Skip to Content
MDPIMDPI

Advanced

by
  • Erdinç Koca 1,*,
  • Sevgi Kutlusoy 2 and
  • Tarkan Koca 4
  • et al.

Reviewer 1: Anonymous Reviewer 2: Ahmed Al Marouf

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a method for automatic assessment of tracheal intubation using DL models. The topic is interesting and important. However, several improvements are needed to improve the paper's scientific quality.

 

  1. A literature review was not conducted, and knowledge gaps in this area were not identified. Therefore, the contribution and novelty of this work are difficult to assess.
  2. It is unclear how the network training and validation process was conducted. Did the set of 16 images represent a single input data vector for each patient? How many samples were in each class? How was the network validation performed? Please describe this procedurę in more detail.
  3. Comparison of the obtained results with those of other authors should be more structured. It should be presented in tabular form and include information about the datasets used, the implemented models, and the quality measures obtained.
  4. The computational complexity of applied the DL models and the reference models should be provided.
  5. The limitations of the proposed method were not discussed.

Author Response

Manuscript Number: diagnostics-4175575

 

Article Title: “Predicting Difficult Tracheal Intubation Using Multi-Angle Photographic Analysis with Convolutional Neural Networks and EfficientNet”

 

Dear Editor,

We would like to sincerely thank the editor and the reviewers for their careful evaluation of our manuscript and for their valuable comments and suggestions. These comments helped us significantly improve the quality, clarity, and scientific rigor of the paper. All reviewer comments have been carefully addressed, and the manuscript has been revised accordingly. All modifications in the revised manuscript have been highlighted in red color for clarity. Below, we provide a detailed response to each reviewer comment.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

REVIEWER 1

Reviewer#1: This paper proposes a method for automatic assessment of tracheal intubation using DL models. The topic is interesting and important. However, several improvements are needed to improve the paper's scientific quality.

 Reviewer#1, Concern# 1: A literature review was not conducted, and knowledge gaps in this area were not identified. Therefore, the contribution and novelty of this work are difficult to assess.

Author response: We thank the reviewer for this important comment. In the revised version of the manuscript, we have substantially expanded the literature review and clearly identified the knowledge gap addressed by this study. Specifically, the Introduction section has been revised to include several additional recent studies related to artificial intelligence–based prediction of difficult tracheal intubation. These studies have been discussed to provide a clearer overview of the current state of research in this area. Furthermore, to better highlight the novelty and scientific contribution of our work, we have added a new paragraph at the end of the Introduction section summarizing the main contributions of the study. These revisions clarify the novelty of the proposed approach and strengthen the positioning of the study within the existing literature. The related modifications can be found in the Introduction section (pages XX–XX) of the revised manuscript.

After Revision

The main contributions of this study can be summarized as follows:

  1. A deep learning–based framework for predicting tracheal intubation difficulty using smartphone-acquired bedside photographs is proposed.
  2. Unlike many previous studies that focus on binary classification, this study introduces a three-class classification framework (easy, medium, difficult), enabling a more detailed evaluation of airway difficulty.
  3. Two deep learning architectures, CNN and EfficientNet, are implemented and comparatively evaluated for predicting intubation difficulty from multi-angle facial and neck photographs.
  4. A multi-angle photographic dataset consisting of 16 images per patient is utilized, allowing the models to capture richer anatomical information related to airway assessment.
  5. The effects of different batch sizes and learning rates are systematically analyzed to evaluate the robustness and stability of the proposed models.

 

 

Reviewer#1, Concern# 2: It is unclear how the network training and validation process was conducted. Did the set of 16 images represent a single input data vector for each patient? How many samples were in each class? How was the network validation performed? Please describe this procedurę in more detail.

Author response: We thank the reviewer for this valuable comment. In the revised manuscript, we have clarified the training and validation procedure and provided additional details regarding the dataset structure and model training process.

After Revision

In this study, each patient contributed 16 photographs captured from different anatomical perspectives during the preoperative evaluation. These photographs were obtained in both sitting and supine positions, including anterior and lateral views, mouth open and closed, neck straight, and neck extended positions. Each of these images represents a different anatomical view that may influence airway assessment. During the training process, these images were treated as independent input samples, while their corresponding labels were assigned according to the clinically determined intubation difficulty level of the patient (easy, medium, or difficult). To ensure a reliable evaluation of the proposed models, the dataset was divided into training, validation, and test sets. The deep learning models were trained using the training set, while the validation set was used to monitor model performance during training and prevent overfitting. The final model performance was evaluated using the test set. Prior to training, all images were subjected to standard preprocessing steps including image resizing, normalization, and data augmentation techniques such as rotation, flipping, and brightness adjustments. These steps were applied to improve the generalization capability of the models. The models were trained using the Adam optimization algorithm and categorical cross-entropy loss function, and the training process was conducted for 20–50 epochs until model convergence was achieved. These additional details describing the dataset structure, training procedure, and validation strategy have now been included in the Materials and Methods section of the revised manuscript.

Table 1. Distribution of patients and corresponding images across the intubation difficulty classes.

Intubation Class

Number of Patients

Number of Images

Easy

38

608

Medium

35

560

Difficult

36

576

 

Erdinç hocam tablodaki değler hem buraya hem makaleye eklenmeli

 

After Revision

Reviewer#1, Concern# 3: Comparison of the obtained results with those of other authors should be more structured. It should be presented in tabular form and include information about the datasets used, the implemented models, and the quality measures obtained.

Author response: We thank the reviewer for this helpful suggestion. In the revised version of the manuscript, we have restructured the comparison with previously published studies and summarized the relevant works in a tabular format.

After Revision

To provide a clearer comparison with existing studies, we have added a new table summarizing related work in the literature. The table includes key information such as:

  • the dataset characteristics used in each study
  • the artificial intelligence or machine learning models implemented
  • the classification approach
  • the reported performance metrics

The studies included in the table are those already discussed in the manuscript, including Connor et al., Kim et al., Cuendet et al., and Tavolara et al.

This structured presentation allows readers to easily compare the proposed method with previously reported approaches and better understand the contribution of the present study. The new comparison table has been added to the Discussion section of the revised manuscript.

After Revision

Table 6. Comparison of previous studies on AI-based prediction of difficult tracheal intubation

Study

Dataset

Method / Model

Classification Type

Performance

Connor et al. [42]

80 patients (facial morphology analysis)

Logistic regression

Binary

Sensitivity 90%, Specificity 85%, AUC 0.899

Cuendet et al. [44]

970 patients (facial images and videos)

Random Forest

Binary

AUC 81%

Tavolara et al. [45]

152 facial images

CNN ensemble

Binary

AUC 71.05%

Kim et al. [43]

Smartphone face and neck images

EfficientNet

Binary

AUC 0.81–0.88

Proposed Study

109 patients (16 images per patient)

CNN + EfficientNet

Three-class (easy, medium, difficult)

Accuracy 88.64%, F1-score 87.28%

As shown in Table 6, most previous studies focused on binary classification of intubation difficulty. In contrast, the proposed study performs a three-class classification using deep learning models (CNN and EfficientNet) based on multi-angle photographic data, providing a more detailed evaluation of airway difficulty.

 

Reviewer#1, Concern# 4: The computational complexity of applied the DL models and the reference models should be provided.

Author response: We thank the reviewer for this valuable suggestion. In the revised manuscript, we have added a discussion regarding the computational complexity and efficiency of the deep learning models used in this study, as well as a brief comparison with previously reported methods.

After Revision

A new paragraph discussing the computational complexity of the proposed models has been added to the Discussion section of the manuscript. In this revision, we describe the architectural characteristics and computational efficiency of the CNN and EfficientNet models used in our study and compare them conceptually with previously reported approaches such as logistic regression, random forest, and CNN-based models used in related studies.

 

The discussion emphasizes that EfficientNet provides a favorable balance between model accuracy and computational efficiency through compound scaling, while the CNN architecture used in this study maintains relatively low computational complexity due to its moderate number of convolutional layers.

Table 6. Comparison of previous studies on AI-based prediction of difficult tracheal intubation

Study

Dataset

Method / Model

Classification Type

Performance

Connor et al. [42]

80 patients (facial morphology analysis)

Logistic regression

Binary

Sensitivity 90%, Specificity 85%, AUC 0.899

Cuendet et al. [44]

970 patients (facial images and videos)

Random Forest

Binary

AUC 81%

Tavolara et al. [45]

152 facial images

CNN ensemble

Binary

AUC 71.05%

Kim et al. [43]

Smartphone face and neck images

EfficientNet

Binary

AUC 0.81–0.88

Proposed Study

109 patients (16 images per patient)

CNN + EfficientNet

Three-class (easy, medium, difficult)

Accuracy 88.64%, F1-score 87.28%

As shown in Table 6, most previous studies focused on binary classification of intubation difficulty. In contrast, the proposed study performs a three-class classification using deep learning models (CNN and EfficientNet) based on multi-angle photographic data, providing a more detailed evaluation of airway difficulty. In addition to classification performance, the computational complexity of the proposed models is an important factor for potential clinical implementation. The CNN architecture used in this study consists of three convolutional layers followed by pooling, flatten, and fully connected layers, which provides relatively low computational complexity while maintaining effective feature extraction capability.  EfficientNet, on the other hand, is designed using a compound scaling strategy that simultaneously balances network depth, width, and input resolution. This design allows EfficientNet to achieve high classification accuracy while maintaining computational efficiency compared to many conventional deep learning architectures. When compared with previously reported approaches in the literature, such as logistic regression models based on facial morphology (Connor et al.) or random forest models using facial measurements (Cuendet et al.), the proposed deep learning models require higher computational resources during training but provide significantly improved feature representation capabilities. However, once trained, the inference stage of CNN and EfficientNet models can be executed efficiently, making them suitable for real-time clinical decision support systems.

 

Reviewer#1, Concern# 5: The limitations of the proposed method were not discussed.

Author response: We thank the reviewer for this important comment. In the revised manuscript, we have expanded the discussion of the limitations of the proposed method in order to provide a more balanced evaluation of the study.

After Revision

A new paragraph discussing the limitations of the study has been added to the Conclusion section of the manuscript. The revised version now explicitly addresses several important limitations of the proposed approach, including the relatively small dataset size, the single-center nature of the study, the limited demographic diversity of the dataset, and potential class imbalance between intubation difficulty categories.

 

In addition, we discuss that although the proposed deep learning models achieved promising performance results, further validation using larger multicenter datasets and more diverse patient populations will be necessary to confirm the generalizability of the proposed approach.

 

  1. Conclusions

In our study, instead of using the Cormack–Lehane classification, which is commonly applied in other studies, we categorized intubation as follows. Intubations that were performed easily by a single clinician were classified as easy. Intubations that resulted from external tracheal interventions or required more than one attempt were classified as medium. Finally, intubations that required more than three attempts, involved the use of additional devices, or could not be performed were classified as difficult. This categorization aimed to support the prediction of patients beyond those in the easy group and to enable precautions to be taken before the procedure. We also believe that, in a further step, bedside photographs obtained using a smartphone-based application may assist clinicians in predicting difficult intubation. Such an approach may contribute to the development of an objective preoperative assessment tool that can support anesthesiologists in identifying potential airway difficulties before anesthesia induction. However, this study has several limitations that should be considered. First, the dataset used in this study consisted of a relatively limited number of patients, which may restrict the generalization capability of the developed models. Second, the data were collected from a single medical center, and therefore the dataset may not fully represent different patient populations or clinical conditions. In addition, the distribution of intubation difficulty classes may not be perfectly balanced, which could potentially influence the training process of the deep learning models. Although data augmentation techniques were applied to improve model robustness, larger and more balanced datasets would allow for a more comprehensive evaluation of the proposed approach. Furthermore, the study did not include patients from different nationalities, which may limit the generalizability of the results across diverse populations. Future studies should therefore include larger multicenter datasets and more diverse patient groups in order to further validate the proposed method and evaluate its potential clinical applicability in real-world settings. The present study has several limitations that should be considered when interpreting the results. First, the dataset used in this study consists of a relatively limited number of patients, which may restrict the generalization capability of the developed models. Second, the data were collected from a single medical center, and therefore the dataset may not fully represent different patient populations or clinical conditions. Another limitation is that the distribution of intubation difficulty classes may not be perfectly balanced, which can potentially influence the training process of deep learning models. Although data augmentation techniques were applied to improve model robustness, further studies using larger and more balanced datasets would provide more reliable evaluations. Additionally, the proposed approach relies on photographic data captured under controlled conditions. Variations in lighting conditions, camera positions, and image quality in real clinical environments may affect model performance. Therefore, future studies should evaluate the robustness of the proposed system under different imaging conditions. Finally, although the proposed deep learning models demonstrated promising classification performance, external validation using multicenter datasets and prospective clinical studies will be necessary to confirm the clinical applicability of the proposed method.

 

 

 

 

 

 

 

 

 

Reviewer 2 Report

Comments and Suggestions for Authors

Address the following comments. 

  1. I would suggest to change the title of the paper. Do not use Artificial Intelligence, use the name of the method in AI that you have used, in your case, CNN and EfficientNet. So update the title, "............... using Convolutional Neural Network and EfficientNet".
  2. The supplementary file contains everything in language other than English. Please add a description at the top of the file, what it represents. 
  3. The main contributions of the paper must be highlighted at the end of the introduction section. 
  4. What the color-code in fig-2 and fig-3 bar plot means. Either explain or reproduce it to a simple box plot. 
  5. The current labels (Easy, Moderate, Difficult) are clinically sound, but "Difficult" cases are notoriously rare in clinical datasets. Explicitly state how you handled class imbalance (e.g., SMOTE, weighted loss functions, or oversampling). Additionally, consider grouping "Moderate" and "Difficult" into a single "Potential Difficult Airway" class to improve the model's clinical utility as a screening tool.
  6. Clarify the architecture. Did you use a Feature Fusion approach (extracting features from each angle and concatenating them) or an Ensemble of 16 sub-networks? Providing a simplified diagram of the data pipeline would significantly improve the "Methods" section.
  7. If possible, include a side-by-side comparison of your CNN’s predictions against the Mallampati Score or Thyromental Distance measurements taken for the same patients. Proving your model has a higher Positive Predictive Value (PPV) than a physical exam is the strongest argument for bedside mobile phone implementation.

Author Response

REVIEWER 2

Reviewer#2, Concern# 1: I would suggest to change the title of the paper. Do not use Artificial Intelligence, use the name of the method in AI that you have used, in your case, CNN and EfficientNet. So update the title, "............... using Convolutional Neural Network and EfficientNet".

Author response: We thank the reviewer for this valuable suggestion. In accordance with the reviewer’s recommendation, we have revised the title of the manuscript to explicitly reflect the deep learning architectures used in the study.

After Revision

The title of the manuscript has been updated by replacing the general term “Artificial Intelligence” with the specific deep learning architectures implemented in this work, namely Convolutional Neural Networks (CNN) and EfficientNet. This change provides a clearer and more accurate description of the methodological approach used in the study.

The original title was:

Predicting Difficult Tracheal Intubation by Photographic Analysis Using Artificial Intelligence

The revised title is:

Predicting Difficult Tracheal Intubation Using Multi-Angle Photographic Analysis with Convolutional Neural Networks and EfficientNet

 

Reviewer#2, Concern# 2: The supplementary file contains everything in language other than English. Please add a description at the top of the file, what it represents.

Author response: We thank the reviewer for this helpful comment. In response to this suggestion, we have revised the supplementary file by adding a brief explanatory note at the beginning of the document.

After Revision

A short description has been added at the top of the supplementary file explaining that the document represents the ethical approval obtained for the study. The original document was issued in Turkish by the relevant ethics committee, and an English translation has been provided for clarity and accessibility to international readers.

The English translation of the ethics committee document has been added to the supplementary section.

Reviewer#2, Concern# 3: The main contributions of the paper must be highlighted at the end of the introduction section.

Author response: We thank the reviewer for this helpful suggestion. In accordance with the reviewer’s recommendation, we have added a paragraph at the end of the Introduction section clearly summarizing the main contributions of the study. This paragraph highlights the key novelties of the proposed approach, including the use of a deep learning framework based on CNN and EfficientNet architectures, the utilization of multi-angle photographic data captured using a smartphone, and the introduction of a three-class classification scheme (easy, medium, difficult) for predicting tracheal intubation difficulty.

 

The newly added paragraph outlining the main contributions of the study is presented at the end of the Introduction section in the revised manuscript.

After Revision

The main contributions of this study can be summarized as follows:

  • A deep learning–based framework for predicting tracheal intubation difficulty using smartphone-acquired bedside photographs is proposed.
  • Unlike many previous studies that focus on binary classification, this study introduces a three-class classification framework (easy, medium, difficult), enabling a more detailed evaluation of airway difficulty.
  • Two deep learning architectures, CNN and EfficientNet, are implemented and comparatively evaluated for predicting intubation difficulty from multi-angle facial and neck photographs.
  • A multi-angle photographic dataset consisting of 16 images per patient is utilized, allowing the models to capture richer anatomical information related to airway assessment.
  • The effects of different batch sizes and learning rates are systematically analyzed to evaluate the robustness and stability of the proposed models.

 

Reviewer#2, Concern# 4: What the color-code in fig-2 and fig-3 bar plot means. Either explain or reproduce it to a simple box plot.

Author response: In the revised manuscript, we have clarified the meaning of the color coding used in Figures 2 and 3. Specifically, the box plots represent the distribution of the evaluation metrics obtained under different training configurations (various batch sizes and learning rates). Each colored box corresponds to a specific performance metric, namely accuracy, precision, recall (sensitivity), and F1-score. To improve clarity, the figure captions have been revised to explicitly explain the meaning of the color coding used in the box plots.

After Revision

The confusion matrix in Figure 2 visualizes the classification performance of the model, trained using the best hyperparameter combination of CNN architecture (a batch size of 32 and a learning rate of 0.1) in three classes (easy, medium, difficult). The box plot in Figure 2 shows the overall distribution of the four performance metrics (learning rate, accuracy, precision, sensitivity, F1 scores) of the CNN model. From this plot, the central tendencies, variances, and outliers of the metrics can be easily observed. In particular, the accuracy and precision scores indicate that the model generally performs consistently well overall. Confusion matrix and performance distribution of the CNN model. The confusion matrix illustrates the classification performance obtained using the best hyperparameter configuration (batch size = 32, learning rate = 0.1). The box plot shows the distribution of performance metrics obtained under different training configurations. Each colored box represents a different evaluation metric: accuracy, precision, recall (sensitivity), and F1-score, summarizing their variability across experiments with different batch sizes and learning rates.

Figure 2. CNN model confusion matrix and distribution of performance metrics across different training configurations.

In Table 5, key performance metrics, namely, accuracy, precision, sensitivity and F1 scores are presented in detail for the classification processes performed using EfficientNet architecture. The table presents a systematic evaluation of the impact of different batch size (32, 64, and 128) and learning rate (0.1, 0.01, and 0.001) combinations on the model. The model performed best with an accuracy of 88.64% and an F1 score of 87.28%, obtained with a learning rate of 0.01 and a batch size of 32. In this setup, the precision (recall) value was also 88.01%, indicating that the model correctly recognizes the classes. When the learning rate was 0.001, the accuracy decreased (86.30%), while the precision (88.82%) and recall (88.40%) values increased. This means that, although the number of correctly classified instances by the model decreased, most of these classifications remained correct. Experiments with a batch size of 64 yielded more stable results. In particular, an accuracy of 87.12% and an F1 score of 87.64% were achieved with a learning rate of 0.001. This result shows that the model achieves stable classification performance and has less instability between classes. In the evaluations with a batch size of 128 and a learning rate of 0.01, although the accuracy values were slightly lower in general, a very high F1 score of 88.53% was obtained. This configuration is considered a preferable option, especially in applications in which a balance between classes is important.

Table 5. Results from EfficientNet.

Model

Batch-Size

Learning rate

Accuracy %

Precision %

Sensitivity %

F1-Scores %

EfficientNet

32

0.1

88.21

87.57

85.93

86.93
   

0.01

88.64

87.15

88.01

87.28
   

0.001

86.30

88.40

88.82

87.23
 

64

0.1

87.14

86.89

88.85

86.24
   

0.01

86.77

88.83

87.53

87.19
   

0.001

87.12

87.68

87.28

87.64
 

128

0.1

86.43

87.91

86.07

86.18
   

0.01

87.02

87.31

86.31

88.53
   

0.001

87.33

87.87

86.65

86.62

The confusion matrix in Figure 3 visualizes the model’s classification performance, trained using the best hyperparameter combination in EfficientNet architecture (a batch size of 32 and a learning rate of 0.01) in three classes. The box plot in Figure 3 depicts the overall distribution of the four performance metrics of the EfficientNet model. This visualization helps easily identify the central tendencies, variances, and outliers of the metrics. The F1 score and precision values, in particular, reveal that the model performs consistently well overall. Confusion matrix and performance distribution of the EfficientNet model. The confusion matrix corresponds to the best-performing configuration (batch size = 32, learning rate = 0.01). The box plot visualizes the distribution of evaluation metrics across different training configurations. Each colored box corresponds to one performance metric (accuracy, precision, recall, and F1-score), illustrating the variability of model performance under different parameter settings.

Figure 3. EfficientNet model confusion matrix and distribution of performance metrics across different training configurations

 

Reviewer#2, Concern# 5: The current labels (Easy, Moderate, Difficult) are clinically sound, but "Difficult" cases are notoriously rare in clinical datasets. Explicitly state how you handled class imbalance (e.g., SMOTE, weighted loss functions, or oversampling). Additionally, consider grouping "Moderate" and "Difficult" into a single "Potential Difficult Airway" class to improve the model's clinical utility as a screening tool.

Author response: We thank the reviewer for this valuable and insightful comment regarding the potential issue of class imbalance and the clinical interpretation of the classification categories. In the revised manuscript, we have clarified how the dataset distribution and potential class imbalance were handled during the model development process. Specifically, data augmentation techniques were applied during preprocessing to increase the diversity of training samples and improve the robustness of the deep learning models. In addition, we have added a discussion in the manuscript acknowledging that difficult intubation cases are relatively rare in clinical datasets and that class imbalance may influence model performance. This limitation is now explicitly mentioned in the revised version. Regarding the reviewer’s suggestion to group the Moderate and Difficult classes into a single "Potential Difficult Airway" category, we appreciate this perspective. However, the primary goal of this study was to develop a model capable of distinguishing between three clinically meaningful levels of intubation difficulty (easy, medium, and difficult) rather than performing binary screening. By maintaining this three-class structure, the proposed model provides a more detailed evaluation of airway difficulty and may offer additional information to clinicians during preoperative assessment. Nevertheless, we acknowledge that binary classification approaches focusing on potential difficult airway screening could be an important direction for future work, and this possibility has been added to the discussion section of the revised manuscript.

After Revision

Discussion

Connor et al. classified different cases of difficult intubation using computerized facial analysis. They aimed to categorize intubation into two classes by employing a logistic regression-based model via numerical analysis of facial proportions. They obtained successful results with 90% sensitivity, 85% specificity, and an area under the curve (AUC) of 0.899. However, their study was conducted with a limited number of male patients (n = 80), and facial morphology was analyzed by reducing it to a 50-dimensional eigenspace [42]. By contrast, EfficientNet architecture, which was used in our study, performs automatic feature extraction directly from face images and performs three-class (easy, medium, and difficult) classification. It should also be noted that difficult intubation cases are relatively rare in clinical datasets, which may introduce potential class imbalance in machine learning models. In this study, data augmentation techniques were applied during preprocessing to increase the diversity of training samples and improve model robustness. Although the current study focuses on a three-class classification framework (easy, medium, and difficult), future studies may also explore binary classification approaches by grouping medium and difficult cases into a single "potential difficult airway" category. Such an approach may further improve the applicability of AI-based systems as clinical screening tools. With an accuracy of 88.64%, a precision of 88.01%, and an F1 score of 87.28% for the best hyperparameter combination, the DL-based approach performed quite well despite distinguishing between more classes. While Connor et al.’s study was performed on a limited sample using traditional statistical methods, the EfficientNet model was successful on more complex classification tasks by using modern AI techniques. This finding suggests that EfficientNet offers a more scalable and powerful solution for clinical decision support systems by enabling a more in-depth structural analysis of facial images…….

 

Reviewer#2, Concern# 6: Clarify the architecture. Did you use a Feature Fusion approach (extracting features from each angle and concatenating them) or an Ensemble of 16 sub-networks? Providing a simplified diagram of the data pipeline would significantly improve the "Methods" section.

Author response: We thank the reviewer for this insightful comment. In the revised manuscript, we have clarified the data processing pipeline and the architecture used in this study. In our approach, the 16 photographs obtained from each patient were not processed using a feature fusion strategy or an ensemble of separate sub-networks. Instead, each image was treated as an independent input sample during model training. A single deep learning model (CNN or EfficientNet) was trained to classify images into three categories: easy, medium, and difficult intubation. Each photograph captured from different anatomical perspectives contributed to the training process by providing additional visual information about airway-related anatomical structures. To further improve the clarity of the methodology, we have added a simplified data processing pipeline description to the Methods section, illustrating the steps from image acquisition to classification.

After Revision

  1. Materials and Methods

The study commenced after approval from the Malatya Turgut Ozal University Clinical Research Ethics Committee (approval no. 2022/07) was obtained. A total of 109 patients who underwent general anesthesia at Malatya Training and Research Hospital Hospital were included in the study. Patients were excluded if they were under 18 years of age, had head and neck tumors, underwent emergency surgeries, had neurological deficits, cervical joint restriction, maxillofacial trauma, cervical vertebra trauma, laryngeal injury, or congenital malformations, were pregnant, or did not provide consent. After written informed consent was obtained from the patients, 16 photographs were taken in the preoperative period in the sitting and lying positions anteriorly, laterally, with the mouth open, with the mouth closed, with the neck straight, and with the neck extended. All images were taken with a smartphone (iPhone 7) (Figure 1). After induction of anesthesia, the patients were positioned on the operating table and intubated using a Macintosh blade size 3–4 by an anesthesiologist. As shown in Table 1, the dataset contains patients categorized into three classes representing different levels of intubation difficulty.

Figure 1. Eight patterns were captured in each of the supine and sitting positions for a total of 16 patterns

Intubations performed without intervention for the first time were considered easy. Intubations with external tracheal intervention and with more than one attempt were evaluated as medium. Intubations requiring more than three attempts; intubation with stylets, fiberoptic bronchoscopes, or video laryngoscopes; or cases in which patients could not be intubated and provided airway with a laryngeal mask were considered difficult. Patients were screened for intubation difficulty using the photographs taken. In this study, the 16 photographs obtained from each patient were treated as independent input samples during model training. The overall data processing workflow used in this study is illustrated in Figure 2.

 

 

Figure 2. Overview of the proposed deep learning pipeline for predicting intubation difficulty.

 A feature fusion strategy combining multiple views into a single feature vector was not applied, and no ensemble of separate sub-networks was used. Instead, a single deep learning architecture (CNN or EfficientNet) was trained to classify the images directly into three intubation difficulty categories (easy, medium, and difficult). Each photograph captured from different anatomical perspectives contributes to the learning process by providing complementary visual information related to airway anatomy. This approach allows the deep learning models to learn discriminative features directly from individual images while benefiting from the diversity of multiple viewpoints. For each patient included in the study, 16 photographs were captured from different anatomical perspectives during the preoperative evaluation. These photographs included anterior and lateral views taken in both sitting and supine positions, with variations such as mouth open, mouth closed, neck in a neutral position, and neck extended. These multiple views were designed to capture different anatomical characteristics relevant to airway assessment. During model training, the images were treated as individual input samples, while the corresponding label for each image was determined according to the clinically observed intubation difficulty level of the patient (easy, medium, or difficult). All images were first resized to a fixed resolution and normalized before being fed into the deep learning models. To improve model robustness and reduce overfitting, data augmentation techniques such as horizontal flipping, rotation, and brightness adjustments were applied.

The dataset was divided into training, validation, and test subsets, enabling reliable performance evaluation of the models. The CNN and EfficientNet architectures were trained using the Adam optimizer with categorical cross-entropy loss, and training was conducted for approximately 20–50 epochs depending on the convergence behavior of the models.

 

Reviewer#2, Concern# 7: If possible, include a side-by-side comparison of your CNN’s predictions against the Mallampati Score or Thyromental Distance measurements taken for the same patients. Proving your model has a higher Positive Predictive Value (PPV) than a physical exam is the strongest argument for bedside mobile phone implementation.

Author response: Thank you for your valuable feedback on our study.

After Revision

We added in material and methods

In our study, patients' Mallampati scores and thyromental distances were recorded. Mallampati scores were found to be 1 and 2 in the easy group, 1, 2, and 3 in the moderate group, and 2 and 3 in the difficult group. Thyromental distances were found to be 5.1 cm to 11.6 cm in the easy group, 5.4 cm to 10.6 cm in the medium group, and 4.9 cm to 9.1 cm in the difficult group.

 

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank for very detailed answers to all my comments. I satisfied with them

Just one issue should be addressed. This paper only considers the advantages of using AI in difficult tracheal intubation prediction. The conclusions should be expanded to include potential general risks and threats of AI in radiology, including those resulting from gender- or race-bias and DL model hallucinations.  I encourage you to consider this topic, citing recent works in this field, e.g.:
10.1038/s41598-025-31231-7
10.3390/diagnostics16030396
10.1186/s40249-025-01339-z

Author Response

Dear Reviewer

I think the source you suggested will contribute to my article.

 

In addition to these studies, the work of Obuchowicz et al. Regarding the use of artificial intelligence has drawn attention to some issues. They stated that in the foreseeable future, artificial intelligence will sup-port physicians rather than replace them, involving the automation of well-defined tasks under human water-pervision, while clinical integration, physical examination, procedural performance, ethical judgment, and accountability will remain dependent on the physician [49].

Reviewer 2 Report

Comments and Suggestions for Authors

Thanks for addressing all the comments.

Author Response

Dear Reviewer,
First of all, thank you for the valuable time you have dedicated to reading my article. Sincerely,