Abstract
Aim
To develop a risk prediction model for unplanned extubation (UE) in invasively ventilated patients in the respiratory intensive care unit (RICU) using nursing-related and ventilation characteristics, and to perform risk stratification.
Methods
This single-center retrospective study included adult patients receiving invasive mechanical ventilation in the RICU. Nursing-related variables (RASS score, delirium, physical restraint use, nurse-to-patient ratio) along with baseline and treatment data were collected. Univariate and multivariate logistic regression was employed to identify UE-associated factors and build a nomogram. Model performance was evaluated using ROC curve, Hosmer-Lemeshow test, calibration curve, decision curve analysis, and bootstrap internal validation. Risk stratification was performed based on predicted probability.
Results
Among 1120 patients, 102 (9.11%) experienced UE. Multivariate analysis identified RASS category, nurse-to-patient ratio, and FiO2 as independent factors. Compared with agitation, awake/mild sedation (OR = 0.262, P = 0.002) and deep sedation (OR = 0.071, P < 0.001) were protective. A nurse-to-patient ratio ≥ 1:4 was an independent risk factor (OR = 3.257, P = 0.001). FiO2 was protective (OR = 0.037, P = 0.009). The model achieved an AUC of 0.781 (95% CI: 0.738–0.824), sensitivity 0.760, specificity 0.700, Brier score 0.076, Hosmer-Lemeshow P 0.301, and calibration slope 1.000. UE incidence in low-, medium-, and high-risk groups was 2.80%, 15.60%, and 25.00%, respectively.
Conclusion
The UE risk prediction model based on nursing-related characteristics has good discrimination and calibration. RASS category, nurse-to-patient ratio, and FiO2 are key independent factors. This model enables effective risk stratification and supports early identification of high-risk patients and optimized nursing management; however, its generalizability should be interpreted cautiously because of the single-center retrospective design and lack of external validation.
Keywords: unplanned extubation, respiratory intensive care unit, mechanical ventilation, nursing factors, risk prediction model, nomogram
Introduction
Unplanned extubation (UE) is defined as the unintended removal of an endotracheal tube in mechanically ventilated patients before the planned time of extubation. It generally includes two major categories: self-extubation, in which the patient actively removes the tube, and accidental extubation, which occurs during nursing care, patient movement, transport, or other unplanned clinical procedures. UE may result in acute airway loss, hypoxemia, aspiration, respiratory or circulatory instability, and emergency reintubation. It can also prolong mechanical ventilation and ICU stay, increase nursing workload and healthcare costs, and adversely affect patient prognosis and ICU safety management. Therefore, it is regarded as one of the key sensitive indicators of ICU nursing quality and patient safety.1,2 Previous studies have shown considerable variation in the incidence of UE among adult ICU patients on mechanical ventilation, but the overall rate remains significant. Moreover, some patients require emergency reintubation after UE, suggesting that the clinical consequences are heterogeneous and potentially severe.3,4
Current research generally agrees that UE is not caused by a single patient factor but rather results from the combined effects of patient consciousness, sedation depth, delirium, physical restraints, intensity of nursing monitoring, and organizational factors.5 The literature indicates that agitation, higher levels of arousal, and inadequate sedation are the most consistent high-risk features for UE. Delirium, use of physical restraints, and inadequate nursing monitoring are also closely associated with increased UE risk.6,7 The Richmond Agitation-Sedation Scale (RASS) has been confirmed as a reliable tool for assessing sedation/agitation in adult ICU patients, and the Confusion Assessment Method for the ICU (CAM-ICU) is a classic method for delirium screening. The Pain, Agitation/Sedation, Delirium, Immobility, and Sleep Disruption (PADIS) guidelines also emphasize that sedation and delirium assessment, along with nursing care, should be incorporated into routine care processes for mechanically ventilated patients.
However, several limitations remain in current UE research. First, most previous studies have focused on general or surgical ICUs, whereas evidence specifically targeting patients in the respiratory intensive care unit (RICU) remains limited. Compared with other ICU populations, RICU patients often have prolonged dependence on mechanical ventilation, complex respiratory conditions, fluctuating oxygenation status, and unique nursing requirements, which may influence UE risk profiles.
Second, previous studies have mainly focused on identifying individual risk factors associated with UE, such as agitation, delirium, restraint use, and inadequate nursing monitoring. Although these findings provide important insights into potential mechanisms, isolated risk factor analysis cannot directly estimate individual UE probability or support bedside risk stratification. Clinical prediction models may overcome this limitation by integrating multiple clinically relevant predictors and evaluating their discrimination, calibration, and clinical utility. Third, although several UE prediction models based on electronic medical record data have recently been developed, many models have relied predominantly on demographic characteristics, disease severity scores, or treatment-related variables. Furthermore, some machine-learning approaches may have limited interpretability and practical implementation in routine ICU care, while external validation remains insufficient. Importantly, nursing-sensitive and bedside-accessible variables, which may provide real-time information for UE risk identification, have been insufficiently incorporated into existing prediction models.8 Therefore, a clinically interpretable prediction model integrating nursing-related characteristics and ventilation-related indicators is needed to facilitate early UE risk identification and individualized prevention strategies among RICU patients.9
Based on this rationale, the present study focused on adult patients receiving invasive mechanical ventilation in the RICU. Rather than performing a conventional risk factor analysis alone, this study aimed to develop and internally validate a clinically interpretable prediction model for UE by integrating nursing-related characteristics, ventilation-related indicators, and disease severity measures. Furthermore, we developed a risk stratification scheme to identify patients with different UE risk profiles and to provide a potential tool for guiding targeted monitoring and preventive nursing management.
Methods
Study Design and Setting
This was a single-center retrospective cohort study. The study included adult patients admitted to the RICU of a tertiary hospital between January 2020 and December 2025 who received invasive mechanical ventilation. The aim was to build a risk prediction model for UE based on nursing-related characteristics, perform risk stratification, and produce a clinically usable nomogram. This study was reported according to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guideline. This study was approved by the hospital’s Ethics Committee (Approval No. 2025SLYS8). Informed consent was waived due to the retrospective design.
Inclusion criteria were: (1) age ≥ 18 years; (2) received invasive mechanical ventilation via endotracheal tube during RICU stay; (3) duration of intubation ≥ 24 hours. Exclusion criteria included: (1) patients with tracheostomy; (2) ICU stay < 24 hours; (3) planned extubation; (4) missing > 20% of key variables or inability to determine the primary outcome; (5) for patients with multiple RICU admissions, only the first admission was included, so as to avoid measurement bias from repeated entries.
Outcome Definition
The primary outcome was the occurrence of UE. UE was defined as the unintended removal of the endotracheal tube, including self-extubation by the patient, accidental dislodgement or slippage of the tube during nursing care or other unplanned procedures. UE events were confirmed by cross-checking nursing records, medical records, and adverse event reporting systems. If records were inconsistent, a third researcher reviewed and adjudicated the case. For patients with UE, the index time was defined as 24 hours before the occurrence of UE. For patients without UE, a random index time was selected during the period of invasive mechanical ventilation to establish a comparable observation window. All candidate predictors were extracted from information available before the index time to ensure temporal precedence and avoid potential information leakage.
Sample Size Calculation
Based on the rule of thumb for logistic regression analysis (events per variable, EPV ≥ 10),10 assuming 12 predictor variables in the model, at least 100 events are required. Previous studies show that the incidence of UE in ICUs is approximately 8–12%.11 Assuming a UE incidence of 10%, the minimum sample size is 100/0.10 = 1000. Therefore, this study planned to include 1000–1200 patients.
Data Collection and Candidate Predictors
All data were extracted from the EMR system by two researchers using a unified data dictionary and cross-checked. During the six-year study period (2020–2025), ICU management protocols, mechanical ventilation strategies, and nursing procedures were continuously standardized according to institutional practice. However, potential temporal variations in staffing patterns and clinical practices during this period cannot be completely excluded. The collected data were divided into four categories:
-
1)
Baseline and disease severity: age, sex, body mass index (BMI), primary admission diagnosis (classified by system), Acute Physiology and Chronic Health Evaluation (APACHE) II score (highest/closest value within 24 hours of RICU admission), and Sequential Organ Failure Assessment (SOFA) score (within 24 hours of RICU admission).
-
2)
Mechanical ventilation and treatment-related variables: duration of mechanical ventilation (days); ventilatory parameters: fraction of inspired oxygen (FiO2) within 24 hours after ICU admission, positive end-expiratory pressure (PEEP; values taken from the most recent 6 hours before the index time, or the corresponding window for the same time point; for patients without UE, values were taken from a randomly selected index time during ventilation); use of sedatives/analgesics (yes/no) and main drug categories (propofol/midazolam/dexmedetomidine, etc; opioids, etc). Index time was defined as the time of UE occurrence for the UE group; for the non-UE group, a matched ventilation day (eg, morning of day 3) or a randomly sampled time point was used as the index time, and exposure variables from the same window before that time point were extracted to reduce time-dependent bias.
-
3)
Nursing-related core predictor variables: Sedation/agitation level was assessed using the RASS with the most recent record before the index time. Categories used for modeling included agitation: RASS ≥ +1; awake/mild sedation: RASS 0 to −2; and deep sedation: RASS ≤ −3. Any positive CAM-ICU result within 24 hours before the index time was defined as delirium. Physical restraint status recorded within 6 hours before the index time was categorized as none/unilateral/bilateral (or dichotomized as present/absent depending on data distribution and model stability). Based on the shift at the index time, nurse-to-patient ratio was recorded from the shift schedule and stratified as ≤ 1:2, 1:3, or ≥ 1:4.
Statistical Analysis
Statistical analysis was performed using R 4.5.2 and SPSS 26. All tests were two-sided, and P < 0.05 was considered statistically significant. Continuous variables were first tested for normality. Normally distributed variables were expressed as mean ± standard deviation and compared using the independent samples t-test. Non-normally distributed variables were expressed as median and interquartile range [M (IQR)] and compared using the Mann–Whitney U-test. Categorical variables were expressed as frequencies and percentages and compared using the χ2-test. Patients with an indeterminable primary outcome or more than 20% missing data for key variables were excluded according to the predefined eligibility criteria. For model development and validation, complete-case analysis was performed using records with available data for the candidate predictors included in the regression models.
Candidate predictors were selected based on previous evidence, clinical relevance, bedside availability, and their potential association with UE. Continuous predictors, including age, BMI, APACHE II score, SOFA score, time from mechanical ventilation to index time, FiO2, and PEEP, were initially retained as continuous variables to avoid unnecessary loss of information. Potential nonlinearity was assessed by comparing models with linear terms and natural spline terms using likelihood ratio tests and Akaike information criterion values. Variables without evidence of meaningful nonlinearity were modeled linearly.
Univariate logistic regression was first performed to explore the association between each candidate predictor and UE. Variables with P < 0.10 in univariate analysis, together with clinically important variables, were considered for the full multivariable logistic regression model. Collinearity among candidate predictors was examined before final model construction, with particular attention to RASS category, delirium, restraint use, and sedative use. Generalized variance inflation factors and tolerance values were used to evaluate multicollinearity, and an adjusted GVIF < 5 was considered to indicate no serious multicollinearity.
The final parsimonious prediction model was selected by considering statistical significance, clinical plausibility, model stability, avoidance of overfitting, and bedside applicability. The nomogram and risk stratification tool were constructed based on the final parsimonious model. Model discrimination was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC) with its 95% CI. Model calibration was assessed using the Hosmer-Lemeshow test and a calibration curve. Clinical utility was evaluated using decision curve analysis (DCA). Model stability was assessed via internal validation using bootstrap resampling (1000 iterations), and overall predictive performance was evaluated using the Brier score. Model optimism was estimated, and optimism-corrected discrimination was evaluated using the corrected AUC. Calibration performance was assessed using the bootstrap-corrected calibration slope. Patients were stratified by risk based on the model-predicted probability, and a risk stratification tool was developed for clinical application.
Results
Comparison of Baseline Characteristics
There were no statistically significant differences between the UE and non-UE groups in age, sex, BMI, diagnostic category, APACHE II score, or SOFA score (all P > 0.05; Table 1).
Table 1.
Comparison of Baseline Characteristics
| Variable | Non-UE Group (n = 1018) | UE Group (n = 102) | Z/χ2 | P |
|---|---|---|---|---|
| Age (years) | 65.00 (56.00, 75.00) | 65.00 (54.00, 76.00) | −0.915 | 0.36 |
| Sex | 0.066 | |||
| Female | 383 (37.62%) | 29 (28.43%) | ||
| Male | 635 (62.38%) | 73 (71.57%) | ||
| BMI (kg/m2) | 23.74 (21.15, 26.57) | 23.81 (20.89, 25.81) | −0.521 | 0.602 |
| Diagnostic category | 1.794 | 0.616 | ||
| Infection/sepsis | 265 (26.03%) | 21 (20.59%) | ||
| Neurological disease | 111 (10.90%) | 13 (12.75%) | ||
| Others | 79 (7.76%) | 7 (6.86%) | ||
| Respiratory diseases | 563 (55.30%) | 61 (59.80%) | ||
| APACHE II score | 19.00 (15.00, 24.00) | 18.00 (14.00, 23.00) | −1.875 | 0.061 |
| SOFA score | 7.00 (5.00, 10.00) | 7.00 (5.00, 9.00) | −0.742 | 0.458 |
Abbreviations: UE, unplanned extubation; BMI, body mass index; APACHE II, Acute Physiology and Chronic Health Evaluation II; SOFA, Sequential Organ Failure Assessment.
Comparison of Nursing-Related Variables
There were statistically significant differences between the UE and non-UE groups in RASS score and category, delirium, use of physical restraints, and nurse-to-patient ratio (all P < 0.05). Compared with the non-UE group, the UE group had higher RASS scores, a significantly higher proportion of agitation, and a significantly lower proportion of deep sedation. In addition, the incidence of delirium was significantly higher in the UE group. Regarding restraint use, the proportion of bilateral restraints was higher in the UE group. Regarding nurse-to-patient ratio, the proportion of high ratios (≥ 1:4) was significantly higher in the UE group, while the proportion of low ratios (≤ 1:2) was lower (Table 2).
Table 2.
Comparison of Nursing-Related Variables
| Variable | Non-UE Group (n = 1018) | UE Group (n = 102) | Z/χ2 | P |
|---|---|---|---|---|
| RASS score | −1.00 (−3.00, 0.00) | 0.00 (−1.00, 2.00) | −7.910 | < 0.001 |
| RASS category | 75.322 | < 0.001 | ||
| Agitation (RASS ≥ +1) | 147 (14.44%) | 47 (46.08%) | ||
| Awake/mild sedation (RASS 0 to −2) | 547 (53.73%) | 49 (48.04%) | ||
| Deep sedation (RASS ≤ −3) | 324 (31.83%) | 6 (5.88%) | ||
| Delirium (CAM-ICU) | 39.293 | < 0.001 | ||
| No | 765 (75.15%) | 47 (46.08%) | ||
| Yes | 253 (24.85%) | 55 (53.92%) | ||
| Use of physical restraints | 13.475 | 0.001 | ||
| Bilateral restraints | 439 (43.12%) | 63 (61.76%) | ||
| No restraints | 349 (34.28%) | 21 (20.59%) | ||
| Unilateral restraints | 230 (22.59%) | 18 (17.65%) | ||
| Nurse-to-patient ratio | 35.535 | < 0.001 | ||
| ≤ 1:2 | 322 (31.63%) | 12 (11.76%) | ||
| 1:3 | 447 (43.91%) | 39 (38.24%) | ||
| ≥ 1:4 | 249 (24.46%) | 51 (50.00%) |
Abbreviations: UE, unplanned extubation; RASS, Richmond Agitation-Sedation Scale; CAM-ICU, Confusion Assessment Method for the Intensive Care Unit.
Comparison of Mechanical Ventilation and Treatment-Related Variables
There were statistically significant differences between the UE and non-UE groups in length of RICU stay, time from mechanical ventilation to index time, FiO2, and PEEP (all P < 0.05). Compared with the non-UE group, the UE group had shorter RICU stay and shorter time from mechanical ventilation to index time, as well as lower FiO2 and PEEP levels. In addition, there was a significant difference between the two groups in the use of sedatives (P < 0.001), with a lower proportion of sedative use in the UE group (Table 3).
Table 3.
Comparison of Mechanical Ventilation and Treatment-Related Variables
| Variable | Non-UE Group (n = 1018) | UE Group (n = 102) | Z/χ2 | P |
|---|---|---|---|---|
| Length of RICU stay (days) | 9.00 (6.00, 13.00) | 8.00 (5.00, 11.00) | −4.080 | < 0.001 |
| Time from mechanical ventilation to index time (days) | 4.00 (2.00, 6.00) | 3.00 (2.00, 5.00) | −3.245 | 0.001 |
| FiO2 | 0.44 (0.37, 0.51) | 0.39 (0.33, 0.44) | −5.089 | < 0.001 |
| PEEP (cmH2O) | 7.50 (6.08, 9.02) | 6.85 (5.69, 8.41) | −2.362 | 0.018 |
| Sedative use | 14.901 | < 0.001 | ||
| No | 326 (32.02%) | 52 (50.98%) | ||
| Yes | 692 (67.98%) | 50 (49.02%) |
Abbreviations: UE, unplanned extubation; RICU, respiratory intensive care unit; FiO2, fraction of inspired oxygen; PEEP, positive end-expiratory pressure.
Univariate and Multivariate Logistic Regression Analysis of UE
Univariate logistic regression analysis showed that time from mechanical ventilation to index time, FiO2, PEEP, RASS category, delirium, use of physical restraints, nurse-to-patient ratio, and sedative use were associated with UE (all P < 0.05). Compared with agitation, awake/mild sedation and deep sedation were protective factors, while delirium and a higher nurse-to-patient ratio (≥ 1:4) significantly increased UE risk.
Before multivariable modeling, continuous predictors were assessed for potential nonlinearity. No meaningful nonlinearity was observed for age, BMI, APACHE II score, SOFA score, time from mechanical ventilation to index time, or PEEP. FiO2 showed evidence of potential nonlinearity in the spline comparison; however, considering model interpretability and bedside applicability, it was retained as a continuous predictor in the parsimonious prediction model. Collinearity diagnostics showed no serious multicollinearity among the candidate predictors. In particular, the adjusted GVIF values for RASS category, delirium, restraint use, and sedative use were 1.409, 1.864, 1.023, and 1.232, respectively.
Variables with P < 0.10 in univariate analysis, together with clinically important variables, were entered into the full multivariable logistic regression model. Model selection was not based on univariate P values alone, but also considered clinical relevance, model stability, avoidance of overfitting, and bedside usability. The final parsimonious prediction model retained RASS category, nurse-to-patient ratio, and FiO2, which were then used for nomogram construction and risk stratification. Compared with agitation, awake/mild sedation (OR = 0.262, P = 0.002) and deep sedation (OR = 0.071, P < 0.001) significantly reduced UE risk. A nurse-to-patient ratio ≥ 1:4 was an independent risk factor (OR = 3.257, P = 0.001). FiO2 showed an inverse association with UE in the multivariable model (OR = 0.037, P = 0.009). Delirium, sedative use, and APACHE II score were not statistically significant in the multivariate analysis (Tables 4 and 5, Figure 1).
Table 4.
Univariate Analysis
| Variable | OR | 95% CI | P |
|---|---|---|---|
| Age | 0.993 | 0.979–1.007 | 0.307 |
| Male sex | 1.518 | 0.980–2.411 | 0.068 |
| BMI | 0.980 | 0.930–1.033 | 0.457 |
| APACHE II score | 0.972 | 0.942–1.003 | 0.078 |
| SOFA score | 0.967 | 0.913–1.023 | 0.254 |
| Time from mechanical ventilation to index time (days) | 0.862 | 0.791–0.933 | < 0.001 |
| FiO2 | 0.005 | 0.000–0.040 | < 0.001 |
| PEEP | 0.887 | 0.800–0.981 | 0.021 |
| RASS category (ref.: agitation) | |||
| Awake/mild sedation | 0.280 | 0.180–0.435 | < 0.001 |
| Deep sedation | 0.058 | 0.022–0.128 | < 0.001 |
| Delirium (yes vs no) | 3.538 | 2.340–5.372 | < 0.001 |
| Restraint (ref.: bilateral) | |||
| No restraint | 0.419 | 0.246–0.689 | 0.001 |
| Unilateral restraint | 0.545 | 0.307–0.924 | 0.030 |
| Nurse-to-patient ratio (ref.: ≤ 1:2) | |||
| ≥ 1:4 | 5.496 | 2.965–11.007 | < 0.001 |
| 1:3 | 2.431 | 1.243–4.735 | 0.012 |
| Sedative use (yes vs no) | 0.453 | 0.300–0.683 | < 0.001 |
Abbreviations: BMI, body mass index; APACHE II, Acute Physiology and Chronic Health Evaluation II; SOFA, Sequential Organ Failure Assessment; FiO2, fraction of inspired oxygen; PEEP, positive end-expiratory pressure; RASS, Richmond Agitation-Sedation Scale; OR, odds ratio; CI, confidence interval.
Table 5.
Multivariate Analysis
| Variable | OR | 95% CI | P |
|---|---|---|---|
| RASS category (ref.: agitation) | |||
| Awake/mild sedation | 0.262 | 0.107–0.574 | 0.002 |
| Deep sedation | 0.071 | 0.019–0.233 | < 0.001 |
| Delirium (yes vs no) | 0.810 | 0.340–1.719 | 0.607 |
| Nurse-to-patient ratio (ref.: ≤ 1:2) | |||
| ≥ 1:4 | 3.257 | 1.696–6.702 | 0.001 |
| 1:3 | 1.878 | 0.974–3.872 | 0.071 |
| APACHE II score | 0.975 | 0.943–1.007 | 0.127 |
| FiO2 | 0.037 | 0.003–0.421 | 0.009 |
| Sedative use (yes vs no) | 1.409 | 0.836–2.414 | 0.204 |
Notes: FiO2 was entered into the regression model as a continuous fractional variable ranging from 0 to 1. Therefore, the reported OR for FiO2 represents the change in odds associated with a 1.0 absolute increase in FiO2; for clinical interpretation, this corresponds approximately to an OR of 0.72 for each 0.10 increase in FiO2.
Abbreviations: RASS, Richmond Agitation-Sedation Scale; APACHE II, Acute Physiology and Chronic Health Evaluation II; FiO2, fraction of inspired oxygen; OR: odds ratio; CI, confidence interval.
Figure 1.
Nomogram for predicting the probability of unplanned extubation in RICU patients receiving invasive mechanical ventilation.
Notes: The nomogram was constructed using the final parsimonious prediction model including RASS category, nurse-to-patient ratio, and FiO2. For each patient, points are assigned according to the value of each predictor on the corresponding axis. The points for all predictors are summed to obtain the total score, which is then projected onto the bottom scale to estimate the individual predicted probability of UE.
Abbreviations: RASS, Richmond Agitation-Sedation Scale; FiO2, fraction of inspired oxygen; UE, unplanned extubation; RICU, respiratory intensive care unit.
These findings indicate that behavioral status, nursing workload, and oxygenation-related treatment status were the main predictors retained in the final model. In particular, agitation-related RASS status identified patients with higher UE probability, whereas a higher nurse-to-patient ratio reflected an important organizational risk signal relevant to bedside monitoring and nursing resource allocation.
Model Performance and Internal Validation
ROC curve analysis showed that the model achieved an AUC of 0.781 (95% CI: 0.738–0.824), indicating good discriminatory ability. Using the optimal cutoff value, the sensitivity was 0.760 and the specificity was 0.700.
The overall predictive performance of the model was good, with a Brier score of 0.076. The Hosmer-Lemeshow goodness-of-fit test showed χ2 = 9.510, P = 0.301, indicating good model fit. The calibration curve showed that the predicted probabilities were generally consistent with the actual probabilities, and the calibration slope was 1.000, indicating good calibration.
DCA showed that the nomogram model achieved a higher net benefit over a wide range of threshold probabilities compared with the “treat all” or “treat none” strategies.
Internal validation using 1000 bootstrap resamples showed stable model performance, with an optimism-corrected AUC of 0.765 compared with an apparent AUC of 0.781. The bootstrap-corrected calibration slope was 0.917, indicating good calibration with limited optimism (Table 6 and Figure 2).
Table 6.
Performance Evaluation and Internal Validation of the Risk Prediction Model
| Metric | Value |
|---|---|
| AUC (95% CI) | 0.781 (0.738–0.824) |
| Sensitivity | 0.760 |
| Specificity | 0.700 |
| Brier score | 0.076 |
| Hosmer-Lemeshow test (χ2, P) | 9.510, 0.301 |
| Apparent calibration slope | 1.000 |
| Bootstrap optimism-corrected AUC (1000 resamples) | 0.765 |
| Bootstrap-corrected calibration slope (1000 resamples) | 0.917 |
Abbreviation: AUC, area under the curve.
Figure 2.
Performance evaluation of the nomogram model for predicting UE.
Notes: (A) Receiver operating characteristic curve showing the discriminatory performance of the model. (B) Calibration curve comparing the predicted probability of UE with the observed probability. The closer the calibration curve is to the reference line, the better the agreement between predicted and observed risk. (C) Decision curve analysis showing the clinical net benefit of the model across different threshold probabilities compared with treat-all and treat-none strategies.
Abbreviations: UE, unplanned extubation; ROC, receiver operating characteristic; DCA, decision curve analysis.
Results of Risk Stratification for UE
Patients were stratified into low-, medium-, and high-risk groups based on the model-predicted probability. The low-risk group included 668 patients, of whom 19 experienced UE (incidence 2.80%). The medium-risk group included 320 patients, of whom 50 experienced UE (incidence 15.60%). The high-risk group included 132 patients, of whom 33 experienced UE (incidence 25.00%). The observed UE incidence increased stepwise across the low-, medium-, and high-risk groups, from 2.80% to 15.60% and 25.00%, respectively. This gradient suggests that the model-based stratification may help distinguish patients with different levels of UE risk and support prioritization of closer observation and targeted nursing attention in routine ICU management (Table 7 and Figure 3).
Table 7.
Results of Risk Stratification for UE
| Risk Stratum | Number of Patients (n) | Number of UE Events (n) | UE Incidence (%) |
|---|---|---|---|
| Low-risk group | 668 | 19 | 2.80 |
| Medium-risk group | 320 | 50 | 15.60 |
| High-risk group | 132 | 33 | 25.00 |
Abbreviation: UE, unplanned extubation.
Figure 3.
Risk stratification of unplanned extubation based on model-predicted probability.
Notes: Patients were categorized into low-, medium-, and high-risk groups according to the predicted probability generated by the final prediction model. The observed incidence of UE increased stepwise across the three strata: 2.80% in the low-risk group, 15.60% in the medium-risk group, and 25.00% in the high-risk group. This plot illustrates the practical use of the model for distinguishing patients with different levels of UE risk.
Abbreviation: UE, unplanned extubation.
Discussion
This study developed a UE risk prediction model based on nursing-related characteristics using real-world data from RICU patients receiving invasive mechanical ventilation. The incidence of UE was 9.11%, which is similar to the range reported in recent systematic reviews and quality improvement studies of adult critically ill patients, indicating that UE remains a non-negligible patient safety event in critical care.1,12 Although reported incidence varies across studies, differences in ICU type, patient characteristics, duration of mechanical ventilation, and UE definitions may contribute to this variability. An important feature of this study is that baseline severity indicators such as age, BMI, APACHE II score, and SOFA score did not differ significantly between groups. In contrast, nursing-related variables and ventilation management indicators within the index time window showed stronger explanatory power. This suggests that UE risk may not be fully explained by underlying disease severity alone and may reflect interactions among patient status, nursing workload, and the care process. This aligns with the emerging view that UE is a nursing-sensitive outcome amenable to intervention.2,12
Our study found that RASS category was one of the most stable independent influencing factors. Compared with agitation, both awake/mild sedation and deep sedation were associated with lower odds of UE, with the strongest inverse association observed in the deep sedation category. This finding supports the clinical observation that agitation may serve as an important proximal warning signal for UE. Agitated patients are more likely to have impulses to pull out the tube, have poor tolerance to the catheter, and resist nursing care, which may increase the likelihood of self-extubation or accidental dislodgement. At the same time, this result suggests that the key to sedation management is not “the deeper, the safer” but rather avoiding unrecognized and untreated agitation. Recent guidelines still emphasize an assessable, goal-directed sedation strategy rather than empirical over-sedation, and recommend incorporating pain, anxiety, delirium, and sleep into the overall management framework.13 Studies on sedation protocols for mechanical ventilation also suggest that standardized, goal-directed sedation management can improve clinical outcomes.14 Therefore, from a nursing practice perspective, our results suggest that continuous bedside RASS assessment and agitation alerts may help identify patients who require closer observation and individualized preventive attention.
It is worth noting that delirium and use of physical restraints were significantly associated with UE in univariate analysis but did not remain independent factors in the multivariate model. This can be understood in two ways. First, there is considerable clinical overlap and mutual influence among delirium, agitation, and sedation status. Patients with a positive CAM-ICU often have concurrent attention disorders, behavioral disturbances, and poor cooperation. When RASS category enters the model, part of the effect of delirium may be captured by the agitation phenotype. Second, physical restraints may partly reflect a response to a high-risk clinical state rather than acting as an independent causal exposure. When a patient is already agitated, delirious, or in a high-risk monitoring situation, restraints are often used reactively. Therefore, the statistical association may reflect risk clustering rather than an independent causal effect. Recent reviews have shown that physical restraints are commonly used in ICUs and are associated with adverse outcomes such as skin injury, worsening delirium, and psychological trauma. The benefits and risks are not balanced, so restraints should not be viewed as a simple substitute for nursing observation and communication interventions.15–17 These findings suggest that restraint use alone should not be considered sufficient for UE prevention, and that active identification of agitation and delirium remains clinically important.
Another clinically valuable finding of this study is that a higher nurse-to-patient ratio (≥ 1:4) was an independent risk factor for UE. This finding is highly consistent with recent evidence on the relationship between nurse staffing and patient safety. Insufficient registered nurse staffing is associated with increased mortality, complications, and delays in care. In the intensive care setting, where patient conditions fluctuate rapidly, device dependence is high, and assessment and intervention are frequent, a higher nurse-to-patient ratio may be associated with reduced opportunities for continuous monitoring of agitation, delirium, restraint status, and other high-risk care periods.18–20 Therefore, this study not only identifies individual-level risk factors but also elevates the issue of UE to the organizational management level: for high-risk patients, relying solely on individual nurse experience may not fully compensate for the challenges associated with high nursing workload. In other words, UE risk should not be understood solely as an individual patient behavior, but also as a signal that may reflect the ability of the nursing system to recognize and respond to high-risk situations in a timely manner.
Regarding ventilation-related indicators, FiO2 showed an inverse association with UE in the multivariable model, while time from mechanical ventilation to index time, PEEP, and sedative use did not retain independence in the multivariate analysis. This finding indicates that lower FiO2 levels were associated with higher odds of UE in our cohort. However, this association should be interpreted cautiously because FiO2 may reflect multiple clinical processes, including respiratory status, disease stage, sedation strategy, and intensity of monitoring, rather than representing an independent causal factor. In other words, One possible explanation is that patients requiring higher oxygen support may receive more intensive monitoring and respiratory management, which could partly contribute to the observed association. Conversely, when FiO2 decreases and patients gradually become awake or even agitated, if nursing monitoring and risk identification are not simultaneously intensified, UE may become more likely. Previous studies on outcomes after UE have also shown that higher FiO2, PEEP, or not having entered the weaning phase is often associated with failed UE and reintubation risk.21,22 Furthermore, recent quality improvement studies have pointed out that current UE prevention still lacks standardized strategies, and most interventions focus on single factors rather than risk identification based on dynamic changes in patient status.23 Studies on weaning failure in recent years have also shown that oxygenation status, delirium, and level of ventilatory support jointly affect reintubation risk, further supporting the stage-specific characteristics of UE risk.24 Therefore, the inverse association between FiO2 and UE should not be interpreted as evidence that higher oxygen administration itself prevents UE but rather reflects differences in the disease stage, monitoring intensity, and airway management strategy. This interpretation also emphasizes that these findings suggest that patients who appear more stable, particularly during the peri-weaning period, sedation reduction, or nursing handoffs, may still require careful UE risk assessment.
From the perspective of model performance, the nomogram model established in this study achieved an AUC of 0.781, a Hosmer-Lemeshow test P value of 0.301, and a Brier score of 0.076, indicating good discrimination, calibration, and overall predictive performance. After risk stratification, the incidence of UE increased stepwise across the low-, medium-, and high-risk groups, further supporting the clinical stratification value of the model. Compared with other recent UE prediction studies, our model does not pursue complex algorithms but rather builds on nursing and ventilation variables that can be readily obtained at the bedside, offering greater interpretability and potential clinical applicability.25,26 Moreover, recent systematic reviews on the implementation of the ABCDEF bundle have shown that standardized sedation, delirium management, and multidisciplinary collaboration can improve mechanical ventilation-related outcomes. This provides a realistic pathway for the implementation of our model: combining the nomogram with RASS/CAM-ICU assessments, handoff alerts, and dynamic allocation of nursing manpower may provide additional support beyond routine clinical judgment for early identification of high-risk patients.27
This study also has several limitations. First, this was a single-center retrospective study, and the conclusions may be affected by the case mix, ward management model, and documentation quality. Because the data were extracted from electronic medical records, potential information bias related to incomplete, delayed, or inconsistent documentation cannot be fully excluded, although UE events were cross-checked using nursing records, medical records, and adverse event reporting systems. External generalizability requires multicenter validation. Second, some exposure variables were extracted based on the index time window. Residual confounding from unmeasured factors, such as dynamic sedation adjustments, nursing experience, and institutional workflow differences, cannot be completely excluded. Although this helps reduce time-dependent bias, it cannot completely replace prospective continuous monitoring. Third, this study did not include finer-grained variables such as tube fixation method, shift timing, nurse experience, or sedative dosage and combination regimens, which may underestimate the impact of some management factors. Fourth, although continuous predictors were assessed for nonlinearity, FiO2 showed evidence of a potential nonlinear association with UE. For clinical interpretability and bedside implementation, FiO2 was retained as a continuous predictor in the final parsimonious nomogram. Future studies with larger external cohorts should further evaluate whether flexible modeling of FiO2 can improve predictive performance without compromising usability. Finally, although internal validation was performed, external validation and prospective clinical application evaluation are still lacking. Future prospective studies in multicenter RICU populations, combined with EMR-based real-time alerts and nursing process redesign, are needed to verify the actual benefits of this model in reducing UE events, optimizing nursing resource allocation, and improving patient outcomes. Furthermore, before clinical implementation, practical issues including integration into electronic medical record systems, availability of real-time predictor updates, staff training, and adaptation across different staffing models should be considered.
Conclusion
This study suggests that UE risk in the RICU may be associated with a combination of patient behavioral status, nursing workload, and care process characteristics. The prediction model incorporating RASS category, nurse-to-patient ratio, and FiO2 may help stratify UE risk and facilitate early identification of patients requiring closer monitoring. However, external validation in independent and preferably multicenter cohorts is required before the model can be recommended for routine clinical use.
Acknowledgments
Wenqing Xu and Ying Yang share first authorship.
Funding Statement
There is no fund support from any institution or individual for this research.
Data Sharing Statement
The de-identified data used and/or analyzed during the current study are available from the corresponding author upon reasonable request, subject to approval by the study institution and compliance with applicable institutional and legal data-protection policies. No directly identifiable patient information will be shared.
Human Ethics and Consent to Participate Declarations
This retrospective study was conducted in accordance with the principles of the Declaration of Helsinki. The study protocol was approved by the Human Ethics Committee of The Second Affiliated Hospital of Naval Medical University (Shanghai Changzheng Hospital) (Ethics Approval Number: 2025SLYS8). Given the retrospective nature of the study and the use of de-identified clinical data, the requirement for informed consent was waived by the ethics committee. To protect patient privacy, all data were de-identified before analysis, and direct personal identifiers were removed or replaced with coded study IDs. Data access was restricted to authorized members of the research team, and all data storage, analysis, and management procedures complied with institutional regulations and applicable data-protection requirements.
Author Contributions
All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.
Disclosure
The authors state that they have no financial or commercial ties to other entities that could be seen as a conflict of interest in the research.
References
- 1.Li P, Sun Z, Xu J. Unplanned extubation among critically ill adults: a systematic review and meta-analysis. Intensive Crit Care Nurs. 2022;70:103219. doi: 10.1016/j.iccn.2022.103219 [DOI] [PubMed] [Google Scholar]
- 2.Wu J, Liu Z, Shen D, et al. Prevention of unplanned endotracheal extubation in intensive care unit: an overview of systematic reviews. Nurs Open. 2023;10(2):392–13. doi: 10.1002/nop2.1317 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Guillemin J, Rieu B, Huet O, et al. Prospective multi-center evaluation of the incidence of unplanned extubation and its outcomes in French intensive care units. The Safe-ICU study. Anaesth Crit Care Pain Med. 2024;43(5):101411. [DOI] [PubMed] [Google Scholar]
- 4.Lin PH, Chen C-F, Chiu H-W, et al. Outcomes of unplanned extubation in ordinary ward are similar to those in intensive care unit: a STROBE-compliant case-control study. Medicine. 2019;98(11):e14841. doi: 10.1097/MD.0000000000014841 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Anis A, Patel R, Tanios MA. Analytical review of unplanned extubation in intensive care units and recommendation on multidisciplinary preventive approaches. J Intensive Care Med. 2024;39(6):507–513. doi: 10.1177/08850666231199055 [DOI] [PubMed] [Google Scholar]
- 6.Cosentino C, Fama M, Foà C, et al. Unplanned extubations in intensive care unit: evidences for risk factors. A literature review. Acta Biomed. 2017;88(5S):55–65. doi: 10.23750/abm.v88i5-S.6869 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Ai ZP, Gao XL, Zhao XL. Factors associated with unplanned extubation in the intensive care unit for adult patients: a systematic review and meta-analysis. Intensive Crit Care Nurs. 2018;47:62–68. doi: 10.1016/j.iccn.2018.03.008 [DOI] [PubMed] [Google Scholar]
- 8.Hur S, Min JY, Yoo J, et al. Development and validation of unplanned extubation prediction models using intensive care unit data: retrospective, comparative, machine learning study. J Med Internet Res. 2021;23(8):e23508. doi: 10.2196/23508 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Inata Y, Nakagami-Yamaguchi E, Ogawa Y, et al. Quality assessment of the literature on quality improvement in picus: a systematic review. Pediatr Crit Care Med. 2021;22(6):553–560. doi: 10.1097/PCC.0000000000002683 [DOI] [PubMed] [Google Scholar]
- 10.Riley RD, Snell KI, Ensor J, et al. Minimum sample size for developing a multivariable prediction model: PART II - binary and time-to-event outcomes. Stat Med. 2019;38(7):1276–1296. doi: 10.1002/sim.7992 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.da Silva PS, Fonseca MC. Unplanned endotracheal extubations in the intensive care unit: systematic review, critical appraisal, and evidence-based recommendations. Anesth Analg. 2012;114(5):1003–1014. doi: 10.1213/ANE.0b013e31824b0296 [DOI] [PubMed] [Google Scholar]
- 12.Gupta K, Espinosa L, Penikilapate S, et al. Impact of performance improvement strategies on unplanned extubation in an inner-city intensive care unit. Ther Adv Respir Dis. 2025;19:17534666251383662. doi: 10.1177/17534666251383662 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Lewis K, Balas MC, Stollings JL, et al. A focused update to the clinical practice guidelines for the prevention and management of pain, anxiety, agitation/sedation, delirium, immobility, and sleep disruption in adult patients in the ICU. Crit Care Med. 2025;53(3):e711–e727. doi: 10.1097/CCM.0000000000006574 [DOI] [PubMed] [Google Scholar]
- 14.Hughes CG, Mailloux PT, Devlin JW, et al. Dexmedetomidine or propofol for sedation in mechanically ventilated adults with sepsis. N Engl J Med. 2021;384(15):1424–1436. doi: 10.1056/NEJMoa2024922 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Berger S, Grzonka P, Amacher SA, et al. Adverse events related to physical restraint use in intensive care units: a review of the literature. J Intensive Med. 2024;4(3):318–325. doi: 10.1016/j.jointm.2023.11.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Kavumpurath J, KC Mani K, Refaat F, et al. An integrative review on physical restraint in adult critical care unit. F1000Res. 2023;12:114. doi: 10.12688/f1000research.127358.1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Zare-Kaseb A, Sarmadi S, Sanaie N, et al. Prevalence and variability in use of physical restraints in intensive care units: a systematic review and meta-analysis. Aust Crit Care. 2025;38(3):101210. doi: 10.1016/j.aucc.2025.101210 [DOI] [PubMed] [Google Scholar]
- 18.Dall’Ora C, Saville C, Rubbo B, et al. Nurse staffing levels and patient outcomes: a systematic review of longitudinal studies. Int J Nurs Stud. 2022;134:104311. doi: 10.1016/j.ijnurstu.2022.104311 [DOI] [PubMed] [Google Scholar]
- 19.Falk AC. Nurse staffing levels in critical care: the impact of patient characteristics. Nurs Crit Care. 2023;28(2):281–287. doi: 10.1111/nicc.12826 [DOI] [PubMed] [Google Scholar]
- 20.McHugh MD, Aiken LH, Sloane DM, et al. Effects of nurse-to-patient ratio legislation on nurse staffing and patient mortality, readmissions, and length of stay: a prospective study in a panel of hospitals. Lancet. 2021;397(10288):1905–1913. doi: 10.1016/S0140-6736(21)00768-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Chang TC, Cheng A-C, Hsing S-C, et al. Risk factors for reintubation and mortality among patients who had unplanned extubation. Nurs Crit Care. 2023;28(1):56–62. doi: 10.1111/nicc.12777 [DOI] [PubMed] [Google Scholar]
- 22.Ali K, Almahdi M, Algarni SS, et al. Trends, risk factors, and outcomes of unplanned extubation in a neonatal intensive care unit: a seven-year retrospective study. Front Pediatr. 2025;13:1593335. doi: 10.3389/fped.2025.1593335 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Tanios M, Nguyen HM, Devlin JW, et al. Impact of adding an airway safety prompt to the A-F bundle on unplanned extubation in the intensive care unit. BMJ Open Qual. 2025;14(3):e003196. doi: 10.1136/bmjoq-2024-003196 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Zheng X, Lu L, Ma M, et al. Effect of prophylactic noninvasive oxygen therapy after planned extubation on extubation failure in high-risk patients: a retrospective propensity score-matched cohort study. Front Med Lausanne. 2024;11:1481083. doi: 10.3389/fmed.2024.1481083 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Liu K, Liu Z, Li L-Q, et al. Predictive value of the unplanned extubation risk assessment scale in hospitalized patients with tubes. World J Clin Cases. 2022;10(36):13274–13283. doi: 10.12998/wjcc.v10.i36.13274 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Mou H, Ergashev A, Zhou B, et al. Risk prediction of unplanned extubation in inpatients using random forest and logistic regression models. J Patient Saf. 2025;21(6):386–392. doi: 10.1097/PTS.0000000000001365 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Moraes FDS, Marengo LL, Moura MDG, et al. ABCDE and ABCDEF care bundles: a systematic review of the implementation process in intensive care units. Medicine. 2022;101(25):e29499. doi: 10.1097/MD.0000000000029499 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The de-identified data used and/or analyzed during the current study are available from the corresponding author upon reasonable request, subject to approval by the study institution and compliance with applicable institutional and legal data-protection policies. No directly identifiable patient information will be shared.