Critical element prediction of tracheal intubation difficulty: Automatic Mallampati classification by jointly using handcrafted and attention-based deep features

• Published in: Computers in biology and medicine 2022-11, Vol.150, p.106182, Article 106182
• Main Authors: Zhang, Fan, Xu, Yuelei, Zhou, Zhaoyun, Zhang, Han, Yang, Ke
• Format: Article
• Language: English
• Subjects: Algorithms; Anesthesia; Artificial Intelligence; Artificial neural networks; Attention mechanism; Attention task; Classification; Clinical medicine; Complications; Datasets; Deep convolutional neural network (DCNN); Deep features; Deep learning; Feature extraction; Handcrafted features; Image classification; Intubation; Intubation, Intratracheal - methods; Laryngoscopy - methods; Machine learning; Mallampati classification; Methods; Mouth; Neural networks; Semantics; Support vector machines
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2022.106182

Preoperative assessment of the difficulty of tracheal intubation is of great importance in anesthesia practice because failed intubation can lead to severe complications and even death. The Mallampati score is widely used as a critical assessment criterion in combination with other measures to assess the difficulty of tracheal intubation. The performance of existing methods for Mallampati classification with artificial intelligence (AI) is unreliable to the extent that the current clinical judgment of the Mallampati score relies entirely on doctors' experience. In this paper, we propose a new method for automatic Mallampati classification. Our method extracts deep features that are more favorable for the Mallampati classification task by introducing an attention mechanism into the basic deep convolutional neural network (DCNN) and then further improves the classification performance by jointly using attention-based deep features with handcrafted features. We conducted experiments on a dataset consisting of 321 oral images collected online. The proposed method has a classification accuracy of 97.50%, a sensitivity of 96.52%, a specificity of 98.05%, and an F1 score of 96.52% after five-fold cross-validation. The experimental results show that our proposed method is superior to other methods, can assist doctors in determining Mallampati class objectively and accurately, and provide an essential reference element for assessing the difficulty of tracheal intubation. Our method extracts deep features that are more favorable for the Mallampati classification task by introducing the coordinate attention (CA) module into the basic deep convolutional neural network, and then further improves Mallampati classification performance by jointly using attention-based deep features with handcrafted features. [Display omitted] •Introducing an attention mechanism to DCNN to extract attention-based deep features.•Training DCNN using a transfer learning approach when extracting deep features.•Extracting two handcrafted features, namely HOG and LBP.•Joint handcrafted and attention-based deep features for the Mallampati classification.•Our proposed method achieves the best results on the larger dataset we constructed.