Editorial Views  |   October 2017
Extubation, Black Boxes, and Ontology
Author Notes
  • From the Division of Pulmonary and Critical Care Medicine, Hines Veterans Affairs Hospital and Loyola University of Chicago Stritch School of Medicine, Hines, Illinois.
  • Corresponding article on page 666.
    Corresponding article on page 666.×
  • Accepted for publication June 26, 2017.
    Accepted for publication June 26, 2017.×
  • Address correspondence to Dr. Tobin:
Article Information
Editorial / Airway Management
Editorial Views   |   October 2017
Extubation, Black Boxes, and Ontology
Anesthesiology 10 2017, Vol.127, 599-600. doi:
Anesthesiology 10 2017, Vol.127, 599-600. doi:

“[T]he primary challenge for clinicians is to take data generated in groups of patients and determine how to best apply the information in the single patient…”

Image: J. P. Rathmell.
Image: J. P. Rathmell.
Image: J. P. Rathmell.
IN the present issue of Anesthesiology, Silva et al.1  report on the ability of thoracic ultrasound to predict the development of respiratory distress in patients extubated after tolerating 60 min of pressure support set at 7 cm H2O. They studied 136 patients, and 18.4% required reintubation. Integrated statistical models based on thoracic ultrasound data, encompassing respiratory, cardiac, and diaphragmatic variables, predicted the development of postextubation distress with remarkable accuracy (receiver operating characteristic curves greater than 0.90). The sonographic data of greatest reliability were signs of pulmonary edema and increased diastolic left-ventricular pressure.
The study deals with an important question: development of respiratory failure after extubation. The data are novel as thoracic ultrasound has not been widely applied in decision-making about weaning/extubation. The findings are biologically plausible: cardiac problems and pulmonary edema can be responsible for respiratory failure after removal of mechanical ventilation.2 
Several aspects of the methodology raise questions. According to the authors’ explicit criteria for postextubation distress, patients were required to have a respiratory rate of more than 25 breaths/min for 2 h (Supplemental Digital Content 2 from Silva et al.,1 This means that a patient reintubated at 119 min (for any reason) was not classified as having respiratory distress. It is unsettling to witness reverence for numbers trouncing common sense. The diagnosis of respiratory distress depends primarily on tacit knowledge gleaned by a clinician standing at the bedside; no set of numbers can substitute.3 
Another concern is the method of data analysis. Silva et al.1  employed machine-learning methods to develop a mathematical model that integrated several sonographic measurements. Specifically, they employed partial least square (PLS) regression together with a bootstrap statistical procedure in developing their model and calculating the predictions. Several proprietary and open-source software packages are available for conducting PLS regression, and the technique is growing in popularity (especially in social-science disciplines).4  Statistical experts, however, have raised serious questions about the reliability of PLS regression.5 
Apart from fundamental mathematical questions about PLS regression, a major problem for clinicians is its black-box character. Combining several sonographic measurements into weighted sums (composites) is troublesome because the generated entities cannot be discerned and are difficult to interpret. (Mathematical quirks, such as “capitalization on chance,”5  can inflate results, and seeing several receiver operating characteristic curves above 0.90 arouses suspicion of such occurrences.) Aggregation of several variables obscures the contribution of any given antecedent in the outcome of a model, making it difficult for physicians to assess the causal contribution of any particular phenomenon to a clinical catastrophe. Black-box techniques raise the prospect of clinicians pushing buttons on digital devices to arrive at decisions based on analytic foundations they do not understand.
The reference standard against which the sonographic measurements were compared was a determination of the causes of postextubation respiratory failure by experts reviewing the medical records. Since the days of David Hume (1711 to 1776), it has been recognized that reaching conclusions on causation is extremely complex, even when rigorous experimentation is employed. Considering the amount of missing information in any medical record, this reference standard is dubitable.
An aligned concern is learning of a technique being commended for its ability to compensate for missing data (Supplemental Digital Content 4 from Silva et al.,1 According to this framework, a doctor is advised to make life-and-death decisions based on data he or she is simply imagining as opposed to measuring. For most of the twentieth century, clinical scientists aspired to the methodology of physics. Science may not always attain ontologic certitude—ontology being the branch of metaphysics concerned with what really exists (as opposed to what appears to exist and does not)6 —but science prides itself, unceasingly, as the polar opposite of fiction. Today’s researchers rely increasingly on paradigms borrowed from social-science disciplines and business. When imaginary data are considered the equal of actual measurements, we have strayed very far from the certainty desired by Thomas the Apostle.
As with all of medicine, the primary challenge for clinicians is to take data generated in groups of patients and determine how to best apply the information in the single patient being managed at a given moment in time.7  In patients recovering from respiratory failure, extubation is fraught with a high risk of death. Most patients (80 to 90%) tolerate extubation without difficulty, but patients developing distress after extubation have a high mortality (sevenfold higher in the present study). Extubation is too great a hazard to doff one’s thinking cap and rely instead on a black box.8 
Contrary to the authors’ claim that pressure support of 7 cm H2O is a low setting, it achieves an average decrease in work of breathing of at least 40%.8,9  In some patients, the decrease exceeds 80%. Before removing an endotracheal tube in a fragile patient, the clinician needs to ensure that the patient can cope with an increase in the work of breathing of this magnitude. In such circumstances, it is imperative to ensure that the patient can breathe without distress in the complete absence of ventilator assistance—with pressure support and positive end-expiratory pressure both set at 0 or with use of a T-piece circuit.8 
The limited differences in respiratory frequency (and total lack of discrimination in tidal volume) between success and failure patients at pressure support of 7 cm H2O provide vivid evidence of the confounding influence of pressure support when making decisions about weaning/extubation (tables in Supplemental Digital Content 6 from Silva et al.,1 The presence of pulmonary edema at the start of the study and its lack of progression over time (Supplemental Digital Content 8 from Silva et al.,1 further highlights how pressure support clouds the view of a clinician trying to reach a clear picture of a patient’s likelihood of tolerating extubation.
As reported by Silva et al.,1  the use of ultrasound to detect pulmonary edema and left-ventricular pressure may help clinicians when making decisions about weaning/extubation. Considering the hazards of mechanical ventilation, it would be imprudent to delay extubation based solely on the mathematical model presently reported. Decisions about weaning/extubation are confusing enough, given the prevailing practice of evaluating patients at high levels of ventilator assistance (namely, pressure support 5 to 7 cm H2O), and further extending the leap into the dark through use of imaginary numbers is best avoided.
Competing Interests
Dr. Tobin receives royalties for two books on critical care published by McGraw-Hill, Inc. (New York, New York). Dr. Laghi declares no competing interests.
Silva, S, Ait Aissa, D, Cocquet, P, Hoarau, L, Ruiz, J, Ferre, F, Rousset, D, Mora, M, Mari, A, Fourcade, O, Riu, B, Jaber, S, Bataille, B . Combined thoracic ultrasound assessment during a successful weaning trial predicts postextubation distress. Anesthesiology 2017; 127:666–74
Tobin, MJ, Laghi, F . Tobin, MJ . Extubation, Principles and Practice of Mechanical Ventilation, 2012, pp 3rd edition. Edited by New York, McGraw-Hill Inc.1221–36
Laghi, F, Tobin, MJ . Tobin, MJ . Indications for mechanical ventilation, Principles and Practice of Mechanical Ventilation, 2012, pp 3rd edition. Edited by New York, McGraw-Hill Inc.129–62
McIntosh, CN, Edwards, JR, Antonakis, J . Reflections on partial least squares path modeling. Organizational Research Methods 2014; 17(2):210–51 [Article]
Rönkkö, M., McIntosh, CN, Antonakis, J, Edwards, JR. Partial least squares path modeling: Time for some serious second thoughts. J Operations Management 2016; 47–8: 9–27
Wulff, HR, Pedersen, SA, Rosenberg, R . Philosophy of medicine: An introduction, 1990, p 2nd edition. Oxford, Blackwell Scientific Publications, 14
Tobin, MJ . Generalizability and singularity: The crossroads between science and clinical practice. Am J Respir Crit Care Med 2014; 189:761–2 [Article] [PubMed]
Tobin, MJ . Extubation and the myth of “minimal ventilator settings.” Am J Respir Crit Care Med 2012; 185:349–50 [Article] [PubMed]
Jubran, A, Van de Graaff, WB, Tobin, MJ . Variability of patient-ventilator interaction with pressure support ventilation in patients with chronic obstructive pulmonary disease. Am J Respir Crit Care Med 1995; 152:129–36 [Article] [PubMed]
Image: J. P. Rathmell.
Image: J. P. Rathmell.
Image: J. P. Rathmell.