and RNA viruses. Specifically, Huaiyangshan banyangvirus was detected exclusively in the survival group, along with coronaviruses, enteroviruses, and orthohantaviruses. In contrast, Human orthopneumovirus and Human metapneumovrus were detected only among non-survivors in this cohort. Among bacteria, Mycobacterium, Mycoplasma, and Chlamydia were more frequently detected in survivors, while Cryptococcus neoformans and Rhizopus delemar were detected only in the survival group (Supplementary Figure 1).

In addition to pathogenic bacteria, we analyzed the bacterial flora of BALF from 119 patients (77 in the survival group and 42 in the death group). A total of 262 species were detected across both groups, with Corynebacterium striatum as the dominant bacterium in the survival group and Pseudomonas aeruginosa as the dominant bacterium in the death group (Figures 6A, B). However, no significant differences were observed between the two groups in terms of α-diversity or β-diversity (Figures 6C–E). LEfSe analysis identified 20 discriminatory biomarkers between the two groups, with 12 species in the survival group (Streptococcus oralis, Pasteurellales, Pasteurellaceae, Haemophilus, and Olsenella uli, etc.) and 8 species in the death group (Stutzerimonas, Stutzerimonas stutzeri, Acinetobacter junii, Roseomonas mucosa and so on) (Figures 6F, G).

To identify the most reliable predictor of patient survival outcomes, we trained and compared four machine-learning classifiers built on differentially abundant species. Classification performance was evaluated with receiver operating characteristic (ROC) curve analyses (Figure 7; Supplementary Table 1). The random-forest (RF) model attained the highest area under the curve (AUC = 0.722), outperforming the generalize linear model (GLM, 0.686), gradient-boosting machine (GBM, 0.672) and deep learning (DP, 0.667) algorithms. These results indicated that RF model provides the strongest discriminative capacity for distinguishing survival from mortality, whereas the remaining models exhibit moderate but still appreciable predictive value.