Machine Learning Revolution in Predicting Difficult Intubation: A Systematic Review

Parisa  Moradimajd; Alireza  Babajani; Fatemeh  Mehdipour; Mahdi  Nazari

doi:10.18502/aacc.v12i2.20958

Parisa Moradimajd Department of Anesthesia, Faculty of Allied Medical Sciences, Iran University of Medical Sciences, Tehran, Iran.
Alireza Babajani Department of Anesthesiology, School of Allied Medical Sciences, Alborz University of Medical Sciences, Karaj, Iran.
Fatemeh Mehdipour Department of Anesthesia, School of Allied Medical Sciences, Iran University of Medical Sciences, Tehran, Iran
Mahdi Nazari Department of Anesthesia, School of Allied Medical Sciences, Iran University of Medical Sciences, Tehran, Iran

DOI: https://doi.org/10.18502/aacc.v12i2.20958

Keywords: Difficult airway; Machine learning; Intubation prediction; Laryngoscopy; Artificial intelligence

Abstract

Background: The presence of a difficult airway (DA) remains a major concern in anesthesia, contributing significantly to patient complications and adverse outcomes. Traditional clinical assessments often fall short in accurately predicting difficult intubation. With the advancement of artificial intelligence, machine learning (ML) has emerged as a promising approach for enhancing airway risk prediction. This systematic review aimed to evaluate current studies that utilize machine learning models for predicting difficult laryngoscopy and intubation and to assess the features, algorithms, and predictive performance of these models.

Methods: Following PRISMA guidelines, a comprehensive search was conducted in seven databases (PubMed, Scopus, Web of Science, Science Direct, Wiley, SID, and Google Scholar) to identify relevant original articles published between 2000 and July 2025. Studies using ML models to predict difficult intubation based on clinical, morphological, or acoustic features were included. A total of nine eligible studies were reviewed.

Results: Various ML algorithms, including KNN, SVM, Random Forest, XGBoost, and decision trees (J48), were applied across studies. Feature inputs ranged from traditional clinical parameters (e.g., Mallampati score, neck circumference) to advanced modalities such as voice analysis and facial image processing. Reported model performance (AUC) ranged from 0.71 to 0.924, indicating generally high predictive accuracy. Models incorporating non-traditional data (e.g., acoustic or imaging features) tended to perform better.

Conclusion: ML-based models show strong potential in improving the prediction of difficult airways and can serve as supportive tools in preoperative assessment. However, standardization of input features, external validation, and enhanced model interpretability are essential for successful clinical implementation.