Pre-operative prediction of difficult tracheal intubation using an explainable multi-pose shared encoder vision...

Difficult tracheal intubation remains poorly predicted by conventional bedside airway tests. Facial photographs analysed by deep learning have been proposed as an alternative, and multi-pose acquisition protocols are increasingly advocated; however, whether acquiring more poses actually improves prediction has not been tested.

In this prospective, single-centre observational study, standardised pre-operative facial photographs were obtained in seven poses from 250 patients undergoing surgery under general anaesthesia with orotracheal intubation. Difficult intubation was defined pragmatically as two or more intubation attempts and/or use of an alternative airway device. A multi-pose shared-encoder Vision Transformer was trained and evaluated under stratified five-fold cross-validation for every one of the 127 possible pose combinations and compared with conventional bedside tests. To quantify post-selection optimism, the pose-selection step was itself nested within the resampling: it was repeated using the training folds only and evaluated in the withheld fold. The calibration of the preferred configuration was assessed under a separate held-out scheme.

Difficult intubation occurred in 70 of 250 patients (28.0%). Conventional bedside tests discriminated poorly: the highest area under the receiver operating characteristic curve (AUC) was 0.66 (95% CI 0.58 to 0.73) for a neck circumference ≥ 400 mm, and the confidence interval of every other test included 0.50. Across all 127 combinations, discrimination was unrelated to the number of poses acquired (Spearman's rho = 0.09, P = 0.34): the seven-pose model (mean AUC 0.666) performed no better than the average single-pose model (0.660). Inter-fold variability fell markedly once more than one pose was acquired (median SD of AUC 0.170 for single-pose versus 0.098 for multi-pose combinations; P < 0.001) and did not fall further thereafter. After this pose-number analysis, upper lip bite plus lateral neutral position was selected as a pragmatic two-pose candidate on the basis of discrimination, stability, anatomical complementarity, and acquisition burden. It ranked first among the 21 two-pose combinations and tied third overall under the four-criterion ranking, achieving a mean AUC of 0.688 (95% CI 0.594 to 0.781), with pooled out-of-fold sensitivity 67.1% and specificity 71.1%; its confidence interval overlapped that of neck circumference.

Acquiring additional facial poses did not improve the prediction of difficult tracheal intubation. A second, anatomically complementary pose stabilised an otherwise erratic estimate, but further poses produced no observed gain in discrimination or stability. The internally cross-validated performance of the pragmatic two-pose candidate was modest and was not demonstrably superior to neck circumference. Nesting the pose-selection step within the training folds lowered discrimination from 0.688 to 0.614, an optimism of 0.074 AUC points, and predicted probabilities carried little information beyond cohort prevalence; the naive estimate is therefore optimistic, and no external validation has been performed.

ClinicalTrials.gov NCT06961240 (registered 29 April 2025).