Application of an artificial intelligence-based airway identification system in tracheal intubation

Abstract

Background

Tracheal intubation is an important component in pre-hospital cardiopulmonary resuscitation and trauma resuscitation, which generally requires higher technical skill and systematic training. Due to the lack of sufficient clinical experience among frontline health care providers and the limited availability of airway management tools, establishing advanced airways promptly remains challenging. To address this challenge, we developed an artificial intelligence model designed for real-time identification of airway structures to assist health care providers in quickly mastering tracheal intubation procedures.

Methods

A total of 3912 airway‑anatomical images derived from 978 patients at the Daping Hospital of Army Medical University (January 2024 – July 2024) were included. Each patient’s single static image was rotated three times to generate three additional augmented views, resulting a final dataset of 3912 images. In Part 1, five artificial intelligence target detection models were trained with vocal fissures and aryepiglottic fold as identification targets, and model performance was evaluated based on the 784-image test set using precision, recall, F1 value, and mAP. In Part 2, some of the models were deployed on the end-side of a mobile smartphone, incorporating 72 trainee physicians with no intubation experience, grouped under a standardized procedure to complete tracheal intubation on a simulator using video laryngoscopy and mobile tools, comparing the time to glottic exposure and the first-attempt success rate.

Results

Among the five deep learning models developed, YOLO was the best-performing model, achieving a precision of 94.3% and 62.4% for identifying glottis fissure and aryepiglottic fold structures, respectively, as well as a recall rate of 87.0% and 55.7%, respectively. Map50 was 0.924. Outcomes of Part 2 showed that trainee physicians using the AI‑assisted intubation system had their average intubation time reduced by 11.1 s, and the first‑attempt success rate increased by 19.4%.

Conclusions

The YOLO model can effectively recognize and label airway anatomical structures, serving as a reliable tool to guide operators in real time to complete tracheal intubation. Its deployment on mobile devices emphasizes its clinical practicality and pre-hospital accessibility.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12871-025-03504-1.

Key points

Question

Can artificial intelligence assist health care providers in recognizing airway anatomy in real time and completing tracheal intubation of patients in pre-hospital cardiac arrest?

Findings

We trained a lightweight AI model and deployed it for mobile smartphones to reduce the difficulty of tracheal intubation.

Meaning

By recognizing and prompting airway anatomy and guiding the operator in exposing the acoustic valve area, artificial intelligence can effectively reduce the time required for tracheal intubation and simplify the procedure; mobile deployment further enhances the utility of pre-hospital emergency care.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12871-025-03504-1.

Introduction

Airway management is a crucial component of pre‑hospital cardiopulmonary resuscitation (CPR) and the resuscitation of patients with traumatic shock. In such critical situations, airway management is recommended to initiate initial bag-mask ventilation [1, 2], followed by advanced airway techniques, including endotracheal intubation (ETI) and supraglottic airway (SGA) devices [3, 4]. Recent investigations have explored the relationship between advanced airway strategies and clinical outcomes, yet the existing studies have reported conflicting findings [5–12]. Consequently, the optimal pre‑hospital advanced airway strategy—particularly the choice between ETI and SGA—remains controversial.

The primary advantage of ETI lies in its high and stable ventilation efficiency, rendering it suitable for a broad spectrum of scenarios, especially for patients at high risk of aspiration, those with difficult airways, or those requiring prolonged ventilation. When performed by well‑trained, experienced emergency personnel, the high success rate and rapid execution of ETI can improve the return of spontaneous circulation (ROSC) and neurologically intact survival rates in witnessed cardiac arrest or trauma patients [10–13]. In contrast, SGA devices and their modified versions allow convenient and rapid insertion, and some studies have shown clinical outcomes comparable to or even superior to ETI [5, 7, 9]. Despite these advantages, their clinical application still has limitations. SGA devices may become displaced, with risks of oropharyngeal leakage and increased aspiration. Therefore, they are contraindicated in patients at high risk of aspiration, such as those with a full stomach, obesity, or elevated intra-abdominal pressure [14]. While ETI continues to offer irreplaceable benefits in airway protection, ventilation reliability, and applicability across diverse clinical contexts, it demands higher technical proficiency and typically requires systematic training and workflow optimization [8, 15, 16]. Studies have shown that only 23.4% of out-of-hospital cardiac arrest patients receive effective ETI [17, 18]. Repeated intubation failures can lead to respiratory-circulatory collapse and severe complications such as laryngeal edema and aspiration pneumonia [19]; the risk of these complications rises markedly when the operator is not an anesthesiologist [20]. Consequently, there is a need for an auxiliary endotracheal intubation tool that can achieve a high success rate, reduce procedure time, and ensure patient safety even with limited training. The key to tracheal intubation is adequate exposure to the glottis [21–23]. The widespread use of video laryngoscopes has improved glottic visualization to some extent [24, 25], but beginners still require a certain level of anatomical knowledge. In recent years, the development of artificial intelligence (AI) technology has provided new ideas for optimizing ETI operations, especially AI-assisted recognition of airway anatomical structures and guiding ETI [26–32]. However, existing research still has limitations: Most studies have only included glottic images of patients [26, 28, 29, 31, 32], but recognizing a single glottic anatomical feature is applicable only under ideal conditions where the glottis is already exposed. This approach overlooks the anatomical heterogeneity commonly present in real patients’ airways, and a difficult airway may result in insufficient glottic exposure [33]. In addition, most current AI airway‑recognition solutions rely on large hospital‑based equipment [26, 27, 29, 30], making it difficult to meet the portability and accessibility requirements of pre‑hospital emergency care. To address these issues, this study proposes a hybrid recognition strategy that combines the glottis with the aryepiglottic fold, aiming to provide less‑experienced users with more comprehensive guidance of airway anatomy.

Therefore, this study leveraged real‑patient airway imaging data to develop an AI‑assisted endotracheal intubation system. The system employs a lightweight model architecture that can be deployed on mobile devices, thereby enhancing its practicality and accessibility in pre‑hospital emergency settings. We also performed a preliminary validation of the system’s efficacy using a manikin‑based ETI simulation conducted by trainee physicians.

Materials and methods

General information

This study was conducted in conformance with the Declaration of Helsinki and approved by the Clinical Research Ethics Committee of Daping Hospital of Army Medical University, Chongqing, China (ratification number: 2024 − 204). Retrospective clinical data collection complies with the criteria for exemption from informed consent. This study has been registered at Chinese Clinical Trial Registry (ChiCTR), No. ChiCTR2400092499 (registration date: 2024-11-18). We analyzed 978 adult patients (18–75 years) undergoing tracheoscopy and intubation between January and June 2024. For each patient, one static airway image frame was extracted during standardized glottic exposure attempts. Patients were positioned supine with full neck extension, and images were captured using a UE Classic reusable video laryngoscope (TD-C model) (Figure S4). This dataset was utilized for model training and validation.

Dataset processing

Static airway images (320 × 320 pixels) were uniformly scaled and annotated using the open-source Computer Vision Annotation Tool (CVAT). Three anatomical structures were delineated by experienced anesthesiologists (> 5 years of intubation expertise): aryepiglottic fold, glottis, and composite structure. The label definitions are as follows: Glottis: the sagittal fissure including the structures of both vocal cords; Aryepiglottic fold: the mucosal fold where the tip of the arytenoid cartilage meets the lateral margin of the epiglottic cartilage; Composite structure: the area where the glottis and aryepiglottic fold are visible simultaneously. All annotations employed rectangular bounding boxes compatible with subsequent object detection modeling.

The primary image dataset was partitioned into training (60%, N = 586), validation (20%, N = 196), and test sets (20%, N = 196) using randomized stratification. This allocation strategy optimizes model learning capacity while maintaining evaluation reliability, consistent with established machine learning practices [34]. Stratification by baseline POGO score (< 50% vs. ≥50%) [35] and Modified Cormack-Lehane grade [36] ensured comparable distributions of vocal fold visibility across sets (Table 1). Annotation distribution analysis revealed the following training set characteristics: glottis (N = 486), aryepiglottic fold (N = 569), and composite structure (N = 477) (Fig. 1). To augment dataset robustness, each baseline image underwent 90-degree rotation enhancement, generating three derivative images per original sample [37]. This data augmentation protocol yielded a final curated dataset of 3,912 images for model development.

Table 1.

Image baseline data and airway classification

	Amount (N = 978)	Modify C-L Grade or POGO score
Glottis
Completely invisible	185 (18.9%)	C-L Grade III/IV or POGO = 0%
Partially visible	157 (16.1%)	C-L Grade II or POGO 1–49%
Full visibility	636 (65.0%)	C-L Grade I or POGO 50–100%
Aryepiglottic fold
Completely invisible	46 (4.7%)	C-L Grade III/IV or POGO = 0%
Partially visible	59 (6.0%)	C-L Grade IIb
Full visibility	873 (89.2%)	C-L Grade I/II
Potentially difficult airway
Glottis completely invisible + aryepiglottic fold visible	139 (14.2%)	C-L Grade IIb (Aryepiglottic fold-guided intubation)
Glottis partially visible + aryepiglottic fold visible	157 (16.1%)	C-L Grade II (Reorientation of the field of view)
Glottis completely invisible + aryepiglottic fold completely invisible	46 (4.7%)	C-L Grade III/IV (Other ways of airway management)

C-L Grade Cormack-Lehane Grade, The structures of the larynx visible under laryngoscope exposure. C-L Grade I The glottis is completely visible, Grade II The arytenoid cartilage (posterior wall of the glottis) and the posterior half of the glottis are visible, IIa The posterior part of the glottis is visible, IIb Only arytenoid cartilage or a very small part of the posterior end of the glottis is visible, Grade III Only the epiglottis is visible, Grade IV The epiglottis is not visible, POGO score Percentage of glottic opening during intubation attempt

Fig. 1.

Image label distribution. A Three labels in fully exposed samples. B One label in incomplete exposed samples. C Label type, quantity, and distribution. “x” and “y” represent the position of the label center point within the entire image. “width” and “height” represent the width and height ratio of the label to the entire image. D (0,0) indicates the distribution of the width of the label, (3,3) indicates the distribution of the height of the label, and the rest of the figure shows the relationship between these four variables

Model development methods

Three lightweight single-stage object detection frameworks were prioritized for developing the mobile-compatible airway recognition system: YOLOv5s (Fig. 2), PP-Picodet-s (Figure S1), and SSD-MobileNet-V2 (Figure S2). This selection was driven by their documented capacity to maintain real-time inference capabilities and sustained accuracy under resource-constrained environments [38, 39]. To establish comprehensive performance benchmarks, comparative models employing VGG-16 and ResNet-50 backbones were implemented within the Faster R-CNN architecture.

Fig. 2.

Network structure frameworks of YOLOv5s. conv: Basic feature extraction unit; Concat: Fusion layer feature module; Upsample: Feature map resolution expansion module; SPPF: Key enhancement module

All models were trained using official open-source implementations with ImageNet pre-trained weight initialization. Training employed the SGD optimizer under the following hyperparameters: initial learning rate 2 × 10⁻³, cosine annealing scheduler, batch size 16, and ≈ 300 epochs (exact epoch count determined by early stopping based on loss curve convergence to prevent overfitting; Figure S3). Preliminary validation identified YOLOv5s as optimal for deployment due to its precision-speed balance. This model was converted to Open Neural Network Exchange (ONNX) format and integrated into an Android prototype application. Real-time video acquisition utilized a compact endoscope (DY68-L1-V26) connected via USB OTG interface.

System implementation

The system is organized into three layers—device capture, inference execution, and visualization—optimized for mobile-embedded platforms and supporting ONNX or Paddle model formats (Figure S5).

The design priorities are low latency, resource reuse, and robust peripheral handling. Layered architecture (capture → pipeline → inference → render) enables decoupling and easy backend replacement. Buffer pooling and native processing minimize allocations and Java Native Interface (JNI) overhead, reducing latency and improving stability. Decoupled inference, post-processing and rendering with robust USB Monitor lifecycle handling ensure real-time performance and resilience in mobile-embedded deployments.

Validation

Part 1

All models underwent standardized evaluation on an independent test set (N = 784). Detection performance was quantified using Intersection over Union (IoU), defined as the ratio of overlap between predicted and ground-truth bounding boxes to their union area [40]. Predictions exceeding IoU thresholds (range: 0.5–0.95) were classified as true positives (TP), otherwise as false positives (FP). Primary performance metrics (Precision, Recall, F1-score) were derived from confusion matrices generated at IoU = 0.5. These metrics were calculated as follows:

The global metrics use two types of mean accuracy averages (mAP): mAP₅₀: the average of the areas under the Precision-Recall curves for each type of category under the IoU = 0.5 threshold; mAP_50:95: the average of the areas under the Precision-Recall curves for each type of category under the IoU threshold from 0.5 to0.95 (in steps of 0.05) at 10 levels [14].

To assess runtime performance on resource-constrained devices, a 20-second intubation video captured at 30 Frames Per Second (FPS) was processed through the Android prototype application on a Qualcomm Snapdragon 865 platform. The test was conducted with the phone stationary, the screen on, and the video stream looped for ten inferences. Frame rate, end-to-end processing latency, and runtime memory usage were monitored and averaged.

Part 2

Seventy-two first-year resident physicians without prior intubation experience were enrolled after applying exclusion criteria (anesthesia/critical care/emergency medicine trainee physicians, specialists, or physicians with > 1-year clinical experience). Participants underwent block randomization via random number table assignment to:

• Group A (N = 36): Video laryngoscopy (UE Classic reusable video laryngoscope TD-C)

• Group B (N = 36): AI-assisted video laryngoscopy (Figure S4)

Sample size was determined a priori using G*Power 3.1 (paired t-test; α = 0.05, power = 0.8, CI = 95%) based on glottic exposure time.

All participants received standardized 30-minute didactic training on intubation fundamentals. Group B received supplementary 10-minute training on AI device operation. Procedures were performed on identical airway manikins (Laerdal Airway Management Trainer), with proficiency assessed pre-intervention using Likert scales.

The primary endpoint was the successful placement of the endotracheal tube into the simulated airway or the number of attempts reaching three. The primary outcome measures were first-attempt success rate (endotracheal tube placement confirmed within ≤ 3 attempts) and total intubation time (laryngoscope insertion → confirmed tube placement), which were compared between groups using chi-square tests and independent samples t-tests, respectively. The secondary outcome was the time to glottic view (laryngoscope insertion → clear vocal cord visualization), which was compared between groups using an independent samples t-test.

Results

Outcomes of the part 1

All results are based on the same test set of images (Fig. 3). The YOLO model showed the highest precision rate (Precision = 0.943) and achieved a better balance between precision and recall (F1 = 0.905) under the target detection task of glottis. Compared to the large-scale model, the lightweight model is more advantageous in such small target detection tasks with fixed features. Under the target detection task of the aryepiglottic fold, the SSD-MobileNet-v2 model had the highest precision rate (Precision = 0.875). However, it had a poor balance between precision and recall (F1 = 0.648), with low feature discrimination. PP-Picodet achieved a better balance of recall (F1 = 0.667) while maintaining a high precision rate (Precision = 0.710). The better performance of the large-scale model in this task may be attributed to the presence of different degrees of anatomical variability in the aryepiglottic fold, as well as the finer granularity of feature extraction in the large-scale model, which resulted in improved performance on the test set. Under the overall target detection task, the YOLO model achieved a higher overall detection accuracy compared to other lightweight models, with an overall mAP₅₀ of 0.924 and a mAP_50:95 of 0.585, resulting in less performance degradation compared to the large-scale model, and specific results are included in the supplementary materials (Table S1).

Fig. 3.

Performance of various models on the test dataset. A Model performance under the glottis target detection task; B Model performance under the aryepiglottic fold target detection task; C Overall average precision. mAP: mean Average Precision, mAP50 represents the average precision of the model when the IoU threshold is set to 0.5. mAP50:95 represents the average precision of the model within the IoU threshold range of 0.5 to 0.95

Although the model produced negative results in some aryepiglottic fold samples, this was mainly due to anatomical variations (edema of the spatulate cartilage folds). However, in actual cases, this only affected a single frame of the video stream, and the impact would be reduced during video detection. Deploying YOLOv5s on a mobile device with a 320 × 320 pixels input resolution, the model achieved an average single‑frame inference latency of 60.4 ± 3.03 milliseconds (16.60 ± 0.94 FPS) and an end‑to‑end latency of 78.6 ± 4.98 milliseconds (12.78 ± 0.88 FPS) when the device temperature was ≤ 40 °C. The memory usage during inference was 134 ± 3.3 MB.

Outcomes of the part 2

The results (Table 2) showed no statistically significant difference in baseline characteristics between the two groups of participants and were comparable (P > 0.05). When performing tracheal intubation on a manikin using different intubation tools (Table 3), group B (AI laryngoscope group) demonstrated superior tracheal intubation performance compared to group A (video laryngoscope group), with significantly shorter average intubation times (43.6 ± 9.1 s vs. 54.7 ± 11.8 s, p < 0.001) and higher first-attempt success rate (94.4% vs. 75%, P < 0.05).

Table 2.

Comparison of preoperative proficiency in tracheal intubation between the two groups of participants

perception	Groups	Strongly disagree	Disagree	Not sure	Agree	Strongly agree	P- value
I was very familiar with tracheal intubation.	A	9(25.0)	24(66.7)	3(8.3)	0(0)	0(0)	0.531
	B	11(30.6)	23(63.9)	2(5.6)	0(0)	0(0)
I knew a lot about airway structure.	A	5(13.9)	18(50.0)	13(36.1)	0(0)	0(0)	0.454
	B	9(25.0)	15(41.7)	12(33.3)	0(0)	0(0)
I can independently perform tracheal intubation.	A	7(19.4)	16(44.4)	13(36.1)	0(0)	0(0)	0.695
	B	3(8.3)	21(58.3)	12(33.3)	0(0)	0(0)
I have got a lot of theoretical knowledge.	A	5(13.9)	21(58.3)	10(27.8)	0(0)	0(0)	0.510
	B	8(22.2)	12(33.3)	16(44.4)	0(0)	0(0)

Table 3.

Performance of two groups of people in operating a tracheal intubation manikin. Group A: tracheal intubation using a video laryngoscope; Group B: tracheal intubation using an assisted laryngoscope with an artificial intelligence program

	Group A (N = 36)	Group B (N = 36)	P-value	RR (95% CI)
First-attempt success rate, n (%)	27 (75.0)	34 (94.4)	0.022	1.259 (1.026, 1.545)
Total time (s)	54.7 ± 11.8	43.6 ± 9.1	<0.001
Glottic exposure time (s)	36.7 ± 12.1	26.5 ± 9.4	<0.001

Discussion

This study demonstrated that the AI model using the combined detection strategy of glottis and aryepiglottic fold can effectively identify airway openings and thus assist in the completion of ETI. The optimal model achieves an average precision of 94.3% for detecting glottis structures, as well as an average precision of 62.4% for detecting aryepiglottic fold structures. Simulation validation showed that the AI‑assisted intubation system can shorten the time required for tracheal intubation and improve the success rate of first-attempt intubation in trainee physicians with no prior intubation experience in a manikin model.

The key points of tracheal intubation are the determination of the glottic opening and the exposure of the glottic field. Kim et al. developed the YOLOv4 model [28] based on 20,161 images from 84 patients to identify vocal folds quickly identify vocal folds when the lens field of view is obstructed due to sputum and blood soiling, with an accuracy of 90.4%. However, in many cases, lens soiling only accounts for a portion of poor vocal fold exposure, and more poor vocal fold exposure in emergencies originate from physiological and anatomical variants [19]. And about 23.7% of the cases of first-time intubation failure originated from the insufficient exposure of the glottic field [33], whereas the incidence of difficult airway in tracheal intubation outside of the operating room ranges from 11% to 50% [41, 42]. In contrast, the aryepiglottic folds can be well exposed under visual laryngoscopy, and sometimes, part of the structure of the aryepiglottic fold can still be seen when the root of the epiglottis cannot be lifted [33]. It constitutes the lower wall of the glottic opening, so operators can judge the approximate direction of the glottic opening by the aryepiglottic folds and adjust the direction of the laryngoscope and the tracheal tube to expose the glottic field adequately. This is more helpful for non-experienced health care providers, as they often have difficulty achieving ideal glottis exposure.

Wu et al. decomposed 30 straightforward intubation videos into 887 static images to develop a YOLOv3 model [27]. By identifying six key anatomical structures (lips, epiglottis, laryngopharynx, glottis opening, tube tip, and the black line on the tracheal tube), they achieved temporal sequence analysis of the intubation process. The overall recognition accuracy of the model was 92.4%, which increased to 99.4% after excluding the poorly recognized tube tip target. This demonstrates the potential application of artificial intelligence in assisting tracheal intubation. However, the study’s small sample size may limit its ability to fully capture the complex anatomical variations encountered in real clinical practice. Furthermore, the model validation was retrospective, performing only “post-hoc” analysis on existing videos; thus, the model’s extrapolation and validity require further verification. Majeedi et al. utilized 1,623 frames from 84 neonatal intubation videos to train a YOLOv8 model specifically designed to identify four key anatomical structures [26]: the uvula, esophagus, epiglottis, and glottis opening. However, their model was targeted solely at the neonatal population and similarly underwent only “post-hoc” validation on intubation videos.

Unlike previous studies, this study developed five artificial intelligence models for the specific pre-hospital emergency scenario and detection structures, and selected the optimal YOLOv5s model. This approach primarily considers that different models may exhibit performance differences due to variations in detection targets and application scenarios, thereby potentially overlooking the superior performance of certain models under specific conditions. This study places particular emphasis on the performance evaluation of lightweight models, as lightweight models are suitable for compact and portable devices, thereby reducing the likelihood of usage restrictions caused by environmental or device-related factors and demonstrating stronger applicability in pre-hospital settings [32, 43]. Part 2 evaluated the feasibility of the algorithm in a simulated environment. Although the models are sufficiently lightweight, conventional video laryngoscopes lack the computational capacity required for inference; therefore, we deployed the model on a mobile device. This deployment not only alleviates the difficulty of large‑scale dissemination but also leverages the smartphone’s capability to transmit video streams, offering potential for tele‑medicine applications. This exploratory validation provides a degree of guidance for future application development.

The study has some limitations. Firstly, the small sample size may limit the generalizability of findings, and the precision of the model in identifying aryepiglottic fold structures needs to be further improved. In Part 1, baseline admission characteristics were not accounted for in the imaging population, potentially overlooking airway anatomy associated with different Clinical data on the patients. In addition, the study was performed only on a manikin airway model that replicates human anatomy but does not account for secretions, tissue mobility, and other factors present in a real airway. All participants lacked laryngoscopy experience, so the results should not be directly extrapolated to experienced clinical operators, and future work should validate the tool in clinical, real‑airway settings to assess its effectiveness for experienced operators.

Conclusions

Among the five AI models evaluated on the clinical intubation image database, YOLOv5s achieved the highest precision in recognizing airway anatomy. This result confirms that YOLOv5s is the most reliable model for real‑time anatomical recognition in actual clinical settings. Simulation-based validation performed on 72 first‑year trainee physicians who had no prior experience with tracheal intubation demonstrated the model’s ability to reduce procedural difficulty for novices. By providing real‑time recognition of airway anatomy, the system assists the operator in exposing the glottic field of view, thereby substantially reducing the procedural difficulty for novices. Deployment of the intelligent program on the mobile side increased the popularity of such a tool, which provides a basis for the AI technology to reduce the threshold of tracheal intubation operation for health care providers in the pre-hospital emergency care scenarios provides a rationale, suggesting its potential value in promoting the popularization of first aid skills.

Authors' contributions

Huanqing Liu and Xu: Formal Analysis, Software, Validation, Visualization, Writing - Original Draft. Duan and Kaijun Liu: Data Curation, Resources. Xiaona Wang and Wu: Investigation. Tang and Zhen Wang: Conceptualization, Funding Acquisition, Methodology, Project Administration, Supervision, Writing - Review & Editing.

Funding

Financial support and sponsorship: Supported by Chongqing Medical Key Discipline Construction Fund (Grant No. zdxk202102).

Data availability

The datasets generated and analysed during the current study are not publicly available due to the special nature and privacy of anatomical images, but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in conformance with the Declaration of Helsinki and approved by the Clinical Research Ethics Committee of Daping Hospital, the Third Affiliated Hospital of the Army Medical University, Chongqing, China (ratification number: 2024 − 204). Since patient data were obtained from medical records and had been de-identified prior to their processing and evaluation, the need for informed consent was waived by the ethics committee of Daping Hospital. Registered at Chinese Clinical Trial Registry (ChiCTR, https://www.chictr.org.cn), No. ChiCTR2400092499 (registration date: 2024-11-18).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.