Abstract
Background
Deep learning integrated with ultrasound systems may assist in predicting difficult airway, a life-threatening complication in anesthesia. The study aimed to assess the feasibility of AI-based ultrasound for predicting difficult video laryngoscopy (DVL); and identify ultrasonographic measurements associated with DVL using AI-generated heatmaps.
Methods
Patients who underwent ultrasonographic airway examinations were enrolled. DVL were classified as Cormack-Lehane grade 3 or 4. The methodology involved two steps: In step 1, ResNet-18 extracted features from ultrasound scans, which were then used to develop predictive models with various machine learning algorithms. The area under the receiver operating characteristic curves (AUROC) were used to assess the performance of models. In step 2, heatmaps identified critical regions for DVL, and multivariable logistic regression was used to assess the association between ultrasonographic measurements and DVL.
Results
1,474 patients were included, of whom 95 had DVL. In step 1, the combined model (including section of the hyoid bone, sagittal section of the tongue and thyromental distance) with LightGBM algorithm showed an AUROC value of 0.804, with a sensitivity of 0.889, a specificity of 0.657 in the test set. In step 2, distance from the skin to the end of the tongue root (DSTR) was significantly associated with DVL [odds ratios (OR): 4.83, 95% CI: 2.99–8.01]. Restricted cubic spline revealed a nonlinear association between DSTR and DVL, with a significant rise in the OR for DVL, particularly at DSTR ≥ 3.89 cm.
Conclusion
AI-based ultrasound technology can predict DVL, and DSTR is associated with DVL.
Trial registration
This study was registered at ClinicalTrials.gov (NCT05207254).
Supplementary Information
The online version contains supplementary material available at 10.1186/s12871-026-03948-z.
Keywords: Artificial Intelligence; Deep learning; Laryngoscopy; Intubation, Intratracheal; Ultrasonography
Introduction
Airway management is an essential component of safe anesthesia [1]. While direct laryngoscopy remains the standard, video laryngoscopy is increasingly used [2], reducing difficult intubations [3]. Recent guidelines have increasingly recommended video laryngoscopy as a routine or first‑line technique for tracheal intubation in many clinical settings [4, 5]. Unlike direct laryngoscopy, video laryngoscopy introduces distinct procedural challenges, particularly related to tracheal tube advancement, which may persist despite an adequate glottic view. Therefore, predictive models derived from direct laryngoscopy may not be directly applicable to video laryngoscopy, and there is a growing need for video laryngoscopy‑specific predictors and assessment strategies [4].
However, the continued use of awake intubation techniques [6] suggests video laryngoscopy is not a solution for all airway challenges [7]. In addition, the independent risk factors in direct and video laryngoscopy are different [8], making direct laryngoscopy-based assessment tools unsuitable for video laryngoscopy.
Ultrasound is a valuable adjunct for predicting difficult airways, offering both qualitative and quantitative insights [9, 10], with advantages such as accessibility, affordability, and absence of radiation risk [9]. A recent multicenter pilot study examined the use of point‑of‑care ultrasound in this setting and reported promising results [10]. However, there is currently no standardized set of ultrasonographic criteria for predicting difficult airways. Advancements in technology, notably within artificial intelligence (AI) and deep learning [11], could potentially improve the application of ultrasound in predicting difficult airways.
This study aimed to assess the feasibility of utilizing AI-based ultrasound techniques for predicting difficult video laryngoscopy (DVL), and to identify ultrasonographic measurements that can predict DVL based on AI-generated heatmaps.
Methods
Study design
This study was approved by the Ethics Committee of Shanghai Ninth People’s Hospital (SH9H-2021-T356-2) and registered at ClinicalTrials.gov (NCT05207254). All participants provided written informed consent.
Inclusion and exclusion criteria
The study was conducted from January 27th, 2022, to November 23rd, 2023, in the Shanghai Ninth People’s Hospital. Individuals were eligible for inclusion if they were adult patients (≥ 18 years of age) with an American Society of Anesthesiologists Physical Status (ASA-PS) classification of I to III, scheduled for elective surgery requiring tracheal intubation using video laryngoscopy. The exclusion criteria were language communication problems, cervical spine fracture or diseases, and allergy to relevant medications.
Primary endpoint
The primary endpoint was difficult video laryngoscopy (DVL). DVL was classified as a Cormack-Lehane grade 3–4, whereas non-difficult video laryngoscopy (non-DVL) was classified as grade 1–2 [12]. During tracheal intubation, the Cormack–Lehane grade was evaluated (grade 1 represents a complete view of the glottis, grade 2 represents a partial view of the glottis or arytenoids, grade 3 represents only the visibility of the epiglottis, and grade 4 represents the absence of epiglottis visibility). Although techniques like external laryngeal manipulation were used to optimize the Cormack-Lehane grade when the initial view was poor, only the initial grade was used as the primary endpoint. While we acknowledge that videolaryngoscopic difficulty may also arise from difficult tube delivery despite an adequate view (the ‘can see, can’t intubate’ scenario) [10], the Cormack‑Lehane grade was chosen as the primary endpoint because it is a standardized, reproducible measure widely used in both clinical practice and previous studies.
Pre-operative care protocol
The baseline characteristics and medical history of each patient (including age, previous anesthesia, complications) were recorded before surgery. A specific airway assessment team performed pre-operative classical bedside examination [13, 14], which included neck mobility, mandibular protrusion, inter-incisor gap, thyromental distance, modified Mallampati test, neck circumference, upper lip bite test, mandibular length, sternomental distance, hyomental distance [13, 15], incisor length [16], thyroid and hyoid distance [17], length of tongue [18], and jaw depth [19]. The definition and classification basis of the examinations are shown in Table S1.
Two experienced researchers conducted all ultrasound examinations, each examining one patient. Both had specialized airway ultrasound training and had demonstrated proficiency in identifying airway structures in over 50 patients before the study. Two ultrasound devices were used (SonoSite SII; SonoSite Corporation, Bothell, WA, USA; Mindray M9; Shenzhen Mindray Bio-Medical Electronics Co, Shenzhen, China), with each patient examined using only one device. In addition, the ultrasound operators were blinded to clinical risk factors. Patients were instructed to keep their mouths closed, ensuring uniform tongue positioning with the tongue tip lightly touching the incisors [20]. All ultrasound images were obtained based on previous studies [17, 20–23], including: (I) sagittal section of the tongue (ST, Fig. 1-a), (II) axial section of the cricothyroid membrane (SCM, Fig. 1-b), (III) axial section of the vocal cord (SVC, Fig. 1-c), (IV) sagittal section of the airway (SA, Fig. 1-d), (V) axial section of the thyroid isthmus (STI, Fig. 1-e), (VI) axial section of the jugular notch (SJN, Fig. 1-f), (VII) section of the hyoid bone (SHB, Fig. 1-g), and (VIII) section of the epiglottis (SE, Fig. 1-h). The methods of obtaining ultrasound images adhered to an indication, acquisition, interpretation, and medical decision making (I-AIM) framework [24], which is shown in Table S2 and Figure S1.
Fig. 1.
Methodological diagram of the study; all ultrasound images in this study, including (I) sagittal section of the tongue (ST, Fig. 1-a), (II) axial section of the cricothyroid membrane (SCM, Fig. 1-b), (III) axial section of the vocal cord (SVC, Fig. 1-c), (IV) sagittal section of the airway (SA, Fig. 1-d), (V) axial section of the thyroid isthmus (STI, Fig. 1-e), (VI) axial section of the jugular notch (SJN, Fig. 1-f), (VII) section of the hyoid bone (SHB, Fig. 1-g), and (VIII) section of the epiglottis (SE, Fig. 1-h)
Intra-operative care protocol
Monitoring consisted of continuous electrocardiography and pulse oximetry, and intermittent non-invasive blood pressure measurement (CARESCAPE B650 Monitor, General Electric HealthCare, USA). General anesthesia was induced with intravenous midazolam (2–3 mg), fentanyl (2–4 µg/kg), propofol (2-2.5 mg/kg), and rocuronium (0.6 mg/kg). Tracheal intubation was performed using a video laryngoscope (McGrath MAC; Aircraft Medical Co., Ltd., Edinburgh, UK) with a Macintosh size 3 blade by anesthesiologists with over 3 years of experience in airway management and more than 3,000 cases of tracheal intubation using video laryngoscopy. The video laryngoscope was chosen according to guidelines [4, 5, 25], as it is the preferred device for routine tracheal intubation in our center.
Pre-processing
To address the heterogeneity caused by two ultrasound devices, ultrasound images were automatically cropped to the same proportions and resolution. Images acquired with the high-frequency probe were cropped to 3.3⋅3.3 cm (Fig. 2A), while images acquired with the low-frequency probe were cropped to 8⋅8 cm (Fig. 2B). During cropping, the upper boundary of the crop was located at the upper boundary of the ultrasound image and the cropped image was in the middle of the ultrasound image. Finally, the images were padded and resized to 224⋅224 pixels.
Fig. 2.
A Pre-processing of the ultrasound images acquired by the high frequency probe; B Pre-processing of the ultrasound images acquired by the low frequency probe; C ResNet-18; D The process of training, validation, and testing; E.The ultrasound measurements for causal analysis in the sagittal section of the tongue (ST); DSTR: distance from the skin to the end of the tongue root; DST: distance from the skin to the tongue surface; DGT: distance from the genioglossus to the tongue surface
Step 1: AI-based ultrasound techniques for predicting DVL
The dataset was divided using stratified sampling with an approximate ratio of 6:2:2 (training/validation/test). The deep learning process was implemented using PyTorch (Meta Platforms Inc., Menlo Park, CA, USA) on a workstation equipped with four NVIDIA GeForce RTX 2080Ti graphics processing units (Nvidia, Santa Clara, CA, USA), 128 GB of memory, and an Intel Xeon Gold 5218 central processing unit (Intel, Santa Clara, CA, USA). We employed a ResNet-18 deep convolutional network (Fig. 2C) for classifying DVL, as it is known for its efficient architecture utilizing residual connections [26]. Our approach utilized a ResNet18 backbone pre-trained on ImageNet-1k. We modified the architecture by replacing the final classification layer with a binary output layer. The entire network was then fine-tuned, with all layers (including both the pre-trained features and new classification head) being updated during training. To mitigate class imbalance, we oversampled the minority class (DVL) by randomly duplicating samples until a 1:1 ratio with the non-DVL class was achieved. All oversampling procedures were confined to the training set to avoid data leakage; performance metrics reported are based on the original (imbalanced) test set, reflecting real-world generalization. Additionally, we applied data augmentation techniques, including random cropping, rotation (− 10° to 10°), and horizontal flipping, to further diversify the dataset. To enhance model performance on the imbalanced data, we employed focal loss (Eq. 1) [27]. Here, α and γ are hyperparameters, and p represents the predicted probability for DVL:
 |
1 |
We evaluated three values for γ (0.8, 1, and 2) and found that γ = 2 yielded the best results. The model was trained with a batch size of 64 using the AdamW optimizer and a warm-up cosine annealing learning rate schedule [28]. The warm-up phase provided stability at the start, while the cosine annealing ensured smooth and efficient convergence as training progressed. The initial learning rate was set to 2.5⋅10− 7 with a maximum learning rate of 1⋅10− 3. The hyperparameter search space for the maximum learning rate included values of 0.001, 0.0008, 0.0007, 0.0006, 0.0005, and 0.0001. The batch size search space included values of 16, 32, and 64. We tested the number of epochs for the warm-up phase with values of 5, 10, 15, and 20. After the warm-up, the total training continued for up to 50 epochs, ensuring comprehensive model training. The AdamW optimizer was chosen for optimization. All feature selection, threshold optimization, and model tuning were performed strictly within the training and validation datasets, with no access to the test set.
To mitigate overfitting, we incorporated L1 regularization into the model. The optimal regularization strength (λ) was determined through 5-fold cross-validation. We selected λ.1se to achieve a more parsimonious model while maintaining competitive performance. We implemented and compared several machine learning algorithms for predictive modeling: Logistic Regression, Random Forest, Support Vector Machine (SVM), XGBoost, Linear Discriminant Analysis (LDA), and LightGBM. Each algorithm underwent comprehensive hyperparameter optimization (Table S3). Model selection was based on the highest Area Under the Receiver Operating Characteristic curve (AUROC) performance. The complete workflow is visually summarized in Fig. 2D.
Step 2: identify ultrasonographic measurement for predicting DVL
Critical regions with high weighting for DVL were visualized using the class activation map (CAM) which was plotted by Grad-CAM + + with critical regions highlighted in red and yellow. Multivariable logistic regression was then used to assess the association between ultrasonographic measurements in identified high-weighted regions and DVL.
The following ultrasonographic measurements were measured: (I) distance from the skin to the end of the tongue root (DSTR), (II) distance from the skin to the tongue surface (DST), and (III) distance from the genioglossus to the tongue surface (DGT). DSTR is an ultrasonographic measurement of a critical region in the sagittal section of the tongue. DST and DGT are ultrasonographic measurements recommended in a previous study [20]. The ultrasonographic measurements are shown in Fig. 2E.
To calibrate the impact of covariates, three distinct models were utilized. Model 1 included only age, sex, and ASA-PS as covariates. Model 2 expanded on model 1, by integrating BMI, neck mobility, mandibular protrusion, inter-incisor gap, thyromental distance, and modified Mallampati test, which are key indicators in the El-Ganzouri risk index [29]. Model 3 further expanded on model 2, by integrating neck circumference, snoring history, upper lip bite test, and mandibular length, which are risk factors for difficult airway [13, 15].
We performed sensitivity analyses by transforming the variables into dichotomous and trichotomous variables. For dichotomous variables, optimal cut-off values were determined using the maximum Youden’s Index. Trichotomous variables were categorized into three levels based on data distribution, with tertiles forming classes 1 (smallest 1/3), 2 (middle 1/3), and 3 (largest 1/3). The motivation of study design is shown in Figure S2.
Statistical analysis
We wrote a data analysis and statistical plan and filed with the institutional review board before data were accessed. Continuous variables are presented as mean ± SD or median (IQR), and categorical variables as n (%). Categorical data were compared using chi-squared test, and continuous data using independent t-test, both two-sided. Effect sizes were assessed using standardized mean differences (Cohen’s d). In step 1, missing values were imputed using the missRanger package in R [30]. In step 2, complete case analysis was used for logistic regression. The classification ability of the model was assessed using AUROC. The optimal cutoff value was determined by the maximum Youden’s Index. Variable importance was ranked using IncNodePurity from the Random Forest model [31] and assessed in the LightGBM model. In step 2, multivariable logistic regression and restricted cubic spline (RCS) were used to explore and validate both linear and nonlinear associations between the exposures and the outcome. Odds ratios (OR) were reported, and RCS plots illustrated nonlinear trends. Post‑hoc analyses indicated that the current sample size exceeded the minimum requirements for model development (Step 1) [32] and for detecting an odds ratio of 1.5 with over 80% power using G*Power (Step 2) [33]. P < 0.05 was considered statistically significant. R version 4.2.1 (https://www.r-project.org/) was used for analysis.
Results
1,706 patients were eligible, but 232 were excluded due to postponed surgeries (113), direct laryngoscopy (51), poor-quality ultrasound images (50), or missing Cormack-Lehane data (18). The final cohort was 1,474 patients, of which 95 (6.45%) had DVL (Fig. 3). Patients with DVL were older, had higher BMI, ASA status, and neck circumference, and were more likely to smoke, snore, have a history of head and neck radiotherapy, multiple intubation attempts, difficult mask ventilation, and device changes during intubation (Table 1). There was no difference in the distribution of DVL and non-DVL patients between different anesthesiologists, assessors and ultrasound devices (Table S4).
Fig. 3.
The flow chart of the study; DVL: difficult video laryngoscopy
Table 1.
Baseline characteristics of included patients
| Number | All | Non-DVL (1379) |
DVL (95) |
P | SMD | |
|---|---|---|---|---|---|---|
| Age (year, SD) | 1474 | 33.55 (14.45) | 32.79 (13.95) | 44.68 (16.94) | < 0.001 | 0.766 |
| Sex (male, %) | 1474 | 565 (38.3) | 524 (38.0) | 41 (43.2) | 0.373 | 0.105 |
| BMI (kg/m2, SD) | 1469 | 21.70 (3.57) | 21.58 (3.52) | 23.36 (3.81) | < 0.001 | 0.485 |
| Non-ASA I (%) | 1473 | 136 (9.3) | 118 (8.7) | 18 (19.4) | 0.001 | 0.325 |
| Alcohol consumption (%) | 1437 | 288 (19.6) | 268 (19.5) | 20 (21.3) | 0.779 | 0.044 |
| Smoking (%) | 1438 | 225 (15.3) | 201 (14.6) | 24 (25.5) | 0.007 | 0.275 |
| History of snoring (%) | 1455 | 604 (41.5) | 540 (39.7) | 64 (67.4) | < 0.001 | 0.577 |
| History of maxillofacial radiotherapy (%) | 1466 | 21 (1.4) | 14 (1.0) | 7 (7.4) | < 0.001 | 0.321 |
| History of surgery (%) | 1470 | 794 (54.0) | 739 (53.7) | 55 (58.5) | 0.425 | 0.097 |
| History of Mandibular Surgery (%) | 1466 | 89 (6.1) | 85 (6.2) | 4 (4.2) | 0.573 | 0.090 |
| Neck circumference (cm, SD) | 1471 | 34.66 (3.84) | 34.56 (3.78) | 36.01 (4.33) | < 0.001 | 0.357 |
| Inter-incisor gap (cm, SD) | 1449 | 4.27 (0.89) | 4.31 (0.87) | 3.67 (0.94) | < 0.001 | 0.704 |
| Upper incisor length (cm, SD) | 1442 | 0.86 (0.13) | 0.86 (0.13) | 0.85 (0.14) | 0.37 | 0.095 |
| Length of tongue (cm, SD) | 1466 | 4.37 (0.99) | 4.39 (0.99) | 4.00 (1.02) | < 0.001 | 0.392 |
| Thyroid and hyoid distance (cm, SD) | 1470 | 4.36 (1.22) | 4.40 (1.23) | 3.86 (0.96) | < 0.001 | 0.482 |
| Hyomental distance (cm, SD) | 1469 | 4.12 (0.76) | 4.13 (0.75) | 3.85 (0.85) | < 0.001 | 0.350 |
| Thyromental distance (cm, SD) | 1468 | 9.45 (1.39) | 9.52 (1.38) | 8.42 (1.08) | < 0.001 | 0.883 |
| Sternomental distance (cm, SD) | 1471 | 17.92 (2.08) | 18.02 (2.05) | 16.42 (1.96) | < 0.001 | 0.801 |
| Mandible length (cm, SD) | 1471 | 9.13 (0.85) | 9.13 (0.84) | 9.18 (0.97) | 0.590 | 0.054 |
| Jaw depth (cm, SD) | 1471 | 3.64 (0.56) | 3.65 (0.56) | 3.51 (0.59) | 0.015 | 0.254 |
| Cervical spine mobility (%) | 1471 | 0.706 | 0.121 | |||
| 1 | 1461 (99.3) | 1366 (99.3) | 95 (100.0) | |||
| 2 | 8 (0.5) | 8 (0.6) | 0 (0.0) | |||
| 3 | 2 (0.1) | 2 (0.1) | 0 (0.0) | |||
| Head and Neck Scar (%) | 1471 | 10 (0.7) | 10 (0.7) | 0 (0.0) | 0.851 | 0.121 |
| Mandibular protrusion (%) | 1463 | < 0.001 | 0.409 | |||
| 1 | 1322 (90.4) | 1250 (91.3) | 72 (76.6) | |||
| 2 | 114 (7.8) | 96 (7.0) | 18 (19.1) | |||
| 3 | 27 (1.8) | 23 (1.7) | 4 (4.3) | |||
| Upper lip bite test (%) | 1460 | < 0.001 | 0.451 | |||
| 1 | 913 (62.5) | 871 (63.8) | 42 (44.7) | |||
| 2 | 414 (28.4) | 382 (28.0) | 32 (34.0) | |||
| 3 | 133 (9.1) | 113 (8.3) | 20 (21.3) | |||
| Modified Mallampati test (%) | 1468 | < 0.001 | 0.697 | |||
| 1 | 407 (27.7) | 399 (29.0) | 8 (8.6) | |||
| 2 | 278 (18.9) | 268 (19.5) | 10 (10.8) | |||
| 3 | 729 (49.7) | 664 (48.3) | 65 (69.9) | |||
| 4 | 54 (3.7) | 44 (3.2) | 10 (10.8) | |||
| DSTR (cm, SD) | 1467 | 3.24 (0.63) | 3.20 (0.61) | 3.86 (0.65) | < 0.001 | 1.047 |
| DST (cm, SD) | 1377 | 5.60 (0.65) | 5.58 (0.64) | 5.97 (0.66) | < 0.001 | 0.608 |
| DGT (cm, SD) | 1377 | 4.06 (1.23) | 4.05 (1.26) | 4.23 (0.49) | 0.218 | 0.187 |
| > 2 attempts (%) | 1474 | 40 (2.7) | 25 (1.8) | 15 (15.8) | < 0.001 | 0.509 |
| Difficult mask ventilation (%) | 1474 | 90 (6.1) | 75 (5.4) | 15 (15.8) | < 0.001 | 0.341 |
| Introducers during intubation (%) | 1474 | 333 (22.6) | 305 (22.1) | 28 (29.5) | 0.126 | 0.169 |
| Device changes during intubation (%) | 1474 | 8 (0.5) | 0 (0) | 8 (8.4) | < 0.001 | 1.146 |
DVL Difficult video-laryngoscopy, SD Standard deviation, ASA the American Society of Anesthesiologists, SMD Standardized mean differences (Cohen’s d), DSTR Distance from the skin to the end of the tongue root, DST Distance from the skin to the tongue surface, DGT Distance from the genioglossus to the tongue surface; Introducers during intubation: including intubating stylet, bougie and cuff inflation; Difficult mask ventilation was defined as the inability to achieve adequate ventilation with one-handed technique by an unassisted anesthesiologist
Step 1: AI-based ultrasound techniques for predicting DVL
The AUROCs of the 8 ultrasound images ranged from 0.550 to 0.783, and sagittal section of the tongue (ST) showed an AUROC value of 0.783 (95% confidence interval (CI): 0.686–0.881) in the test set. After feature selection, SHB, ST and thyromental distance were finally screened out. Various algorithms were employed (Table 2 and Table S5), and the combined model with LightGBM demonstrated the best performance, achieving an AUROC value of 0.804 (95% CI: 0.713–0.894), with a sensitivity of 0.889 (95% CI: 0.722-1.000), a specificity of 0.657 (95% CI: 0.600, 0.709), and an accuracy of 0.671 (95% CI: 0.615, 0.724). Specificity and accuracy were calculated when sensitivity was 0.9 to provide greater transparency and flexibility for clinical interpretation (Table S6). Sensitivity analysis using class weights [34] showed a performance to our original 1:1 oversampling (Table S7).
Table 2.
The results of predictive performances for all models
| AUROC (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Accuracy (95% CI) | |
|---|---|---|---|---|
| Single image | ||||
| SCM | 0.645 (0.500, 0.791) | 0.611 (0.389, 0.889) | 0.781 (0.434, 0.860) | 0.770 (0.459, 0.841) |
| SE | 0.550 (0.406, 0.694) | 0.611 (0.222, 1.000) | 0.640 (0.076, 0.913) | 0.643 (0.131, 0.873) |
| STI | 0.581 (0.454, 0.709) | 0.722 (0.444, 1.000) | 0.615 (0.204, 0.777) | 0.622 (0.247, 0.760) |
| SJN | 0.561 (0.419, 0.703) | 0.556 (0.278, 1.000) | 0.728 (0.094, 0.909) | 0.728 (0.152, 0.873) |
| SA | 0.539 (0.384, 0.693) | 0.611 (0.111, 0.944) | 0.626 (0.170, 0.985) | 0.622 (0.219, 0.933) |
| SHB | 0.718 (0.593, 0.844) | 0.833 (0.500, 1.000) | 0.608 (0.468, 0.925) | 0.622 (0.502, 0.901) |
| SVC | 0.574 (0.441, 0.708) | 0.833 (0.167, 1.000) | 0.392 (0.117, 0.977) | 0.419 (0.170, 0.933) |
| ST | 0.783 (0.686, 0.881) | 0.778 (0.611, 1.000) | 0.762 (0.483, 0.876) | 0.767 (0.512, 0.869) |
| Combined model | ||||
| LR | 0.777 (0.682, 0.873) | 0.833 (0.667, 1.000) | 0.702 (0.645, 0.755) | 0.710 (0.657, 0.760) |
| RF | 0.782 (0.671, 0.893) | 0.833 (0.667, 1.000) | 0.638 (0.581, 0.694) | 0.650 (0.597, 0.703) |
| SVM | 0.770 (0.642, 0.898) | 0.833 (0.667, 1.000) | 0.694 (0.638, 0.751) | 0.703 (0.650, 0.756) |
| XGBoost | 0.783 (0.685, 0.880) | 0.778 (0.556, 0.944) | 0.660 (0.600, 0.717) | 0.668 (0.611, 0.721) |
| LDA | 0.535 (0.459, 0.611) | 0.111 (0.000, 0.278) | 0.958 (0.932, 0.981) | 0.905 (0.880, 0.926) |
| LightGBM | 0.804 (0.713, 0.894) | 0.889 (0.722, 1.000) | 0.657 (0.600, 0.709) | 0.671 (0.615, 0.724) |
ST Sagittal section of the tongue, SCM Horizontal section of the cricothyroid membrane, SVC Horizontal section of the vocal cord, SA Sagittal section of the airway, STI Horizontal section of the thyroid isthmus, SJN Horizontal section of the jugular notch, SHB Section of the hyoid bone, SE Section of the epiglottis, AUROC Area under the receiver operating characteristic curve, LR Logistic Regression, RF Random Forest, SVM Support Vector Machine, XGBoost eXtreme Gradient Boosting, LDA Linear Discriminant Analysis, LightGBM Light Gradient Boosting Machine
We calculated the confusion matrix for all models and found a relatively high level of false positives across all, while false negatives were very low (Table S8). The Brier scores were 0.049 in the validation set, and 0.055 in the test set for the combined model with LightGBM algorithm, which suggests that our model was well calibrated (Figure S3A and S3B). The area under the precision-recall curve was 0.157 in the validation set, and 0.233 in the test set (Figure S3C and S3D). Compared between the combined ultrasound-AI model (Table 2) and bedside tests (Table S9), the combined ultrasound-AI model had a higher AUROC value.
The result of variable ranking showed that the most important variable was ST (Figure S4). A heatmap analysis was conducted (Fig. 4A). For the ST, the critical region was concentrated around the end of the tongue root, which differs from a previous study [20].
Fig. 4.
The results of heatmap analysis, ROC analysis and RCS analysis; A The results of heatmap analysis; ST: sagittal section of the tongue; SCM: axial section of the cricothyroid membrane; SVC: axial section of the vocal cord; SA: sagittal section of the airway; STI: axial section of the thyroid isthmus; SJN: axial section of the jugular notch; SHB: section of the hyoid bone; SE: section of the epiglottis. B ROC analysis of DSTR, DST, and DGT; C RCS analysis between DSTR and difficult video laryngoscopy; D RCS analysis between DST and difficult video laryngoscopy; E RCS analysis between DGT and difficult video laryngoscopy; F ROC analysis of DSTR/Height, DST/Height, and DGT/Height; G RCS analysis between DSTR/Height and difficult video laryngoscopy; H RCS analysis between DST/Height and difficult video laryngoscopy; I RCS analysis between DGT/Height and difficult video laryngoscopy. RCS was calibrated using model3 and covariates included age, sex, American Society of Anesthesiologists Physical Status, BMI, neck mobility, mandibular protrusion, inter-incisor gap, thyromental distance, modified Mallampati test, neck circumference, snoring history, upper lip bite test, and mandibular length. ROC: receiver operating characteristic; RCS: restricted cubic spline; DSTR: distance from the skin to the end of the tongue root; DST: distance from the skin to the tongue surface; DGT: distance from the genioglossus to the tongue surface
Step 2: identify ultrasonographic measurement for predicting DVL
Critical regions with high weighting were screened according to the heatmaps. Ultrasonographic measurements in the ST included DSTR, DST, and DGT. DSTR is an ultrasonographic measurement in a critical region, and DST and DGT are ultrasonographic measurements recommended in a previous study [20].
In multivariable logistic regression (Table 3), DSTR was significantly associated with DVL, with ORs of 5.81 (95% CI: 3.82–9.03, P < 0.001), 4.75 (95% CI: 2.95–7.83, P < 0.001), and 4.83 (95% CI: 2.99–8.01, P < 0.001) in models 1, 2, and 3, respectively. Similar findings were noted in sensitivity analyses. DST was also significantly associated, but with lower ORs: 2.09 (95% CI: 1.42–3.08), 1.76 (95% CI: 1.12–2.76), and 1.65 (95% CI: 1.04–2.61). DGT showed no association, with ORs close to 1 in all models. Similar results were obtained when analyzing DSTR/Height, DST/Height, and DGT/Height.
Table 3.
The results of multivariable regression
| Model 1 OR (95% CI) |
Model 2 OR (95% CI) |
Model 3 OR (95% CI) |
|
|---|---|---|---|
| DSTR (cm) | |||
| Continuous variable | 5.81 (3.82, 9.03) * | 4.75 (2.95, 7.83) * | 4.83 (2.99, 8.01) * |
| Binary variable | |||
| < 3.89 cm | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| ≥ 3.89 cm | 6.38 (4.02, 10.17) * | 5.00 (2.98, 8.43) * | 5.13 (3.05, 8.70) * |
| Tertile | |||
| Tertile1 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| Tertile2 | 2.00 (0.95, 4.53) | 1.64 (0.76, 3.76) | 1.55 (0.71, 3.58) |
| Tertile3 | 5.31 (2.74, 11.32) * | 3.37 (1.65, 7.50) * | 3.38 (1.64, 7.57) * |
| DST (cm) | |||
| Continuous variable | 2.09 (1.42, 3.08) * | 1.76 (1.12, 2.76) * | 1.65 (1.04, 2.61) * |
| Binary variable | |||
| < 5.54 cm | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| ≥ 5.54 cm | 2.65 (1.50, 4.87) * | 2.17 (1.17, 4.19) * | 2.01 (1.07, 3.90) * |
| Tertile | |||
| Tertile1 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| Tertile2 | 2.43 (1.17, 5.43) * | 1.87 (0.83, 4.51) | 1.77 (0.78, 4.30) |
| Tertile3 | 3.47 (1.70, 7.70) * | 2.83 (1.28, 6.79) * | 2.60 (1.17, 6.31) * |
| DGT (cm) | |||
| Continuous variable | 1.01 (0.80, 1.12) | 1.00 (0.72, 1.12) | 1.00 (0.69, 1.12) |
| Binary variable | |||
| < 4.01 cm | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| ≥ 4.01 cm | 1.72 (1.02, 2.95) * | 1.58 (0.90, 2.82) | 1.47 (0.83, 2.65) |
| Tertile | |||
| Tertile1 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| Tertile2 | 1.58 (0.82, 3.14) | 1.37 (0.67, 2.88) | 1.30 (0.63, 2.76) |
| Tertile3 | 1.93 (1.00, 3.87) | 1.80 (0.89, 3.79) | 1.67 (0.82, 3.54) |
| DSTR/Height (per 0.001 value change) | |||
| Continuous variable | 1.33 (1.24, 1.42) * | 1.28 (1.19, 1.39) * | 1.28 (1.19, 1.39) * |
| Binary variable | |||
| < 21.8 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| ≥ 21.8 | 4.84 (3.03, 7.86) * | 3.47 (2.08, 5.86) * | 3.68 (2.19, 6.25) * |
| Tertile | |||
| Tertile1 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| Tertile2 | 1.37 (0.62, 3.19) | 1.08 (0.47, 2.55) | 1.02 (0.44, 2.42) |
| Tertile3 | 5.32 (2.74, 11.36) * | 3.18 (1.55, 7.05) * | 3.21 (1.57, 7.16) * |
| DST/Height (per 0.001 value change) | |||
| Continuous variable | 1.14 (1.07, 1.22) * | 1.10 (1.02, 1.18) * | 1.09 (1.01, 1.17) * |
| Binary variable | |||
| < 34.9 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| ≥ 34.9 | 3.82 (2.25, 6.65) * | 2.96 (1.66, 5.44) * | 2.82 (1.57, 5.20) * |
| Tertile | |||
| Tertile1 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| Tertile2 | 1.49 (0.63, 3.80) | 1.55 (0.59, 4.53) | 1.52 (0.58, 4.47) |
| Tertile3 | 4.84 (2.32, 11.40) * | 3.93 (1.66, 10.93) * | 3.70 (1.54, 10.33) * |
| DGT/Height (per 0.001 value change) | |||
| Continuous variable | 1.00 (0.97, 1.02) | 1.00 (0.95, 1.02) | 1.00 (0.94, 1.02) |
| Binary variable | |||
| < 24.7 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| ≥ 24.7 | 2.03 (1.21, 3.42) * | 1.90 (1.10, 3.36) * | 1.83 (1.05, 3.24) * |
| Tertile | |||
| Tertile1 | 1.00 (reference) | 1.00 (reference) | 1.00 (reference) |
| Tertile2 | 1.19 (0.59, 2.46) | 0.86 (0.40, 1.89) | 0.84 (0.39, 1.86) |
| Tertile3 | 1.99 (1.05, 3.95) * | 1.43 (0.71, 3.00) | 1.35 (0.67, 2.85) |
Model 1 was only simply corrected and its covariates included only age, gender, and ASA-PS. Model 2 expanded on model 1, by integrating additional covariates, including BMI, neck mobility, mandibular protrusion, inter-incisor gap, thyromental distance, and modified Mallampati test. Model 3 further expanded on model 2, by integrating additional covariates, including neck circumference, snoring history, upper lip bite test, and mandibular length
DSTR Distance from the skin to the end of the tongue root, DST Distance from the skin to the tongue surface, DGT Distance from the genioglossus to the tongue surface, OR Odds ratio, CI Confidence interval
For dichotomous variables, optimal cut-off values were determined using the maximum Youden’s Index. Trichotomous variables (tertiles) were categorized into three levels based on data distribution
* P < 0.05
Compared to DST and DGT, DSTR had a higher AUROC value (Fig. 4B) of 0.766 (95% CI: 0.714, 0.819). Additionally, compared to DST/Height and DGT/Height, DSTR/Height had a higher AUROC value (Fig. 4F) of 0.776 (95% CI: 0.725, 0.827). The results of predictive performance for the ultrasonographic measurements are shown in Table S10.
The RCS revealed a nonlinear association between DSTR and DVL, with a significant rise in the OR for DVL, particularly at a DSTR ≥ 3.89 cm (maximum Youden’s Index, Fig. 4C). In contrast, an increase in DST and DGT did not significantly increase the risk of DVL (Fig. 4D and E). Additionally, RSC showed similar results in DSTR/Height, DST/Height, and DGT/Height (Fig. 4G, H and I).
We categorized anesthesiologists into moderate-experience (3–5 years of experience in airway management) and extensive-experience (> 5 years of experience in airway management) groups. The results of this sub group analysis were consistent with the main findings of our study (Table S11 and Table S12).
Discussion
This study is the first to explore the potential of AI, specifically deep learning-based ultrasound analysis, as a valuable adjunct to clinical assessment to predict DVL. In Step 1, the model achieved an AUROC of 0.804 for DVL. Heatmap analyses highlighted a strong association between tongue root thickness and DVL—an observation that contrasts with earlier studies which emphasized maximum tongue thickness as the key predictor [20]. In Step 2, this finding was further validated, identifying tongue root thickness as an independent and significant predictor of DVL. RCS showed a nonlinear relationship, with the risk rising at DSTR ≥ 3.89 cm.
The final ultrasound model demonstrated a higher predictive performance than the individual bedside tests and composite scores. Among the bedside tests, only the thyromental distance had a reasonable AUROC of 0.738. The Wilson score and EGRI showed AUROC values of 0.571 and 0.683, respectively. This likely due to the high-quality, diverse dataset and advanced preprocessing techniques which improved robustness and reduced overfitting. In addition, ResNet18’s deep layers and skip connections effectively captured intricate features in the ultrasound images, essential for identifying DVL [35].
Heatmap analysis showed that image of ST displayed centralized focus of critical areas for DVL around the tongue root. Previous research has indicated an association between DVL and maximum tongue thickness [20], rather than tongue root thickness. These distinctions require further investigation and, therefore, we performed a post-hoc analysis. The results showed that DSTR was significantly associated with DVL, which is significantly higher than the corresponding ORs for DST and DGT. Sensitivity analysis results showed a significant increase in ORs for the posterior 1/3 distribution compared to the anterior 2/3 distribution. The RCS results revealed the risk of DVL increased significantly with rising DSTR, particularly at DSTR ≥ 3.89 cm.
Video laryngoscopy differs from direct laryngoscopy in the approach to visualization [8]. When using direct laryngoscopy, the operator’s sight needs to pass through the oral cavity, laryngeal cavity, and vocal cords in sequence. The sightline is angled, and the greatest interference with visualization is in the middle region of the tongue, which is usually its maximum thickness. However, when using video laryngoscopy, the sightline is horizontal because the camera is positioned at the front of the blade. The greatest interference with visualization is the thickness at the end of the tongue root. In addition, unlike direct laryngoscopy, video laryngoscopy involves distinct challenges beyond glottic visualization, particularly related to tracheal tube advancement. The Videolaryngoscopic Classification of Intubation (VCI) explicitly distinguishes visualization from tube delivery, recognizing that good views may still be associated with difficult intubation [36]. This framework highlights that predictive models should address both dimensions to fully capture videolaryngoscopic complexity.
Our study has several limitations. Firstly, the study was conducted using a single video laryngoscope. Video laryngoscopic performance and difficulty vary according to device design and blade geometry; therefore, our results may not be directly generalizable to other video laryngoscopes, particularly hyperangulated devices, where tube delivery challenges are often more pronounced. Secondly, the definition of difficult video laryngoscopy in this study relies predominantly on glottic visualization assessed by the Cormack-Lehane classification. Although widely used [37], this scale was originally developed for direct laryngoscopy and does not fully capture the procedural complexity of video laryngoscopy. In video laryngoscopy, difficulty also involves tube delivery, exemplified by the “can see, can’t intubate” scenario [10]. Future studies should incorporate video laryngoscopy‑specific frameworks, such as the VCI, which explicitly distinguish visualization from tube delivery. Thirdly, this study was conducted at a general hospital renowned for the department of oral and maxillofacial surgery, which resulted in a slightly higher incidence of difficult airway. Previous study with large samples also showed that patients in the maxillofacial (8.9%) and ENT (7.4%) departments have the highest rates of difficult video laryngoscopy [38], and such results are consistent with the results in our center. Fourthly, the initial C-L grade may not reflect the best achievable view and could be influenced by operator experience, though this choice was intentional to maintain reproducibility and clinical relevance. Finally, the relative homogeneity of our study sample may limit generalizability to other populations. In the future, a randomized controlled trial may help validate the role of AI in airway management.
Conclusion
AI-based ultrasound technology may assist in identifying difficult video laryngoscopy. Tongue root thickness is associated with difficult video laryngoscopy.
Supplementary Information
Acknowledgements
We thank all the participants and research staff who took part in the cohort for their contributions. Special thanks to Ming Xia from Yangpu Hospital, Tongji University for dedicated support and expertise throughout the study. We sincerely thank Prof. Michael G. Irwin (Department of Anesthesiology, The University of Hong Kong), Prof. Vladimir Cerny (Department of Anesthesiology and Intensive Care Medicine, University Hospital Hradec Kralove), Prof. Giustino Varrassi (Paolo Procacci Foundation), Dr. Denise Battaglini (Anaesthesia and Intensive Care, IRCCS Ospedale Policlinico San Martino) and Dr. Rita Cataldo (Research Unit of Anesthesia, Intensive Care and Pain Management, Department of Medicine and Surgery, Università Campus Bio-Medico di Roma) for their valuable assistance with editing and revising the manuscript. Their feedback greatly improved the clarity and quality of the final version. We thank Limei Zhang, Peipei Su and Xueting Liu from Shanghai Ninth People’s Hospital for assistance with literature search.
Abbreviations
- DVL
Difficult video laryngoscopy
- AUROC
Area under the receiver operating characteristic curve
- DSTR
Distance from the skin to the end of the tongue root
- OR
Odds ratio
- AI
Artificial intelligence
- ASA-PS
American Society of Anesthesiologists Physical Status
- SCM
Axial section of the cricothyroid membrane
- ST
Sagittal section of the tongue
- SVC
Axial section of the vocal cord
- SA
Sagittal section of the airway
- STI
Axial section of the thyroid isthmus
- SJN
Axial section of the jugular notch
- SHB
Section of the hyoid bone
- SE
Section of the epiglottis
- I-AIM
Indication, acquisition, interpretation, and medical decision making
- SVM
Support Vector Machine
- LDA
Linear Discriminant Analysis
- DST
Distance from the skin to the tongue surface
- DGT
Distance from the genioglossus to the tongue surface
Authors’ contributions
*Study design: * Hong Jiang, Chenyu Jin.*Supervision and coordination of team* : Jia Yan, Hong Jiang.*Data collection* : Chenyu Jin, Bei Pei, Shuang Cao.*Data cleaning* : Chenyu Jin, Bei Pei.*Data analysis* : Chenyu Jin, Bei Pei, Ren Zhou, Yu Sun, Zhiliang Lin, Mingxu Zhao, Yaokun Zheng, Qian Wang, Maosong Cao.*Writing and revisions of the manuscript* : Chenyu Jin, Bei Pei, Ren Zhou, Jia Yan.*Literature search* : Chenyu Jin, Bei Pei, Shuang Cao.
Funding
This study was supported by the Emerging Interdisciplinary Research Specialized Project of Shanghai Municipal Health Commission (No. 2022JC024), Technology Standardization and Promotion Project of Shanghai Hospital Development Center (No. SHDC22023219), the Clinical Research Program of Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. JYLJ202308), the Health Development Promotion Project-Anesthesia and Critical Care Research Project of Beijing Health Alliance Charitable Foundation (No. MZ-20240924-03), and Research and Translational Project on a Big Data-Driven Multimodal Device for Pre-anesthesia Planning and Risk Assessment (the Yangtze River Pharmaceutical Group, No. JYHX2025028).
Data availability
The data supporting this study’s findings are available from the corresponding author upon reasonable request.
Declarations
Ethics approval and consent to participate
This study was approved by the Ethics Committee of Shanghai Ninth People’s Hospital (SH9H-2021-T356-2) and registered at ClinicalTrials.gov (NCT05207254). All participants provided written informed consent.
Consent for publication
Not applicable.
Competing interests
The authors have received funding from Yangtze River Pharmaceutical Group and Beijing Health Alliance Charitable Foundation for this research project. The funder had no role in the study design, data collection, analysis, interpretation, or writing of the manuscript.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Chenyu Jin and Bei Pei contributed equally to this work.
Jia Yan and Hong Jiang are jointly supervised this work.
Contributor Information
Jia Yan, Email: mzkyanj@163.com.
Hong Jiang, Email: jianghongjiuyuan@163.com.
References
- 1.Cumberworth A, Lewith H, Sud A, Jefferson H, Athanassoglou V, Pandit JJ. Major complications of airway management: a prospective multicentre observational study. Anaesthesia. 2022;77(6):640–8. [DOI] [PubMed] [Google Scholar]
- 2.Cook TM, Kelly FE. A national survey of videolaryngoscopy in the United Kingdom. Br J Anaesth. 2017;118(4):593–600. [DOI] [PubMed]
- 3.Nalubola S, Jin E, Drugge ED, Weber G, Abramowicz AE. Video Versus Direct Laryngoscopy in Novice Intubators: A Systematic Review and Meta-Analysis. Cureus. 2022;14(9):e29578. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Gómez-Ríos M, Van Zundert AAJ, McNarry AF, Law JA, Higgs A, De Jong A, Jaber S, Karamchandani K, Hansel J, Saracoglu KT, et al. Guidelines on strategies for the universal implementation of videolaryngoscopy. Eur J Anaesthesiol. 2025;42(10):872–88. [DOI] [PubMed] [Google Scholar]
- 5.Law JA, Duggan LV, Asselin M, Baker P, Crosby E, Downey A, Hung OR, Kovacs G, Lemay F, Noppens R, et al. Canadian Airway Focus Group updated consensus-based recommendations for management of the difficult airway: part 2. Planning and implementing safe management of the patient with an anticipated difficult airway. Can J Anaesth = J canadien d’anesthesie. 2021;68(9):1405–36. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Vora J, Leslie D, Stacey M. Awake tracheal intubation. BJA Educ. 2022;22(8):298–305. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Ahmad I, Bailey CR. Time to abandon awake fibreoptic intubation? Anaesthesia. 2016;71(1):12–6. [DOI] [PubMed] [Google Scholar]
- 8.Díaz-Gómez JL, Satyapriya A, Satyapriya SV, Mascha EJ, Yang D, Krakovitz P, Mossad EB, Eikermann M, Doyle DJ. Standard clinical risk factors for difficult laryngoscopy are not independent predictors of intubation success with the GlideScope. J Clin Anesth. 2011;23(8):603–10. [DOI] [PubMed] [Google Scholar]
- 9.Carsetti A, Sorbello M, Adrario E, Donati A, Falcetta S. Airway Ultrasound as Predictor of Difficult Direct Laryngoscopy: A Systematic Review and Meta-analysis. Anesth Analg. 2022;134(4):740–50. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Fernández-Vaquero MA, De Luis-Cabezón N, García-Aroca MA, Álvarez-Avello JM, Vives-Santacana M, Greif R, Martinez-Hurtado ED, Ly-Liu D. Pilot multicenter study to determine the utility of point-of-care ultrasound to predict difficulty of tracheal intubation using videolaryngoscopy with the McGrath™ Mac videolaryngoscope. Front Med. 2024;11:1406676. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Connor CW. Artificial Intelligence and Machine Learning in Anesthesiology. Anesthesiology. 2019;131(6):1346–59. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Cormack RS. Cormack-Lehane classification revisited. Br J Anaesth. 2010;105(6):867–8. [DOI] [PubMed] [Google Scholar]
- 13.Detsky ME, Jivraj N, Adhikari NK, Friedrich JO, Pinto R, Simel DL, Wijeysundera DN, Scales DC. Will This Patient Be Difficult to Intubate? The Rational Clinical Examination Systematic Review. JAMA. 2019;321(5):493–503. [DOI] [PubMed] [Google Scholar]
- 14.Xia M, Ma W, Zuo M, Deng X, Xue F, Battaglini D, Aggarwal V, Varrassi G, Cerny V, Di Giacinto I, et al. Expert consensus on difficult airway assessment. Hepatobiliary Surg Nutr. 2023;12(4):545–66. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Roth D, Pace NL, Lee A, Hovhannisyan K, Warenits AM, Arrich J, Herkner H. Airway physical examination tests for detection of difficult airway management in apparently normal adult patients. Cochrane Database Syst Rev. 2018;5(5):Cd008874. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Seo SH, Lee JG, Yu SB, Kim DS, Ryu SJ, Kim KH. Predictors of difficult intubation defined by the intubation difficulty scale (IDS): predictive value of 7 airway assessment factors. Korean J anesthesiology. 2012;63(6):491–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Alessandri F, Antenucci G, Piervincenzi E, Buonopane C, Bellucci R, Andreoli C, Alunni Fegatelli D, Ranieri MV, Bilotta F. Ultrasound as a new tool in the assessment of airway difficulties: An observational study. Eur J Anaesthesiol. 2019;36(7):509–15. [DOI] [PubMed] [Google Scholar]
- 18.Chen YY, Lin JC, Tsai MJ, Cheng KY. Retromolar intubation with video intubating stylet in difficult airway: A randomized crossover manikin study. Am J Emerg Med. 2022;54:212–20. [DOI] [PubMed] [Google Scholar]
- 19.Molins G, Valls A, Guijarro R, Hernández-Alfaro F. Mandibular hypoplasia and narrow airway in goldenhar syndrome: Anticipation of difficult intubation with cone-beam computed tomography. J Clin Anesth. 2016;34:1–2. [DOI] [PubMed] [Google Scholar]
- 20.Yao W, Wang B. Can tongue thickness measured by ultrasonography predict difficult tracheal intubation? Br J Anaesth. 2017;118(4):601–9. [DOI] [PubMed]
- 21.Falcetta S, Cavallo S, Gabbanelli V, Pelaia P, Sorbello M, Zdravkovic I, Donati A. Evaluation of two neck ultrasound measurements as predictors of difficult direct laryngoscopy: A prospective observational study. Eur J Anaesthesiol. 2018;35(8):605–12. [DOI] [PubMed] [Google Scholar]
- 22.Bianchini A, Nardozi L, Nardi E, Scuppa MF. Airways ultrasound in predicting difficult face mask ventilation. Minerva Anestesiol. 2021;87(1):26–34. [DOI] [PubMed] [Google Scholar]
- 23.Abdallah FW, Yu E, Cholvisudhi P, Niazi AU, Chin KJ, Abbas S, Chan VW. Is Ultrasound a Valid and Reliable Imaging Modality for Airway Evaluation? An Observational Computed Tomographic Validation Study Using Submandibular Scanning of the Mouth and Oropharynx. J ultrasound medicine: official J Am Inst Ultrasound Med. 2017;36(1):49–59. [DOI] [PubMed] [Google Scholar]
- 24.Bahner DP, Hughes D, Royall NA. I-AIM: a novel model for teaching and performing focused sonography. J ultrasound medicine: official J Am Inst Ultrasound Med. 2012;31(2):295–300. [DOI] [PubMed] [Google Scholar]
- 25.Buwei Yu XW, Mingzhang Zuo X, Deng X, Gao. Ming Tian: Guideline for Difficult Airway Management. J Clin Anesthesiology (China). 2013;29(1):93–8. [Google Scholar]
- 26.He F, Liu T, Tao D. Why ResNet Works? Residuals Generalize. IEEE Trans neural networks Learn Syst. 2020;31(12):5349–62. [DOI] [PubMed] [Google Scholar]
- 27.Lin TY, Goyal P, Girshick R, He K, Dollar P. Focal Loss for Dense Object Detection. IEEE Trans Pattern Anal Mach Intell. 2020;42(2):318–27. [DOI] [PubMed] [Google Scholar]
- 28.Loshchilov I, Hutter F. Sgdr: Stochastic gradient descent with warm restarts. In: arXiv https://arxiv.org/abs/1608.03983
- 29.el-Ganzouri AR, McCarthy RJ, Tuman KJ, Tanck EN, Ivankovich AD. Preoperative airway assessment: predictive value of a multivariate risk index. Anesth Analg. 1996;82(6):1197–204. [DOI] [PubMed] [Google Scholar]
- 30.Stekhoven DJ, Bühlmann P. MissForest–non-parametric missing value imputation for mixed-type data. Bioinf (Oxford England). 2012;28(1):112–8. [DOI] [PubMed] [Google Scholar]
- 31.Dunne R, Reguant R, Ramarao-Milne P, Szul P, Sng LMF, Lundberg M, Twine NA, Bauer DC. Thresholding Gini variable importance with a single-trained random forest: An empirical Bayes approach. Comput Struct Biotechnol J. 2023;21:4354–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Riley RD, Ensor J, Snell KIE, Harrell FE Jr., Martin GP, Reitsma JB, Moons KGM, Collins G, van Smeden M. Calculating the sample size required for developing a clinical prediction model. BMJ (Clinical Res ed). 2020;368:m441. [DOI] [PubMed] [Google Scholar]
- 33.Faul F, Erdfelder E, Lang AG, Buchner A. G*Power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behav Res Methods. 2007;39(2):175–91. [DOI] [PubMed] [Google Scholar]
- 34.Razali MN, Arbaiy N, Lin P-C, Ismail S. Optimizing Multiclass Classification Using Convolutional Neural Networks with Class Weights and Early Stopping for Imbalanced Datasets. 2025;14(4):705.
- 35.Xia M, Jin C, Zheng Y, Wang J, Zhao M, Cao S, Xu T, Pei B, Irwin MG, Lin Z, et al. Deep learning-based facial analysis for predicting difficult videolaryngoscopy: a feasibility study. Anaesthesia. 2024;79(4):399–409. [DOI] [PubMed] [Google Scholar]
- 36.Fernández-Vaquero M, Charco-Mora P, Sastre-Rincon JA, Gomez-Rios M, Cardenas-Gomez J, Martinez-Hurtado E, Aracil-Escoda N, Vallejo-Tarrat A, Thion-Soriano I, Peral-Sanchez D, et al. VCISpain: protocol for a prospective multicenter observational study to validate a standardized classification tool for tracheal intubation using videolaryngoscopy. Brazilian J anesthesiology (Elsevier). 2025;75(5):844653. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Apfelbaum JL, Hagberg CA, Connis RT, Abdelmalak BB, Agarkar M, Dutton RP, Fiadjoe JE, Greif R, Klock PA, Mercier D, et al. 2022 American Society of Anesthesiologists Practice Guidelines for Management of the Difficult Airway. Anesthesiology. 2022;136(1):31–81. [DOI] [PubMed] [Google Scholar]
- 38.Heinrich S, Birkholz T, Irouschek A, Ackermann A, Schmidt J. Incidences and predictors of difficult laryngoscopy in adult patients undergoing general anesthesia: a single-center analysis of 102,305 cases. J Anesth. 2013;27(6):815–21. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The data supporting this study’s findings are available from the corresponding author upon reasonable request.