Neural Network Detection of Esophageal Intubation in Humans
Mauricio A. Leon, MD and Jukka Resnen, MD
Department of Anesthesiology, University of South Florida
Introduction: An experimental study1 demonstrated that a neural network was capable of identifying a tracheal or esophageal airway placement using gas flow and pressure waveforms. This study was designed to test whether esophageal intubation could be accurately detected in anesthetized patients.
Methods: After Institutional Review Board approval and written informed consent, patients scheduled to receive general anesthesia for elective operations, not involving the gastrointestinal system, were included in the study. Pneumotachographs were connected to the inspiratory and expiratory limbs of a standard anesthesia breathing circuit. Differential and absolute pressures from each pneumotachograph were transduced, demodulated into analog voltages, and sampled with an A/D converter installed in a 486 AT compatible computer. After induction of anesthesia and tracheal intubation, a tube, similar to the one placed in the trachea, was placed in the esophagus. After assuring stable circulatory function, ventilation and oxygenation, the inspired tidal volume was lowered to 5 ml/kg, and respiratory rate was changed to 15 bpm. Thereafter, gas flow and pressure resulting from ventilation of the lungs were recorded for 15 seconds. Ventilation of the lungs was then discontinued, and the breathing circuit was attached to the esophageal tube without changing the ventilator settings. A 15-second recording of circuit flow and pressure resulting from ventilation of the esophagus was made.
Subsequently, ventilation of the lungs was resumed with the original ventilator settings, the stomach was aspirated, and the esophageal tube was withdrawn. During off-line analysis, individual breaths were separated, and the number of data points was reduced by window-averaging. Single breaths were labeled with a score of one (1), if they resulted from tracheal ventilation, or minus one (-1), if they were esophageal in origin. A neural network simulator was defined to identify the ventilation target using flow and pressure data points as input.
Results: Recordings from 46 patients were analyzed. Data from 25 patients, 100 tracheal and 94 esophageal waveforms, were used for training, and data from the remaining 21 subjects were used for testing the neural network. The neural network identified correctly all tracheal breaths (n=84) with scores ranging from 0.71 to 1.01 (0.99 0.05; mean SD). The esophageal waveforms (n=76) were identified with equal accuracy in a range of -0.77 to -1.02 (-0.99 0.03) (Figure).
Conclusion: Although inadvertent esophageal intubation is a rare complication in anesthesia and critical care, it usually has devastating consequences to both the patient and to those responsible for airway management. Several methods and devices to detect esophageal intubation have been proposed but they all require specific action or interpretation on part of the provider or do not work in all clinical circumstances. The results of this study show that direct recognition of the mechanical characteristics of the ventilated structure allows accurate, reliable differentiation between tracheal and esophageal intubation in anesthetized patients. Whether a neural network could be configured to continuously monitor airway pressure and flow patterns on-line and alert for misplaced tracheal tube, requires further study.
Anesth Analg 78:548-53, 1994