Towards Artificial Intelligence-enabled Medical Pre-operative Airway Assessment

For surgeries which require general anesthesia, airway management is imperative. Difficult airway, which inhibits proper intubation, can be fatal. As such, pre-operative airway assessments are conducted by clinicians to determine the ease of intubation as well as to identify patients with difficult airway. To improve the process, artificial intelligence (AI) methods can be employed to predict such difficult airway situations so that suitable preparations can be made beforehand. However, due to the need for explainability of AI models required by healthcare regulations, typical black box models which work best with most data-driven AI methods cannot be used. Therefore, in the current work, a machine learning model has been established to predict the specific medical facial landmarks that are currently used by clinicians. These include the eyes, mentum, thyroid notch, suprasternal notch, forehead, tragus and radix. The model is based on convolutional neural network and a practical facial landmark detector concept. Furthermore, k-fold cross-validation sampling and the Adabelief optimizer have been utilized. The model prediction results display accurate prediction of the features, with the testing loss exhibiting good stability and maintaining well below 0.01 throughout. Attributed to that, the current model can lead to meaningful diagnosis of difficult airway during airway assessments.