Of 336 consecutive patients who were eligible for the study, 36 were excluded (). Vital status one year after enrollment was confirmed by telephone follow-up or medical record review in 263 patients and by NDI review in 25 patients. One-year mortality was unknown for 12 patients, for an overall follow up rate of 96%.
Patient demographics and outcomes for the development and validation sets are presented in and . ICU admission diagnoses are shown in the table in the Appendix
. The groups were mostly similar, but patients in the validation set had higher admission APACHE II scores and higher hospital and 3-month mortality. Thirty-seven of the patients who survived hospitalization were not confirmed to have died by review of medical records or telephone follow-up. NDI records were available for 25 of them. All were noted to have survived the year. The remaining 12 patients for whom NDI records were not yet available were counted as survivors based upon survival of the 25 patients lost to telephone follow-up who did have NDI data available.
Results of bivariate analyses are presented for descriptive purposes in . All predetermined predictor variables were included in the initial maximal logistic regression model. Requirement of vasopressors, platelets ≤150 x109
/L, age ≥50, requirement of hemodialysis, and upper extremity weakness were independent predictors of death at one year in a reduced model. Clinically, upper extremity weakness was considered difficult to reproduce since the use of sedatives, which strongly affected this measurement, varied significantly between patients. This issue has affected the reliability of other illness severity models. (19
) Therefore models with and without this variable were compared. The area under the ROC curve for the final reduced model shown in (vasopressors, platelets ≤150 x109
/L, age ≥50, requirement of hemodialysis) was 0.82 (se 0.03). This compares to 0.84 (se 0.02) for the model with the final four variables plus upper extremity weakness and 0.85 (se 0.03) for the maximal model, (p=0.46 for comparison of all three). For the final reduced model, sensitivity for identifying patients at ≥50% risk of death was 0.58 (se 0.16), and specificity was 0.91 (se 0.16). Sensitivity for identifying patients at ≥90% risk of death was 0.42 (se 0.12), and specificity was 0.99 (se 0.01). The model had good fit based on its non-significant GoF test (χ210df
= 6.72, p = 0.75).
Bivariate analysis of associations between predetermined predictive variables and one-year mortality in development set
Using values measured in patients from the validation set, the same model had an area under the ROC curve of 0.82 (se 0.05), (p=0.93 compared to development set), and again demonstrated good fit (GoF χ216df = 18.31, p = 0.31). Comparisons of observed to predicted values for development and validation sets are shown in . Reliability of the model was very consistent in the validation set. These four variables were also independent predictors of 3-month mortality in a separate model (GoF χ210df = 11.39, p = 0.33), and showed consistent performance in the validation set (GoF χ216df = 23.99, p = 0.09). ()
Comparison of observed and predicted one-year mortality for patients divided into 5 equal sized groups from the Development set (2A) and Validation set (2B).
As a sensitivity analysis, patients in the validation set who were lost to follow up were assumed to have all died (rather survive, as was the case in the development set, confirmed by the NDI). The area under the ROC curve for that model was 0.76 (se 0.05). Sensitivity was 0.37 (se 0.18) and specificity remained high at 0.93 (se 0.02) for ≥90% likelihood of death.
In order to create a prognostic scoring system that could ultimately be used by clinicians in daily practice, we assigned points to each of the four predictive variables in proportion to the regression coefficients from the development model. The regression coefficients were of similar magnitude, so we assigned one point for each risk factor resulting in a range of scores from 0 to 4. Performance of the 4-point prognostic scoring system (Prognosis for Prolonged Ventilation (ProVent) score) is shown in . Predicted and observed mortality for patients in the development set and observed mortality for patients in the validation set are included. In the development set, patients with the ProVent score of 0, representing no risk factors (n=41, 21%) had a 1-year mortality of only 15%. Patients with the score of 1, representing 1 risk factor (n=98, 50%) had a 1-year mortality of 42%. Patients with the score of 2, representing 2 risk factors (n=26, 13%), had mortality of 77% at 3 months and 88% at 1 year. Patients with the 3 or 4 risk factors had 3-month mortality of 90% and 1-year mortality of 97%. This highest risk group (scores 3 or 4) represents 16% of the development set and 24% of the validation set. The area under the ROC curve for ProVent score for the combined cohort is 0.82 (se 0.03; 95% C.I. 0.75, 0.88). For patients with ≥50% risk of death, sensitivity is 0.58 (se 0.12) and specificity is 0.95 (se 0.07). For patients with ≥ 90% risk of death, sensitivity is 0.32 (se 0.20) and specificity is 0.99 (se 0.01). Survival according to ProVent score risk group is shown in .
Prognosis for Prolonged Ventilation (ProVent) Score
Figure 3 Kaplan Meier Curves by Risk group for combined cohort: Low = ProVent score 0 (no risk factors), n = 55 (18% of cohort); Intermediate = ProVent score 1 (1 risk factor), n = 137 (47% of cohort); High = Provent score 2 (2 risk factors), n = 47 (16% of cohort); (more ...)
Data on functional status were available for 57% of 1-year survivors. There were no differences between patients with and without available data for age (p=0.4), SOFA score at day 21 (p=0.30), Charlson Score (p=0.88) or premorbid independence in ADLs (p=0.68). Only 24% of survivors were independent in all ADLs after 1 year. Thirty-nine percent of survivors with ProVent scores of 0, and 18% of survivors with ProVent scores of 1 were independent in all ADLs. None of the patients with ProVent scores of 2 or greater were both alive and independent in all ADLs after 1 year.