This study was a retrospective cohort analysis using the open-source critical care database MIMIC-IV (Version 1.0) (Goldberger et al., 2000; Johnson et al., 2011). The MIMIC-IV database consists of the clinical information of patients who were admitted to the ICU at the Beth Israel Deaconess Medical Center between 2008 and 2019. One author, Huan Yang, completed the Collaborative Institutional Training Initiative examination (certification number: 45308698) to achieve access to the database for data extraction.

Extubation failure was defined as the need for ventilatory support [noninvasive ventilation (NIV) or reintubation] or death within 48 h following extubation (Boles et al., 2007). We extracted variables with PostgreSQL (Version 14.0; PostgreSQL Global Development Group) and pgAdmin 4 (Version 6.11). The inclusion criteria were as follows: i) age > 18 years, ii) height > 152.4 cm, iii) with extubation records. The exclusion criteria were as follows: i) not the first time of extubation during a hospital stay, ii) unplanned extubation, iii) ICU length of stay (LOS) ≤ 48 h, iv) main data missed: MV settings and P_aCO₂ 4 hours before extubation.

Respiratory parameters, arterial blood gas (ABG), and vital signs were extracted 4 hours before extubation. To reduce missing data, other clinical and laboratory variables were extracted within 24 h before extubation, including Glasgow Coma Scales (GCS), the Sequential Organ Failure Assessment (SOFA), blood routines, liver and kidney functions, coagulations, cardiac markers, etc. Reasonable ranges were set for filtering out extreme values (Supplementary Table S10). Observed values outside these ranges were treated as missing data. For some variables with multiple measurements, the worst values were assessed based on clinical relevance. In addition, features (e.g., age, sex, comorbidities) were selected within 24 h after ICU admission. Finally, NIV use, reintubation, and death within 48 h after extubation, in-hospital mortality, ICU LOS, and hospital LOS were also assessed.

Predicted body weight was calculated using the NHLBI ARDS network formula (Acute Respiratory Distress Syndrome Network et al., 2000). f e m a l e = 45.5 + 0.91 ∗ h e i g h t   i n   c m − 152.4 m a l e = 50 + 0.91 ∗ h e i g h t   i n   c m − 152.4

VR was calculated as follows (Sinha et al., 2009): V R = m i n u t e      v o l u m e   m L × P a C O 2   m m H g / p r e d i c t e d   b o d y   w e i g h t   k g / 100   m L / k g × 37.5   m m H g

Continuous variables are presented as medians with their interquartile ranges (IQRs), and categorical variables are presented as total numbers and percentages. Normal distribution for continuous variables was determined using the Shapiro‒Wilk test (Razali and Wah, 2011). Continuous variables (median) were compared using Student’s test (normal distribution) (Mishra et al., 2019) or the Wilcoxon rank sum test (nonnormal distribution) (McKnight et al., 2010), and proportions were compared using χ2 or Fisher exact tests when the numbers were small (Hess and Hess, 2017). Statistical significance was considered to be at two-sided p < 0.05. The MIMIC-IV dataset was first randomly split into the training set (70%) and the internal validation set (30%). In the training set, a receiver operating characteristic (ROC) curve analysis was used to confirm the best cut-off values of VR, RSBI, and the partial pressure of oxygen to the fraction of inspired oxygen (P_aO₂/F_iO₂) at 4 h before extubation. The predictive ability of cut-off values was assessed for discrimination using C-statistics in the validation set. Variables with a two-sided p-value of < 0.05 from the independent Student’s t-test/Wilcoxon rank sum test and χ2/Fisher exact tests were considered for multivariate logistic regression. Lasso regression was performed for automatic feature selection. Multicollinearity was calculated through the analysis of the variance inflation factors. Thereafter, factors with clinical plausibility were selected and further analyzed in stepwise multiple logistic regression to identify the association between extubation failure and VR. To test the robustness of any association between extubation failure and VR, interactions with other covariates [i.e., age, RSBI, P_aO₂/F_iO₂, sex, positive end-expiratory pressure (PEEP), SOFA] that increased extubation failure were tested (Thille et al., 2013). Furthermore, a sensitivity analysis was conducted by excluding patients who were admitted for surgical reasons. The results were summarized with the odds ratio (OR), 95% confidence interval (CI), and p-value. To avoid bias introduced by missing data, variables with more than 10% missing data were excluded, and the analysis of the primary outcome was imputed using multiple imputation chain equations using the “MICE” package in R. All statistical analyses were performed with R version 4.0.5 (http://www.R-project.org) and Stata/SE 15.1.

As shown in Figure 1, a total of 76,540 records were obtained from patients who were admitted to the ICU. Of those, 23,499 records were extracted from patients (over 18 years old) who had extubation records. 3,569 patients who underwent planned extubation were ultimately extracted from the MIMIC-IV database according to prespecified inclusion criteria. Overall, the majority of patients (67.1%) were male, and the mean age was 65.29 ± 13.98 years (Supplementary Table S1). The Charlson comorbidity index was 6 (IQR, 4–7) and the median SOFA score was 6 (IQR, 5–9) before extubation. The detailed clinical information for the included patients before extubation is listed in Supplementary Table S1, and the baseline characteristics on ICU admission are presented in Supplementary Table S2. 1839 (56.0%, n = 3,284) patients had VR > 2 on admission, and 496 patients (13.9%, n = 3,569) had VR > 2 before extubation. The median duration of ICU LOS was 4 days (IQR, 3–7) and that of hospital LOS was 10 days (IQR, 7–16). There were 454 (12.7%) patients who failed extubation within 48 h, and the overall in-hospital mortality was 7.3%. The extracted dataset was randomly divided into the training set (n = 2,498) and the internal validation set (n = 1,071). The baseline characteristics were similar between the training set and validation set (Supplementary Table S3). For sensitivity analysis, 2,870 non-surgical patients were obtained. The rate of extubation failure was 13.6% (391) and the rate of mortality was 8.7% (250) in this population (see Supplementary Table S4).

In the training set, patients in the extubation failure group had significantly higher VRs (1.71 [1.36, 2.10] vs. 1.41 [1.19, 1.71], p < 0.001). Moreover, the extubation failure group exhibited a greater prevalence of comorbidities, required higher ventilation settings, had higher mean values for heart rate, GCS, SOFA, P_aCO₂, anion gap, and a lower mean value of pH, P_aO₂/F_iO₂, and peripheral arterial oxygen saturation (S_pO₂). Notable differences in laboratory variables included higher blood urea nitrogen level (BUN), creatinine level, and platelet count among the extubation failure group (all p < 0.05), as shown in Table 1.

Increased VR was an independent risk factor for extubation failure after multivariable adjustment (OR 2.14 [1.65–2.72], p < 0.001). Other factors, including elevated heart rate (OR 1.01 [1.00–1.02]), lower tidal volume (OR 1.00 [1.00–1.00]), higher PEEP (OR 1.08 [1.01–1.15]), increased BUN (OR 1.01 [1.01–1.02]), increased platelet count (OR 1.00 [1.00–1.00]), greater SOFA (OR 1.12 [1.07–1.17]), presence of chronic pulmonary disease (OR 1.33 [1.01–1.74]), presence of paraplegia (OR 2.44 [1.29–4.41]), presence of metastatic solid tumor (OR 2.27 [1.12–4.33]), and decreased pH (OR 0.12 [0.02–0.79]), were also significantly associated with an increased risk for extubation failure (Table 2).

Figure 2 shows ROC curves for VR at 4 hours before extubation (as a continuous variable) in the training set. The area under the ROC curve (AUC) maximized was 0.669 with a cut-off value of 1.595 using the Youden index (Supplementary Table S5). VR ≥ 1.595 remained an independent factor of extubation failure even after adjusting for age, different age groups, type of admission, P_aO₂/F_iO₂, RSBI, sex, PEEP, and SOFA (Supplementary Table S6), which was confirmed in the validation set (C-index: 0.585 [0.542, 0.628]) and non-surgical group (C-index: 0.632 [0.605–0.658]). Compared with patients with VR < 1.595, patients with VR ≥ 1.595 had prolonged ICU LOS (4 [3, 8] vs. 4 [2, 7] days, p = 0.033), a higher risk of extubation failure [78 (19.2%) vs. 70 (10.5%), p < 0.001], and higher risk of in-hospital mortality (37 (9.1%) vs. 36 (5.4%), see Table 3). Similar results were found in non-surgical patients (see Supplementary Table S7).

We also tested the predictive ability of RSBI and P_aO₂/F_iO₂ in this population, and the AUC of VR was significantly higher than that of RSBI (0.510 [IQR, 0.476–0.545]) and P_aO₂/F_iO₂ (0.586 [IQR, 0.551–0.621]) at 4 hours before extubation for predicting reintubation (all p-value < 0.05, Supplementary Table S5). In addition, significant differences were found between survivors and nonsurvivors in VR (1.44 [1.20, 1.74] vs. 1.49 [1.19, 1.96], p = 0.032, see Supplementary Table S8).