A practical nomogram for predicting multidrug-resistant organism infection in a tertiary county-level hospital to guide antimicrobial therapy

Abstract

Background

The prevalence of multidrug-resistant organisms (MDROs) varies geographically, and uniform standards for its management are lacking.

Methods

This retrospective single-center study included 1045 unique patients after excluding duplicate isolates from 1304 non-repetitive strains collected between January and December 2023. Multivariable logistic regression was used to identify MDRO predictors and construct a nomogram. Model performance was assessed by ROC curve, calibration curve, and decision curve analysis, with bootstrap internal validation (1000 resamples). A sensitivity analysis excluding ESBL was also performed.

Results

Six independent risk factors were identified: age ≥60 years (OR = 1.47), urinary tract infection(UTIs) (OR = 1.79), ESBL production(OR = 15.18), Pseudomonas aeruginosa (OR = 3.66), S. aureus (OR = 29.23), and E. faecium (OR = 55.48). The bootstrap-corrected AUC was 0.82 (95%CI: 0.79–0.85), with a sensitivity of 85%, specificity of 70%, PPV of 47%, and NPV of 94%. Calibration was satisfactory (slope = 1.000, intercept = 0.000, Brier = 0.136, HL P = 0.926). After ESBL exclusion, the AUC decreased to 0.70, suggesting that other variables retained independent predictive value.

Conclusions

Age ≥60 years, urinary tract infection, ESBL production, P. aeruginosa, S. aureus, and E. faecium were independent risk factors for MDRO infection. The nomogram demonstrated good discrimination and calibration in this cohort and may serve as a reference for MDRO risk stratification in county-level hospitals, though external validation is warranted.

1. Introduction

The prevalence of multidrug-resistant organisms (MDROs) has steadily increased in recent years, resulting in a notable rise in the incidence of MDRO infections among patients [[1], [2], [3]]. This phenomenon can be attributed to a multitude of factors, including the frequent use of antibiotics, regular medical consultations, and exposure to health care settings. Furthermore, elderly patients who suffer from multiple chronic diseases and who are frequently hospitalized or visit day hospitals, who receive long-term immunosuppressive therapy, and who use invasive devices such as ventilators, urinary catheters, or intravascular catheters are at increased risk of infection with MDROs [[4], [5], [6], [7]]. In addition, recent reports have indicated that changes in metabolic pathways are a predominant strategy used by bacteria to achieve resistance, persistence, and drug resistance [[8], [9], [10]]. A previous study demonstrated that the inappropriate use of antibiotics has resulted in increased mortality rates in patients with multidrug-resistant microbial infections, particularly in cases of shock [[11], [12], [13], [14]]. Therefore, empirical selection of antibiotic treatment based on the source and severity of the infection, local epidemiological conditions, and risk factors for multidrug-resistant bacteria is also important [15].

The presence of MDRO infections has been demonstrated to have a dual effect on antibiotic resistance and health care expenditures. First, it has been shown to increase antibiotic resistance; second, it has been demonstrated to result in unnecessary medical costs. Therefore, early identification of MDRO infections is vital for clinicians. Nevertheless, accurately assessing the risk of MDRO infections remains challenging. Numerous risk factors associated with MDRO infections have been identified; however, these risk factors frequently lack specificity [4,6,7]. Consequently, continuous updates of local epidemiology and reporting of the most common risk factors for MDRO infections are crucial for strengthening appropriate empirical antibiotic treatment.

In recent years, numerous studies have been conducted to develop MDRO prediction models in different populations. For example, Wang [16] identified several risk factors for pulmonary infections in 494 elderly ICU patients, including length of stay, COPD, cerebrovascular disease, tracheotomy, and prior carbapenem exposure. The area under the curve (AUC) for this model was 0.748. Similarly, Zhang [17] found that pneumonia type, COPD, invasive procedures, ICU admission, combination therapy, carbapenem use, and hypoalbuminemia were risk factors for MDRO infection in severe pneumonia patients (training AUC = 0.910). In another study, Li [18] reported that age, duration of antibiotic use, and central venous catheterization were independent risk factors in 476 cancer hospital patients (AUC = 0.880). Likewise, Yang [19] compared five machine learning models in 791 neurocritical care patients with HAP, with logistic regression performing the best (AUC = 0.805); risk factors included length of NICU stay, number of antibiotics, diabetes, and urea level. Taken together, these studies provide important references for MDRO risk prediction. However, most have focused on ICU or specialty populations, with limited research targeting general ward patients in county-level general hospitals. Moreover, significant differences in MDRO infections exist across different countries and hospitals, further highlighting the need for hospital-specific prediction tools [4,6,7].

Therefore, this study aimed to develop a risk prediction model for MDRO infections based on patient-level data from our county-level hospital, to assist clinicians in risk stratification and early targeted antimicrobial strategy adjustment when preliminary microbiological identification and ESBL phenotype results become available.

2. Materials and methods

2.1. Patients and study design

A retrospective study was performed to investigate specimens from inpatients at Taihe County People's Hospital who were admitted between January and December 2023. The samples included sputum, blood, secretions, urine, pus, and other sample types, which were subsequently dispatched to our microbiology laboratory for analysis. A total of 1860 non-repetitive bacterial strains were detected. After excluding duplicate isolates from the same patient, 1045 unique patients were included in the primary analysis. The patient-level dataset comprised demographics (age, sex), clinical variables (ward of admission, infection site, length of hospital stay before culture, and antibiotic exposure within 90 days), and microbiological data (ESBL phenotype and species identification). This retrospective study was approved by the Ethics Committee of Taihe County People's Hospital (approval number: PJ2025-01-25) and was exempt from the requirement for informed consent, as the study involved only anonymized data extracted from existing medical records, posed no additional risk to participants, and was conducted in accordance with the principles of the Declaration of Helsinki.

2.2. Bacterial identification and drug sensitivity testing

Bacterial identification and antimicrobial susceptibility testing were performed via the VITEK2 compact fully automated bacterial identification system. The results were interpreted according to the drug breakpoints specified in the Clinical Laboratory Standards Institute (CLSI) M100-S31 document [20]. The quality control strains included Escherichia coli ATCC 25922, Klebsiella pneumoniae ATCC 700603, Pseudomonas aeruginosa ATCC 27853, Staphylococcus aureus ATCC 25923, and Enterococcus faecalis ATCC 29212.

2.3. Bacterial isolates and inclusion criteria

To exclude repetitive isolates from the same patient, the following criteria were used to define non - repetitive bacterial strains:(1) Time criterion: For the same patient and the same infection site, only the first isolate obtained within 7 days was included. If the same bacterial species with an identical antibiogram was isolated from the same site within 7 days, subsequent isolates were considered repetitive and excluded. (2) Site criterion: Isolates of the same species from different infection sites (e.g., blood, sputum, urine) in the same patient were considered independent infection events and were all included.(3) Species criterion: Isolates of different species from the same patient were all included.(4) Antibiogram change criterion: If the same bacterial species was isolated from the same site after an interval of more than 7 days but with a significant change in the antibiogram (e.g., from susceptible to resistant), it was considered a new isolate and included. These criteria were established in accordance with guidelines from the Clinical and Laboratory Standards Institute (CLSI) and the CDC/National Healthcare Safety Network (NHSN). All predictors in the model are available at the prediction time point (after positive culture, species identification, and ESBL phenotype are known). Species identification and ESBL phenotype are reported by the microbiology laboratory within 48–72 h of positive culture; clinical variables (age, diagnosis of urinary tract infection, length of hospital stay, prior antibiotic exposure, etc.) are extractable from the electronic medical record system at the time of specimen submission.

The six bacterial species selected for the primary analysis (E. coli, K. pneumoniae, P. aeruginosa, S. aureus, E. faecium, and E. faecalis) were the most prevalent in our hospital, accounting for 1304 of the 1860 non-repetitive isolates (70.1%), representing the predominant pathogens in our institution. The remaining 556 strains comprised a diverse array of rare species, each with very low individual frequencies. Including them would have led to unstable effect estimates due to sparse data and compromised model parsimony. This selection strategy may introduce selection bias, as the excluded rare species could have different resistance patterns and clinical implications (see Discussion 4.6).

2.4. Definition of events

Infection confirmation and classification: Only patients with confirmed infection were included in this study. Confirmed infection was defined as the isolation of pathogenic bacteria from a normally sterile site (e.g., blood, cerebrospinal fluid, pleural fluid) or a clinically significant non-sterile site (e.g., sputum, urine, wound discharge), accompanied by corresponding clinical manifestations of infection (e.g., fever, local inflammatory signs, imaging changes). Infection types were classified according to the CDC/National Healthcare Safety Network (NHSN) surveillance definitions.

MDRO infection definition: Multidrug-resistant organism (MDRO) infection is defined as an infection caused by microorganisms that are resistant to at least three different antibiotic families. All culture-positive infections that did not meet the MDRO definition (i.e., resistance to fewer than three antibiotic families) were classified as non-MDRO.

2.5. Statistical analyses

Statistical analyses were performed using SPSS version 26.0 and R version 4.2.1 (R Foundation, Vienna, Austria). WHONET 5.6 was used for susceptibility data management. Continuous variables were tested for normality using the Shapiro–Wilk test. Normally distributed variables were presented as mean ± standard deviation (SD) and compared using the independent-samples t-test; non-normally distributed variables were presented as median (interquartile range) [M (Q1, Q3)] and compared using the Mann–Whitney U test. Categorical variables were presented as counts (percentages) and compared using the χ² test or Fisher's exact test. Multivariable logistic regression was used to identify independent risk factors for MDRO infection. All variables had complete data; therefore, no imputation was performed. The optimal cut-off value (0.194) for risk stratification was determined by the maximum Youden index, i.e., maximizing (sensitivity + specificity − 1). Collinearity was assessed using the car package (version 3.1-2) in R to calculate the variance inflation factor (VIF) to detect potential multicollinearity among predictors in the multivariable logistic regression model. VIF values < 5 were considered indicative of no significant collinearity. The VIF values for all variables ranged from 1.034 to 2.479, all below the pre-specified threshold, suggesting no significant multicollinearity in the model (Supplementary Table S3). To address potential sparse-data bias, we performed a sensitivity analysis using Firth penalized logistic regression. The Firth method introduces a Jeffreys prior penalty into the likelihood function to reduce finite-sample bias in logistic regression [21]. The analysis was conducted using the logistf package (version 1.26.0) in R.The Firth penalized regression was performed using default settings of the logistf package, with no additional tuning parameters required. The penalized likelihood approach automatically addresses separation and sparse-data bias without the need for manual specification of penalty weights.

2.6. Nomogram development and validation

A nomogram was constructed based on the independent predictors identified in the multivariable logistic regression model. Each predictor was assigned a corresponding score according to its regression coefficient; the sum of individual scores yielded the total points, which projected to the predicted probability of MDRO infection. The nomogram was drawn using the rms package in R software.For model performance assessment, internal validation was performed using the bootstrap method with 1000 resamples. Discrimination was evaluated by the ROC curve and the area under the curve (AUC). Calibration was assessed by the calibration curve and the Hosmer-Lemeshow goodness-of-fit test. Clinical utility was examined by decision curve analysis (DCA), which quantifies net benefit across different threshold probabilities.

2.7. Sensitivity analysis

To assess the potential overlap between ESBL and the MDRO outcome definition, we performed a sensitivity analysis excluding the ESBL variable from the multivariable model, comparing AUC, calibration, and variable effects before and after exclusion. In addition, to address potential sparse-data bias, we performed a sensitivity analysis using Firth penalized logistic regression, as described in Section 2.5.

3. Results

3.1. Basic characteristics of MDROs and NMDROs

We retrospectively enrolled inpatients from Taihe County People's Hospital between January and December 2023 (Fig. 1). From 3151 specimens, 1860 non-duplicate strains were isolated. After excluding repeated isolates from the same patient, 1045 unique patients were included, with 253 MDRO cases (24.21%) and 792 controls (75.79%). Table 1 shows baseline comparisons. The MDRO group had more patients aged ≥60 years (64.03% vs 56.82%, P = 0.043), while sex distribution was similar (P = 0.495). For microbiological features, ESBL positivity was higher in the MDRO group (45.85% vs 20.96%, P < 0.001). The MDRO group had higher detection rates of S. aureus (29.25% vs 8.08%) and E. faecium (9.49% vs 1.01%) (both P < 0.001), but lower rates of K. pneumoniae (9.49% vs 30.93%) and P. aeruginosa (7.91% vs 15.66%) (P < 0.01). E. coli showed no difference between groups (P = 0.614).

Fig. 1.

Flowchart of patient selection. A total of 3151 specimens yielded 1860 non-repetitive strains. The six most prevalent species (1304 strains) were selected, with 556 other strains excluded. After excluding duplicate isolates, 1045 unique patients were included in the primary analysis, comprising 253 MDRO cases (24.21%) and 792 non-MDRO cases (75.79%).

Table 1.

Baseline characteristics of patients in the primary analysis (patient-level analysis).

Variables	Total (n = 1045)	NMDOR (n = 792)	MDOR (n = 253)	Statistic	P
Age, n(%)				χ2 = 4.11	0.043
<60	433 (41.44)	342 (43.18)	91 (35.97)
≥60	612 (58.56)	450 (56.82)	162 (64.03)
Sex, n(%)				χ2 = 0.47	0.495
Female	431 (41.24)	322 (40.66)	109 (43.08)
Male	614 (58.76)	470 (59.34)	144 (56.92)
ESBL, n(%)				χ2 = 60.29	<0.001
-	763 (73.01)	626 (79.04)	137 (54.15)
+	282 (26.99)	166 (20.96)	116 (45.85)
HepatobiliarySurgery, n(%)				χ2 = 1.69	0.194
No	928 (88.80)	709 (89.52)	219 (86.56)
Yes	117 (11.20)	83 (10.48)	34 (13.44)
AnalSurgery, n(%)				χ2 = 2.33	0.127
No	988 (94.55)	744 (93.94)	244 (96.44)
Yes	57 (5.45)	48 (6.06)	9 (3.56)
BurnSurgery, n(%)				χ2 = 9.86	0.002
No	987 (94.45)	758 (95.71)	229 (90.51)
Yes	58 (5.55)	34 (4.29)	24 (9.49)
Respiratory, n(%)				χ2 = 0.59	0.444
No	1008 (96.46)	762 (96.21)	246 (97.23)
Yes	37 (3.54)	30 (3.79)	7 (2.77)
Urology, n(%)				χ2 = 3.74	0.053
No	991 (94.83)	757 (95.58)	234 (92.49)
Yes	54 (5.17)	35 (4.42)	19 (7.51)
Rehabilitation, n(%)				χ2 = 1.40	0.237
No	1015 (97.13)	772 (97.47)	243 (96.05)
Yes	30 (2.87)	20 (2.53)	10 (3.95)
OBGYNdepartment, n(%)				χ2 = 1.22	0.269
No	999 (95.60)	754 (95.20)	245 (96.84)
Yes	46 (4.40)	38 (4.80)	8 (3.16)
ICU, n(%)				χ2 = 14.41	<0.001
No	772 (73.88)	562 (70.96)	210 (83.00)
Yes	273 (26.12)	230 (29.04)	43 (17.00)
Paediatriciand, n(%)				χ2 = 0.69	0.407
No	998 (95.50)	754 (95.20)	244 (96.44)
Yes	47 (4.50)	38 (4.80)	9 (3.56)
OtherInternal, n(%)				χ2 = 1.14	0.286
No	839 (80.29)	630 (79.55)	209 (82.61)
Yes	206 (19.71)	162 (20.45)	44 (17.39)
OtherSurgical, n(%)				χ2 = 10.84	<0.001
No	903 (86.41)	700 (88.38)	203 (80.24)
Yes	142 (13.59)	92 (11.62)	50 (19.76)
UTIs, n(%)				χ2 = 13.23	<0.001
No	932 (89.19)	722 (91.16)	210 (83.00)
Yes	113 (10.81)	70 (8.84)	43 (17.00)
BTIs, n(%)				χ2 = 1.50	0.22
No	972 (93.01)	741 (93.56)	231 (91.30)
Yes	73 (6.99)	51 (6.44)	22 (8.70)
BSIs, n(%)				χ2 = 1.39	0.239
No	845 (80.86)	634 (80.05)	211 (83.40)
Yes	200 (19.14)	158 (19.95)	42 (16.60)
RTIs, n(%)				χ2 = 20.70	<0.001
No	727 (69.57)	522 (65.91)	205 (81.03)
Yes	318 (30.43)	270 (34.09)	48 (18.97)
E coli, n(%)				χ2 = 0.25	0.614
No	609 (58.28)	465 (58.71)	144 (56.92)
Yes	436 (41.72)	327 (41.29)	109 (43.08)
K pneumonia, n(%)				χ2 = 46.15	<0.001
No	776 (74.26)	547 (69.07)	229 (90.51)
Yes	269 (25.74)	245 (30.93)	24 (9.49)
P aerugiosa, n(%)				χ2 = 9.70	0.002
No	901 (86.22)	668 (84.34)	233 (92.09)
Yes	144 (13.78)	124 (15.66)	20 (7.91)
S aureus, n(%)				χ2 = 74.96	<0.001
No	907 (86.79)	728 (91.92)	179 (70.75)
Yes	138 (13.21)	64 (8.08)	74 (29.25)
E faecium, n(%)				χ2 = 46.41	<0.001
No	1013 (96.94)	784 (98.99)	229 (90.51)
Yes	32 (3.06)	8 (1.01)	24 (9.49)
E faecalis, n(%)				χ2 = 3.96	0.046
No	1019 (97.51)	768 (96.97)	251 (99.21)
Yes	26 (2.49)	24 (3.03)	2 (0.79)

Notes: UTIs, urinary tract infections; BSIs, bloodstream infections; RTIs, respiratory tract infections; ICU, intensive care unit; ESBL, extended-spectrum β-lactamase.χ²: Chi-square test.

Regarding ward distribution, the MDRO group had fewer ICU admissions (17.00% vs 29.04%, P < 0.001), but more from other surgical departments (19.76% vs 11.62%, P < 0.001) and burn surgery (9.49% vs 4.29%, P = 0.002). For infection sites, UTIs were more common in the MDRO group (17.00% vs 8.84%, P < 0.001), while RTIs were less so (18.97% vs 34.09%, P < 0.001). No significant differences were found for other departments or bacterial species.

3.2. Predictors of infections caused by multidrug-resistant organisms

Multivariable logistic regression analysis identified six independent predictors of MDRO infection: age ≥60 years, urinary tract infection, ESBL production, P. aeruginosa, S. aureus, and E. faecium. Among these, E. faecium showed the strongest association (OR = 55.48, 95%CI: 21.66–142.15), followed by S. aureus (OR = 29.23, 95%CI: 16.41–52.07) and ESBL production (OR = 15.18, 95%CI: 9.04–25.51) (Table 2). A nomogram was constructed based on these six variables (Fig. 2). Internal validation via 1000 bootstrap resamples yielded a corrected AUC of 0.82 (95%CI: 0.79–0.85), indicating good discriminative ability (Fig. 3). The model was well calibrated, with a calibration slope of 1.000 and intercept of 0.000 after bootstrap correction (95%CI: 1.000–1.000 and 0.000–0.000, respectively), and the calibration curve closely aligned with the 45° reference line (Fig. 4). The Hosmer-Lemeshow test (χ² = 1.9305, df = 6, P = 0.926) and Brier score (0.136) further supported acceptable calibration and predictive accuracy. Decision curve analysis demonstrated a positive net benefit within the threshold range of approximately 20%–45% (Fig. 5). Key performance metrics are summarized in Table 3. These findings preliminarily suggest that the nomogram has satisfactory discriminative and calibration performance in this cohort; however, its clinical utility requires further evaluation through external validation in diverse populations.

Table 2.

Multivariable logistic regression results in the primary analysis.

Variables	β	S.E	Z	P	OR (95% CI)
Age ≥60 years	0.39	0.18	2.21	0.027	1.47 (1.04–2.08)
Urinary tract infection	0.58	0.26	2.21	0.027	1.79 (1.07–3.00)
ESBL production	2.72	0.26	10.27	<0.001	15.18 (9.04–25.51)
P. aeruginosa	1.3	0.34	3.83	<0.001	3.66 (1.89–7.12)
S. aureus	3.38	0.29	11.46	<0.001	29.23 (16.41–52.07)
E. faecium	4.02	0.48	8.37	<0.001	55.48 (21.66–142.15)

NOTES: OR, odds ratio; CI, confidence interval. β, regression coefficient; S.E., standard error; Z, Wald statistic. ESBL, extended-spectrum β-lactamase.

Fig. 2.

Nomogram for predicting MDRO infection risk. The nomogram incorporates six predictors, with each variable assigned a score; the total score projects to the predicted probability.

Fig. 3.

ROC curve of the nomogram for predicting MDRO infection. The bootstrap-corrected AUC was 0.82 (95% CI: 0.79–0.85), indicating good discriminative ability.

Fig. 4.

Calibration curve of the nomogram for predicting MDRO infection. The calibration curve closely aligns with the 45° ideal reference line, with a Hosmer-Lemeshow test P = 0.926, indicating good calibration.

Fig. 5.

Decision curve analysis (DCA) of the nomogram for predicting MDRO infection. Within the approximate threshold range of 20%–45%, the net benefit curve of the nomogram exceeded both the “treat all” and “treat none” strategies, suggesting positive clinical net benefit in this range.

Table 3.

Model performance metrics in the primary analysis(n = 1045 patients).

Metric	Value (Apparent)	Bootstrap-corrected (95% CI)	Ideal value
AUC (C-statistic)	0.80	0.82 (0.79–0.85)	>0.7
Accuracy	0.73 (0.71–0.76)	—	—
Sensitivity	0.85 (0.81–0.89)	—	—
Specificity	0.70 (0.66–0.73)	—	—
PPV	0.47 (0.43–0.52)	—	—
NPV	0.94 (0.92–0.96)	—	—
Cut-off	0.194	—	—
Calibration slope	1	1.000 (1.000–1.000)	1
Calibration intercept	0	0.000 (0.000–0.000)	0
Brier score	0.136	—	<0.25

NOTES: AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value; CI, confidence interval. Cut-off determined by the maximum Youden index. Apparent values were derived from the training set. Bootstrap-corrected estimates were calculated from 1000 resamples. “—” indicates not applicable. Brier score <0.25 indicates acceptable predictive accuracy. Calibration slope and intercept are shown as apparent values with bootstrap-corrected estimates and 95% CIs.

3.3. Sensitivity analysis

To assess the independent contribution of the ESBL variable to model performance and to address the potential risk of incorporation bias, we performed a sensitivity analysis excluding ESBL from the multivariable model. After ESBL removal, the bootstrap-corrected AUC decreased from 0.82 (95% CI: 0.79–0.85) in the full model to 0.70 (95% CI: 0.66–0.74). Notably, P. aeruginosa shifted from a significant risk factor in the full model (OR = 3.66, P < 0.001) to a non-significant factor in the ESBL-excluded model (OR = 0.74, P > 0.05), suggesting that its effect estimate was partially dependent on adjustment for ESBL. In contrast, S. aureus (OR = 6.01) and E. faecium (OR = 10.64) remained statistically significant (P < 0.001) after ESBL removal, indicating that their predictive effects were largely independent of ESBL. Model calibration remained satisfactory, with a Hosmer-Lemeshow test P-value of 0.977. These findings suggest that while ESBL is an important contributor to model discrimination, other variables—particularly S. aureus and E. faecium—retained independent predictive information, supporting the overall robustness of the model structure (Supplementary Table S2).

To address potential sparse-data bias due to rare events (e.g., E. faecium, n = 47), we further performed Firth penalized logistic regression as an additional sensitivity analysis. Results were consistent with the primary model: E. faecium remained the strongest predictor (penalized OR = 51.40, 95% CI: 21.24–135.52), and all variables retained statistical significance with effect estimates in the same direction as the primary analysis (Supplementary Table S4). These findings further support the robustness of our primary model.

4. Discussion

4.1. Summary of main findings

In this study, six independent risk factors for MDRO infection were identified: age ≥60 years, urinary tract infection (UTI), ESBL production, P. aeruginosa, S. aureus, and E. faecium. The predictive model incorporating these six predictors demonstrated good discriminatory ability, with a bootstrap-corrected AUC of 0.82 (95% CI: 0.79–0.85), a sensitivity of 85%, specificity of 70%, positive predictive value of 47%, and negative predictive value of 94%. Calibration was satisfactory, with a bootstrap-corrected calibration slope of 1.000 (95% CI: 1.000–1.000), calibration intercept of 0.000 (95% CI: 0.000–0.000), and Brier score of 0.136. The Hosmer-Lemeshow test (χ² = 1.9305, df = 6, P = 0.926) further supported acceptable model calibration. Decision curve analysis demonstrated that the nomogram provided a positive net benefit across threshold probabilities ranging from approximately 20% to 45%, suggesting potential clinical utility. Patient-level sensitivity analysis confirmed the robustness of these findings, and a sensitivity analysis excluding ESBL indicated that while ESBL contributed substantially to model discrimination, other variables retained independent predictive value.

4.2. Comparison with previous studies

4.2.1. Age as a risk factor for MDRO infection

In this study, age ≥60 years was identified as an independent risk factor for MDRO infection (OR = 1.47, 95% CI: 1.04–2.08, P = 0.027). This finding is consistent with previous studies. Theodorakis et al. reported that elderly patients are at increased risk due to immunosenescence, frailty, and multimorbidity [22]. Okoye et al. demonstrated that incorporating frailty and MDR status significantly improved mortality prediction in older sepsis patients [23]. Chen et al. also confirmed age >60 years as a significant risk factor in Taiwan [24]. It is worth noting that the significance of age differed between the isolate-level screening and the final patient-level analysis. In the isolate-level data, age did not reach statistical significance for entry into the multivariable model (P = 0.07). However, after removing duplicate isolates from the same patient and retaining only the first isolate per patient (1045 unique patients), age became significant (P = 0.027). This discrepancy may be attributed to repeated specimens from the same patient diluting the true effect of age. The patient-level analysis, which accounted for within-patient clustering by retaining only unique patients, likely provides a more accurate estimate of the age-related risk. These findings suggest that age ≥60 years is an independent risk factor for MDRO infection, reinforcing the importance of age consideration in risk stratification.

4.2.2. Urinary tract infection as a risk factor for MDRO infection

In this study, urinary tract infection (UTI) was identified as an independent risk factor for MDRO infection (OR = 1.79, 95%CI: 1.07–3.00, P = 0.027). This finding is consistent with previous reports that UTIs are among the most common infection sites associated with multidrug-resistant organisms [25,26], and that gram-negative bacteria—particularly E. coli, P. aeruginosa, and Enterococcus spp.—are the predominant uropathogens Gram-negative bacteria, particularly Escherichia coli and Pseudomonas aeruginosa, as well as Enterococcus spp., are the predominant uropathogens [27,28]. However, the empirical use of broad-spectrum antibiotics for UTIs without prior urine culture has contributed to the emergence of antimicrobial resistance worldwide [29]. Given that antibiotic susceptibility patterns vary geographically, egional surveillance of urinary pathogen resistance has been recommended [30,31]. In our cohort, UTIs were frequently identified in patients with MDRO infections, further supporting the clinical significance of this risk factor. These findings underscore the importance of urine culture-guided antibiotic therapy and local resistance surveillance in antimicrobial stewardship, particularly in patients with UTIs.

4.2.3. ESBL-producing bacteria

In this study, ESBL production was identified as a strong independent risk factor for MDRO infection (OR = 15.18, 95%CI: 9.04–25.51, P < 0.001). ESBLs are enzymes that inactivate most penicillins, cephalosporins, and amikacin, and ESBL-encoding genes are frequently associated with concurrent resistance mutations, leading to multidrug resistance [32]. While ESBL genes can be carried by various gram-negative organisms, they are most commonly found in Escherichia coli, Klebsiella pneumoniae, Klebsiella oxytoca, and Proteus mirabilis [[33], [34], [35]]. In our cohort, ESBL production was predominantly detected in E. coli and K. pneumoniae, and was most frequently isolated from blood, sputum, and urine specimens, with the highest prevalence in the hepatobiliary surgery, ICU, and urology departments. Invasive procedures in ICU settings, such as mechanical ventilation and urinary catheterization, may further facilitate the spread of ESBL-producing organisms, although the strong independent effect observed in our multivariable analysis (OR = 15.18) suggests that ESBL production itself is a major driver of MDRO risk, regardless of other factors. In addition, biofilm formation has been associated with ESBL-producing organisms and may further complicate clinical outcomes [36]. These findings highlight the clinical importance of rapid ESBL detection and targeted antibiotic therapy in patients with suspected ESBL-producing infections.

4.2.4. Pseudomonas aeruginosa

Pseudomonas aeruginosa is a Gram-negative, aerobic bacterium capable of causing various infections, including urinary tract and respiratory tract infections [37]. It is a primary pathogen responsible for hospital-acquired pneumonia, commonly affecting patients with structural pulmonary changes, impaired immune function, or other specific risk factors. Previous antibiotic use and prior hospitalization or ICU stay are the most significant risk factors for acquiring drug-resistant P. aeruginosa, highlighting the importance of antimicrobial stewardship and infection control measures [38]. In the present study, P. aeruginosa strains were predominantly isolated from sputum samples of ICU patients, consistent with previous findings [[39], [40], [41]], and were associated with high mortality rates [[42], [43], [44], [45], [46]]. Furthermore, a multicenter cohort study involving 7384 clinical isolates collected over 7 years identified ICU admission, immunosuppressive therapy, and carbapenem exposure as independent risk factors for MDRO infection [47]. Another retrospective study reported that elevated ICU occupancy rates, prolonged ICU stay, and prolonged use of invasive devices were independent predictors of healthcare-associated infections, including those caused by multidrug-resistant and extensively drug-resistant (XDR) pathogens [48]. Collectively, these findings support the role of P. aeruginosa as a significant risk factor for MDRO infection.

4.2.5. Staphylococcus aureus and Enterococcus faecium

In this study, S. aureus was predominantly isolated from secretions in the burn department, consistent with previous findings that staphylococcal infections are common in burn wound patients [49]. S. aureus was identified as one of the independent risk factors for MDRO infection in our cohort (OR = 29.23, 95%CI: 16.41–52.07), suggesting that patients with S. aureus isolates should be carefully monitored for the development of multidrug resistance. Similarly, E. faecium isolates in this study were predominantly recovered from urine and bile cultures from the hepatobiliary surgery and urology departments, consistent with previous reports [[50], [51], [52]]. E. faecium is a commensal organism of the human gastrointestinal tract but has been increasingly recognized as a major cause of multidrug-resistant infections, including urinary tract infections, hepatobiliary infections, catheter-related bloodstream infections, and surgical wound infections [[53], [54], [55], [56]]. The strong association observed in our study (OR = 55.48, 95%CI: 21.66–142.15) underscores the clinical significance of this pathogen in MDRO risk assessment.

4.3. Clinical implications of the nomogram

The nomogram developed in this study provides a quantitative tool for risk stratification of MDRO infection in this cohort. Clinicians can estimate individualized MDRO risk probability based on six readily available predictors: age ≥60 years, urinary tract infection, ESBL production, P. aeruginosa, S. aureus, and E. faecium. Internal validation demonstrated good discrimination (AUC = 0.82) and reliable calibration, and decision curve analysis suggested a positive net benefit within the approximate threshold range of 20%–45%.

It should be noted that the prediction time point of this model is defined as the stage when microbiological identification and ESBL phenotype are available—i.e., the phase of partial susceptibility results—rather than the empirical treatment stage before culture results are obtained. Therefore, the model is primarily intended to assist clinicians in risk assessment for early antimicrobial strategy adjustment when preliminary microbiological information is available. In clinical practice, the risk probability provided by this model should be interpreted in conjunction with the patient's overall clinical status, infection type, and local resistance epidemiology. Until multicenter external validation is obtained, this model is suggested as a supplementary reference tool for clinical decision-making, rather than a substitute for comprehensive clinical judgment.

4.4. Consideration of the overlap between ESBL production and MDRO

ESBL positivity was included as a predictor, while MDRO was defined as resistance to ≥3 antibiotic families, raising concern for incorporation bias. We performed a sensitivity analysis excluding ESBL. The AUC decreased from 0.82 (95%CI: 0.79–0.85) to 0.70 (95%CI: 0.66–0.74) after ESBL removal (drop of 0.12); both models maintained adequate calibration (HL P = 0.926 vs 0.977). Notably, P. aeruginosa was a significant risk factor in the full model (OR = 3.66, 95% CI: 1.89–7.12, P < 0.001) but became non-significant after ESBL removal (OR = 0.74, 95% CI: 0.43–1.21, P = 0.242), while other variables remained stable. It is worth noting that the loss of statistical significance for P. aeruginosa after ESBL removal may reflect a combination of both factors. On one hand, ESBL status may indeed mediate part of the association between P. aeruginosa and MDRO—as ESBL-producing P. aeruginosa strains are more likely to carry additional resistance determinants. On the other hand, the reduction in statistical power due to the removal of a strong predictor may also contribute to the non-significant P-value. However, the fact that S. aureus and E. faecium remained highly significant (P < 0.001) after ESBL exclusion suggests that the effect estimates for these species were not materially affected by variable removal, supporting the robustness of their independent predictive effects. This suggests ESBL not only improves overall discrimination but also facilitates species-level risk stratification.

ESBL-positivity ≠ MDRO—the former indicates a single resistance mechanism, while the latter indicates resistance to ≥3 classes. Many ESBL-positive isolates remain susceptible to aminoglycosides, fluoroquinolones, and carbapenems, thus not meeting MDRO criteria. The incremental value of the nomogram after ESBL is known lies in: (1) integrating multidimensional information into a unified risk probability; (2) providing continuous risk estimation rather than a binary result; and (3) filling the decision gap (approximately 24–48 h) when ESBL is known but other susceptibilities are pending. Our model is intended for use when ESBL phenotype is known, but full susceptibility is unavailable, and the AUC of 0.82 should be interpreted within this context. External validation in diverse populations is warranted before clinical implementation.

4.5. Patient-level sensitivity analysis

To assess the potential impact of repeated specimens from the same patient, we performed a patient-level analysis by retaining only the first isolate per patient. After deduplication, the 1304 isolates corresponded to 1045 unique patients. All six risk factors remained directionally and statistically consistent with the primary analysis, with similar effect sizes. Model performance and calibration were also stable (AUC = 0.82; Hosmer-Lemeshow P = 0.926), and decision curve analysis confirmed clinical utility, consistent with the primary findings.

Notably, age was not significant in the isolate-level data (P = 0.07) but became significant in the patient-level analysis (P = 0.027), suggesting that repeated specimens may have diluted the true effect of age. This finding supports the use of the patient-level analysis as the primary analytical approach, as it likely provides a more accurate estimate of the age-related risk. Overall, these results indicate that our conclusions are robust to the choice of analytical unit.

4.6. Limitations

This study has several limitations that should be acknowledged. First, as a retrospective single-center study, the findings may reflect regional and hospital-specific characteristics, limiting generalizability to other settings. External validation in multicenter cohorts is needed. Second, several important confounders (e.g., prior antibiotic exposure, invasive devices, comorbidities, and immunosuppression) could not be fully captured due to the retrospective data source. However, E-value analysis (E-value = 110 for E. faecium) suggested that unmeasured confounding is unlikely to fully explain the observed associations, supporting the robustness of our main findings. Third, only the six most common bacterial species were included, excluding over 500 rare strains. These rare species may have different resistance patterns and clinical implications; therefore, extrapolation of our findings to rare species should be made with caution. Fourth, sparse data for E. faecium (only 47 isolates, of which 38 were MDRO) may have led to unstable or inflated odds ratios. To address this, we performed Firth-penalized logistic regression, which yielded a penalized OR of 51.40 (95% CI: 21.24–135.52), consistent with the primary analysis. The patient-level analysis further supported robustness (OR = 55.48, 95% CI: 21.66–142.15). These counts reflect the isolate-level background data that informed the Firth-penalized regression sensitivity analysis; the primary analysis was based on the 1045-patient dataset.As discussed in Section 4.4, the loss of significance for P. aeruginosa after ESBL removal may reflect a combination of true mediation by ESBL status and reduced statistical power after variable removal.Fifth, the prediction model demonstrated good performance only in internal validation and has not yet been validated in independent external cohorts. Additionally, some variables (e.g., prior carbapenem exposure) relied on medical record completeness and may be subject to information bias. Consequently, the findings of this study should be considered exploratory. Future prospective, multicenter, large-sample studies incorporating a more comprehensive set of confounders are needed to further validate and optimize the prediction model.

5. Conclusions

This study developed and internally validated a nomogram incorporating six predictors for MDRO infection risk based on patient-level data from a county-level hospital: age ≥60 years, urinary tract infection, ESBL production, P. aeruginosa, S. aureus, and E. faecium. The model demonstrated good internal discriminative ability (AUC = 0.82, 95%CI: 0.79–0.85) and satisfactory calibration, with a high negative predictive value (94%), suggesting potential utility in ruling out MDRO infection. However, due to the single-center retrospective design, its external generalizability and clinical safety require confirmation through prospective multicenter studies. Until higher-level evidence is obtained, this model should be used only as a supplementary reference tool within comprehensive clinical assessment.

CRediT authorship contribution statement

Lianxiu Han: Writing – review & editing, Writing – original draft, Visualization, Data curation. Haoyu Zhao: Software, Resources. Jun Cheng: Methodology, Investigation. Yufeng Gao: Funding acquisition, Writing – review & editing, Project administration.

Informed consent

The requirement for informed consent was waived by the ethics committee due to the retrospective nature of the study.

Consent for publication

Not applicable.

Ethical approval

This study was approved by the Ethics Committee of Taihe County People's Hospital (approval number: PJ2025-01-25). All procedures were performed in accordance with the Declaration of Helsinki.

Data availability

The data supporting this study are available from the corresponding author upon reasonable request.

Funding

The study was supported by the National Natural Science Foundation of China (No. 82370608) and the Youth Project of the Fuyang Municipal Health Commission (FY2023--122).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

List of abbreviations:

MDROs

multidrug-resistant organisms

ESBLs

extended-spectrum beta-lactamases

UTIs

urinary tract infections

BTIs

Biliary Tract Infections

BSIs

biofilm-associated bloodstream infections

RTIs

Respiratory Tract Infections

ICU

intensive care unit

AUC

area under the curve

OR

odds ratio

ROC

receiver operating characteristic curve

Appendix A. Supplementary data

The following is the Supplementary data to this article: