Carlson JN, Das S, De la Torre F, et al. A novel artificial intelligence system for endotracheal intubation. Prehosp Emerg Care. Mar 17, 2016. [Epub ahead of print.]
In the 1983 science-fiction film WarGames, Matthew Broderick believed he was going head-to-head with a random simulation program. However, he soon learned he’d accidentally hacked into a government artificial intelligence (AI) computer system programmed to repeatedly run thermonuclear simulations until it learned the best strategy for winning a war.
Over 30 years later, AI is no longer the stuff of science fiction. In fact, AI algorithms are found in thousands of widely used applications. But is AI technology advanced enough to assist in achieving successful endotracheal intubations (ETI) with 100% accuracy?
Background: Alternatives to ETI have become popular, as have the development of electronic tools to make ETI attempts more successful.
Recognizing that the inability to identify landmarks within the airway has been a fundamental obstacle to successful ETI, video laryngoscopy (VL) has become popular for its ability to more easily visualize the glottic opening.
More recently, VL recordings have been used retrospectively for a comparative analysis. Analyzing VL videos frame-by-frame can reveal distinct differences in successful ETI attempts vs. those that aren’t. (Think last week’s football films used to identify the exact characteristics of all the plays that led to a successful touchdown.)
Noting that video analysis was already being used to optically detect physical features, the authors considered whether AI could be used to digitally identify specific airway structures during real-time VL. The clinician could be guided on the best approach to finding the glottic opening based on data collected from hundreds or thousands of previous ETI attempts.
Methods: Using a convenience sample of just seven clinical providers—three experienced and four novices—the authors had each provider attempt 10 total intubations on the same manikin using VL. The first five attempts by each provider were used to train the AI software, while the last five were used to test the software. Each attempt was allowed to continue until ETI was successful.
The recording of each ETI attempt was divided into one-second intervals that were examined to determine if any part of the glottic opening was visible. The data was then entered into four commonly used AI algorithms to see if it could learn to recognize the glottic opening when it came into sight.
Although the authors’ ultimate objective will likely be to develop an algorithm that will have 100% sensitivity and specificity for detecting the glottic opening, for their initial study they were simply hoping the algorithm could detect the glottic opening at the same success rate as ETI in the prehospital field, which they determined to be approximately 77% of the time. Ultimately, the AI algorithms were able to identify the glottic opening with nearly 80% accuracy overall.
Limitations: The authors note that their small sample size and the lack of variation of clinicians are serious limitations. Additionally, the use of one manikin far from mimics the disparities that would be realized with the infinite variations of airway structures and their varying amounts of saliva, emesis, blood and trauma.
The authors acknowledged the adaptability of AI technology and its need to experience numerous exposures of actual ETIs before realizing the success needed to have practical application.
Discussion: As the associated costs of VL become more practical for prehospital agencies, its use will become more commonplace. With many studies attesting to the overall success of ETI using VL, successful ETI in the prehospital environment is quite likely to rise.
Using evolving technologies to continuously improve upon that success is certainly appealing. However, the idea that the current AI technology is capable of providing 100% accuracy of ETI is still perhaps a lofty goal.
True AI algorithms rely on the input of numerous combinations of variables in order to truly provide the precise expected outcomes. Although not impossible, the logistics to capture this data would be a difficult undertaking.
This study certainly provides for the first step in what appears to be a promising logical and potentially successful marriage of existing clinical hardware and evolving learning software. Although developing it for use in the clinical field would require considerable data collection and processing from various medical disciplines and settings, the potential outcome is a product that could perhaps result in an excellent training application if not an effective tool for real-world clinical practice.
What we already know: Video laryngoscopy is an effective tool for successful prehospital endotracheal intubation, significantly reducing both the number of intubation attempts, as well as many of the associated adverse factors.
What this study adds: In conjunction with video laryngoscopy, artificial intelligence algorithms could perhaps provide more effective endotracheal intubation techniques both in training, as well as in real-life scenarios.
Visit www.pcrfpodcast.org for additional podcasts and further audio commentary.