Secondary Logo

Journal Logo

Automatic Computerized Endotracheal Tube Position Verification: An Animal Model Evaluation

Shamir, Micha Y. MD*,†; Lederman, Dror PhD, EMT-P; Gravenstein, Dietrich MD§

doi: 10.1213/ANE.0b013e31822e17a4
Technology, Computing, and Simulation: Research Reports
Free
SDC

BACKGROUND: Improper endotracheal tube positioning carries a high risk for morbidity and mortality; verification and confirmation of correct placement is necessary. We propose a computer-automated identification of endotracheal tube positioning using image analysis. The end product will not retain a monitor; rather, the acquired image will be automatically analyzed by a mini electronic processor.

METHODS: An algorithm that automatically analyzes images has been developed: it classifies images into esophagus, trachea, and carina. Image processing includes converting the image to grayscale and extracting and classifying into 1 class, on the basis of similarity to pretrained patterns. A prototypical video sensor mounted on an intubating stylet has also been assembled. This stylet was introduced into 10 bovine throats, and video images were gathered. Videos were analyzed and classified as carina, trachea, or esophagus. The videos were then introduced to the new algorithm. In each test cycle, 9 videos were used to train the algorithm, and the 10th was used as a benchmark. This procedure was repeated 10 times so that each video was used 9 times for teaching and 1 time for testing.

RESULTS: Ten videos were recorded, of which 1600 images were extracted (trachea: 490 images; carina: 550 images; and esophagus: 560 images). Only 1 esophageal image was classified as trachea (false positive 0.001%). Two carinal images and 22 tracheal images were recognized as esophagus (false negative 0.041%), sensitivity 0.98 and specificity 0.99. Twenty images of the carina were identified as trachea, and 25 images of the trachea were identified as the carina (false positive 0.045%, false negative 0.041%, sensitivity 0.96 and specificity 0.95).

CONCLUSION: A potential tube position verification system was assessed. High accuracy of the analysis algorithm was shown using nonperfused biological tissue, justifying further research.

Published ahead of print September 29, 2011 Supplemental Digital Content is available in the text.

From the *Department of Anesthesiology, Perioperative Medicine and Pain Management, University of Miami Miller School of Medicine, Miami, Florida; Department of Anesthesiology and Critical Care Medicine, Hadassah Hebrew University Medical Center, Jerusalem, Israel; Department of Radiology, University of Pittsburgh, Pittsburgh, Pennsylvania; and §Department of Anesthesiology, University of Florida, Gainesville, Florida.

Funding: None.

Conflict of Interest: See Disclosures at the end of the article.

Reprints will not be available from the authors.

Address correspondence to Micha Y. Shamir, University of Miami Miller School of Medicine, 1554 NW 179th Ave., Pembroke Pines, FL 33029, Phone: 954 234 2475, Cell: 954 614 6440. Address e-mail to shamir61@gmail.com.

Accepted July 12, 2011

Published ahead of print September 29, 2011

There are 2 stages in the process of endotracheal intubation: the first stage is glottic opening visualization and endotracheal tube (ETT) insertion; the second stage is confirmation of proper tracheal ETT positioning, a mandatory process required by current practice guidelines.12

Many new instruments were introduced over the last decade to improve laryngeal visualization (i.e., GlideScope, Airtraq, Trueview, Storz, etc.). These devices do not preclude the need for proper ETT position confirmation. Capnography and stethoscope auscultation offer ETT position confirmation and are considered by many as the “gold standard.” However, literature review suggests numerous clinical scenarios in which there have been capnography failures (cardiac arrest, difficult mask-bag ventilation plugged tube, etc).36 In addition, outside the operating room (prehospital, intensive care units, emergency departments, and hospital floors), even trained practitioners may experience higher failure rates and may have difficulty confirming ETT position.1,711 Despite these limitations, no new devices for confirmation of proper tracheal ETT positioning have been introduced recently.

In this study, we introduce a new concept for ETT position confirmation: identifying airway or esophageal anatomy using computerized analysis of visual images. The system we propose acquires a picture from a video camera at the tip of an intubating stylet, analyzes patterns found within the image, compares it with a predetermined definition of an airway, and deduces the ETT position. It relies on the identification of the unique anatomy of the airway and esophagus. Computerized analysis of medical images has been extensively studied in various clinical settings.1213 Examples include algorithms to detect breast masses in mammograms or chest computed tomography1415 and coronary artery disease diagnosed in standard magnetic resonance imaging by identifying the pattern of the coronary arteries.16 The key role in these algorithms is to automatically identify the unique predetermined visual patterns inside images. These algorithms are commonly used as “computer-aided diagnosis” systems, i.e., assisting, not replacing, humans in clinical decision-making.13

We previously tested our algorithm using a simulation mannequin,17 and in this study we tested the algorithm using biological (bovine) throat specimens.

Back to Top | Article Outline

METHODS

We assembled a prototype of an intubating stylet comprising a complementary metal oxide silicon video imaging sensor (Logitech C210® 640X480 pixels resolution sensor, Apples, Switzerland). We added an accessory lumen for air/saline spray to clean the camera surface if needed and wires to transfer the image to a digital signal processor/PC.

A supervised classifier was trained to classify each image as 1 of 3 anatomical structures: esophagus, trachea, or carina. The colored images were converted to a gray scale, represented in a predefined 5-dimensional feature space and classified into 1 of the anatomical structures on the basis of the similarity between the image pattern and the pretrained patterns of the anatomical structures. A detailed explanation of the algorithm can be found elsewhere.17

We trained our supervised classifier by presenting images from 10 fresh adult bovine throats purchased from a local butcher. The prototype apparatus was inserted into the esophagus and then into the trachea while recording videos at 25 frames per second. The authors visually examined the videos offline, frame by frame, and each frame was classified by into 1 of 3 categories: trachea (an image showing tracheal rings and only a single lumen), carina (an image of tracheal rings and 2 lumens), and esophagus (an image without rings and 1 lumen). Each frame, and its classification, was automatically stored by MATLAB® (MathWorks, Inc., Natick, MA) software.

The supervised system was implemented and trained using MATLAB to compare, frame by frame, the initial classification made by the algorithm to the position determined previously by the authors. The system is based on a probabilistic framework, according to which the images are represented in a compact feature space using statistical models called Gaussian mixture models (GMMs).18 These statistical models are estimated using an algorithm called expectation–maximization (EM),18 which was implemented by the authors in MATLAB. The EM is used until the convergence criterion is satisfied.

We tested the supervised system by presenting images to be classified using a “leave-one-case-out” validation method. (Images extracted and classified from the videos from 9 bovine throats were used to train the model, and the images from the remaining throat were used to test system performance.) This method is a commonly accepted method of testing performance of machine-learning algorithms.19 This process was repeated 10 times, such that the videos from each throat were used only once in the testing phase.

The supervised system's classification was compared with the human classification. Using a bivariate χ2 test (implemented using MATLAB's statistical toolbox), P ≤ 0.005 was considered statistically significant.

Back to Top | Article Outline

RESULTS

We extracted 1600 images from 10 bovine throat videos. The authors classified 490 images as trachea, 550 as carina, and 560 as esophagus. Figure 1 shows images from each class. Table 1 shows that the supervised system classified correctly 95.6% of the images (1530 of 1600 images).

Figure 1

Figure 1

Table 1

Table 1

Discrimination between esophageal and tracheal intubation (Table 2): In one case, of 560 esophageal images, an esophageal image was mistakenly classified as a trachea. This is a false positive rate of 1 in 580 (0.001). No esophageal image was classified as carina. In 2 cases, of 550 carinal images, a carinal image was mistakenly classified as an esophagus, and in 22 cases, of 490 tracheal images, a tracheal image was recognized as esophagus. These 2 results together are a false negative rate of 0.041. Calculated sensitivity was 0.98 and specificity 0.99.

Table 2

Table 2

Identification of correct intrarespiratory tract positioning (Table 3): 20 images of the carina were identified as the trachea (3.64%), and 25 images of the trachea were identified as the carina (5.34%). This is a false positive rate of 0.045 and false negative of 0.043 (Table 3). Calculated sensitivity was 0.96 and specificity 0.95. Figure 2 is a sample of frames incorrectly classified by the algorithm.

Table 3

Table 3

Figure 2

Figure 2

Back to Top | Article Outline

DISCUSSION

The worst mistake in endotracheal intubation is not recognizing esophageal intubation, because the result might be severe hypoxia leading to anoxic brain damage or death. Capnography and stethoscope auscultation offer ETT position confirmation by demonstrating indirect characteristics of respiration baring significant false positive and false negatives.5,2024 In this research we studied the ability of a computerized algorithm to successfully differentiate an intraluminar image of an esophagus from that of the trachea/carina as a new tool to verify proper ETT position. It was our hypothesis that image analysis is accurate because it uses direct airway property (i.e., visual anatomy).

In our study only 1 esophageal video image (of 560) was mistakenly recognized as airway; this made for a false positive rate of 0.001. Similarly, 24 images (of 1040) of trachea/carina were recognized as esophagus, yielding a 0.041 false negative rate. Of interest is the fact that the algorithm identified carina as esophagus in 2 instances, and trachea was recognized 22 times as esophagus. This 10-fold difference can be attributed to the anatomic similarity of trachea and esophagus, namely, having 1 lumen.

Currently, carbon dioxide (CO2) in exhaled air (capnography) serves as the confirmation method of choice for proper ETT positioning. This is, however, not failsafe because the air in the stomach may contain CO2.6 Keller et al. found that there is a significant potential for false-positive colorimetric capnometric results even in the presence of small amounts of carbonated beverages.25 In a different clinical scenario, that of cardiac arrest, Kramer-Johansen et al. compared the values of end-tidal (ETCO2) measured from an ETT located in the trachea and in the esophagus. The ETCO2 values obtained were 11.3 mm Hg and 3.8 mm Hg, respectively. These low values are the result of low cardiac output and cannot be distinguished clinically, nor were they statistically different.26 A meta-analysis study demonstrated capnography sensitivity and specificity to be 93% and 97%, respectively.5 In our study the algorithm differentiated esophagus from airway images (carina and trachea together) with 98% sensitivity and 99% specificity. It should be noted that the results of the meta-analysis were derived from emergency intubation trials, understandably difficult clinical scenarios, yet it gives an objective value that puts our results in perspective.

There are other methods to confirm ETT proper position, but they are even less accurate than is capnography: auscultation uses the sound of airflow. One assumes that if the stethoscope diaphragm is positioned over the chest, the sound of airflow will originate from the airways. However, airflow sound from the stomach can be transmitted to the chest, yielding a false-negative finding.4,27 Other methods have been suggested over the years to verify ETT location; none of them was widely adopted.26,2831 For example, Kelly et al. studied in canines the reliability of water condensation seen on the inner walls of an ETT upon ventilation. They found that condensation was seen in all tubes located in the trachea; however, it was also seen in 83% of tubes placed in the esophagus.32

The aforementioned ETT placement confirmation methods (fogging, auscultation, capnography, etc.) suffer significant false positives and negatives because they rely upon interpretation of secondary physical products of ventilation. Bronchoscopy through the ETT is more sensitive because it uses the primary characteristic of ETT position, visualization of anatomical image of tracheal rings. Theoretically, we could have confirmed each ETT position using a bronchoscope. However, bronchoscopes are large, fragile, and expensive. Economically speaking, we cannot equip every remote anesthesia location, disaster field hospital, military forward surgical team, and advance life support ambulance with a bronchoscope. Our proposed system mimics the bronchoscope by acquiring a direct video image, with subsequent analysis and interpretation.

The system described herein, as we see it in the future, should not be cumbersome: a video camera is mounted on the tip of an intubating stylet and connected to a small processor (personal digital assistant size). The stylet will be used inside an elastic condom (like vaginal ultrasound) and is therefore for multiple uses. While introducing the ETT, video images are acquired and transferred to the processor while the algorithm analyzes the ETT's location. Upon identification of an esophageal image, an audio alarm sounds, alerting the operator to an esophageal intubation location, allowing the operator to maintain visualization of the anatomy at all times. A different and unique sound is chimed once a carinal image replaces a tracheal image, suggesting proper ETT positioning. The video camera lens possesses a focal distance, and therefore the distance from the carina at which the image will change is about 2 cm. For noisy environments (such as in a helicopter), light signals might replace audio signals.

The video camera will be a part of an intubating stylet and will allow ETT shape manipulation. It allows repeated confirmations of ETT location without stopping chest compressions as requested by the 2010 American Heart Association cardiopulmonary resuscitation algorithm.1 It enables quick verification after and during patient transport. If needed, the system can be modified for continuous ETT position monitoring by connecting the processor to ETView® (Misgav, Israel). This product is a US Food and Drug Administration–approved ETT that has a video camera embedded at its end. If the medical team finds it appropriate, the ETT can be replaced for continuous monitoring.

The end product will not contain a monitor, although some anesthesiologists presented with the concept stated that they would rather trust their eyesight. There are 3 reasons for the lack of a monitor: first, avoiding the need to disconnect the user's continuous visualization of the upper airway; second, eliminating the need for bronchoscopy operator training, because it is not meant to be a mini-bronchoscope; and finally, cost of the end product. Another concern of many was the use of a “black box” in the clinical setting, i.e., a monitor without a known method of operation. The algorithm used here is part of a family of “image pattern analysis systems” also commonly known as “biometric recognition systems.” This technology is widely used by security agencies to identify potential terrorists, identifying people at border controls as well as limiting access to restricted areas where it was proved to carry 99.9% accuracy.12 We therefore believe that it is a safe method.

We “taught” the algorithm to differentiate tracheal from carinal image (in addition to esophageal intubation exclusion). In 971 (of 1016) airway images the algorithms correctly differentiated trachea from carinal image, i.e., 96% accuracy (0.96 sensitivity, 0.95 specificity). The system will identify correct intratracheal positioning by finding the first transition from 1 lumen (trachea) to 2 lumens (carina). This potential ability to prevent endobronchial intubation is an additional merit of the system, not present in any commercial monitor. Intratracheal positioning is a secondary benefit, not the main purpose of the system.

This study suffers a few limitations mandating further investigation. The study design was aimed at simulating resuscitation, a clinical scenario with a high rate of intubation failures, done in austere locations and under tremendous stress.8,9 In fact 10% of ETT misplacements were in cardiac arrest scenarios.5 The use of fresh bovine specimens was therefore chosen as a model for unperfused biological tissue, mandating repeating the study in a human model. Substantial additional research is needed involving the performance of the algorithm under inferior visual conditions (e.g., pulmonary edema, pulmonary bleeding, copious bronchial secretions). This comparison will be a challenge because of the difficulty performing such a prospective study. It will, however, be very interesting because it is these clinical scenarios in which capnography might fail secondary to sampling line obstruction and auscultation is limited because alveoli are filled with fluid. Our statistical evaluation tool (leave-one-case-out method) is a common acceptable method to evaluate performance of supervised systems. However, it is known to suffer from a tendency to have optimistic results on laboratory studies. A larger clinical study is therefore desirable.

In summary, a potential novel tube position verification system incorporating visual pattern recognition algorithms was assessed. High accuracy of the analysis algorithm was shown using nonperfused biological tissue, justifying further research.

Back to Top | Article Outline

DISCLOSURES

Name: Micha Y. Shamir, MD.

Contribution: This author helped design the study, analyze the data, and write the manuscript. The author approved the last manuscript.

Conflicts of Interest: This author has no conflict of interest to report.

Name: Dror Lederman, PhD, EMT-P.

Conflicts of Interest: Dr. Lederman is the inventor of the system presented in this paper and founder and owner of Tube-Eye Medical Ltd., which aims to commercialize the invention.

Contribution: This author helped design the study, conduct the study, analyze the data, and write the manuscript. The author approved the last manuscript.

Name: Dietrich Gravenstein, MD.

Contribution: This author helped analyze the data and write the manuscript. The author approved the last manuscript.

Conflicts of Interest: This author has no conflict of interest to report.

This manuscript was handled by: Dwayne R. Westenskow, PhD.

Back to Top | Article Outline

REFERENCES

1. Neumar RW, Otto CW, Link MS, Kronick SL, Shuster M, Callaway CW, Kudenchuk PJ, Ornato JP, McNally B, Silvers SM, Passman RS, White RD, Hess EP, Tang W, Davis D, Sinz E, Morrison LJ. Adult advanced cardiovascular life support: 2010 American Heart Assosciation guidelines for cardiopulmonary resuscitation and emergency cardiovascular care. Circulation 2010;122(Suppl 3):5729–67
2. Rudraraju P, Eisen LA. Confirmation of endotracheal tube position: a narrative review. J Intens Care Med 2009;24:283–92
3. Bhende SM, Thompson AE. Evaluation of an end-tidal CO2 detector during pediatric cardiopulmonary resuscitation. Pediatrics 1995;95:395–9
4. Greenberg SB, Murphy GS, Vender JS. Standard monitoring technique. In Barash PG, Cullen BF, Stoelting RK, Cahalan M, Stock MC, eds. Clinical Anesthesia. 6th ed. Philadelphia: Lippincott Williams & Wilkins, 2011:699
5. Li J. Capnography alone is imperfect for endotracheal tube placement confirmation during emergency intubation. J Emerg Med 2001;20:223–9
6. Bjoraker DG. Airway management in the operating room. In: Gravenstein JS, Jaffe MB, Paulus DA, eds. Capnography, Clinical Aspects. Edinburgh: Cambridge University Press, 2004:40–1
7. Tovar E, Abshier S, Borsari A, Patel R. Mainstem bronchial intubation: an unrecognized source of hypotension during coronary artery bypass surgery. J Cardiothor Vasc Anesth 1997;11:615–8
8. Griesdale DE, Bosma TL, Kurth T, Isac G, Chittock DR. Complications of endotracehal intubation in the critically ill. Intens Care Med 2008;34:1835–42
9. Sakles JC, Laurin EG, Rantapaa A, Panacek EA. Airway management in the emergency department: a one-year study of 610 tracheal intubations. Ann Emerg Med 1998;31:325–32
10. Jaber S, Amraoui J, Lefrant JY, Arich C, Cohendy R, Landreau L, Calvet Y, Capdevila X, Mahamat A, Eledjam JJ. Clinical practice and risk factors for immediate complications of endotracheal intubation in the intensive care unit: a prospective, multiple-center study. Crit Care Med 2006;34:2355–61
11. Benedetto WJ, Hess DR, Gettings E, Bigatello LM, Toon H, Hurford WE, Schmidt U. Urgent tracheal intubation in general hospital units: an observational study. J Clin Anesth 2007:1920–4
12. Jain K. Biometric recognition. Nature 2007;449:38–40
13. Wei G. Recent advances in mathematical methods for the analysis of biomedical image. Int J Biomed Imaging 2006;Article 70578:1
14. Wang X, Lederman D, Tan J, Wang XH, Zheng B. Computerized detection of breast tissue asymmetry depicted on bilateral mammograms: a preliminary study of breast risk stratification. Acad Radiol 2010;17:1234–41
15. Abdel-Qader I, Shen L, Jacobs L, Abu Amara F, Pashaie-Rad S. Unsupervised detection of suspicious tissue using data modeling and PCA. Int J Biomed Imaging 2006;Article 57850:1
16. Bitter I, Van Uitert R, Wolf I, Tzatha E, Gharib AM, Summers R, Meinzer HP, Pettigrew R. Virtual contrast for coronary vessel based on level set generated subvoxel accurate centerlines. Int J Biomed Imaging 2006;Article 94025:1–8
17. Lederman D, Shamir MY. Novel automatic endotracheal position confirmation system: Mannequin model algorithm evaluation. J Clin Monit Comp 2010;24:335–40
18. Rose RC, Reynolds DA. Text-independent speaker identification using automatic acoustic segmentation. IEEE Int Conf Acoust Speech Signal Proc 1990;1:293–6
19. Li Q, Doi K. Reduction of bias and variance for evaluation of computer-aided diagnostic schemes. Med Phys 2006;33:868–75
20. Brunel W, Coleman DL, Schwartz DE, Peper E, Cohen NH. Assessment of routine chest roentograms and the physical examination to confirm endotracheal tube position. Chest 1989;96:984–6
21. Bair AE, Smith D, Lichty L. Intubation confirmation techniques associated with unrecognized non-tracheal intubations by pre-hospital providers. J Emerg Med 2005;28:403–7
22. Birmingham PK, Cheney FW, Ward RJ. Esophageal intubation: a review of detecting techniques. Anesth Analg 1986;65:886–91
23. Salem MR, Wafai Y. Practical confirmation of endotracheal intubation in the trauma patient. Anesthesiol News 1997; 4–32
24. Grmec Š. Comparison of three different methods to confirm tracheal tube placement in emergency intubation. Intensive Care Med 2002;28:701–4
25. Keller WR, Biehler J, Linares MY, Garcia-Pena BM. False-positive colorimetric capnometry after ingestion of carbonated beverages. Pediatr Emerg Care 2009;25:69–73
26. Kramer-Johansen J, Eilevstjønn J, Olasveengen TM, Tomlinson AE, Dorph E, Steen PA. Transthoracic impedance changes as a tool to detect malpositioned tracheal tubes. Resuscitation 2008;76:11–6
27. Andersen KH, Hald A. Assessing the position of the tracheal tube. The reliability of different methods. Anaesthesia 1989;44:984–5
28. Wee M. The esophageal detector device. Anaesthesia 1991;46:1086–7
29. Tanigawa K, Takeda T, Goto E, Tanaka K. The efficacy of esophageal detector devices in verifying tracheal tube placement: a randomized cross-over study of out-of-hospital cardiac arrest patients. Anesth Analg 2001;92:375–8
30. Raphael DT. Acoustic reflectometry profiles of endotracheal and esophageal intubation. Anesthesiology 2000;92:1293–9
31. Tejman-Yarden S, Lederman D, Weksler N, Gurman GM. Acoustic monitoring of double lumen ventilated lungs for the detection of selective unilateral lung ventilation. Anesth Analg 2006;103:1489–14
32. Kelly JJ, Eynon CA, Kaplan JL, de Garavilla L, Dalsey WC. Use of tube condensation as an indicator of endotracheal tube placement. Ann Emerg Med 1998;31:575–8
© 2011 International Anesthesia Research Society