Machine learning predictive models with voice parameters for prediction of difficult laryngoscopies: a prospective cohort

Abstract

Background: Voice has shown association with difficult laryngoscopies though only with poor predictive performance. As machine learning (ML) algorithms are advocated to improve predictive performance as compared to regular logistic regression models, we aimed to evaluate predictive values of models from both ML algorithms and regular logistic regression including voice parameters for prediction of difficult laryngoscopies. Methods: A prospective cohort with patients undergoing general anesthesia for surgical procedures. We preoperatively collected data on sex, age, weight, height, body mass index and recorded audios from 5 different phonemes: /a/, /e/, /i/, /o/, and /u/. The main outcome was difficult laryngoscopies defined as Cormack and Lehane's classes 3 or 4. Posteriorly, the formant frequencies were extracted from the recorded audios. Five formants were extracted from each phoneme, yielding 25 formants for each patient: 5 formants from each of the 5 phonemes. Uni and multivariable analyses were performed to evaluate association between variables and to build the predictive models. Results: From a total of 453 patients, 29 (6.4%) presented difficult laryngoscopies. Sex (p=0.015), weight (p=0.009), height (p=0.001), and the formants eF2(p<0.001), eF3(p=0.002), iF2(p=0.011), and oF2(0.018) were associated with difficult laryngoscopies. Age (p=0.212) and BMI (p=0.108) did not present significant difference between the 2 groups. Six predictive models were built including the 25 voice formants. The ROC curves for the predictive models are presented in Figure 1 and their AUC were as follows: Logistic Regression model - LR (60.24%; 95% CI: 37.96-82.52%); Linear Discriminant Analysis - LDA (66.44%; 95% CI: 44.11-88.76%); Classification and Regression Trees - CART (50%; 95% CI: 50-50%); k-Nearest Neighbors - kNN (59.4%; 95% CI: 36.94-81.86%); Random Forest - RF (50.3%; 95% CI: 21.36-79.23%); Generalized Boosted Model - GBM (45.87%; 95% CI: 20.73-71%). Conclusions: ML algorithms did not improve airway prediction performed by voice parameters as compared to regular LR model.