Nomogram for predicting prognostic risk in severe pulmonary tuberculosis: a retrospective analysis from the MIMIC-IV database

Introduction

Tuberculosis (TB) is a chronic infectious disease caused by Mycobacterium tuberculosis infection. It is primarily transmitted through the respiratory tract and features a prolonged incubation period, subtle initial symptoms, and high transmissibility (1). In 2024, TB stands as the leading cause of death worldwide from a single infectious agent, with a mortality toll nearly twice that of acquired immunodeficiency syndrome (AIDS) (2). Concurrently, the alarming projection of approximately 10.7 million new TB cases and 1.23 million related deaths in 2024 underscores the urgent need for comprehensive strategies to address its impact (2). Recognizing the multifaceted challenges posed by TB, it is crucial to highlight the critical role of intensive care units (ICU) in managing severe TB.

Patients afflicted with TB may require admission to the ICU for various reasons, including respiratory failure, miliary TB, multiorgan failure, and diminished consciousness associated with central nervous system disorders (3). Among these, respiratory failure and miliary TB emerge as particularly common catalysts for ICU admission, underscoring the pivotal role of these specialized units in addressing the complex clinical presentations of severe TB cases.

Pulmonary tuberculosis (PTB) constitutes >80% of all TB cases, making it the predominant form of the disease. However, few studies have explored the mortality risk factors among critically ill PTB patients in the ICU, and there are currently no dedicated models designed to predict inpatient ICU mortality in individuals affected by PTB. Hence, the objective of this study was to delineate the risk factors associated with mortality within 365 days after ICU admission in patients with PTB, with the aim of constructing a nomogram model for predicting patient mortality based on these identified risk factors. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2603/rc).

Methods

Data source

This retrospective analysis utilized data from the Medical Information Mart for Intensive Care IV (MIMIC-IV v2.2) database, compiling clinical information on patients admitted to Beth Israel Deaconess Medical Center (BIDMC) between 2008 and 2019 (4). The database encompasses information from 299,712 hospitalized patients, among whom 73,181 were in the ICU. It includes comprehensive data such as vital signs, laboratory test results, comorbidities, nursing records, and medication administration details (5). The researcher obtained access to the database for data extraction upon successful completion of the requisite course and examination. The MIMIC-IV database ensures the anonymization of patients’ basic personal information, adheres to rigorous medical ethics standards, does not pose ethically relevant concerns, and has received prior approval from the Institutional Review Board (6). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Patients and variables

We employed Structured Query Language (SQL) programming within Navicat Premium (version 16.1.12) for data extraction. The International Classification of Diseases (ICD)-9/ICD-10 codes were used to retrieve data on ICU patients diagnosed with PTB from the MIMIC-IV database. The exclusion criteria were as follows: (I) patients under 18 years of age; (II) patients with ICU stays of less than 24 hours; and (III) patients with more than 20% missing data. In cases of multiple hospitalizations for the same patient, only information from the initial admission was considered to prevent duplication of data and ensure accuracy in the analysis.

After identifying eligible subjects, we utilized their subject_id, hadm_id, and icustay_id parameters to extract comprehensive information from the relevant tables. This included demographic details such as age, gender, marital status, race, comorbidities, vital signs, laboratory parameters, renal replacement therapy (RRT) use, mechanical ventilation use, severity scoring system, and survival information. Comorbidities included AIDS, diabetes, and gastroesophageal reflux disease (GERD). Post-admission vital signs were averaged over the initial 24 h of ICU stay, including heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), mean blood pressure (MBP), respiratory rate (RR), and body temperature. Laboratory parameters were analyzed based on the first obtained post-ICU admission, including pH, red blood cells (RBC), white blood cells (WBC), platelets (PLT), lymphocytes (LYM), hematocrit (HCT), lactate (LAC), creatinine (Cr), blood urea nitrogen (BUN), albumin (ALB), hemoglobin (Hb), glucose (GLU), sodium, potassium, calcium chloride, partial pressure of oxygen (PO₂), partial pressure of carbon dioxide (PCO₂), and transcutaneous oxygen saturation (SpO₂). Severity scoring systems, such as Sequential Organ Failure Assessment (SOFA) and Acute Physiology Score III (APS III), were incorporated, and survival information, including length of stay (LOS) in the ICU and hospital as well as mortality outcomes, were included in the analysis.

Marital status was classified into three categories: single, married, and other (divorced and widowed). Ethnicity was stratified into the White, Black, and other categories. Additionally, the mode of ventilation was categorized as none, non-invasive ventilation, or invasive ventilation.

The primary outcome was time-to-event all-cause mortality, and the prediction model was used to estimate survival probabilities at 28, 180, and 365 days after ICU admission.

Statistical analysis

Variables with missing values exceeding 20% were excluded, and multiple imputations were used to estimate the remaining missing data. We randomly allocated 70% of the patients in the MIMIC-IV database to the training cohort and 30% to the validation cohort. The training cohort was utilized for constructing the nomogram, while the validation cohort was employed for validation purposes (7). Categorical variables were characterized using frequency and percentage, and group differences were determined using the chi-square test or Fisher’s exact test. Continuous variables were first assessed for normality using the Shapiro-Wilk test. For variables that followed a normal distribution, the mean and standard deviation values were reported, and Student’s t-test was used to evaluate group differences. For non-normally distributed continuous variables, the median and interquartile range (IQR) were reported, and the Mann-Whitney U test was applied to assess differences between groups.

Least absolute shrinkage and selection operator (LASSO) Cox regression serves as a technique for variable selection and shrinkage within the Cox proportional hazards model (8,9). This analysis method constructs a penalty function, enhances the model refinement, and contributes to a more precise selection of variables (10). The LASSO regression analysis conducted in R software involved 10 iterations of K-fold cross-validation for the centralization and normalization of the included variables. Subsequently, the optimal λ value was selected. “Lambda.1se” signifies a model that demonstrates exceptional performance while utilizing a minimal number of independent variables, effectively steering clear of overfitting concerns (11). The variables identified through LASSO were subsequently subjected to further validation via Cox regression to ascertain statistically significant predictors. Prognostically significant factors identified in the Cox regression analysis were employed to develop predictive models, which were graphically represented using a nomogram.

The nomogram’s effectiveness was assessed using the concordance index (C-index) and area under the curve (AUC). The predictive accuracy was determined by calculating the net reclassification improvement (NRI) and integrated discrimination improvement (IDI). Calibration plots were used to evaluate the consistency between the nomogram’s prognosis prediction and the observed outcomes. Lastly, the net clinical benefit of the predictive model developed in this study was evaluated using decision-curve analysis (DCA) (12).

Results

Study population characteristics

We screened 1,105 patients who satisfied the inclusion criteria (Figure 1). The baseline characteristics of the patients in the two randomly assigned groups upon admission are presented in Table 1. The enrolled patients exhibited a median age of 65.8 (IQR: 54.3, 77.2) years, comprising 581 women and 524 men. Notably, comorbid diabetes mellitus was more prevalent (33.6%) than comorbid AIDS (3.4%) and GERD (19.1%). The median SOFA, APS III, and Simplified Acute Physiology Score (SAPS) II were 6 (IQR: 3, 8), 56 (IQR: 41, 74), and 39 (IQR: 31, 49), respectively. Regarding physiological parameters, the median PO₂ for both groups was 97 (IQR: 64, 177) mmHg and 98 (IQR: 67, 171) mmHg, while the median MBP was 75.0 (IQR: 68.5, 82.4) and 74.6 (IQR: 68.3, 82.4) mmHg, respectively. Laboratory parameters revealed a median RBC of 3.40 (IQR: 2.90, 3.94) and 3.35 (IQR: 2.92, 3.89) ×10⁶/µL, a median PLT of 197 (IQR: 119, 273) and 202 (IQR: 135, 286) ×10³/µL, and a median ALB of 2.9 (IQR: 2.4, 3.4) and 3.0 (IQR: 2.5, 3.4) g/dL, respectively. In addition, the 365-day mortality rates for the training and validation cohorts were 57.8% and 55.4%, respectively.

Figure 1 Flowchart of study population selection. ICD, International Classification of Diseases; ICU, intensive care unit; LOS, length of stay; MIMIC-IV, Medical Information Mart for Intensive Care IV; PTB, pulmonary tuberculosis.

Selected predictors

Ten potential predictors were identified through LASSO regression analysis from the initial 40 independent variables in the training cohort (Figure 2): age, APS III, SAPS II, PO₂, HR, temperature, PLT, ALB, RBC, and BUN. Subsequently, based on the 10 predictors from the LASSO regression analysis, validation was conducted using multivariate Cox proportional hazard model analysis. Two independent variables, SAPS II (P=0.78) and BUN (P=0.08), were excluded because they demonstrated a P value >0.05. The finalized model comprised eight variables: age, APS III, PO₂, HR, temperature, PLT, ALB, and RBC.

Figure 2 Screening variables through LASSO regression. (A) Illustration of the variability in coefficients for the variables; (B) optimal parameter λ selection process in the LASSO regression model utilizing cross-validation. LASSO, least absolute shrinkage and selection operator.

Establishment of a nomogram using the training cohort

We developed a nomogram to predict 28-day, 180-day, and 365-day overall survival (OS) using the variables identified in the preceding discussion (Figure 3). According to the nomogram, the patients’ mean body temperature exerted the most significant impact on survival, followed by age, APS III, mean HR, ALB, PLT, PO₂, and RBC. Each component was delineated as a line segment on the nomogram, with its numerical scale representing the associated risk level. The cumulative scores across all criteria for each patient corresponded to their respective probability of death at 28, 180, and 365 days.

Figure 3 Nomogram for predicting prognostic survival in patients with PTB in the ICU. ALB, albumin; APS III, Acute Physiology Score III; ICU, intensive care unit; PLT, platelet count; PO₂, partial pressure of oxygen; PTB, pulmonary tuberculosis; RBC, red blood cell count.

Validation of the nomogram using the validation cohort

The C-index and AUC values provide a comprehensive evaluation of the model’s predictive performance. In the training and validation cohorts, the C-index was 0.711 and 0.708, respectively. The receiver operating characteristic (ROC) curves are shown in Figure 4. In the training cohort, the predictive model exhibited a 28-day AUC of 0.747, 180-day AUC of 0.747, and 365-day AUC of 0.755. Similarly, in the validation cohort, the model achieved 28-day, 180-day, and 365-day AUC values of 0.748, 0.763, and 0.750, respectively. Both C-index and AUC metrics suggest that the nomogram exhibits good predictive accuracy and acceptable discriminatory power for survival outcomes in ICU-admitted PTB patients. These results underscore the model’s potential for clinical use while highlighting the importance of cautious interpretation.

Figure 4 Nomogram depicting ROC curves, indicating AUCs for 28-, 180-, and 365-day survival. (A) Training cohort; (B) validation cohort. AUC, area under the curve; ROC, receiver operating characteristic.

Calibration plots were used to assess the agreement between the nomogram’s predicted survival probabilities and the actual observed outcomes. The calibration plot closely aligned with the standard line (Figure 5), suggesting reasonable calibration performance. The DCA curve in Figure 6 further demonstrates the clinical applicability of the nomogram, suggesting the overall advantage of using the new model to predict severe PTB. However, it is important to note that the net benefit demonstrated by the DCA does not specify the exact clinical interventions influenced by the nomogram. For instance, clinicians could potentially use the nomogram to identify high-risk patients who may benefit from closer monitoring, early ICU interventions, or aggressive treatment strategies. These interventions, informed by the nomogram, could theoretically improve patient outcomes by reducing mortality or preventing complications. Future studies are needed to explore the impact of nomogram-driven interventions on clinical outcomes in real-world settings.

Figure 5 Calibration curves for 28-, 180- and 365-day survival in the training cohort (A-C) and validation cohort (D-F).

Figure 6 Decision-curve analysis of the training cohort (A-C) and validation cohort (D-F) for 28-, 180-, and 365-day survival. OS, overall survival; PTB, pulmonary tuberculosis.

Considering the pivotal role of the APS III in critical illness assessment, we evaluated the discriminative capability of the nomogram by comparing it with the APS III using the NRI and IDI metrics. The computed NRI values for the training cohort were 0.507 (95% CI: 0.376–0.732), 0.608 (95% CI: 0.490–0.755), and 0.654 (95% CI: 0.537–0.801) for days 28, 180, and 365, respectively, while the corresponding values in the validation cohort were 0.348 (95% CI: 0.094–0.667), 0.523 (95% CI: 0.346–0.775), and 0.560 (95% CI: 0.242–0.754). Similarly, the IDI values in the training cohort were 0.080, 0.127, and 0.136 for days 28, 180, and 365, and 0.072, 0.105, and 0.115 in the validation cohort. All corresponding P values were less than 0.001, indicating statistically significant improvements in predictive performance. The positive values demonstrate that the nomogram improved the accuracy of patient classification compared to the APS III. Additional analyses demonstrated that the model excluding APS III remained superior to APS III alone, while incorporation of APS III into the full model yielded modest but statistically significant improvements in discrimination, supporting the complementary prognostic contribution of APS III rather than redundancy.

Discussion

TB, traditionally considered a chronic disease, exhibits specific acute manifestations such as miliary TB, PTB, and certain extrapulmonary forms. Because of the rapid progression of acute TB, patients frequently necessitate treatment in the ICU (13). Given the limited existing literature on the prognostic survival of PTB patients in the ICU, this study addresses this gap by offering clinicians an effective predictive model. By leveraging a substantial dataset from the MIMIC-IV public database, this study provides valuable insights for predicting the prognosis of severe PTB.

This retrospective cohort analysis employed LASSO-Cox regression analysis to identify independent risk factors associated with survival after ICU admission in PTB patients from the MIMIC-IV database. The final model comprised eight variables: age, APS III, PO₂, HR, temperature, PLT, ALB, and RBC count. A nomogram was subsequently constructed to estimate 28-, 180-, and 365-day survival probabilities. The nomogram demonstrated good calibration and improved clinical utility compared with APS III.

Age is a pivotal risk factor for adverse prognoses across various diseases, given the inevitable decline in immunity associated with aging (14,15). Furthermore, the compromised nutritional status of the elderly, coupled with their heightened susceptibility to other comorbidities, contributes to the unfavorable prognosis observed in this demographic. The Acute Physiology and Chronic Health Evaluation III (APACHE III) score serves as a reliable predictor of ICU mortality; the APS III as its component, is particularly applicable for clinical utility (16). A higher APS III signifies a more severe condition and poorer prognosis, consistent with the conclusions drawn from the nomogram. It is well known that PO₂ represents the tension of oxygen dissolved in the blood. In PTB patients, lung damage contributes to a reduction in PO₂, resulting in hypoxia (17). The degree of this reduction is intricately linked to the patient prognosis. Elevated respiratory and HRs are significant contributors to unfavorable prognosis in patients. Although composite oxygenation indices are frequently applied in critical care, fraction of inspired oxygen (FiO₂) data in the present database had a substantial proportion of missing values and therefore could not be reliably incorporated into the model. Consequently, PO₂ was retained as a directly available and clinically relevant indicator of oxygenation status.

In addition to their conventional hemostatic function, PLTs also assume a crucial immunologic role in combating Mycobacterium (18). Studies have found that PLT drive monocyte differentiation into epithelioid-like multinucleated giant foam cells with suppressive capacity upon mycobacterial stimulation (19). PLT counts undergo alterations in response to acute and critical illnesses, and the extent of these changes significantly affects the patient’s prognosis. ALB has emerged as an independent risk factor predicting an unfavorable prognosis in patients with PTB in the ICU, aligning with findings from a previous study (20). ALB, serving as a vital nutrient, plays a crucial role in maintaining colloid osmotic pressure. Hypoalbuminemia results in reduced blood volume, subsequently leading to inadequate perfusion of vital organs and exacerbation of clinical symptoms. Anemia is linked to subclinical chronic inflammation (21). Anemia may contribute to nutritional disorders and compromised immune function, presenting as risk factors for the prognosis of patients with PTB (22).

Although it has been suggested that comorbidities, such as diabetes and AIDS, increase the risk of active TB, the risk factor analysis performed in this study did not show that combining these diseases had any significant impact on patient prognosis (23,24). RR and mode of ventilation were also not significant in this study.

A nomogram is a widely employed approach for presenting a model that integrates key prognostic factors and specific endpoints, enabling a quantitative assessment of the prognostic risk for an individual patient (25). In this study, a nomogram was employed to visually represent the developed predictive model. Through calculations of the C-index, AUC, and DCA, the model demonstrated superior predictive capability. Additionally, NRI and IDI calculations verified its superiority over the APS III. The developed predictive model may serve as a potential tool to assist clinicians in predicting the prognostic risk of patients with PTB in the ICU, thereby enabling early intervention to enhance clinical outcomes.

We acknowledge several limitations in this study. First, as a retrospective analysis based on data from a single center within the MIMIC-IV database, the single-center design may introduce selection bias, potentially limiting the generalizability of our findings to other clinical settings. Second, certain disease-specific variables, such as TB drug-resistance status and the extent of extrapulmonary TB involvement, were not available in the database and therefore were not included in the model. Although septicemia could be identified from the database, it was not modeled independently in this analysis. Moreover, missing data precluded the inclusion of potentially relevant variables—such as body mass index and oxygenation index—which may have affected the model’s predictive performance. As a result, residual confounding cannot be ruled out. Third, the study cohort was restricted to patients with PTB who were admitted to the ICU, which limits the applicability of our model to patients with less severe disease. Prospective multicenter studies are warranted to further validate and refine this prognostic model.

Conclusions

Our study highlights that age, APS III, PO₂, HR, temperature, PLT, ALB, and RBC count serve as independent predictors of prognostic risk for patients with PTB in the ICU. Additionally, we demonstrated that the resultant nomogram exhibits commendable predictive performance, although further external validation is needed to confirm its accuracy and applicability.

Acknowledgments

We would like to express our gratitude to the participants, developers, and investigators associated with the MIMIC-IV database (https://mimic.mit.edu/).

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2603/rc

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2603/prf

Funding: This work was supported by the National Key Research and Development Program of China (No. 2022YFC2304800), and Guangzhou Science and Technology Planning Project (Nos. 2023A03J0539, 2023A03J0992, 2024A03J0580, 2024B03J1345).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2603/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Trajman A, Campbell JR, Kunor T, et al. Tuberculosis. Lancet 2025;405:850-66. [Crossref] [PubMed]
2. Global tuberculosis report 2025. Geneva: World Health Organization; 2025. Available online: https://www.who.int/publications/i/item/9789240116924
3. Atiya N, Hasan MS, Wong CK, et al. Prevalence of Culture-Confirmed Pulmonary Tuberculosis Requiring ICU Admission and the Trend in Prevalence Before and During the COVID-19 Pandemic in a Tertiary Academic Medical Center in Kuala Lumpur, Malaysia: A 7-Year Retrospective Cohort Study. Asia Pac J Public Health 2025;37:265-71. [Crossref] [PubMed]
4. Johnson AE, Pollard TJ, Shen L, et al. MIMIC-III, a freely accessible critical care database. Sci Data 2016;3:160035. [Crossref] [PubMed]
5. Johnson AEW, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data 2023;10:1. [Crossref] [PubMed]
6. Han YQ, Yan L, Zhang L, et al. Red blood cell distribution width provides additional prognostic value beyond severity scores in adult critical illness. Clin Chim Acta 2019;498:62-7. [Crossref] [PubMed]
7. Xu Y, Zheng X, Li Y, et al. Exploring patient medication adherence and data mining methods in clinical big data: A contemporary review. J Evid Based Med 2023;16:342-75. [Crossref] [PubMed]
8. Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med 1997;16:385-95. [Crossref] [PubMed]
9. Che WQ, Li YJ, Tsang CK, et al. How to use the Surveillance, Epidemiology, and End Results (SEER) data: research design and methodology. Mil Med Res 2023;10:50. [Crossref] [PubMed]
10. Greenwood CJ, Youssef GJ, Letcher P, et al. A comparison of penalised regression methods for informing the selection of predictive markers. PLoS One 2020;15:e0242730. [Crossref] [PubMed]
11. Waldmann P, Mészáros G, Gredler B, et al. Evaluation of the lasso and the elastic net in genome-wide association studies. Front Genet 2013;4:270. [Crossref] [PubMed]
12. Vickers AJ, van Calster B, Steyerberg EW. A simple, step-by-step guide to interpreting decision curve analysis. Diagn Progn Res 2019;3:18. [Crossref] [PubMed]
13. Jacob JT, Mehta AK, Leonard MK. Acute forms of tuberculosis in adults. Am J Med 2009;122:12-7. [Crossref] [PubMed]
14. Palatella M, Guillaume SM, Linterman MA, et al. The dark side of Tregs during aging. Front Immunol 2022;13:940705. [Crossref] [PubMed]
15. Okuyama H, Weyand CM, Goronzy JJ. Generation and durability of immune memory in older adults. J Allergy Clin Immunol 2023;152:601-3. [Crossref] [PubMed]
16. Cox EGM, Wiersema R, Eck RJ, et al. External Validation of Mortality Prediction Models for Critical Illness Reveals Preserved Discrimination but Poor Calibration. Crit Care Med 2023;51:80-90. [Crossref] [PubMed]
17. Belton M, Brilha S, Manavaki R, et al. Hypoxia and tissue destruction in pulmonary TB. Thorax 2016;71:1145-53. [Crossref] [PubMed]
18. Donohue S, Leisching G, Keane J. The Emerging Role of Immunothrombosis in the Control and Pathogenesis of Mycobacterium tuberculosis. J Infect Dis 2026;233:e596-9. [Crossref] [PubMed]
19. Feng Y, Dorhoi A, Mollenkopf HJ, et al. Platelets direct monocyte differentiation into epithelioid-like multinucleated giant foam cells with suppressive capacity upon mycobacterial stimulation. J Infect Dis 2014;210:1700-10. [Crossref] [PubMed]
20. Chai B, Wu D, Fu N, et al. Evaluation of prognostic inflammatory and systemic inflammatory response indices in auxiliary diagnosis of bacteria-negative pulmonary tuberculosis: A diagnostic accuracy study. Medicine (Baltimore) 2023;102:e33372. [Crossref] [PubMed]
21. Wacka E, Wawrzyniak-Gramacka E, Tylutka A, et al. The Role of Inflammation in Age-Associated Changes in Red Blood System. Int J Mol Sci 2023;24:8944. [Crossref] [PubMed]
22. Gelaw Y, Getaneh Z, Melku M. Anemia as a risk factor for tuberculosis: a systematic review and meta-analysis. Environ Health Prev Med 2021;26:13. [Crossref] [PubMed]
23. Smith SG, Bowness R, Cliff JM. Host-directed therapy in diabetes and tuberculosis comorbidity toward global tuberculosis elimination. Int J Infect Dis 2025;155:107877. [Crossref] [PubMed]
24. Prasanna P, Herrera B, Schlesinger LS, et al. Advances in host-directed therapy for tuberculosis and HIV coinfection: enhancing immune responses. Trends Microbiol 2025;33:961-75. [Crossref] [PubMed]
25. Park SY. Nomogram: An analogue tool to deliver digital knowledge. J Thorac Cardiovasc Surg 2018;155:1793. [Crossref] [PubMed]