Development of machine learning system for airway prediction from facial image with mobile device
Abstract
Goals: A reliable prognostic tool for a difficult airway (DA) may enhance patients’ safety during orotracheal intubation by decreasing
unanticipated DAs. We aim to examine the applicability of an Artificial Intelligence-Deep Learning (AI-DL) algorithm to measure airway’s anatomy, and to predict DA based on published models.
Materials and methods: Observational prospective cohort study with n=503 patients recruited at Galdakao-Usansolo and Basurto University Hospitals (Biscay, Spain) between 2018 and 2020. Two pre-operative photos for each patient were collected: a frontal view, in which patients were instructed to open their mouth completely; and a lateral view, with head in vertical ex tension.
Smartphones with general-purpose cameras were used, and a cue card was added to the scene as reference. Patients’ medical records were logged. After intubation, HAN score and IDS-ASA criteria for intubation difficulty [1] were collected.
Our anaesthesiology team defined a set of relevant orofacial landmarks, whereas our data-science team developed an AI-DL algorithm, trained to identify locate them automatically within the images. In a previous evaluation, the system achieved an accuracy comparable to the consensus of two human annotators [2]. Landmark positions output by the AI-DL method were subsequently used by the system to ex tract two anatomical measurements: thyromental distance and interincisor gap. Finally, these two were integrated into a published model for DA prognosis: Naguib et al. 2006 [3], which also employed patients’ height and Mallampati score.
Results and discussion: The estimated incidence of DA was 6.36% (32 out of 503 patients) according to the IDS-ASA criteria. Naguib’s
model, when used in combination with our automatic AI-DL based measurements, achieved 53.12% sensitivity and 79.83% specificity; compared to clinicians’ subjective assessment, who obtained 25.00% sensitivity and 93.63% specificity.
Conclusion(s): In this work, we evaluated an AI-DL method to predict DA for intubation, with two pre-operative photos and Naguib’s model. Our results complemented expert judgements’ predictive ability in terms of sensitivity, substantially lowering false negatives; at the expense of a restrained loss in specificity (false positives). Thus, our proposal may provide anaesthesiologists with an automatic, objective and accessible decision support tool for the prognosis of DAs.