Development and validation of the Tracheostomy‑ProVent score to predict long-term mortality in patients on ventilator care > 21 days: a retrospective cohort study

Abstract

Background

Patients who undergo tracheostomy and remain ventilator-dependent for over 21 days (defined as prolonged mechanical ventilation, PMV) represent a critically ill subgroup with high long-term mortality. Existing prognostic models such as the ProVent score may not fully capture the specific risk profile of these patients. This study aimed to develop and validate a novel prognostic model, the Tracheostomy-ProVent score, to predict 3-, 6-, and 12-month mortality using clinical variables measured on day 21 of mechanical ventilation.

Methods

This retrospective cohort study included 213 tracheostomized adult patients who received invasive mechanical ventilation for more than 21 consecutive days in a tertiary care hospital between February 2011 and February 2023. Multivariate logistic regression was used to identify predictors of mortality and construct a scoring model. Model performance was assessed using area under the receiver operating characteristic curve and calibration analyses.

Results

The cumulative mortality rates in our patient cohort were 55.9%, 62.5%, and 66.7% at the 3-month, 6-month, and 1-year time points, respectively. The Tracheostomy-ProVent score showed strong predictive accuracy, with areas under the receiver operating characteristic curves of 0.871, 0.886, and 0.875 for 3-, 6-, and 12-month mortality, respectively. Notably, the new score significantly outperformed the original ProVent score in predicting 1-year mortality (AUC 0.875 vs. 0.768). The key clinical variables that were consistently associated with mortality included the requirement for vasopressors, platelet count of < 150 × 10³/µL, PaO₂/FiO₂ ratio of < 200 mmHg, and serum albumin level of ≤ 2.8 g/dL on day 21 of MV.

Conclusion

The Tracheostomy-ProVent score demonstrates potential as a valuable tool for advanced risk stratification in tracheostomized patients. However, external validation is required to confirm its generalizability prior to widespread clinical application.

Introduction

Caring for patients who require prolonged mechanical ventilation (PMV) poses significant challenges, as these patients often progress to a state of chronic critical illness. This condition is characterized by persistent organ dysfunction, recurrent infections, and heightened vulnerability to complications such as sepsis, muscle atrophy, and ventilator-associated pneumonia, and ventilator-induced lung injury—an often unseen but critical challenge in managing these patients [1, 2]. Owing to the complexity of the medical needs experienced by this patient population and the prolonged nature of their care, patients requiring PMV frequently experience poor long-term outcomes such as increased mortality and diminished quality of life. As a result, predicting long-term prognoses in this patient population is particularly difficult, although it is crucial for guiding clinical decisions and resource allocation.

Tracheostomy is often an essential intervention for patients who require PMV to effectively manage their long-term care needs. This procedure offers several advantages, such as reducing the need for deep sedation, improving patient comfort, facilitating oral hygiene, and lowering the risk of complications associated with prolonged endotracheal intubation—such as ventilator-associated pneumonia and tracheal injury [3, 4]. While it is important to acknowledge that the procedure carries risks and that the choice between percutaneous and surgical techniques is a critical consideration [5], tracheostomy is crucial for improving the immediate management of patients who require PMV and has implications for long-term outcomes. Several studies have reported associations between tracheostomies and improved patient prognoses, particularly in terms of reducing mortality and aiding weaning from mechanical ventilation [6–8]. In addition, recent evidence from a systematic review and meta-analysis of randomized controlled trials suggests that the timing of tracheostomy significantly impacts the outcomes of critically ill patients [9].

Prognostic models such as the ProVent score have been widely studied and used to predict outcomes in patients requiring PMV [10]. The ProVent score, initially developed to estimate 1-year mortality risk in patients who require PMV, has demonstrated utility across various clinical settings, and its effectiveness has been validated in multiple studies [11–13]. This model incorporates factors such as advanced age, vasopressor use, and the need for hemodialysis—all of which have reliably predicted poor outcomes in these patients​. Importantly, patients who undergo tracheostomy within the PMV pathway constitute a physiologically distinct subgroup because the procedure itself fundamentally alters their clinical trajectory. As noted earlier, tracheostomy confers significant therapeutic benefits—such as reduced sedation requirements, improved airway clearance, and facilitation of rehabilitation—which mitigate the immediate risks associated with prolonged translaryngeal intubation [3, 4, 6–8]. Consequently, the determinants of long-term survival in this cohort differ from those in the general PMV population. This modified risk profile implies that general ProVent models may not accurately capture the specific prognosis of tracheostomized patients, highlighting the need for a tailored prognostic instrument. Moreover, it is essential to develop a model that accounts for clinical variables and underlying comorbidities that significantly influence long-term outcomes based on reported data [14–16]. Conditions such as chronic organ dysfunction, immunocompromised status, and other comorbidities may profoundly affect patient prognoses and should be integrated into predictive models to improve their accuracy. By including clinical and comorbidity-related factors, we aimed to provide a more comprehensive tool for predicting long-term mortality in patients with tracheostomies requiring PMV.

This study aimed to create and validate a prognostic model specifically tailored to patients who have undergone tracheostomies and require PMV. This model incorporates clinical variables measured during MV, as well as underlying comorbidities, to predict mortality at 3 months, 6 months, and 1 year following the initiation of MV. Our goal was to improve the accuracy of long-term prognostication in this patient population to assist clinicians with making more informed decisions.

Materials and methods

Study design and patient selection

This retrospective observational study was conducted in the adult respiratory intensive care unit (ICU), comprising 12 beds, at a 1,100-bed university-affiliated tertiary care hospital. Also, our study was reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines. The ICU was fully equipped with cardiovascular support and advanced airway monitoring systems, with a nurse-to-bed ratio of 1:3. Over the study period, all participating patients were managed according to standardized therapeutic guidelines, including lung-protective ventilator strategies [17].

We analyzed data from adult patients aged ≥ 18 years who were admitted to the respiratory ICU between February 1, 2011 and February 28, 2023. Those eligible for screening had undergone MV for > 21 days following initial endotracheal intubation. Those with pre-existing chronic conditions requiring invasive MV before ICU admission, as well as those aged < 18 years, were excluded from the analysis. Further exclusion criteria included patients with irreversible brain damage or neuromuscular diseases (regardless of length of ICU stay) as well as those with incomplete medical records lacking essential clinical information. These inclusion and exclusion criteria were based on previously published protocols [10, 13, 14]. We ultimately included only patients undergoing tracheostomy procedures within the first 21 days of receiving MV. The attending physicians decided to perform tracheostomies for patients who were anticipated to require PMV.

Outcome and data collection

The primary outcomes of this study were mortality rates at 3 months, 6 months, and 1 year after the 21st day of MV. Demographic and clinical data—including age, sex, and body mass index—were retrospectively collected from each patient’s electronic medical records (EMRs). Illness severity was assessed using the Acute Physiology and Chronic Health Evaluation II score, while organ failure was evaluated using the Sequential Organ Failure Assessment score, calculated from both laboratory and clinical data obtained on the day of each patient’s admission to the ICU [18, 19].

Comorbidities present at the time of admission and pre-existing conditions were also extracted from the patients’ EMRs and quantified using the Charlson Comorbidity Index [20]. Additionally, data were retrospectively gathered regarding the primary diagnoses requiring MV, duration of MV, length of stay (LOS) in both the ICU and hospital, and in-hospital mortality. Mortality at 3 months, 6 months, and one year after the 21st day of MV was verified using the National Health Insurance Service database.

To ensure the reliability of the prognostic model, we implemented strict data completeness criteria. Patients with missing values for essential clinical variables on day 21 of MV (specifically, laboratory parameters and physiological data required for scoring) were excluded during the initial screening process. Consequently, the final analytic cohort (n = 213) contained no missing data for the candidate predictors included in the logistic regression models, and no statistical imputation methods were employed.

ProVent score and clinical variables on day 21 of MV

The patients’ ProVent scores were calculated based on clinical variables obtained on day 21 of MV, which included platelet count, need for hemodialysis, and requirement for vasopressors [10]. In addition to these parameters, other laboratory and clinical variables identified as prognostic factors for long-term MV in previous relevant studies were collected on day 21 of MV [13–15, 21]. These included hemoglobin and leukocyte counts, PaO₂/FiO₂ ratio, serum albumin, and bilirubin levels, and the need for any analgesic and sedative agents. A requirement for hemodialysis was defined as any renal replacement therapy administered on day 21 or within 48 h of day 21 of MV.

Statistical analysis

Continuous variables were tested for normality using the Shapiro-Wilk test. Continuous variables that followed a normal distribution were expressed as means ± standard deviations (SD) and compared using the Student’s t-test. Non-normally distributed continuous variables were reported as medians (ranges) and analyzed using the Mann-Whitney U test (Wilcoxon rank-sum test). Categorical variables were expressed as numbers (percentages) and compared using either χ² or Fisher’s exact tests, as appropriate.

Candidate predictors for 3-month, 6-month, and 1-year mortality were first evaluated using univariate logistic regression analysis. Continuous variables were categorized based on clinically relevant cutoffs or determined using the Youden index prior to inclusion. Variables with a p-value < 0.10 or of known clinical relevance were considered for inclusion in the multivariate model. Backward stepwise logistic regression was then performed to identify independent predictors, with retention criteria based on likelihood ratio tests. Since the goal was to create a unified prognostic tool applicable across all time points, we identified ‘common predictors’ that showed statistical significance or strong trends across the 3-month, 6-month, and 1-year models. To develop the practical ‘Tracheostomy-ProVent score’, the ẞ-coefficients of these final common predictors were simplified into integer values. Each independent predictor was assigned an integer point value proportional to its regression coefficient (rounded to the nearest whole number), following previously established methods [10, 14, 15, 22–24]. Specifically, variables with ẞ-coefficient ≥ 1.5 were assigned 2 points, while those with values between 0.5 and 1.49 were assigned 1 point. Variables with coefficients < 0.5 were not assigned points. Multicollinearity was assessed using the variance inflation factor (VIF), and all included predictors had VIF < 5. The final Tracheostomy-ProVent score was then derived by summing these points.

Kaplan–Meier survival analysis, log-rank tests, receiver operating characteristics (ROC) curve analysis, area under the ROC curve (AUC) calculations, and model calibration were used to evaluate each prognostic model. Model calibration was further assessed using the Hosmer–Lemeshow test [25]. DeLong’s test was used to compare the AUCs between the ProVent and Tracheostomy-ProVent scores in terms of predicting 1-year mortality [26]. The optimal cutoff values were determined by maximizing the Youden index [27].

We considered an AUC of 0.90–1.00 to indicate excellent accuracy; 0.80–0.90 as good; 0.70–0.80 as fair; 0.60–0.70 as poor; and 0.50–0.60 as a failure of the model [28, 29]. The sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, positive predictive value, and negative predictive value of the Tracheostomy-ProVent score in terms of predicting 3-month, 6-month, and 1-year cumulative mortality were calculated.

To evaluate internal validity, we performed non-parametric bootstrap resampling with 1,000 iterations for each outcome. In each iteration, a bootstrap sample of equal size was drawn with replacement, and the AUC was recalculated using the Tracheostomy-ProVent score as a single continuous predictor. The 2.5th and 97.5th percentiles of the bootstrap AUC distribution were used to derive 95% confidence intervals, which are reported as the primary internal validation metrics.

All tests were two-tailed, and p-values of < 0.05 were considered statistically significant. All statistical analyses were performed using SPSS version 25.0 (IBM Corp., Armonk, NY, USA), MedCalc version 23.0.2 (MedCalc Software, Ostend, Belgium), and R version 4.3.1 (The R Foundation for Statistical Computing, Vienna, Austria; http://cran.r-project.org).

Results

Patient characteristics

Over the study period, 2,635 patients received ventilation, of which 242 (9.18%) required MV for > 21 days. Three were excluded from this analysis because of irreversible brain injury, resulting in a cohort of 239 patients who received MV for > 21 days (Fig. 1).

Fig. 1.

Study flow diagram detailing patient selection: initial cohort, inclusion/exclusion criteria, final analysis cohort of tracheostomized patients requiring prolonged mechanical ventilation (PMV). From an initial cohort of 242 patients requiring mechanical ventilation for > 21 days, patients meeting exclusion criteria (e.g., irreversible brain injury) were removed. The final analysis included 213 tracheostomized patients requiring PMV

From an initial cohort of 239 patients who required PMV, 213 (89.1%) underwent tracheostomy procedures within the first 21 days of MV. A comparison of the clinical characteristics and outcomes between the tracheostomized and non-tracheostomized patients revealed several key differences. The tracheostomized patients had longer durations of both MV and hospital LOS than the non-tracheostomized ones. The non-tracheostomized patients exhibited significantly higher 3-month (55.9% vs. 88.5%, p = 0.001), 6-month (62.9% vs. 88.5%, p = 0.008), and 1-year (66.8% vs. 88.5%, p = 0.024) cumulative mortality rates than the ones who had undergone tracheostomies (Table 1). This disparity confirms that tracheostomy patients follow a unique and lower-risk clinical trajectory.

Table 1.

Baseline demographic and clinical characteristics of the patients who did and did not undergo tracheostomy during their hospital stay

	Total (n = 239)	Tracheostomy (n = 213)	Non-tracheostomy (n = 26)	p-value
Age (years)	67.9 ± 12.2	68.4 ± 12.3	63.7 ± 11.2	0.057
Male	157 (65.7)	143 (67.1)	14 (53.8)	0.194
BMI (kg/m2 )	22.7 ± 4.4	22.5 ± 3.7	23.9 ± 8.0	0.412
MV LOS	30 (21–383)	31 (21–383)	25 (21–47)	0.003
ICU LOS	31 (7–126)	31 (7–126)	28 (21–64)	0.119
Hospital LOS	46 (21–383)	48 (21–383)	31 (21–148)	0.006
APACHE II score a	21.0 ± 7.3	20.9 ± 7.3	21.9 ± 7.7	0.544
SOFA score a	7.3 ± 3.6	7.2 ± 3.5	7.9 ± 4.2	0.478
Major diagnoses leading to MV
Pulmonary including pneumonia	174 (72.8)	157 (73.7)	17 (65.4)	0.360
Neurologic	20 (8.4)	19 (8.9)	1 (3.8)	0.706
Infectious diseases other than pneumonia	21 (8.8)	17 (8.0)	4 (15.4)	0.260
Postoperative state	6 (2.5)	5 (2.3)	1 (3.8)	0.503
Trauma	6 (2.5)	6 (2.8)	0 (0.0)	> 0.999
Underlying comorbidities
Diabetes mellitus	81 (33.9)	70 (32.9)	11 (42.3)	0.382
Chronic cardiovascular diseases b	61 (25.5)	56 (26.3)	5 (19.2)	0.634
Chronic lung diseases c	58 (24.3)	53 (24.9)	5 (19.2)	0.633
Chronic neurologic diseases d	57 (23.8)	53 (24.9)	4 (15.4)	0.339
Hemato-oncologic diseases	57 (23.8)	50 (23.5)	7 (26.9)	0.807
Immunocompromised e	38 (15.9)	30 (14.1)	8 (30.8)	0.043
Chronic kidney diseases	31 (13.0)	27 (12.7)	4 (15.4)	0.756
Chronic liver diseases	25 (10.5)	21 (9.9)	4 (9.9)	0.492
In-hospital mortality	121 (50.6)	111 (52.1)	10 (38.5)	0.216
3-month cumulative mortality	142 (59.4)	119 (55.9)	23 (88.5)	0.001
6-month cumulative mortality	155 (64.9)	132 (62.9)	23 (88.5)	0.008
1-year cumulative mortality	165 (69.0)	142 (66.7)	23 (88.5)	0.024

Continuous variables are reported as means ± standard deviations (SD) or medians (ranges) depending on their distribution, and categorical variables are reported as numbers (%)

^a All clinical data were calculated or obtained from medical records on the day of ICU admission

^b Chronic cardiovascular diseases include old myocardial infarction, heart failure, arrhythmia, cardiomyopathy, and valvular heart disease

^c Chronic lung disease includes emphysema, chronic obstructive pulmonary disease, interstitial lung disease, and chronic empyema

^d Chronic neurological diseases include old cerebrovascular accidents, history of brain damage or spinal injury, status epilepticus, dementia, and Parkinson’s disease

^e Immunocompromised includes patients receiving therapy that suppresses resistance to infection (e.g., immunosuppression, chemotherapy, radiation, long-term or recent high-dose steroids) or have a disease that is sufficiently advanced to suppress resistance to infection (e.g., leukemia, lymphoma, AIDS)

Abbreviations: BMI Body mass index, MV mechanical ventilation, ICU intensive care unit, LOS Length of stay, APACHE Acute Physiology and Chronic Health Evaluation, SOFA Sequential Organ Failure Assessment, AIDS acquired immunodeficiency syndrome

Development of prognostic models for tracheostomized patients

Multivariate logistic regression analysis identified several significant predictors of lower 3-month, 6-month, and 1-year mortality rates (Table 2). The β-coefficients of each selected predictor were simplified into integer points. The Tracheostomy-ProVent scores for predicting 3-month, 6-month, and 1-year mortality were calculated as the sum of these points, with total possible scores ranging from 0 to 10, respectively (Table 2).

Table 2.

Multivariate logistic regression analysis of multiple risk variables for 3-month, 6-month, and 1-year mortality

Variable	OR (95% CI)	p-value	β value	Points
3-Months
Requirement of sedatives on day 21 of MV	2.473 (1.121–5.456)	0.025	0.905	1
Requirement of vasopressors on day 21 of MV	3.061 (1.335–7.016)	0.008	1.119	1
Platelet count < 150 × 103/uL on day 21 of MV	3.334 (1.581–7.030)	0.002	1.204	1
PaO2/FiO2 < 200 mmHg on day 21 of MV	2.429 (1.041–5.664)	0.040	0.887	1
Albumin ≤ 2.8 g/dL on day 21 of MV a	6.552 (2.634–16.297)	< 0.001	1.880	2
Hemato-oncologic malignancies as underlying comorbidities	2.849 (1.030–7.877)	0.044	1.047	1
6-Months
Requirement of vasopressors on day 21 of MV	2.568 (1.069–6.167)	0.035	0.943	1
Requirement of hemodialysis on day 21 of MV	5.130 (1.202–21.895)	0.027	1.635	2
Platelet count < 150 × 103/uL on day 21 of MV	3.854 (1.733–8.570)	0.001	1.349	1
PaO2/FiO2 < 200 mmHg on day 21 of MV	3.422 (1.437–8.145)	0.005	1.230	1
Albumin ≤ 2.8 g/dL on day 21 of MV a	7.694 (3.218–18.397)	< 0.001	2.040	2
Chronic cardiovascular diseases as underlying comorbidities	2.878 (1.112–7.451)	0.029	1.057	1
Hemato-oncologic malignancies as underlying comorbidities	4.316 (1.436–12.969)	0.009	1.462	1
1-Year
Age ≥ 70 b	2.847 (1.304–6.215)	0.009	1.046	1
Requirement of vasopressors on day 21 of MV	3.290 (1.312–8.252)	0.011	1.191	1
Requirement of hemodialysis on day 21 of MV	9.762 (2.087–45.672)	0.004	2.279	2
Platelet count < 150 × 103/uL on day 21 of MV	2.787 (1.304–6.215)	0.012	1.025	1
PaO2/FiO2 < 200 mmHg on day 21 of MV	2.432 (1.012–5.844)	0.047	0.889	1
Albumin ≤ 2.8 g/dL b on day 21 of MV a	6.480 (2.834–14.815)	< 0.001	1.869	2
Immunocompromised c	11.747 (2.092–65.959)	0.005	2.464	2

^a Cut-off value was based on the maximum Youden’s index. For 3-month mortality, the AUC was 0.767 (95% CI 0.705–0.822, p < 0.001; sensitivity 58.5%, specificity 76.5%); for 6-month mortality, the AUC was 0.787 (95% CI 0.725–0.840, p < 0.001; sensitivity 56.8%, specificity 87.1%); and for 1-year mortality, the AUC was 0.765 (95% CI 0.702–0.820, p < 0.001; sensitivity 56.3%, specificity 83.0%)

^b The cutoff value was based on the maximum Youden’s index. For 1-year mortality, AUC was 0.572 (95% CI 0.503–0.639, p = 0.099; sensitivity 62.0%, specificity 56.3%)

^c Immunocompromised patients include those receiving therapies or having conditions that suppress infection resistance (e.g., immunosuppression, chemotherapy, radiation, high-dose steroids, leukemia, lymphoma, and AIDS)

Abbreviations: OR odds ratio, CI confidence interval, MV mechanical ventilation, AIDS acquired immunodeficiency syndrome

The Tracheostomy-ProVent scores for predicting 3-month, 6-month, and 1-year mortality, calculated from their β-coefficients, demonstrated good discriminative ability and excellent calibration, with Hosmer–Lemeshow p-values of 0.701, 0.733, and 0.959, respectively, for each time point (Table 3). The designated cutoff values for the 3-month, 6-month, and 1-year mortality predictions are shown in Table 3. The Kaplan–Meier survival curves for all the predictive models are expressed in Fig. 2. The AUC of the Tracheostomy-ProVent score demonstrated significant superiority over the original ProVent score, which was calculated from reported data (Fig. 3) [10].

Table 3.

Sensitivity, specificity, and positive and negative predictive values of prognostic models (Tracheostomy-ProVent scores for 3-, 6-, and 12-month mortality)

Model	Areaa (95% CI)	Cut-off value	SS (%)	SP (%)	PPV (%)	NPV (%)	PLR	NLR
Tracheostomy-ProVent score : 3-months	0.871 (0.818–0.913)	3	79.8	80.9	84.1	76.0	4.17	0.25
Tracheostomy -ProVent score : 6-months	0.886 (0.835–0.925)	3	69.7	88.9	91.1	64.3	6.27	0.34
Tracheostomy -ProVent score : 1-year	0.875 (0.823–0.916)	3	74.7	88.7	93.0	64.6	6.63	0.29
ProVent score b	0.768 (0.706–0.823)	2	62.0	78.9	85.4	50.9	2.93	0.48

^a Area under the receiver operating characteristic curve

^b Based on reported data from Crit Care Med (2008;36:2061–2069) and Crit Care Med (2012;40:1171–1176)

Abbreviations: OR odds ratio, MV mechanical ventilation, SS sensitivity, SP specificity, PPV positive predictive value, NPV negative predictive value, PLR Positive likelihood ratio, NLR Negative likelihood ratio

Fig. 2.

Kaplan–Meier survival curves comparing mortality outcomes across Tracheostomy-ProVent score strata at 3 months (A), 6 months (B), and 1 year (C). Patients were stratified according to their Tracheostomy-ProVent scores. Higher scores were significantly associated with increased mortality rates across all time points (Log-rank test, p < 0.001)

Fig. 3.

Receiver operating characteristic (ROC) curves demonstrating discrimination performance of the Tracheostomy-ProVent score versus the original ProVent model in predicting 1-year mortality. The Tracheostomy-ProVent score (red line, AUC = 0.875) demonstrated significantly higher discriminative accuracy compared to the original ProVent score (blue line, AUC = 0.768). The difference between the curves was statistically significant (DeLong’s test, p < 0.05)

Development of prognostic models using common prognostic factors

To create a consistent and clinically applicable prognostic score for 3-month, 6-month, and 1-year mortality, we identified four common variables that were statistically significant across 3-month, 6-month, and 1-year time points: vasopressor requirement, platelet count < 150 × 10³/µL, PaO₂/FiO₂ ratio < 200 mmHg, and serum albumin ≤ 2.8 g/dL.

Based on these variables, the definitive Tracheostomy-ProVent score was constructed. 1 point was assigned for vasopressor requirement, thrombocytopenia, and low PaO₂/FiO₂ ratio, while hypoalbuminemia was assigned 2 points (Table 4). The final scoring system, ranging from 0 to 5 points, is presented in Table 4. A cutoff of ≥ 3 was selected using the Youden Index to maximize predictive accuracy. We compared the performance of this simplified integer-based score against the full logistic regression model derived from the original β-coefficients. While the full model demonstrated slightly higher discrimination (AUCs: 0.871, 0.886, and 0.875 for 3-, 6-, and 12-month mortality, respectively), the simplified Tracheostomy-ProVent score maintained robust accuracy with AUCs of 0.839, 0.851, and 0.827 (Table 5). Given this minimal trade-off in performance for a significant gain in ease of use, the simplified model was designated as the definitive tool for clinical application. Also, Bootstrap internal validation using 1,000 resamples yielded 95% confidence intervals of 0.783–0.885 for 3-month mortality, 0.798–0.895 for 6-month mortality, and 0.768–0.876 for 12-month mortality. These findings indicate stable model performance and minimal overfitting of the simplified score. The corresponding discrimination metrics are summarized in Table 5.

Table 4.

Calculation of the final Tracheostomy-ProVent score

Prognostic variable on day 21 of MV	Clinical criterion	Point assigned
Vasopressors requirement	Required	1
Platelet count	< 150 × 10³/µL	1
PaO₂/FiO₂ ratio	< 200 mmHg	1
serum albumin	≤ 2.8 g/dL	2
Total score		Range 0–5

Abbreviations: MV mechanical ventilation

Table 5.

Sensitivity, specificity, and positive and negative predictive values of final Tracheostomy-ProVent score ^a

Mortality	Apparent AUCb (Bootstrap 95% CI)	Cutoff value	SS (%)	SP (%)	PPV (%)	NPV (%)	PLR	NLR
3-months mortality	0.839 (0.783–0.885)	3	81.5	70.2	77.6	75.0	2.74	0.26
6-months mortality	0.851 (0.798–0.895)	3	79.5	75.3	84.0	69.3	3.22	0.27
1-year mortality	0.827 (0.768–0.876)	3	76.1	76.1	86.4	61.4	3.18	0.32

^a Common prognostic factors include the requirement for vasopressors, a platelet count < 150 × 10³/µL, a PaO₂/FiO₂ ratio < 200 mmHg, and serum albumin ≤ 2.8 g/dL on day 21 of MV. When creating the equation, serum albumin ≤ 2.8 g/dL is assigned double weight

^b Area under the receiver operating characteristic curve

Abbreviations: CI confidence interval, SS sensitivity, SP specificity, PPV positive predictive value, NPV negative predictive value, PLR Positive likelihood ratio, NLR Negative likelihood ratio, MV mechanical ventilation

Discussion

This study highlights the need to develop a tailored prognostic model specifically for tracheostomized patients undergoing PMV. Our findings revealed clear differences in long-term outcomes between tracheostomized and non-tracheostomized patients on PMV, with the former exhibiting significantly lower mortality rates at 3 months, 6 months, and 1 year after day 21 of MV. These differences underscore the importance of creating prognostic tools that account for the unique clinical challenges and treatment paths experienced by tracheostomized patients. Focusing on this specific subgroup can enhance predictive accuracy, thereby improving patient management and optimizing resource allocation.

In this study, we observed that clinical variables measured on day 21 of MV, demographic data, and underlying comorbidities significantly affected long-term mortality outcomes. These findings are consistent with recent research emphasizing the importance of clinical and non-clinical factors when conducting prognoses in this patient group [14, 15]. Moreover, in contrast to the original ProVent score [10]—which relies on age, platelet count, vasopressor requirement, and hemodialysis—our new model distinguishes itself by explicitly incorporating underlying comorbidities and nutritional status (serum albumin). These additional factors reflect the chronic physiological depletion characteristic of tracheostomized patients, thereby offering a more comprehensive and accurate assessment of patient risk. We also considered the utility of general severity scoring systems, such as APACHE II and SOFA. While these models are robust for predicting outcomes based on acute physiological derangement at ICU admission, their predictive value often diminishes as patients transition into the chronic phase of critical illness. In contrast, our score utilizes variables measured on day 21 of mechanical ventilation, allowing for a dynamic assessment of the patient’s physiological reserve. Given the unique clinical pathways of tracheostomized patients, our findings suggest that they benefit from a prognostic scoring system specifically tailored to their needs—one that differs from models used for the general or non-tracheostomized PMV population. By integrating these additional factors, our model provides a more holistic risk assessment for tracheostomized patients who require PMV.

Our data identified several key prognostic factors that consistently predicted mortality at 3, 6, and 12 months on day 21 of MV—including the requirement for vasopressors, platelet count of < 150 × 10³/µL, PaO₂/FiO₂ ratio of < 200 mmHg, and serum albumin level of ≤ 2.8 g/dL. The sustained predictive power of these four common factors reflects persistent organ dysfunction and systemic compromise. Specifically, the identified predictors reflect distinct pathophysiological domains: vasopressor requirement and low PaO₂/FiO₂ ratio indicate persistent organ failure, while hypoalbuminemia and thrombocytopenia signal chronic catabolism and immune exhaustion. The consistent significance of these factors suggests that long-term survival in tracheostomized patients depends on the interplay between residual physiological reserve and the resolution of acute organ dysfunction.

These shared clinical variables emerged as consistent predictors of mortality across all timeframes, allowing us to develop a unified prognostic model that simplifies the predictive process while maintaining strong clinical accuracy. Although predicting long-term outcomes based solely on these clinical indicators on day 21 has certain limitations, these metrics can provide valuable insights for ICU physicians assessing long-term patient prognoses. This, in turn, may help guide critical decision-making and patient care planning, offering practical utility despite the inherent constraints of using these markers in isolation. However, further research is warranted to verify the clinical applicability of this model and confirm its usefulness across diverse patient populations.

We also acknowledge that long-term outcomes in this population are influenced by factors beyond clinical variables, including patient frailty, the availability of specialized early rehabilitation, and ethical decision-making regarding life-sustaining treatment [30–36]. While our model focuses on objective clinical metrics, future research should aim to integrate these holistic elements with clinical scores to further refine risk stratification and personalize care strategies.

Several key limitations were associated with this study despite its notable strengths. First, while the single-center, retrospective design is an inherent limitation, our findings provide a robust internal validation and establish a crucial foundation. Future large-scale, multicenter cohort studies are imperative to conduct rigorous external validation of the Tracheostomy-ProVent score and confirm its generalizability across diverse ICU settings and patient populations. Second, its retrospective nature may have introduced selection bias, as some patients with incomplete medical records or those who did not survive long enough to undergo tracheostomy procedures were excluded. Third, variability in weaning practices and tracheostomy timing may have influenced patient outcomes. Since this study was retrospective, there was no standardized MV weaning protocol, and the specific timing of tracheostomy (within the 21-day window) was determined by attending physicians. This heterogeneity in clinical management practice is an inherent limitation and should be addressed in future prospective studies with standardized protocols. Fourth, given the relatively long study period (2011–2023), therapeutic and diagnostic protocols for critical care have inevitably evolved over time. This temporal heterogeneity suggests that our results cannot be considered conclusive but rather highlight a strong correlation between the proposed score and patient outcomes. Thus, these associations should be interpreted with caution regarding causality.

Conclusions

This study highlights the need for a specific prognostic model for tracheostomized patients who require PMV, given the distinct outcomes observed in this patient group compared to their non-tracheostomized counterparts. Our findings suggest that the Tracheostomy-ProVent score may outperform the original ProVent score in terms of predicting 1-year mortality in this specific demographic. While the model shows promise as a comprehensive risk assessment tool, future studies focusing on multicenter external validation are imperative to confirm its robustness and generalizability across diverse ICU settings.

Abbreviations

MV

Mechanical ventilation

PMV

Prolonged MV

ICU

Intensive care unit

IRB

Institutional Review Board

LOS

Length of stay

ROC

Receiver operating characteristics

AUC

Area under the ROC curve

Authors’ contributions

HJ and KL conceived and designed the study. HJ, WY, MKL, HL, JH, and YK were responsible for the acquisition of data. HJ performed the statistical analysis and interpretation of the data. HJ wrote the first draft of the manuscript. KL was responsible for project administration, supervision, and funding acquisition. All authors (HJ, WY, MKL, HL, JH, YK, KL) critically reviewed and edited the manuscript for important intellectual content and approved the final version to be published.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Biomedical Research Institute, Pusan National University Hospital (Grant No. 202400650001).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. However, these data are not publicly accessible due to privacy or ethical constraints.

Declarations

Ethics approval and consent to participate

This study’s protocol was approved by the institutional review board of Pusan National University Hospital (approval code: 2411-016-145, approval date: November 27, 2024), which waived the requirement for informed consent from the enrolled patients, their families, or surrogates, owing to its observational nature. This study had no impact on the treatment of the enrolled patients.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.