NEURAL-NETWORK-BASED DETECTION OF ESOPHAGEAL INTUBATION IN ANESTHETIZED PATIENTS

Objective. To test whether a neural network-based method could differe ntiate between tracheal and esophageal intubation in anesthetized pati ents by recognizing breathing circuit pressure and flow waveform patte rns. Methods. Tracheal tubes were placed in the trachea and in the eso phagus of adult patients undergoing elective operations. After ensurin g for proper oxygenation, ventilator settings were changed to 5 ml/kg tidal volume (VT) and 15 cpm and circuit pressure and flow were record ed for 15 seconds. Then, the breathing circuit was switched to the tub e placed in the esophagus, and signals were recorded for an additional Ii-second period. During offline analysis, individual waveforms were separated. Tracheal breaths were labeled with a score of 1 while esoph ageal ''breaths'' were labeled with -1. A neural network was defined t o learn to associate waveforms to their corresponding scores. Data fro m 54% of the patients were used to train the neural network. Data from the remaining subjects were used for testing. Results. Forty-six pati ents were studied. Neural network training was achieved with 100 trach eal and 94 esophageal waveforms from 25 patients. Neural network perfo rmance was tested on 84 tracheal and 76 esophageal waveforms from 21 s ubjects. The neural network assigned scores of 0.99 +/- 0.05 (mean +/- SD) to tracheal waveforms and -0.99 +/- 0.03 to esophageal waveforms. The difference between mean esophageal and tracheal scores was -1.99 with a 99.999% confidence range of -2.01 to -1.96. Any arbitrary cutof f threshold, ranging between -0.76 and 0.7, separated tracheal and eso phageal score regions, yielding no false positive or negative results. Conclusion. A neural network differentiated consistently tracheal fro m esophageal intubation when the ventilation test-mode was used. The v entilation mode employed is feasible in most adult patients undergoing elective procedures under general anesthesia. Further research is req uired to train neural networks to recognize esophageal intubation in d ifferent age groups and when different ventilation modes are applied.