A fully-automatic semi-supervised deep learning model for difficult airway assessment.

Background: Difficult airway conditions represent a substantial challenge for clinicians. Predicting such conditions is essential for subsequent treatment planning, but the reported diagnostic accuracies are still quite low. To overcome these challenges, we developed a rapid, non-invasive, cost-effective, and highly-accurate deep-learning approach to identify difficult airway conditions through photographic image analysis. Methods: For each of 1000 patients scheduled for elective surgery under general anesthesia, images were captured from 9 specific and different viewpoints. The collected image set was divided into training and testing subsets in the ratio of 8:2. We used a semi-supervised deep-learning method to train and test an AI model for difficult airway prediction. Results: We trained our semi-supervised deep-learning model using only 30% of the labeled training samples (with the remaining 70% used without labels). We evaluated the model performance using metrics of accuracy, sensitivity, specificity, F1-score, and the area under the ROC curve (AUC). The numerical values of these four metrics were found to be 90.00%, 89.58%, 90.13%, 81.13%, and 0.9435, respectively. For a fully-supervised learning scheme (with 100% of the labeled training samples used for model training), the corresponding values were 90.50%, 91.67%, 90.13%, 82.25%, and 0.9457, respectively. When three professional anesthesiologists conducted comprehensive evaluation, the corresponding results were 91.00%, 91.67%, 90.79%, 83.26%, and 0.9497, respectively. It can be seen that the semi-supervised deep learning model trained by us with only 30% labeled samples can achieve a comparable effect with the fully supervised learning model, but the sample labeling cost is smaller. Our method can achieve a good balance between performance and cost. At the same time, the results of the semi-supervised model trained with only 30% labeled samples were very close to the performance of human experts. Conclusions: To the best of our knowledge, our study is the first one to apply a semi-supervised deep-learning method in order to identify the difficulties of both mask ventilation and intubation. Our AI-based image analysis system can be used as an effective tool to identify patients with difficult airway conditions. Clinical trial registration: ChiCTR2100049879 (URL: http://www.chictr.org.cn).