A decision-tree model for predicting extubation outcome in elderly patients after a successful spontaneous breathing trial

Anesth Analg. 2010 Nov;111(5):1211-8. doi: 10.1213/ANE.0b013e3181f4e82e. Epub 2010 Sep 14.


Background: The commonly used single tests, based on a 1-time measurement of a physiologic variable, are often poorly predictive of tracheal extubation outcome because they examine only a single aspect of physiological function that affects the extubation outcome. We hypothesized that the construction of a decision-tree model, which includes multiple variables and considers the changes of these variables, may more accurately predict successful extubation.

Methods: This was a prospective observational study. From 2007 to 2008, 113 elderly patients in the medical intensive care unit on ventilation for >48 hours were enrolled. All patients underwent a 60-minute spontaneous breathing trial (SBT) [positive end-expiratory pressure of 5 cm H(2)O; automatic tube compensation, 100%]. Patients tolerating the trial were extubated immediately. The mouth occlusion pressure (P(0.1)), rapid shallow breathing index (RSBI,) and their combination (P(0.1) × RSBI) were recorded at the first, 30th, and 60th minute of the SBT. The changes in RSBI, which were determined at the 30th and 60th minute of the SBT (ΔRSBI30, ΔRSBI60), were assessed as the ratio (of RSBI30 or RSBI60) to RSBI at the first minute of the SBT.

Results: Twenty-two patients (19.5%) failed the SBT and were not included in the analysis, and 91 tolerated the trial and were extubated. At 48 hours, 73 (80.2%) remained extubated (successful extubation), and 18 (19.8%) required reintubation (extubation failure). Although theΔRSBI(30) was significantly higher in the extubation failure patients (118% ± 34%) than that in the successful extubation patients (93% ± 35%, P = 0.01), the receiver operating characteristic (ROC) analysis demonstrated that this index, with the threshold of <98%, presented poor performance in predicting successful extubation with area under the ROC curve (AUC) of only 0.76. The classification and regression-tree analysis selected 3 variables (P(0.1) × RSBI(30), RSBI(1), ΔRSBI(30)) and began with P(0.1) × RSBI(30). For patients with P(0.1) × RSBI-(30) >474 cmH(2)O*breaths/min/L, ΔRSBI(30) >98% defined a group including all failure patients but no success patients, whereas ΔRSBI(30) ≤98% included all success patients with no failure patients. For patients with P(0.1) × RSBI(30) ≤474 cm H(2)O*breaths/min/L, the combination of both a P(0.1) × RSBI(30) >328 cm H(2)O*breaths/min/L and RSBI(1) >112 breaths/min/L also defined a group including all success patients but no failure patients. Indeed, the diagnostic accuracy (DA) of the tree model, which was 89.1% with only the P(0.1) × RSBI(30) included, increased to 94.5% when both the P(0.1) × RSBI(30) and ▵RSBI(30) were included. The final tree model with the inclusion of all 3 discriminators could capture the successful extubation with diagnostic accuracy of 96.7%, AUC of 0.94 (95% confidence interval [CI], 0.87 to 0.98).

Conclusion: If the current tree model is confirmed by a prospective study with a larger sample size, it would be useful in guiding physicians making extubation decisions in elderly medical intensive care unit patients.

MeSH terms

  • Aged
  • Algorithms
  • China
  • Decision Support Techniques*
  • Decision Trees*
  • Female
  • Humans
  • Intensive Care Units
  • Intubation, Intratracheal*
  • Lung / physiopathology*
  • Male
  • Patient Selection
  • Predictive Value of Tests
  • Prognosis
  • Prospective Studies
  • Pulmonary Ventilation*
  • Respiration, Artificial*
  • Respiratory Function Tests*
  • Time Factors
  • Ventilator Weaning*