Prediction of difficult intubation using a neural network

Abstract

Background and goal of study: The prediction of difficult intubation to prevent unexpected results is a major concern of anaesthesiologists. A neural system could be used to develop a model to predict who is likely to be difficult in undergoing laryngeal intubation. The purpose of this study was to assess and to compare the usefulness of a neural network to predict difficult intubation by combining a patient's anatomical features.

Materials and methods: After receiving approval from our ethical subcommittee, 462 ASA I-II patients (242 men, 220 women) aged 16 or older were prospectively studied. Mallampati's method which was modified by Samsoon and Young was carried out [1]. The airway was classified (oropharyngeal classification; OPC). In a supine position, hyo-mental distance (HD) with head in neutral, thyro-mental distance (TD), and interincisors distance (DI) on mouth opening with head fully extended, were all measured. We defined the cases of intubation trial more than 3 times in patients with laryngoscopic grade 3 or 4 as difficult intubation. Multilayer perceptron was used in the neural network. The number of nodes in a layer was selected automatically and the learning algorithm was the conjugate gradient method. Normalization of data was performed. Equalization of data was performed for the skewed class in a target variable.

Results and discussion: Assignment to single signs had relatively high sensitivity and specificity. In various combinations, OPC and DI/HD resulted in increasing sensitivity. The neural network showed improved sensitivity compared to the ordinary single and combined performance of anatomical signs.