Hospital acquired drug resistant pathogens infections in patients with viral respiratory tract infections: a retrospective study

Abstract

Background

Viral respiratory infections (VRTIs) caused by influenza (Flu) and COVID-19 pose significant global health challenges. Clinical outcomes are further exacerbated by infections with hospital acquired drug resistant pathogens (DRPs).

Methods

A retrospective analysis was conducted on the data of 1,051 hospitalized patients with VRTIs from 2018 to 2024 at Beijing Ditan Hospital. Firstly, 280 drug-resistant strains were isolated from 185 patients with hospital-acquired DRPs infections for extended antibiogram analysis. Secondly, Interpretable machine learning (ML) was employed to predict the risk factors for hospital acquired DRPs infections in patients with VRTIs. Using the optimal feature subset, seven ML prediction models were developed. Parameter tuning was performed via 10-fold cross-validation and grid search. Model performance was evaluated using area under the curve (AUC), sensitivity, specificity, precision, and F1 score. SHAP (SHapley Additive exPlanations) was used to interpret the optimal model.

Results

Pathogen distribution in 280 clinical samples revealed sputum (65.36%) as themainsource, followed byblood (15.36%), urine (11.43%), and lavage fluid (5.00%). In all clinical specimens, Pseudomonas aeruginosa, Staphylococcus hominis, Escherichia coli, and Acinetobacter baumannii predominated in sputum, blood, urine, and lavage fluid, respectively. In terms of overall detection counts, the most frequently isolated strains were P. aeruginosa, Klebsiella pneumoniae, and A. baumannii. The drug resistance rate of P. aeruginosa to third-generation cephalosporins (such as ceftriaxone and cefotaxime) exceeds 89%, but it has relatively higher sensitivity to ceftazidime (71.7%) and cefepime (69.6%). Its drug resistance rates to imipenem and meropenem reach 45.7%. Although amikacin shows 100% sensitivity, combination with β-lactam antibiotics is recommended to reduce mortality. K. pneumoniae shows resistance rates of 53.3% to imipenem and 46.7% to meropenem, with over 50% resistance to levofloxacin and ciprofloxacin. Effective agents include sulfamethoxazole (68.9% susceptible), tigecycline (64.4%), chloramphenicol (62.2%), and amikacin (62.2%). Tigecycline combined with aminoglycosides has synergistic effects and inhibits resistant strains. A. baumannii was highly resistant to nearly all tested antibiotics, showing only partial susceptibility to minocycline (59.5%) and trimethoprim-sulfamethoxazole (38.1%). Among the seven ML models, the neural network (NN) achieved the best predictive performance. The SHAP method revealed the top 15 predictive variables by importance ranking, including length of stay (LOS), cholinesterase (CHE), age, albumin (ALB), etc.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-025-11455-8.

Introduction

Respiratory tract infections (RTIs) remain one of the most prevalent causes of global mortality and morbidity, with viruses being the predominant pathogens identified in acute respiratory diseases [1]. In recent years, viral respiratory tract infections (VRTIs), particularly emerging strains of influenza (Flu) and COVID-19, have escalated into significant public health challenges due to their rapid global transmission [2, 3]. COVID-19 and Flu exhibit high infectivity rates, efficient transmission dynamics, and widespread prevalence patterns [4]. According to World Health Organization (WHO) estimates, seasonal epidemics caused by influenza A virus and influenza B virus lead to 3–5 million hospitalizations and 290,000–650,000 respiratory-related deaths annually worldwide [5, 6]. COVID-19, caused by severe acute respiratory syndrome coronavirus 2, continues to have substantial impacts on global health systems, with the WHO reporting 776,754,317 confirmed cases and 7,073,466 deaths as of October 27, 2024 (https://covid19.who.int). The recent surge in hospitalizations has exposed a critical clinical challenge: a concerning rise in drug-resistant pathogens (DRPs) infection rates among patients with VRTIs [7, 8]. Infections caused by DRPs not only complicate clinical management by delaying diagnosis and prolonging treatment duration but also amplify the risks of nosocomial transmission.

DRPs infections represent a frequent complication in viral pneumonia cases, particularly in critically ill patients, serving as a key contributor to clinical deterioration and mortality in VRTIs [9, 10]. This phenomenon may be attributed to multiple risk factors, including compromised host immune responses in severely ill patients and the widespread use of high-dose glucocorticoids in clinical management. In addition, current clinical practice and available studies suggest a high rate of antimicrobial drug use in patients with COVID-19 and the Flu [11, 12]. Inappropriate antimicrobial utilization, especially broad-spectrum antimicrobials, may give rise to adverse antimicrobial reactions and greatly increase antimicrobial resistance. Accordingly, a comprehensive analysis of the extended antibiogram profiles of drug-resistant bacterial isolates from VRTIs patients with hospital-acquired DRPs infections, along with precise identification of the relevant risk factors for these nosocomial DRPs infections, is crucial for optimizing treatment strategies and improving clinical outcomes.

Artificial intelligence (AI) aided disease prediction has undergone considerable development in recent years. Since machine learning (ML) algorithms are inherently powerful in capturing nonlinear relationships. Consequently, more researchers are advocating using new ML-based predictive models to support the appropriate treatment of patients [13, 14]. Furthermore, there is a lack of studies on risk factors for infections with DRPs in the current study, as well as an insufficient sample size, inadequate variables included, and a lack of analysis of pathogen resistance characteristics.

For one thing, this study conducted an extended antibiogram analysis of drug-resistant strains isolated from hospitalized patients with VRTIs complicated by DRPs infections, aiming to provide decision support for precise anti-infective therapy. In another dimension, interpretable ML was used to predict DRPs infections risk factors. Seven ML models were developed, with the optimal one explained via the SHAP (SHapley Additive exPlanations) framework.

Materials and methods

Research participants

This study retrospectively enrolled eligible cases from Beijing Ditan Hospital, Capital Medical University from 2018 to 2024. We initially screened 2,199 patients with VRTIs, of whom 1,151 met the inclusion criteria and were subsequently included in the final analysis (Fig. 1). Flu was diagnosed according to the Infectious Diseases Society of America (IDSA) Influenza Practice Guidelines [15]. The diagnostic criteria for COVID-19 are based on the IDSA Guidelines [16].

Fig. 1.

Flowchart of eligible patient screening and statistical analysis

Inclusion criteria [1] Patients with virological confirmation of FLU (PCR/antigen positive) or COVID-19 (PCR positive) according to IDSA guidelines; [2] Minimum hospitalization duration ≥ 48 h; [3] Patients aged ≥ 18 years with at least one microbial culture performed during hospitalization; [4] Missing data were less than 5% of the total medical records.

Exclusion criteria: [1] Pathogen cultures and resistance test results not available; [2] Corynebacterium diphtheriae, Propionibacterium, Bacillus, and Micrococcus were directly recognized as contaminating bacteria and were not counted in this study; [3] DRPs detected before admission or within the first 48 h of hospitalization.

Definitions

The study outcomes comprise two components: the primary outcome and the secondary outcome. The primary outcome is an extended antibiogram analysis of bacterial isolates from VRTIs patients with hospital-acquired DRPs infections. The secondary outcome involves the development of an interpretable ML predictive model for identifying risk factors associated with hospital-acquired DRPs infections. Clinical specimens, including sputum, urine, blood, bronchoalveolar lavage fluid (BAL fluid), pleural fluid, ascites, and central venous catheter tips (CVC tips), were collected under standardized protocols for microbiological processing. Bacterial identification and phenotypic confirmation of drug resistance were conducted using the BD Phoenix™ and VITEK^® 2 automated systems for antimicrobial susceptibility testing (AST). If the microorganisms have reduced or lost their sensitivity to one or more of the key drugs used in clinical treatment, then DRPs infections are considered to exist. DRPs infections are defined as the massive growth and reproduction of DRPs, whose toxins and metabolites cause damage to the organism, resulting in local or systemic symptoms of infection [17]. Patients with DRPs infections constituted the DRPs group, and patients without DRPs infections constituted the non-DRPs group. Following the acquisition of AST results, all bacterial isolates were categorically stratified as susceptible or resistant. In addition, drug resistance is categorized as multidrug-resistant (MDR), extensively drug-resistant (XDR), or pan-drug-resistant (PDR). MDR is defined as isolates not susceptible to at least 1 agent from ≥ 3 antimicrobial categories, XDR is defined as susceptible to only 1 or 2 categories of antimicrobials, and PDR is defined as non-susceptibility to all drugs in all antimicrobial categories [18].

Data collection

Demographic and clinical parameters were extracted by three board-certified infectious disease specialists (each with over 5 years’ experience) via EMR review. Using a standardized CRF, they performed dual-phase verification: initial extraction and independent validation of all parameters.

Demographic data includes age, sex, pulse, blood pressure, length of stay (LOS), history of diabetes, hypertension, ischemic heart disease, intracerebral hemorrhage, cerebral infarction, rheumatoid arthritis, chronic kidney disease, leukemia, structural lung disease, thyroidectomy, and others. Clinic data included the following: [1] Routine blood test: white blood cell (WBC) count, neutrophil (NE) count, lymphocyte (LY) percent, monocyte (MO) count, eosinophil (EO) count, basophil (BA) count, red blood cell (RBC) count and others; [2] Myocardial enzymes test: Creatine kinase (CK), C-reactive protein (CRP), D.dimer and others; [3] Serum biochemistry tests: serum potassium (K), serum sodium (Na), serum Urea (UR), serum uric acid (UA) and others; [4] coagulation function test: prothrombin time (PT), prothrombin activity (PTA) and others; [5] Liver function indexes: Prothrombin time (TP), alanine aminotransferase (ALT), cholinesterase (CHE), albumin (ALB) and others; [6] therapy-related factors: indwelling urinary catheter (IUC), mechanical ventilation (MV), indwelling central venous catheterization (ICVC) and others [7]. Pathogen-related test: culture results of pathogenic microorganism, results of antimicrobial susceptibility testing.

Data preprocessing and feature selection

To address the common challenge of missing data in clinical research, which may introduce bias if ignored, we employed the k-nearest neighbor (KNN) method via the knnImputation function from the R package DMwR to impute missing values. Prior to performing KNN imputation, we reported the number and ratio of each missing feature, as shown in Supplementary Table 1. The original dataset was randomly split into a training cohort and a validation cohort at a 7:3 ratio. To mitigate class imbalance in the training cohort, we applied the Synthetic Minority Oversampling Technique (SMOTE). Subsequently, feature dimensionality was reduced by systematically combining the Random Forest (RF) algorithm with Recursive Feature Elimination (RFE), eliminating redundant features and creating a compact feature subset to enhance model efficiency.

Statistical analysis and model development

All statistical analyses and calculations were performed using R software (version 4.4.2). Continuous variables were expressed as mean and standard deviation (mean ± SD) and compared between the two groups using an independent-sample t-test. For non-normal distributions, non-parametric tests were used. Categorical variables were reported as frequencies and percentages and were contrasted using a chi-squared test.

In this study, 7 ML algorithms, including extreme gradient boosting (XGBoost), support vector machine (SVM), neural network (NN), logistic regression (LR), gradient boosting machine (GBM), random forest (RF) and k-nearest neighbor (KNN) were adopted to construct the prediction model.

The variables selected were included in the model. Parameter tuning for each model was performed using 10-fold cross-validation with grid search. Receiver operating characteristic (ROC) curve, decision curve analysis (DCA), and calibration curve analysis were performed to assess the models’ performance. The predicted outcome probability was converted into a binary outcome using a threshold of 0.5. Accuracy, sensitivity, specificity, and other metrics were calculated and reported for both the training and test cohorts.

Machine learning explainable tool

SHAP analysis was employed for the interpretation of the optimal model. The SHAP values can show how much each predictor contributes, either positively or negatively, to the target variable. Besides, each observation in the data set could be interpreted by the particular set of SHAP values.

Result

Extended antibiogram of drug-resistant pathogens

Distribution characteristics of drug-resistant strains in clinical specimens

This study analyzed the distribution characteristics of bacterial pathogens in 280 clinical samples (Table 1). Gram-negative bacteria predominated significantly (75.73%), while Gram-positive bacteria accounted for only 24.27% of isolates. Specimen sources were stratified as follows: sputum (n = 183, 65.36%), followed by blood (n = 43, 15.36%), urine (n = 32, 11.43%), and BAL fluid (n = 14, 5.00%). Rare isolates were obtained from pleural effusion (n = 4, 1.43%), CVC tip (n = 2, 0.71%), and ascites (n = 2, 0.71%). The most commonly isolated microorganism from sputum was Pseudomonas aeruginosa (P. aeruginosa), from blood Staphylococcus hominis (S. hominis), from urine Escherichia coli (E. coli) and from BAL fluid Acinetobacter baumannii (A. baumannii).

Table 1.

Distribution of bacterial isolates across specimen types

Bacteria	Blood [N (%)]	Pleural fluid [N (%)]	Sputum [N (%)]	Urine [N (%)]	CVC tip [N (%)]	BALF [N (%)]	Ascites [N (%)]	Total [N (%)]
Gram-negative bacteria
Pseudomonas aeruginosa	0	0	43 (23.5)	2 (6.25)	0	1 (7.14)	0	46 (16.43)
Klebsiella pneumoniae	1 (2.33)	0	37 (20.22)	5 (15.63)	0	2 (14.29)	0	45 (16.07)
Acinetobacter baumannii	2 (4.65)	0	35 (19.13)	1 (3.13)	0	4 (28.57)	0	42 (15)
Escherichia coli	6 (13.95)	0	9 (4.92)	10 (31.25)	0	1 (7.14)	0	26 (9.29)
Stenotrophomonas maltophilia	2 (4.65)	0	21 (11.48)	0	0	2 (14.29)	0	25 (8.93)
Achromobacter xylosoxidans	0	0	1 (0.55)	0	0	0	0	1 (0.36)
ABC complex	0	0	2 (1.09)	0	0	0	0	2 (0.71)
Acinetobacter pittii	1 (2.33)	0	0	0	0	0	0	1 (0.36)
Burkholderia cepacia	0	0	4 (2.19)	0	0	0	0	4 (1.43)
Chryseobacterium gleum	0	0	1 (0.55)	0	0	0	0	1 (0.36)
Elizabethkingia anophelis	0	0	0	0	0	1 (7.14)	0	1 (0.36)
Enterobacter cloacae	0	0	5 (2.73)	2 (6.25)	0	0	0	7 (2.5)
Haemophilus influenzae	0	0	5 (2.73)	0	0	0	0	5 (1.79)
Morganella morganii	0	0	0	1 (3.13)	0	0	0	1 (0.36)
Proteus mirabilis	0	0	1 (0.55)	0	0	0	0	1 (0.36)
Providencia rettger	0	0	1 (0.55)	0	0	0	0	1 (0.36)
Pseudomonas putida	0	0	1 (0.55)	1 (3.13)	0	0	0	2 (0.71)
Serratia marcescens	0	0	1 (0.55)	0	0	0	0	1 (0.36)
Gram-positive bacteria
Staphylococcus aureus	1 (2.33)	0	6 (3.28)	0	0	2 (14.29)	0	9 (3.21)
Staphylococcus capitatus	2 (4.65)	0	0	0	0	0	0	2 (0.71)
Staphylococcus haemolyticus	5 (11.63)	0	0	0	0	0	0	5 (1.79)
Streptococcus pneumoniae	0	0	1 (0.55)	0	0	0	0	1 (0.36)
Enterococcus faecalis	2 (4.65)	0	1 (0.55)	1 (3.13)	0	0	1 (50)	5 (1.79)
Corynebacterium striatum	0	0	3 (1.64)	0	0	1 (7.14)	0	4 (1.43)
Enterococcus faecium	3 (6.98)	2 (50)	2 (1.09)	8 (25)	0	0	1 (50)	16 (5.71)
Staphylococcus epidermidis	8 (18.6)	1 (25)	2 (1.09)	1 (3.13)	1 (50)	0	0	13 (4.64)
Staphylococcus hominis	10 (23.26)	1 (25)	1 (0.55)	0	1 (50)	0	0	13 (4.64)
Total	43 (100)	4 (100)	183 (100)	32 (100)	2 (100)	14 (100)	2 (100)	280 (100)

Abbreviation: CVC tip central venous catheter tip, BAL fluid bronchoalveolar lavage fluid

Antimicrobial resistance profiles of major strains

Among 185 patients with DRPs infections, a total of 280 drug-resistant strains were isolated. The strain distribution analysis revealed the following pathogens ranked by detection frequency: P. aeruginosa (n = 46, 16.43%); Klebsiella pneumoniae (K. pneumoniae, n = 45, 16.07%); A. baumannii (n = 42, 15.00%); E. coli (n = 26, 9.29%); Stenotrophomonas maltophilia (S. maltophilia, n = 25, 8.93%) and Enterococcus faecium (E. faecium, n = 16, 5.71%).

Antibiotic resistance profiles are shown in Fig. 2. The data indicate that P. aeruginosa exhibits high resistance rates to β-lactam antibiotics such as Ceftriaxone (44/46, 95.7%), Ampicillin-Sulbactam (43/46, 93.5%), Cefotaxime (41/46, 89.1%), and Amoxicillin-Clavulanic Acid (44/46, 95.7%). It is noteworthy that among carbapenem antibiotics, except for Ertapenem (39/46, 84.8%), the resistance rates of Imipenem and Meropenem (each 21/46, 45.7%) have exceeded 45%, highlighting the need for continued clinical vigilance regarding carbapenemase prevalence trends. P. aeruginosa demonstrates high susceptibility to aminoglycosides such as Amikacin (46/46, 100%) and Tobramycin (40/46, 87%). Additionally, the pathogen retains relatively high susceptibility to certain β-lactam agents: Piperacillin-Tazobactam (30/46, 65.2%), Ceftazidime (33/46, 71.7%), and Cefepime (32/46, 69.6%). These drugs, especially Ceftazidime and Piperacillin-Tazobactam, remain clinically valuable as critical options for empirical treatment. Notably, fluoroquinolones like Ciprofloxacin (37/46, 80.4%) also exhibit favorable susceptibility.

Fig. 2.

Comparative analysis of antibiotic resistance and susceptibility profiles across clinically relevant bacterial pathogens. A. antibiotic resistance profiles of bacterial pathogens; (B). antibiotic susceptibility profiles of bacterial pathogens. cell values indicate the number of strains exhibiting resistance (A) or susceptibility (B). columns represent bacterial species, and rows list antibiotics. total isolates tested per species are shown in the “number of pathogens” row

A total of 45 K. pneumoniae isolates were analyzed, revealing alarming resistance rates to β-lactams, quinolones, and carbapenems. Among penicillins, ticarcillin (43/45, 95.6%) and ampicillin (40/45, 88.9%) exhibited the highest resistance. Resistance to cephalosporins was also prominent, with ceftriaxone, cefuroxime (each 29/45, 64.4%), and cefazolin (27/45, 60%) all exceeding 60%. For quinolones, resistance rates surpassed 50% for levofloxacin (26/45, 57.8%) and ciprofloxacin (25/45, 55.5%). Notably, carbapenems showed significantly elevated resistance rates: imipenem (24/45, 53.3%) and meropenem (21/45, 46.7%). Regarding susceptibility, K. pneumoniae displayed the highest sensitivity to sulfamethoxazole (31/45, 68.9%), followed by tigecycline (29/45, 64.4%), chloramphenicol, and amikacin (each 28/45, 62.2%). However, carbapenems such as meropenem (21/45, 46.7%) and imipenem (20/45, 44.4%) demonstrated sensitivity in only approximately half of the isolates.

A total of 42 A. baumannii isolates were analyzed, demonstrating extensive resistance to most antimicrobial agents. The bacterium exhibited extremely high resistance rates to β-lactams and β-lactam/β-lactamase inhibitor combinations, including amoxicillin-clavulanate (39/42, 92.9%), ampicillin (39/42, 92.9%), ampicillin-sulbactam (30/42, 71.4%), aztreonam (39/42, 92.9%), and amoxicillin (37/42, 88.1%). Carbapenem resistance was alarmingly high, with resistance rates of 76.2% (32/42) for imipenem, 71.4% (30/42) for meropenem, and 90.5% (38/42) for ertapenem. Among cephalosporins, resistance rates exceeded 70% for third- and fourth-generation agents: ceftriaxone (33/42, 78.6%), cefepime (32/42, 76.2%), and ceftazidime (30/42, 71.4%). Resistance to quinolones was also significant, with ciprofloxacin (32/42, 76.2%) and levofloxacin (27/42, 64.3%) showing high resistance rates. Other antibiotics, including chloramphenicol (38/42, 90.5%), trimethoprim (36/42, 85.7%), and fosfomycin (35/42, 83.3%), also demonstrated substantial resistance. In susceptibility testing, A. baumannii was most sensitive to minocycline (25/42, 59.5%). Trimethoprim-sulfamethoxazole (16/42, 38.1%) followed as the second most effective option.

This study analyzed 26 E. coli isolates, demonstrating a multidrug-resistant (MDR) trend. Among quinolones, the isolates exhibited high resistance rates to levofloxacin (14/26, 53.9%) and ciprofloxacin (12/26, 46.2%). Resistance to third-generation cephalosporins was also notable, with ceftriaxone (12/26, 46.2%) and cefuroxime (13/26, 50.0%) showing elevated resistance. Notably, aztreonam resistance reached 50.0% (13/26). For aminoglycosides, partial resistance was observed to tobramycin (10/26, 38.5%), while amikacin retained high susceptibility (20/26, 76.9%). Carbapenem susceptibility testing revealed preserved activity for both imipenem and meropenem (65.4% susceptibility rate for each, 17/26).

S. maltophilia exhibits broad-spectrum resistance to numerous clinically used antibiotics, including most β-lactams and carbapenems, resulting in extremely limited clinical therapeutic options. Antimicrobial susceptibility testing revealed high resistance rates to β-lactam antibiotics: ampicillin (25/25, 100%), amoxicillin (25/25, 100%), amoxicillin-clavulanate (25/25, 100%), cefotaxime (25/25, 100%), aztreonam (24/25, 96%), ampicillin-sulbactam (24/25, 96%), and piperacillin-tazobactam (23/25, 92%). Significant resistance was also observed to carbapenems: imipenem (24/25, 96%), meropenem (24/25, 96%), and ertapenem (23/25, 92%). Among aminoglycosides, resistance rates reached 100% for amikacin, gentamicin, and streptomycin. Notably, the highest susceptibility was observed to minocycline (25/25, 100%), followed by trimethoprim-sulfamethoxazole (23/25, 92%) and levofloxacin (16/25, 64%), while susceptibility to chloramphenicol was relatively low (11/25, 44%).

In the analysis of 16 E. faecium isolates, all strains exhibited complete resistance to clindamycin (16/16, 100%) and penicillin (16/16, 100%). Additionally, extremely high resistance rates were observed for aztreonam, trimethoprim-sulfamethoxazole, trimethoprim, fusidic acid, and polymyxin B-colistin (15/16, 93.8% for each). Notably, vancomycin (11/16, 68.8%), linezolid (13/16, 81.3%), and streptomycin (10/16, 62.5%) demonstrated higher susceptibility compared to other agents.

Hospital acquired dprs infections risk prediction via interpretable ML

Basic characteristics of patients

Of the 2199 patients diagnosed with VRTIs, 1151 patients were finally included in the statistical analysis after screening. The dataset was randomly divided into 2 parts: 70% (n = 806) of the data was used for model training, and 30% (n = 345) of the data was used for model validation. In the training cohort, 129 patients (16%) formed the DPRs group and 677 patients (84%) formed the non-DPRs group. In the validation cohort, 56 patients (16.2%) formed the DPRs group and 289 patients (83.8%) formed the non-DPRs group (Table 2).

Table 2.

Comparison of demographic and clinical characteristics between training cohort and validation cohort

Variables	The training cohort (n = 806)	The validation cohort (n = 345)	P
General information, median (IQR) or n (%)
Age (years)	51(30–71)	47 (29–66)	0.082
Length of stay	15 (7–28)	17(7–28)	0.840
Gender			0.870
Male	449 (55.7)	194(56.2)
Female	357 (44.3)	151(43.8)
Result			0.923
Yes	129 (16)	56 (16.2)
No	677 (84)	289 (83.8)
Treatment-related factors, n (%)
Mechanical ventilation			0.993
Yes	126 (15.6)	54 (15.7)
No	680 (84.4)	291 (84.3)
Indwelling urethral catheter			0.885
Yes	98 (12.2)	43(12.5)
No	708 (87.8)	302 (87.5)
Indwelling central venous catheterization			0.590
Yes	87(10.8)	41(11.9)
No	719 (89.2)	304 (89.2)
Abdominal puncture and drainage			0.489
Yes	8 (1)	2 (0.6)
No	798 (99)	343 (994)
Closed thoracic drainage			0.887
Yes	15 (1.9)	6 (1.7)
No	791 (98.1)	339 (98.3)
Renal replacement therapy			0.487
Yes	22 (2.7)	7 (2)
No	784 (97.3)	338 (98)
Extracorporeal membrane oxygenation			0.166
Yes	9 (1.1)	1 (0.3)
No	797 (98.9)	344 (99.7)
Immunosuppressive therapy within 1 months			0.173
Yes	56 (6.9)	32 (9.3)
No	750 (93.1)	313 (90.7)
History of underlying disease, n (%)
Diabetes			0.012
Yes	165 (20.5)	49 (14.2)
No	641 (79.5)	296 (85.8)
Hypertension			0.570
Yes	221 (27.4)	89 (25.8)
No	585 (72.6)	256 (74.2)
Chronic ischemic heart disease			0.297
Yes	90 (11.2)	46 (13.3)
No	716 (88.8)	299 (86.7)
Intracerebral haemorrhage			0.100
Yes	7 (0.9)	7 (2.0)
No	799 (99.1)	338 (98.8)
Lung cancer			0.456
Yes	4 (0.5)	3 (0.9)
No	802 (99.5)	342 (99.1)
Cerebral infarction			0.545
Yes	55 (6.8)	27 (7.8)
No	751 (93.2)	318 (92.2)
Rheumatoid arthritis			0.798
Yes	8 (1)	4 (1.2)
No	798 (99)	341 (98.8)
Thyroidectomy			0.063
Yes	8 (1)	0 (0)
No	798 (99)	345 (100)
Splenectomy			0.624
Yes	3(0.4)	2 (0.6)
No	803 (99.6)	343 (99.4)
Chronic Kidney Disease			0.282
Yes	45 (5.6)	14(4.1)
No	761 (94.4)	331 (95.9)
Leukemia			0.456
Yes	4 (0.5)	3 (0.9)
No	802 (99.5)	342 (99.1)
AIDS			0.908
Yes	10 (1.2)	4 (1.2)
No	796 (98.8)	341(98.8)
Structural lung disease			0.467
Yes	51(6.3)	18 (5.2)
No	755 (93.7)	327 (94.8)
Chronic liver disease			0.442
Yes	36 (4.5)	12 (3.5)
No	770 (95.5)	333 (96.5)
Laboratory variables, median (IQR)
White blood cell (10^9/L)	6.01 (4.44–8.12)	5.93(4.31–8.58)	0.750
Neutrophil (10^9/L)	3.97 (2.62–6.10)	3.89(2.57–6.37)	0.944
Lymphocyte (10^9/L)	1.16 (0.74–1.72)	1.31(0.77–1.84)	0.145
Lymphocyte (%)	20.5 (11.58–32.43)	23.4 (11.4–35)	0.121
Monocyte (10^9/L)	0.42 (0.27–0.6)	0.40(0.28–0.55)	0.117
Eosimophil (10^9/L)	0.02 (0.00-0.06)	0.02(0-0.06)	0.572
Basophil (10^9/L)	0.01 (0.00-0.02)	0.01(0-0.02)	0.379
Red blood cell (10^12/L)	4.43 (3.98–4.88)	4.52(4.07–4.88)	0.530
Hemoglobin (g/L)	132 (117–147)	132(119–146)	0.686
Hematocrit (%)	39.3 (35.18–43.1)	39.40(35.70–43.10)	0.562
Mean corpusular volume (fl.)	89.10 (86.00-92.43)	89.30(85.90–92.90)	0.626
Mean corpusular hemoglobin concerntration (g/L)	337(328–345)	336(327–345)	0.220
Red cell distribution width-coefficient of variation (%)	12.7 (12.3–13.5)	12.8(12.3–13.5)	0.408
Mean platelet volume (fl.)	10.3(9.7–11)	10.3(9.7–11)	0.396
Platelet distribution width (%)	11.85 (10.68–13.50)	11.9(10.8–13.6)	0.266
Platelet -larger cell ratio (%)	26.9 (22.3–32.4)	27.8(22.8–32.6)	0.363
Lactate dehydrogenase (U/L)	216.55(180.68–283.3)	223.53(178-303.9)	0.673
Creatine kinase (U/L)	80.6 (52.65–132.1)	78.6(53.2-135.3)	0.900
Creatine kinase isoenzymes (U/L)	15.7(12.4–20.3)	16(12.45–21.1)	0.672
α-Hydroxybutyric dehydrogenase (U/L)	177(148-225.05)	182(148-240.5)	0.211
C-reactive protein (mg/L)	13.20(2.33–50.23)	11.10(1.75–51.45)	0.385
Serum potassium (mmol·L)	3.75 (3.45–4.03)	3.67(3.38–3.99)	0.017
Serum sodium (mmol·L)	139(135.9-141.3)	139.4(136.5-141.45)	0.093
Serum chlorine (mmol·L)	102.7(99.7-105.1)	103.1(100.4-105.3)	0.077
Serum calcium (mmol·L)	2.21(2.08–2.31)	2.20(2.07–2.3)	0.523
Serum phosphorus (mmol.L)	1.06(0.88–1.23)	1.02(0.84–1.23)	0.066
Urea (mmol/L)	4.70(3.62–6.52)	4.67(3.61–6.14)	0.394
Creatinine(µmol/L)	65.75(51.88–82.43)	65.5(51.10-78.55)	0.253
Uric acid (µmol/L)	293(226-365.25)	275(217–352)	0.049
Blood glucose (mmol·L)	6.11(5.19–7.81)	5.90(5.07–7.94)	0.263
PaCO2 (mmHg)	23.9(21.9–25.9)	24(21.9-26.05)	0.770
Prothrombin time (s)	12.1(11.4–13.1)	12.1(11.39–12.9)	0.263
Activeated partial thromboplasting time (s)	31.2(28.9–34)	31.1(28.8-33.55)	0.297
Prothrombin time ratio (%)	1.11(1.05–1.2)	1.11(1.05–1.19)	0.356
International normalized ratio (%)	1.11(1.05–1.2)	1.11(1.05–1.19)	0.388
Thrombin time (s)	14.6(13.5–15.9)	14.6(13.5–16)	0.712
D-dimer (mg/L)	0.54(0.27–1.12)	0.41(0.25–0.97)	0.029
Alanine amiotransferase (U/L)	21.4(13.78–33.5)	19.6(13.7-31.45)	0.220
Aspartate aminotransferase (U/L)	24.35(18.28–34.7)	23.5(17.75–33.75)	0.364
Total bilirubin (µmol·L)	9.2(6.60–12.8)	9.1(6.55–12.4)	0.801
Direct bilirubi (µmol·L)	3.4(2.3–4.9)	3.3(2.2–12.4)	0.454
Total protein (g/L)	70(64.88–74.5)	69.9(64.85–74.1)	0.534
Albumin (g/L)	41.1(35.38–45.03)	41(36.55–44.55)	0.961
Globulin (g/L)	29.1(26.08–32.5)	28.7(25.8-31.71)	0.134
Cholinesterase (U/L)	6902.5 (5165.75-8236.04)	6975(5525–8109)	0.786

Quantitative data with non-normal distributions were presented as median (25%−75%)

Categorical data were expressed as n (%); Result

Whether there is a drug-resistant bacterial infection, with “Yes” indicating infection and “No” indicating non-infection

Features selection

The RFE method combined with RF was used to select the optimal sequence of features in our study. The performance of each feature subset was assessed through 5-fold cross-validation, and the 15 features that made the model perform best were finally selected. These characteristics include ALB, MV, LOS, CHE, D-dimer, Urea, Age, IUC, LY, Ca, CRP, TP, ICVC, NE, and RBC. These variables showed the strongest association with the outcome variable, ensuring model parsimony and mitigating overfitting concerns.

Model building and evaluation

Among the seven ML models evaluated for predicting infection risk of DRPs in VRTI patients, the NN demonstrated superior overall performance. The NN achieved an AUC of 0.938 (95% CI: 0.914–0.963) in the training cohort and 0.907 (95% CI: 0.854–0.960) in the validation cohort, maintaining robust diagnostic efficacy despite not being the highest among all models (Fig. 3A, B). Validation cohort evaluations revealed that the NN model significantly outperformed other models in specificity (0.882) and precision (0.585), confirming its excellent ability to discriminate negative samples and its high positive predictive value. Furthermore, the NN attained an F1-score of 0.701, indicating an optimal balance between minimizing false positives and false negatives. The model exhibited well-matched sensitivity (0.873), reflecting balanced classification performance. In contrast, logistic regression (LR), while achieving the highest sensitivity (0.945), showed relatively lower specificity (0.775) and precision (0.444), suggesting potential overfitting or elevated false-positive rates (Table 3). The calibration curves for both the training and validation sets exhibited some degree of deviation, but remained within acceptable limits (Fig. 3C, D). DCA further confirmed the clinical utility of the NN, as it provided superior net benefit across a wide range of threshold probabilities, supporting its practical applicability in clinical settings (Fig. 3E, F).

Fig. 3.

The receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA) for both the training set and validation set. (A) ROC curves of the models in the training set. (B) ROC curves of models in the validation set. (C) Calibration curves of the models in the training set. (D) Calibration curves of the models in the validation set. (E) DCA of the models in the training set. (F) Clinical Decision Curve of the models in the validation set

Table 3.

The performance of each model was evaluated in the validation cohort.

Models	AUC	Sensitivity	Specificity	Precision	F1 score
LR	0.910	0.945	0.775	0.444	0.605
SVM	0.912	0.855	0.862	0.54	0.662
GBM	0.914	0.927	0.803	0.472	0.626
RF	0.891	0.855	0.824	0.48	0.614
NN	0.907	0.873	0.882	0.585	0.701
XGBoost	0.903	0.836	0.879	0.568	0.676
KNN	0.893	0.8	0.865	0.53	0.638

Abbreviation: AUC, area under the curve; LR, logistic regression; SVM, support vector machine; GBM, gradient boosting machine; RF, random forest; NN, neural network; XGBboost, extreme gradient boosting; KNN, k-nearest neighbor

Explanation of the model with the SHAP method

The SHAP algorithm is employed to determine the importance of each predictive variable in the NN model’s predictions. The included features are ranked by their mean absolute SHAP values in descending order (Figure. 4 A). The ranking from highest to lowest is as follows: LOS, CHE, Age, ALB, NE, Urea, MV, CRP, IUC, D-dimer, ICVC, RBC, TP, LY, and Ca. Additionally, to examine the positive and negative associations of predictive factors with the target outcome, we applied SHAP values to identify risk factors for hospital-acquired DPRs infection in VITRs patients, as illustrated in Figure. 4B. To demonstrate how the NN model evaluates the contributions of individual patient features, we utilize SHAP force plots to interpret individual predictions for two patients (Figure. 4 C, D). Red indicates that the feature negatively contributes to the prediction (the arrow points to the left, decreasing the SHAP value), while yellow indicates that the feature has a positive impact on the prediction (the arrow points to the right, increasing the SHAP value). The length of the color bar indicates the strength of the contribution, and E[f(x)] indicates the SHAP reference value, which is the mean predicted by the model. f(x) represents the predicted value. For patients in the “true positive” group, a predicted value (f(x)) of 1 (> reference value) indicates the presence of DPRs infection (Figure. 4 C). For patients in the “true negative” group, a predicted value (f(x)) of 0.000000000000000333 (< reference value) indicates the absence of DPRs infection or colonization (Figure. 4D).

Fig. 4.

The weights of variable importance, Shapley Additive exPlanations (SHAP) values, and SHAP force plots. (A) The weights of variables importance; (B) The Shapley Additive explanation (SHAP) values; are (C), and (D) SHAP force plor for two cases; LOS, length of stay; ALB, albumin; CHE, cholinesterase; NE, neutrophil count; MV, mechanical ventilation; CRP, C-reactive protein; IUC, indwelling urinary catheter; ICVC, indwelling central venous catheterization; RBC, red blood cell; TP, total protein; LY, lymphocyte percentage; Ca, Serum calcium

Discussion

Principal findings

This study first comprehensively characterized the extended antibiogram of the major detected bacterial strains. Secondly, it systematically identified the key risk factors for hospital-acquired DRPs infections in patients with VRTIs using interpretable ML methods, thereby establishing a novel decision support tool and theoretical framework for the clinical prevention and control of DRPs infections.

The data from this study revealed that P. aeruginosa exhibited resistance rates exceeding 89% to third-generation cephalosporins such as ceftriaxone and cefotaxime, which aligns with prior research findings [19]. The resistance mechanisms are likely closely linked to extended-spectrum β-lactamases (ESBLs) or impaired drug permeability resulting from the loss of the porin protein [20]. Although P. aeruginosa is highly resistant to β-lactam antibiotics due to ESBLs, this study demonstrates that certain drugs within this class, such as ceftazidime (33/46, 71.7%) and cefepime (32/46, 69.6%), still exhibit relatively high susceptibility against P. aeruginosa. A study by Mehrnoush Afsharipour et al. on hospital-acquired infections demonstrated that the susceptibility of P. aeruginosa to ceftazidime exceeded 80%, which aligns with the findings of our research [21]. It is noteworthy that over the past two decades, the resistance of P. aeruginosa to carbapenems has significantly increased. A six-month study conducted across four hospitals in Benghazi, Libya, revealed that the resistance rate of isolated P. aeruginosa strains to imipenem reached 19.8% [22]. In this study, P. aeruginosa exhibited high resistance to imipenem (21/46, 45.7%), highlighting the need for vigilance against the prevalence of carbapenemases. This study analysis indicates that although amikacin demonstrated 100% susceptibility, research by Minyon L. Avent et al. suggests that amikacin is not recommended as monotherapy. However, when combined with β-lactam antibiotics, it significantly reduces patient mortality, highlighting the importance of combination therapy [23].

The analysis of 45 K. pneumoniae isolates in this study highlights the critical severity of multidrug-resistant phenotypes. The findings indicate that traditional penicillins and most cephalosporins have lost clinical efficacy, which is strongly associated with the pathogen’s intrinsic resistance mechanisms [24]. Notably, an epidemiological survey by Yan Li et al. involving 26 hospitals in China revealed that the resistance rate of K. pneumoniae to imipenem and meropenem reached 20.1% during 2019–2020 [25]. This study analysis reveals that the resistance rates to imipenem (53.3%) and meropenem (46.7%) are approaching or exceeding 50%, signaling a critical escalation in the spread of carbapenem-resistant Klebsiella pneumoniae (CRKP) over recent years. This trend underscores the erosion of carbapenems’ role as “last-resort” antibiotics in clinical practice. Among quinolones, resistance rates to levofloxacin and ciprofloxacin exceeded 50%. Quinolones are commonly used for empirical therapy of community-acquired infections, and their high resistance rates may lead to outpatient treatment failure and drive the dissemination of resistance genes both within and beyond hospitals. Our data suggest that sulfamethoxazole (68.9% susceptibility rate), tigecycline (64.4%), chloramphenicol (62.2%), and amikacin (62.2%) remain the primary effective agents against this pathogen. Previous studies have demonstrated that the combination of tigecycline with aminoglycosides (e.g., amikacin) exhibits significant synergistic effects against K. pneumonia, while this regimen also effectively suppresses the selection of tigecycline-resistant strains [26].

Our data reveal that A. baumannii isolates exhibit substantial resistance to nearly all tested antibiotics, with only partial susceptibility to minocycline (59.5% susceptibility) and trimethoprim-sulfamethoxazole (38.1% susceptibility). The reasons are likely attributable to the species’ exceptional capacity to acquire resistance genes and its ability to persist and proliferate in hospital environments. Previous research analyses have shown that colistin is the most effective drug for treating A. baumannii. A five-year antimicrobial susceptibility study conducted by Biagio Santella and colleagues demonstrated that A. baumannii exhibited susceptibility rates nearing 100% to colistin. The study also emphasized the need to rely on combination therapies such as polymyxins or novel antimicrobial agents in clinical practice [27]. In the present study, colistin was not routinely included as a test antibiotic in susceptibility assays against this bacterial strain, precluding a reliable assessment of its antimicrobial susceptibility profile.

E. coli exhibits high resistance rates to fluoroquinolones, such as levofloxacin, with 53.9% resistance, and third-generation cephalosporins, exemplified by ceftriaxone with 46.2% resistance. Compared with previous studies, E. coli has demonstrated a rising trend in resistance to quinolones and third-generation cephalosporins [28]. This necessitates strict limitations on fluoroquinolone use as empirical therapy, prioritizing their application in confirmed susceptible infections to delay resistance evolution. Although carbapenems, including imipenem with 65.4% susceptibility, retain partial effectiveness, their susceptibility rates have declined compared to historical data [29]. Amikacin demonstrates 76.9% susceptibility and serves as a viable agent for combination therapy in severe infections.

Our findings demonstrate that S. maltophilia exhibits markedly high resistance rates to most β-lactam and carbapenem antimicrobials, resulting in significantly elevated infections or colonization risk among critically ill patients undergoing broad-spectrum antibiotic therapy. In this study, minocycline demonstrated a 100% susceptibility rate. Trimethoprim-sulfamethoxazole exhibited a notably high susceptibility rate of 92% and could serve as a viable alternative combination agent. Notably, levofloxacin demonstrated a susceptibility rate of 64% in the current study, which was notably lower than the susceptibility profiles observed in other antimicrobial agents. For comparison, a study by Hsiu Wu et al. analyzing 377 S. maltophilia isolates collected between 1998 and 2008 reported susceptibility rates of 24.4% for levofloxacin, 82.5% for trimethoprim-sulfamethoxazole, and 82.5% for minocycline [30]. Notably, a marked discrepancy in levofloxacin susceptibility was observed between our findings (64%) and those reported by Hsiu Wu et al. (24.4%), This divergence may reflect temporal shifts in antimicrobial resistance patterns or geographic variability in strain prevalence.

The NN model demonstrated superior performance among the seven evaluated ML algorithms. SHAP analysis further enhanced the model’s interpretability, identifying LOS, CHE, Age, ALB, NE count, Urea, MV, CRP, IUC, D-dimer, ICVC, RBC, TP, LY, and Ca as the most influential predictors of hospital-acquired DRPs infections.

The use of invasive catheterization procedures, including MV, ICVC, and IUC, has been well-established as an independent risk factor for hospital-acquired DRPs infection, as such requiring no extensive elaboration [31, 32]. In this study, the LOS was confirmed as the most critical predictive variable. Prolonged LOS may lead to extended courses of broad-spectrum antibiotic therapy, which suppresses the normal microbiota in the gastrointestinal and respiratory tracts, thereby triggering selective proliferation of DRPs [33, 34]. Furthermore, prolonged hospitalization increases exposure frequency to medical devices, creating transmission pathways where DRPs can spread through lapses in hand hygiene compliance or via contaminated environmental surfaces [35, 36]. It is noteworthy that CHE, as a novel predictive indicator, its reduction showed a significant association with DRPs infection or colonization, which might be linked to the cholinergic anti-inflammatory pathway. CHE participates in the cholinergic anti-inflammatory pathway by breaking down acetylcholine (ACh). A decrease in CHE may lead to ACh accumulation, which could excessively activate the α7 nicotinic acetylcholine receptor, potentially triggering immune paralysis and thereby increasing the risk of infection by DRPs [37, 38]. As anticipated, advanced age was closely associated with infection by DRPs. Multiple factors may contribute to the increased risk of infection in elderly populations, including immune system alterations, malnutrition, and socioeconomic determinants. Decreased albumin levels may heighten the risk of DRPs infection via dual pathways: by inducing tissue edema that weakens physiological barrier defenses through hypoalbuminemia, and by elevating free concentrations of protein-bound β-lactam antibiotics such as ceftriaxone due to altered drug binding, which could promote resistance development [39, 40]. Elevated urea levels demonstrated a strong predictive capacity for clinical outcomes, likely mediated through pathogen metabolism of urea into ammonia and carbon dioxide, which increases local pH to create an alkaline microenvironment conducive to the proliferation of specific drug-resistant bacterial populations. CRP levels rise in response to infectious and inflammatory diseases and demonstrate a more pronounced elevation in severe bacterial infections [41]. We first identified elevated D-dimer levels as significantly associated with infection DRPs. Prior retrospective studies have demonstrated that severe infections such as pneumonia and bloodstream infections are frequently complicated by sepsis, with elevated D-dimer levels strongly correlating with both sepsis severity and poor prognosis. Notably, infections caused by DRPs such as carbapenem-resistant Enterobacteriaceae and methicillin-resistant Staphylococcus aureus are more prone to trigger severe sepsis, which may further activate the coagulation cascade and consequently drive D-dimer elevation [42]. In our cohort, we also confirmed that sepsis is a relatively common complication among patients with VRTIs who are co-infected or colonized with DRPs.

Limitations

This study has limitations. First, our NN model was trained and validated using data from a single center, and the regional characteristics of the patient population, clinical management strategies, and microbial epidemiology may limit generalizability. Second, while we maximized variable collection, potential confounders such as pre-admission antibiotic exposure, specific treatment durations, and environmental factors were not fully incorporated, which might indirectly influence colonization or infection risks. Future studies should integrate multicenter prospective cohorts, comprehensive variable collection, and external validation to optimize models, combined with molecular approaches to elucidate resistance mechanisms and advance clinical research.

Conclusion

Hospital-acquired DRPs infections have emerged as a critical global public health challenge requiring urgent resolution. It is imperative to expedite the identification of high-risk population characteristics and critical transmission pathways of pathogens, while concurrently establishing microbiology-guided precision antimicrobial stewardship strategies.

Abbreviations

RTIs

Respiratory Tract Infections

VRTIs

Viral Respiratory Tract Infections

Flu

Influenza

DRPs

Drug-Resistant Pathogens

AI

Artificial Intelligence

ML

Machine Learning

NN

Neural Network

SHAP

Shapley Additive exPlanations

IDSA

Infectious Diseases Society of America

BAL fluid

Bronchoalveolar lavage fluid

CVC tips

Central Venous Catheter Tips

AST

Antimicrobial Susceptibility Testing

MDR

Multidrug-Resistant

XDR

Extensively Drug - Resistant

PDR

Pan-Drug-Resistant

LOS

Length of Stay

WBC

White Blood Cell

NE

Neutrophil

LY

Lymphocyte

MO

Monocyte

EO

Eosinophil

BA

Basophil

RBC

Red Blood Cell

CK

Creatine Kinase

CRP

C-Reactive Protein

K

Serum Potassium

Na

Sodium

UR

Urea

UA

Uric Acid

PT

Prothrombin Time

PTA

Prothrombin Activity

ALT

Alanine Aminotransferase

CHE

Cholinesterase

ALB

Albumin

IUC

Indwelling Urinary Catheter

MV

Mechanical Ventilation

ICVC

Indwelling Central Venous Catheterization

ROC

Receiver Operating Characteristic

DCA

Decision Curve Analysis

ESBLs

Extended-Spectrum β - Lactamases

CRKP

Carbapenem-Resistant Klebsiella Pneumoniae

Authors’ contributions

Zibo Fan: conceptualization of the study, analysis, and interpretation of data, writing, and editing of the manuscript. Xinmin Xu: data provision and manuscript revision. Qun Li: conceptualization of the study, analysis, and interpretation of data, writing and editing of the manuscript. Tong Zhou: data entry and collation. Aibin Wang: data entry and collation. Chengjie Ma: data entry and collation. Zhihai Chen: review and editing. Lianhe Lu: review and editing. Yuanyuan Zhang: Participated in data analysis. Yajie Wang: verification of data accuracy. Wei Zhang: conceived and designed the protocol. All authors read and approved the final version of the manuscript.

Funding

National Key R&D Program of China (2023YFC2308800); The Training Programme for Beijing High-level Public Health Technical Talent Construction Project (Backbone of a discipline-02-31).

Data availability

The data supporting the views of this analysis are available from the corresponding author on request.

Declarations

Ethics approval and consent to participate

The research was conducted in accordance with the Declaration of Helsinki. Ethics approval was obtained from the Institutional Review Board of the Beijing Ditan Hospital, Capital Medical University (No DTEC-KY2025-049-02). As this retrospective analysis utilized routinely collected data, we submitted an application to our institutional review board for a waiver of informed consent and it was granted by the Institutional Review Board of the Beijing Ditan Hospital, Capital Medical University.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.