NEURAL-NETWORK-BASED DETECTION OF ESOPHAGEAL INTUBATION

To improve the accuracy of early detection of inadvertent esophageal intubation, we designed, trained, and tested a neural network-based computer system to detect the mechanical differences between lung and esophagogastric ventilation. Ten 25 to 30-kg anesthetized swine were sequentially ventilated with tidal volumes of 9, 12, and 15 mL/kg, using tubes placed in the trachea and in the esophagus, while flow and pressure waveforms were collected for 9-10 breaths. Gas remaining in the stomach was aspirated after each period of gastric ventilation. A computer program identified each mechanical inspiration, extracted the first 37 flow and pressure data points from each record, and normalized them to an equal amplitude. A backpropagation single-hidden-layer neural network was trained to identify the origin of flow and pressure waveforms as tracheal or esophageal. Ten different training and testing groups were assembled. In each group, data from nine subjects were used for training and data from the remaining subjects were used for testing. A total of 291 esophageal and 300 tracheal flow and pressure waveforms were analyzed by the network. The network identified esophageal intubation correctly during the first five breaths of all esophageal recordings. In one subject, the network identified the eighth esophageal breath as tracheal and could not identify three breaths. All tracheal intubations were identified correctly. Flow and pressure ''signatures'' of pulmonary and gastric ventilation are easily learned by a neural network. Therefore, neural-network recognition of esophageal intubation from flow and pressure signals is possible, and the development of an on-line detector for tracheal tube misplacement seems feasible.