ABSTRACT
PURPOSE: To demonstrate the utility of machine learning models for predicting difficult airways using clinical and ultrasound parameters. METHODS: This is a prospective non-consecutive cohort of patients undergoing elective surgery. We collected as predictor variables age, sex, BMI, OSA, Mallampatti, thyromental distance, bite test, cervical circumference, cervical ultrasound measurements, and Cormack-Lehanne class after laryngoscopy. We univariate analyzed the relationship of the predictor variables with the Cormack-Lehanne class to design machine learning models by applying the random forest technique with each predictor variable separately and in combination. We found each design's AUC-ROC, sensitivity, specificity, and positive and negative predictive values. RESULTS: We recruited 400 patients. Cormack-Lehanne patients≥III had higher age, BMI, cervical circumference, Mallampati class membership≥III, and bite test≥II and their ultrasound measurements were significantly higher. Machine learning models based on physical examination obtained better AUC-ROC values than ultrasound measurements but without reaching statistical significance. The combination of physical variables that we call the "Classic Model" achieved the highest AUC-ROC value among all the models [0.75 (0.67-0.83)], this difference being statistically significant compared to the rest of the ultrasound models. CONCLUSIONS: The use of machine learning models for diagnosing VAD is a real possibility, although it is still in a very preliminary stage of development. CLINICAL REGISTRY: ClinicalTrials.gov: NCT04816435.
