By Sydney Hyder, MD, MS
Division of Pulmonary and Critical Care Medicine, Stroger Hospital of Cook County, Chicago
SYNOPSIS: Authors used a machine learning framework to show that delaying intubation in critically ill patients when compared to early intubation does not result in a greater 30-day mortality.
SOURCE: Wanis KN, Madenci AL, Hao S, et al. Emulating target trials comparing early and delayed intubation strategies. Chest 2023;164:885-891.
While randomized controlled trials (RCTs) often are regarded as the gold standard in clinical research, they often leave a gap in knowledge concerning interventions that cannot be tested in a realistic manner due to safety and ethical concerns. Wanis et al hoped to overcome this barrier using a machine learning framework to answer the question of whether to intubate critically ill patients early in their intensive care unit (ICU) admission or to delay. Machine learning frameworks identifying causal inferences have been used successfully in other scientific fields, but they only recently have been applied to the medical field to identify causal effects in complex phenomena.1
Wanis et al used the concept of targeted trials to answer the question of whether 30-day mortality is improved by early intubation vs. delayed intubation.2 The targeted trials used a nonrandomized dataset, Medical Information Mart for Intensive Care-IV, to create three different structural models between an exposure (early vs. delayed intubation) and outcome (30-day mortality). The data applied to the causal framework ultimately were analyzed through an advanced statistical method created by the research team to measure the primary outcome of 30-day mortality with each treatment strategy and the risk difference. The three target trials varied in flexibility of patient eligibility criteria and treatment strategy:
• Target Trial 1 was the broadest in eligibility criteria, including any ICU patient admitted without defining the respiratory status, but rigid in terms of treatment strategies, requiring a patient to be intubated at eight hours vs. 48 hours irrespective of their respiratory status.
• Target Trial 2 was broad in eligibility criteria, as defined earlier, but flexible in treatment strategies. This allowed providers to either initiate or withhold intubation for patients who met clinical criteria, such as FiO2 ≥ 40% or if the respiratory rate was ≥ 30 breaths/minute, irrespective of the time since admission in both early and delayed intubation categories.
• Target Trial 3 restricted patient eligibility, only including patients admitted for respiratory failure at baseline, but allowed for the same flexible treatment strategy as Trial 2.
Patients were divided equally in each target trial into either early or late intubation categories. Early intubation was defined as intubation at eight hours, while delayed intubation was defined as delaying intubation until 48 hours.
A total of 5,893 patients met the inclusion criteria for Target Trials 1 and 2, and 1,281 patients met the inclusion criteria for Target Trial 3. In Target Trial 1, the estimated 30-day mortality was 18.7% (95% confidence interval [CI], 18.3% to 19.0%) with the intubate early strategy vs. 11.6% (95% CI, 10.7% to 12.5%) with the delayed intubation strategy, with a risk difference of 7.1 percentage points (95% CI, 6.2-7.9). In Target Trial 2, the estimated 30-day mortality was 12.5% (95% CI, 11.7% to 13.2%) with the intubate early strategy vs. 12.1% (95% CI, 11.2% to 12.9%) with the delayed intubation strategy, with a risk difference of 0.4 percentage points (95% CI, -0.1 to 0.9). Lastly, Target Trial 3’s estimated 30-day mortality was 19.3% (95% CI, 17.7% to 20.9%) with the intubate early strategy vs. 20.2% (95% CI, 18.0% to 22.5%) with the delayed intubation strategy, with a risk difference of -0.09 percentage points (95% CI, -2.5 to 0.7). The results from the three trials showed that when a patient was managed with a “realistic treatment strategy” allowing for a flexible response to the patient’s evolving respiratory status, there was no difference in 30-day mortality when intubation was delayed compared to early intubation. Therefore, if clinical status allows, providers should provide strategies that delay intubation in patients as tolerated.
COMMENTARY
Wanis et al hoped to answer the complex question of whether a patient should undergo early vs. late intubation on arrival to the ICU. While this question has been investigated under the lens of observational data, the ability to answer it using an RCT has left researchers defeated, citing patient safety and ethical concerns. By using Targeted Trials, Wanis and his team hoped to investigate a wide range of clinical situations to examine the effect of intubation timing on mortality without causing patient harm in the process. The concept of applying machine learning to better understand the application of medical strategies is enticing, since it opens the doors to investigation of endless possibilities of clinical vignettes and clinical questions that cannot be answered with conventional research methods.
Limitations of this paper include both the research team’s and computer’s inability to differentiate the cause of respiratory failure and the lack of physical exam findings that can be crucial in evaluating respiratory failure. Wanis and his team congregated respiratory failure into one category without differentiating the underlying pathology, such as infection or cardiac disease. The heterogeneity in respiratory failure etiology is notable because different underlying clinical diseases respond differently to various methods of respiratory support. As a result, this makes the study difficult to apply to all clinical settings. Secondly, the inclusion criteria for “respiratory distress” in Trial 3 are defined as an increased oxygen requirement and respiratory rate but do not consider the respiratory effort a patient may show physically on exam, which can play a crucial role in the decision to intubate. Machine learning is a powerful analytical tool, but it is limited by the data provided in its dataset and cannot replace human intuition, which helps shape the art of practicing medicine.
Wanis et al took a large step in demystifying the question around timing of intubation using targeted trials. As we can begin to expect a greater presence of machine learning in medical research, discussions should start to focus on guidelines to ensure quality and standardization, similar to expectations held for other research methodologies. Future work should be done to better understand how research using targeted trials should affect evidence-based medicine, protocols, and clinical decision-making.
REFERENCES
- Wang SV, Schneeweiss S, Franklin JM, et al; RCT-DUPLICATE Initiative. Emulation of randomized clinical trials with nonrandomized database analyses: Results of 32 clinical trials. JAMA 2023;329:1376-1385.
- Hernán MA, Robins JM. Using big data to emulate a target trial when a randomized trial is not available. Am J Epidemiol 2016;183:758-764.
