Emerging technologies in airway management: a narrative review of intubation robotics and anatomical structure recognition algorithms

Abstract

In recent years, the medical field has seen significant advancements in the field of robotics and artificial intelligence (AI). However, many healthcare professionals still find these technologies unfamiliar and complex, especially regarding their use during airway management. This review covers the current capabilities of robots and AI in tracheal intubation (TI), providing new insights that advocate for the broader adoption of these technologies to improve airway management. A literature review on robotics and AI in TI was conducted through searches in the PubMed, Web of Science, and IEEE Xplore databases. Drawing on a classification framework derived from expert opinions and existing literature, these studies are categorized into six key stages. Most of these technologies remain in the testing and validation phases, with only a few having reached commercialization. The primary goal of these robotic and AI systems is to enhance the success rate and operational efficiency of intubation while mitigating the persistent shortage of medical resources and supporting telemedicine. However, ongoing attention is required to address challenges such as high costs, a shortage of interdisciplinary talent, and ethical concerns related to medical bias and data security. Robots and AI are beginning to play a significant role in TI. Although many of these technologies remain in the theoretical stage of clinical application, their potential to enhance clinical practice is substantial, provided they are implemented as complementary tools that support rather than substitute the expertise of healthcare professionals. AI-powered robots show great potential as assistive tools for optimizing intubation maneuvers, whereas clinical decision-making (e.g., determining the necessity of intubation) remains under the supervision of physicians.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12938-025-01408-2.

Introduction

The Fourth Industrial Revolution (4IR) is characterized by groundbreaking technological innovations, such as robotics, artificial intelligence (AI), big data, the Internet of Things, and blockchain [1]. This revolution is defined by the convergence of mathematics, biology, and physics, shaping the operations of various industries and affecting nearly every aspect of life [2]. During the COVID-19 pandemic, the 4IR has been crucial in transforming the medical field, primarily through advancements in robotics and AI technologies [3]. Robots, which are mechanical devices that automatically perform human functions, have assisted in surgeries, support rehabilitation, and deliver medications [4]. The medical robotics market is projected to reach $43.32 billion by 2028 [5]. AI, defined as the capability of machines to mimic human cognitive functions, has become widely integrated into healthcare for applications like object recognition, natural language processing, and clinical decision support [6]. The AI healthcare market is projected to reach $188.0 billion by 2030 [7].

Each year, more than 320 million patients worldwide undergo surgery, with approximately 30% of these procedures in the United States involving tracheal intubation (TI) [8]. This significant number highlights the critical role of TI. In addition to clinical diagnostics, robotics and AI technologies have gained increasing traction in airway management, especially in TI [9, 10]. Properly controlled robotic arms can provide more stability and precision in operations compared to human hands, this precise control enhances the accuracy and flexibility of TI, thus reducing associated risks [11, 12]. Moreover, robots are immune to fatigue and capable of performing repetitive tasks indefinitely, thereby helping to mitigate the ongoing shortage of medical personnel [13]. AI algorithms can effectively identify key anatomical structures along the intubation pathway, providing operators with real-time guidance and feedback. This enhances the ability to manage complex airway, improving both intubation success rates and operational efficiency. Furthermore, AI can detect tracheal tube misplacements, thereby reducing the risk of complications like asphyxia, hypoxemia, and pulmonary aspiration [14–16].

This review aims to examine advancements and achievements in robotics and AI technologies in TI, which are essential for updating healthcare professionals’ and researchers’ knowledge, integrating technological innovations, and guiding industry development [17, 18]. Based on an analysis of existing literature, this review categorizes the development of intubation robots and anatomical structure recognition algorithms in TI into three stages (Table 1), providing a clearer summary of the current status of robotic and AI-assisted TI.

Table 1.

Developmental stages in intubation robotics and anatomical structure recognition algorithms

Developmental stage	Robot	AI algorithms
Stage 1	Combining commercial robotic arms with airway management tools to achieve robot-assisted TI	Using stable airway environment (such as mannequin) to obtain datasets and train AI algorithms to recognize key airway anatomical structures
Stage 2	Independently developing components of the robots with customization and flexibility, aiming for a more compact structure	Using complex airway environment (such as emergency, pre-hospital, and difficult airways) to obtain datasets and develop anatomical structure recognition algorithms
Stage 3	Integrating airway recognition, intelligent navigation, and automatic intubation into the robot to create an intelligent intubation system	Researching the hardware deployment of anatomical structure recognition algorithms developed in both stable and complex airway environment

Methods

We conducted a comprehensive literature search across PubMed, Web of Science, and IEEE Xplore using the retrieval strategy described by Hashimoto et al. [19]. The search terms included: ((robotics) OR (artificial intelligence)) AND ((tracheal intubation) OR (airway management) OR (vocal cord recognition)). The search was limited to English-language articles published between January 1, 2010, and December 31, 2024. Studies were eligible for inclusion if they specifically examined the design or application of robotic intubation systems or anatomical recognition algorithms in the context of airway management. Eligible study types included systematic reviews, clinical trials, animal studies, observational studies, and peer-reviewed conference proceedings. We excluded editorials, letters to the editor, and studies available only in abstract form. Additionally, studies focusing exclusively on airway anatomical recognition without direct relevance to airway management, such as those analyzing quantitative vocal cord motion or glottic midline detection, were also excluded. A single reviewer (W.C.) initially screened titles to remove irrelevant articles. The same reviewer then assessed abstracts of potentially relevant studies, with uncertainties resolved through discussion with a second reviewer (Y.T.). To supplement the database search, manual citation tracking was performed by reviewer W.C. to identify additional relevant studies. From an initial pool of 1534 potentially relevant titles (PubMed: 942; Web of Science: 466; IEEE Xplore: 126), eight additional studies were identified through manual citation tracking. After title screening, 566 abstracts were reviewed, and 188 full-text articles were assessed for eligibility. Ultimately, 28 manuscripts met all inclusion criteria and were included in the final analysis (Fig. 1). A detailed list of included studies, their classifications, and those identified through manual searches is provided in the supplementary material 1.

Fig. 1.

Literature search flowchart

The advancement of tracheal intubation robots

In developing robot-assisted airway management, research has primarily focused on intubation robots. These robots capitalize on their precision, stability, automation, and visualization capabilities, integrating these strengths with the clinical expertise of medical personnel to facilitate TI. This integration paves the way for safer, more efficient airway management (Fig. 2). This section reviews and analyzes the development of robot-assisted TI through the stages of technological evolution (Supplementary Table 1).

Fig. 2.

Workflow for utilizing intubation robots to improve the safety and efficiency of airway management practices

Stage one: integrated tracheal intubation robot

Four studies assessed robotic-assisted or semi-automated TI techniques, with a primary focus on enhancing procedural success rates, minimizing operator dependency, and facilitating remote or semi-automated execution. In 2010, Tighe et al. first demonstrated the feasibility of first robotic-assisted intubation using the da Vinci robot (Type S). In this study, two separate robotic arms held a Karl Storz fiberoptic bronchoscope and a camera, while the other two robotic arms the bronchoscope. They successfully completed oral intubation on the mannequin, in an impressive time duration of 75 s likewise the nasal intubation took 67 s [20]. In 2012, Hemmerling et al. developed the Kepler Intubation System (KIS), which consists of a ThrustMaster T.Flight Hotas X joystick, a JACO robotic arm, a Pentax AWS videolaryngoscope, and a software control system. They conducted 90 intubation experiments on mannequins and clinical trials with 12 patients with normal airways. The team was able to achieve 100% success rate in a average time duration of 46.7 s, while clinical trials averaged 57 s with a 91% success rate (one failure). Unlike Tighe’s setup, the KIS was specifically designed for TI and separated the physician's control end from the execution end, demonstrating the feasibility of remote robot-assisted intubation [21, 22]. In 2016, Leng et al. developed a TI robot similar to the KIS, using the Omega.7 master hand for the physician's end and integrating a 6-degree-of-freedom (DOF) commercial UR robotic arm and laryngoscope for the execution end. Their research focused on optimizing parameters related to joint motion interference [23], this study emphasized kinematic analysis and highlighted several issues that need to be addressed before clinical application, including workspace limitations, flexibility, and the safety of human–machine interactions.

In the early development of TI robots, designs often integrated robotic arms with airway management tools, such as laryngoscopes and bronchoscopes, and were tested in structured airway environment. However, due to their large size and limited flexibility, these robots were less efficient than physicians, making clinical application challenging.

Stage two: modular tracheal intubation robot

Five studies investigated robotic-assisted TI technologies, emphasizing improved intubation success rates, operational efficiency, and enhanced safety for both patients and clinicians. In the second phase of TI robot development, modularity and miniaturization emerged as prominent trends. In 2018, Pan's team developed the Remote Robot-Assisted Intubation System (RRAIS), measuring 8 inches long, 8 inches wide, and 4 inches high. This system adjusted the position and orientation of the tracheal catheter using a self-developed 3-DOF component, it also included a tongue depressor with rotational functions for better glottis exposure and catheter insertion. Researchers recruited 10 medical students to perform basic experiments on 20 Bama miniature pigs, achieving an intubation time of 74.6 ± 2.3 s, a first-attempt success rate of 80%, and an overall success rate of 90% [24], marking the beginning of mini customization TI robots. Similarly, in 2018, Cheng et al. developed a TI robot named IntuBot, featuring three active DOF for controlling the propulsion, bending, and rotation of a custom-made flexible segment. The robot used a Raspberry Pi as the main control computing unit and integrated all electronic components via a custom-printed circuit board, researchers used CT scan images to create 3D silicone airway models for performance testing [25]. However, this device only underwent prototype testing, with no further publication. For manual TI scenarios, Hernandez-Hinojosa et al. developed a bionic exoskeleton intubation assist device that extends finger length, with the distal bending angle of the robotic fingers ranging from 0 to 90 degrees. This device reduced the normal force applied to the patient during intubation, decreasing the incidence of complications such as soft tissue trauma and dental damage [26]. However, with the rapid development of airway management tools like video laryngoscopes and video bronchoscopes, manual TI has become rare in clinical practice, indicating the limited potential and application prospects of this robot. In 2023, Liu et al. proposed a compact nasotracheal intubation robot system, characterized by miniaturization, lightweight design, and low power consumption. The system includes a quick-release structure, reducing usage costs and enabling sterile intubation. It also features a flexible segment with an endoscope integrated at the tip, capable of manually adjusting its angle and curvature based on visual feedback, enabling it to navigate the complex curvatures of the nasal cavity. The system is driven by a Jetson Nano core control unit, which provides graphics processing unit acceleration and forms the foundation for the development and deployment of AI algorithms [11]. In 2024, Qi et al. developed a master–slave robotic system for endotracheal intubation, employing an Omega3 platform as the master console and a dual-slide rail mechanism as the slave manipulator, with a research focus on motion stability and safety [27].

At this stage, researchers recognized the limitations of large robot sizes in clinical applications. To enhance flexibility and convenience, the structure of the robot was predominantly self-designed, resulting in prototypes significantly smaller than in the previous phase.

Stage three: intelligent tracheal intubation robot

Seven studies explored robotic-assisted or automated TI technologies integrating soft robotics, visual servo control, and AI. These studies highlighted the potential for precise autonomous navigation, decreased human intervention, and enhanced clinical safety. In the third phase, integrating AI algorithms has become crucial for achieving high precision and intelligent operation. For instance, AI can accurately identify key anatomical structures such as the uvula, epiglottis, and glottis, enhancing airway visualization and helping the robot precisely locate the glottis. Additionally, AI empowers the robot with visual navigation capabilities, enabling it to quickly determine the correct intubation path in complex airway environment, thereby reducing intubation time and enhancing its intelligence. In 2020, Biro et al. developed the automated TI robot, REALITI. This robot featured a 2-DOF endoscope with 24 interlocking joints, where the bending angle between joints ranges from 2.3° to 18.2°, with curvature gradually increasing toward the tip to allow flexible navigation. The robot offered both manual and automatic control modes. In manual mode, the user controlled the endoscope's direction, while in automatic mode, the robot adjusted the bending direction of the endoscope tip based on identified key anatomical structures, guiding it toward the glottis [28, 29]. The significance of this work lies in it being the first automated TI robot, addressing the limitations of traditional intubation methods that heavily rely on the physician's experience and skills. In 2023, Deng et al. introduced the RNIS TI robot, a 3-DOF nasotracheal intubation device capable of manipulating a commercial flexible endoscope to perform advancing, rotating, and bending movements. The RNIS robot integrated a lumen center detection method based on the Kalman filter, enabling precise navigation while effectively filtering out detection errors caused by noise, occlusion, and other interferences. This allowed the RNIS to autonomously perform nasotracheal intubation in complex and dynamic environments [30, 31]. The key innovation of this system lies in its "model-free" fuzzy logic control strategy, which dynamically adapts control parameters in real-time to handle uncertainties in the airway environment, enhancing robustness and pioneering new directions in control strategies. In the same year, Luo et al. developed a 3-DOF orotracheal intubation robot that uses a commercial bronchoscope as its end effector. The robot featured force feedback functionality enabled by a magnetic powder clutch and force sensor, ensuring the endoscope tip does not collide with the airway wall. Additionally, the robot is equipped with real-time visual feedback, combined with depth estimation and trajectory planning algorithms. These capabilities allow the robot to calculate the optimal intubation path in airway environment, automatically identify and mark key anatomical structures, thus assisting automatic intubation. All volunteers successfully performed intubation on an mannequin within 2 min, meeting clinical operation time requirements [32]. In 2024, Liu et al. developed a TI robot powered by a hydraulic actuation system. The researchers designed a 2-DOF silicone actuator reinforced with fibers and spiral steel wire, ensuring improved flexibility and safety. The robot utilized a classical neural network to identify anatomical features, achieving a precision rate of 96.67% [10]. Despite its innovative structural design, the robot’s algorithm exhibited limited generalization capability. In 2024, Lai introduced a two-segment cable-driven intubation robot that incorporated a novel Sim-to-Real method to guide its actuator. This approach allowed the robot to autonomously navigate to target locations while minimizing environmental collisions [33]. Lai's research underscores the significant integration of robotic technology in TI, presenting a feasible solution for safe human–robot interaction during such procedures.

At this stage, TI robots incorporate advanced control algorithms, such as visual feedback and fuzzy logic control strategy. Most robots now feature visual navigation capabilities, supported by the development of anatomical structure recognition algorithms. As a result, TI robots are progressively advancing toward higher levels of intelligence.

Development of anatomical structure recognition algorithms for tracheal intubation

AI is extensively applied in endoscopic recognition, particularly in tasks such as classification, detection, and segmentation [34]. AI algorithms can extract video or image data during TI, develop anatomical structure recognition algorithms (Fig. 3), and offer prompts and directional guidance to physicians. This aids novice doctors in overcoming hand–eye coordination challenges due to inexperience and enhances senior physicians' capabilities in difficult intubation cases. This section reviews and discusses the development of anatomical structure recognition algorithms for assisting TI, following the three defined stages (Supplementary Table 2).

Fig. 3.

This neural network diagram depicts a hypothetical system intended to demonstrate the process of recognizing key anatomical structures in the airway from laryngoscope images (Given the thousands of nodes in each layer, the diagram does not display all nodes and connections)

Stage one: anatomical structure recognition algorithms based on stable airway environment

Three studies evaluated the application of AI for real-time identification and labeling of airway anatomical structures, including the epiglottis, glottis, vocal cords (VC), and tracheal rings, in video footage from laryngoscopy and bronchoscopy procedures. In 2016, Carlson et al. pioneered the development of an AI algorithm for recognizing glottis. By training four different machine learning models on a video laryngoscopy (VL) dataset from a mannequin, all four algorithms achieved over 70% accuracy in identifying the glottis. Among them, the SVM model performed best, with an accuracy of 80%, sensitivity of 70%, and specificity of 90% [35], marking a significant milestone in developing anatomical structure recognition algorithms. In 2020, Matava et al. advanced this work by validating the capability of convolutional neural networks (CNNs) to classify and recognize VC and tracheal rings. They collected VL and bronchoscopy recordings from 775 patients aged 1 to 23 years to train three CNNs: ResNet, Inception, and MobileNet, and introduced frames per second (FPS) as a key metric for evaluating performance. The study found that the Inception model performed the best, processing real-time video at 10 FPS and accurately classifying and recognizing VC and tracheal rings, with sensitivity, specificity, and overall confidence scores of 0.865, 0.971, and 0.802, respectively [36]. Despite its strong performance, the Inception model's complexity and high resource demands render it less suitable for lightweight, low-computation environments. For low-power portable devices, the more lightweight MobileNet model may be a better choice. In 2024, Masumori et al. proposed a novel cascade architecture that combines YOLO’s real-time object detection with U-Net’s medical image segmentation capabilities to develop an AI-assisted system for VL. This system facilitated rapid localization and precise segmentation of the epiglottis, VC, and glottis during simulated endotracheal intubation, achieving an Intersection over Union (IoU) exceeding 95% on a mannequin [37]. The YOLO module enabled real-time processing (0.14 s per image) for initial anatomical localization, while the subsequent U-Net module refined segmentation to compensate for YOLO’s limitations in capturing fine anatomical details. This hierarchical approach optimized computational efficiency by circumventing resource-intensive full-image segmentation at high resolutions. However, despite the accurate and comprehensive recognition of airway anatomical structures achieved by the YOLO-UNet cascade architecture, the study is limited by its small sample size of 920 bronchoscopic images derived from a single mannequin. This dataset does not sufficiently represent the anatomical variability and challenging conditions encountered in real-world clinical settings, including oral secretions, bleeding, tumors, and light reflections during procedures. Furthermore, as the video data were acquired from bronchoscopy rather than VL, the intended application, the discrepancy between the validation environment and real-world clinical conditions may lead to an overestimation of the model’s performance.

In the initial stages of developing anatomical structure recognition algorithms, videos or images from a stable airway environment were used as inputs. These data, sourced from mannequins or normal airways, were characterized by clearly defined anatomical structures, uniform morphology, and easy accessibility. While using such data may limit the model's generalization and robustness, it allows for efficient algorithm iteration and validation during the initial development phase. This approach helps establish performance benchmarks for future applications in more complex airway environment.

Stage two: anatomical structure recognition algorithms based on complex airway environment

Three studies examined advanced CNN models for robust identification of airway anatomical landmarks under complex clinical conditions, including emergency scenarios with blood, secretions, or challenging patient positioning. In the second stage of developing anatomical structure recognition algorithms, AI models trained on stabilizing airways environment showed limitations in scenarios like difficult airways, prehospital care, and emergency departments (EDs), For example, in cases involving difficult airways and emergency patients, airway abnormalities, pathological changes, and obstructions can cause models to struggle with accurately identifying anatomical structures. To overcome these limitations, three key studies emerged. In 2021, Yoo et al. developed an AI algorithm designed to identify the anatomical locations of the carina and bronchi. To enhance the algorithm's robustness to rotation and occlusion, they randomly rotated and cropped bronchoscopic images into circular patterns, simulating complex airway environments. After evaluating ten lightweight deep learning (DL) models, EfficientNet-B1, the model with the highest accuracy, was selected as the final choice, achieving an accuracy of 0.84, comparable to that of respiratory specialists [38]. However, despite the enhanced dataset, excluding cases with foreign bodies, tumors, excessive mucus, and bleeding may reduce the accuracy and robustness in such complex scenarios. In 2023, Choi et al. evaluated the performance of three image segmentation models on the tongue, epiglottis, VC, and corniculate cartilage (CC). A key strength of this study was its collection of VL data from an ED, which included images with blurred views, motion blur, blood, vomit, saliva, and mucus, reflecting the unpredictable conditions during TI. The study found that the DeepLabv3 + and U-Net models, when combined with EfficientNet-B5, performed well in segmenting the tongue, VC, and CC, achieving Dice Similarity Coefficients (DSCs) of 0.3351 (DeepLabv3 +), 0.766 (DeepLabv3 +), and 0.6906 (U-Net), respectively. For epiglottis segmentation and FPS, Mask R-CNN achieved the highest values, with a DSC of 0.7677 and an FPS of 32. Due to its higher FPS, Mask R-CNN is more practical for future integration into VL systems compared to the other models [39]. This study was the first to develop a DL algorithm for segmenting airway anatomical structures in emergency scenarios and the first to use masks for labeling airway structures, ensuring accurate IoU calculations by avoiding unnecessary area interference. Also in 2023, Kim et al. utilized YOLOv4, based on the Darknet framework, to identify VC structures in clinical emergency scenarios. The research team collected VL data from the ED and enhanced the dataset using random probability techniques to better reflect clinical practice. The test set was divided into four groups: cardiopulmonary resuscitation (CPR), visual difficulty (VD), CPR with VD, and others, to thoroughly evaluate the model's performance under different challenging conditions. The algorithm achieved a sensitivity, specificity, positive predictive value, negative predictive value, and F1 score of 0.963, 0.842, 0.958, 0.902, and 0.906, respectively. Further analysis revealed that the algorithm performed worst in the CPR group but showed the highest performance in the other groups [40]. Compared to Choi et al.'s study, this study used bounding boxes to label VC structures: a method more prone to interference, resulting in less accurate IoU calculations. Additionally, the lack of direct comparison with existing algorithms makes it difficult to fully assess this algorithm's strengths and weaknesses relative to other AI models.

At this stage, researchers started incorporating more complex airway factors, such as blurred vision, airway stenosis, anatomical abnormalities, and large volumes of mucus. This allowed the models to adapt to and manage a variety of complex scenarios, thereby enhancing their generalization and robustness. Furthermore, since 2021, the development of anatomical structure recognition algorithms has adopted more advanced models like EfficientNet-B5, Mask R-CNN, and YOLOv4, which feature greater depth and more complex architectures, enhancing performance while also reducing computational resource consumption.

Stage three: deployment of anatomical structure recognition algorithms onto hardware

Three studies investigated the integration of anatomical recognition algorithms into portable intubation devices to enable real-time procedural guidance during airway management, demonstrating feasibility and potential clinical benefits. In the third stage, deploying AI on intubation devices has not only enhanced visualization and decision-making capabilities but also allowed for testing and validating these models' accuracy and robustness in real-world environments. This process has accelerated the maturation and application of AI in this field. In 2020, DK et al. were the first to deploy a DL algorithm for VC recognition on hardware, creating a Raspberry Pi-based prototype for real-time VC recognition [41]. While this prototype marked an important step toward intelligent intubation devices, it was limited to a camera, Raspberry Pi, and display screen, and did not constitute a complete intelligent intubation system. That same year, Boehler et al. were the first to deploy a machine learning algorithm on an intubation robot. This device used a Haar Cascade Classifier to automatically recognize four key anatomical structures in the airway and employed visual servo control, enabling automatic alignment with the glottis center. To validate the reliability of the feature recognition, the research team conducted TI on a mannequin. The results showed that the feature recognition algorithm achieved a recall rate of 79%, specificity of 100%, and a false negative rate of 21%, ensuring proper alignment of the endoscope in automatic mode and reducing the risk of catheter misplacement and associated complications [29]. While the Haar Cascade Classifier is a lightweight and efficient machine learning algorithm with low hardware requirements, its accuracy, robustness, and generalization ability in recognizing anatomical structures in complex airway environment are significantly lower than those of DL algorithms. In an unpublished study, our team designed an intelligent video laryngoscope and developed a CNN model with an attention mechanism, deploying it on the Orange Pi to recognize key anatomical structures (tongue, uvula, epiglottis, and glottis). The purpose of sequentially recognizing these key anatomical structures is to guide operators step-by-step in performing intubation, establishing a standardized process aimed at increasing success rates and reducing operation time. Our latest results indicate that the model's recall and specificity both exceed 90%.

At this stage, the focus shifts to ensuring that algorithms are lightweight, operate in real-time, and have low latency. Unlike previous stages, hardware deployment introduces more stringent requirements, including efficient operation within limited computing resources, storage space, and power consumption. Algorithms must also process continuous data streams quickly to support real-time visual navigation and decision-making. This requires close collaboration between hardware design and algorithm development teams to ensure both performance and robustness in real-world environments.

Discussion

Robotic and AI technologies have already demonstrated their value in airway management, particularly in TI. This review examines three key developmental stages of intubation robots and the anatomical recognition algorithms used in TI (Fig. 4). These stages are not strictly sequential; instead, they evolve concurrently and influence each other.

Fig. 4.

Developmental stages of intubation robots and anatomical recognition algorithms in TI. A Integrated tracheal intubation robot.

Reproduced with permission from Hemmerling et al. [21], Copyright 2012, Elsevier. B Anatomical structure recognition algorithms based on stable airway environments. C Modular tracheal intubation robot. D Anatomical structure recognition algorithms based on complex airway environments. E Intelligent tracheal intubation robot. F Deployment of anatomical structure recognition algorithms onto hardware

The development of TI robots reveals that successful clinical application requires adherence to several key criteria: First, the robot must feature a compact design and efficient operation. The robot should be small enough to fit alongside the operating table without occupying excessive space, and it should be easy to assemble and disassemble. This ensures that the robot does not interfere with other surgical procedures. Moreover, the intubation process should be completed within 2 min, aiming to match or exceed the efficiency of a moderately experienced physicians, thus preventing delays caused by the robot. Second, the robot should integrate intelligent navigation and safety assurance strategies. This includes a visual navigation system capable of accurately identifying key anatomical structures, a reliable force feedback system, precise control strategies, and robust safety protocols for emergencies. These features are crucial for enhancing the robot's reliability and intelligence in clinical settings. Finally, the robot should be equipped with multi-modal sensory capabilities. In challenging intubation scenarios, such as involving excessive secretions or large masses, the visual navigation system alone may be insufficient. Therefore, integrating additional navigation methods, such as end-tidal CO₂ concentration monitoring or infrared imaging, is crucial for effectively managing these difficult cases.

Recognition algorithms for anatomical structures are influenced by various factors, including individual variability, racial differences, anatomical variations in patients with different pathologies, and imaging quality across different brands of endoscopes. To achieve high accuracy and robustness in clinical settings, these algorithms require datasets that are both comprehensive and diverse. Therefore, creating a multi-center, multi-tool database with precise annotations is crucial for the success of recognition algorithms in TI. Moreover, it is important to include key anatomical structures beyond the glottis, such as the tongue, uvula, epiglottis, and tracheal rings, as their identification during intubation is crucial for clinical practice. When deploying the algorithm on hardware, a careful balance must be struck between lightweight design and model depth, reducing computational demands while still effectively capturing image features. Furthermore, the FPS metric is crucial for evaluating an algorithm's performance during hardware deployment. Algorithms with low FPS may fail to provide real-time video analysis on intubation devices, which is crucial for effective clinical use [42].

Robotics and AI are distinct yet interconnected fields [43]. Robots generate vast amounts of operational and sensory data during tasks, creating an ideal environment for training AI models [44]. In turn, AI enhances a robot's ability to perceive and interact with its surroundings, enabling tasks ranging from simple pick-and-place operations to complex activities in unstructured environments [45]. Looking ahead, the integration of robotics and AI is poised to transform TI. AI-driven robotic systems can process real-time data streams, recognize anatomical structures, and optimize intubation positioning [46]. To translate these advancements into clinical practice, multicenter randomized controlled trials comparing intelligent intubation robots with conventional techniques should prioritize outcomes such as first-pass success rates, procedure time, and complication rates across diverse clinical scenarios, including difficult airway management, intensive care units, and EDs. In China, compliance with regulatory frameworks established by the National Medical Products Administration, particularly the 2022 Registration Review Guideline on AI Medical Devices (No. 8-2022), is essential for establishing robust safety and performance evaluation criteria and facilitating market authorization. Concurrently, interdisciplinary training programs incorporating high-fidelity virtual reality simulations and hands-on workshops should be implemented to ensure clinicians achieve operational proficiency and can promptly assume manual control in unexpected situations. Despite these technological advances, autonomous intubation systems must operate under physician supervision to ensure patient safety, particularly in complex cases where robotic intervention may be contraindicated or require expert clinical judgment. Furthermore, these robots could enable remote guidance and monitoring, allowing experts to provide support remotely. This would reduce the skill demands on frontline personnel and facilitate telemedicine, making advanced care accessible in economically disadvantaged or remote areas [47, 48]. Although robotic and AI systems may enhance procedural accuracy in airway management, they still cannot replicate the contextual clinical judgment required for complex decision-making. For instance, determining whether to intubate high-risk patients with severe airway edema or hemodynamic instability requires nuanced evaluation of dynamic physiological and anatomical factors that exceed algorithmic capabilities [49]. Current evidence highlights that AI lacks tactile feedback, real-time adaptability, and holistic understanding of patient comorbidities, all of which are essential for managing time-sensitive airway emergencies [50, 51]. Moreover, overreliance on AI systems may lead to clinician deskilling and raise ethical concerns regarding algorithmic transparency and accountability in adverse outcomes [52–54]. Therefore, successful integration of these technologies should prioritize synergistic collaboration between human expertise and machine efficiency, ensuring AI functions as a decision-support tool rather than a substitute for clinician judgment [55].

Despite notable advancements in robotic and AI-assisted intubation, several challenges and limitations remain. The high cost of these robotic systems continues to be a significant barrier to commercialization. Among the 28 included studies, only 3 (11%) explicitly proposed cost reduction strategies, underscoring the urgent need to develop economically viable systems. Expensive robotic systems do not necessarily improve surgical efficiency or shorten the learning curve [56], making it challenging to justify the initial investment in the short term. Furthermore, safety concerns remain, as any malfunction could lead to irreversible patient harm. Notably, only 9 studies (32%) conducted dedicated safety testing, whereas all 28 studies (100%) performed universal technical validation (see supplementary materials 2). These risks contribute to skepticism and resistance among patients and healthcare professionals, thereby hindering adoption in clinical practice. Moreover, AI algorithms can perpetuate existing biases in healthcare data, leading to unequal treatment of different patient groups. For example, research by Obermeyer et al. found that AI algorithms in healthcare decision-making were less likely to recommend advanced care for Black patients compared to White patients with the same medical conditions [57]. Furthermore, the integration of AI in healthcare has created a high demand for multidisciplinary professionals who possess both medical knowledge and AI expertise. However, current training models in higher education remain largely traditional, leading to a shortage of such professionals. Addressing cost issues, enhancing technical reliability, and developing comprehensive training programs are essential to overcoming these challenges and maximizing the benefits of robotic and AI technologies in TI. Another challenge is the pivotal role of enterprises in accelerating the clinical application of robotics and AI. However, this involvement often raises concerns about data breaches and ethical risks [58]. The rapid pace of technological development, coupled with outdated laws and regulations, underscores the need for precautionary measures. These measures should include encrypting and securely storing patient data, such as privacy, health status, and medical history, to protect patient rights and ensure ethical handling of medical information.

This narrative review examines the emerging applications of robotics and AI in TI, emphasizing their significance and the need for further development. Unlike systematic reviews, which address specific questions regarding clinical effectiveness or feasibility, this review discusses technological advancements, implementation challenges, and future directions in intelligent airway management. However, several limitations may influence the interpretation and applicability of these findings. Publication bias is a recognized concern in scientific literature, as studies with negative or null results are less likely to be published, potentially leading to an overrepresentation of positive findings in this review. Additionally, no universally accepted standard defines the developmental stages of robotics and AI in TI. To address this, we developed a classification framework based on expert consensus and existing literature to provide a clear developmental overview. Despite these efforts, given the rapid technological advancements and the broad applications in TI, some omissions may remain.

Conclusion

Robotics and AI technologies for TI have evolved from conceptual prototypes to systems demonstrating feasibility in simulation studies and preliminary clinical evaluations. Systems such as the KIS, REALITI, and other emerging modular or intelligent intubation robots have demonstrated performance metrics, including intubation time and success rates, comparable with those of experienced clinicians in controlled environments. With further refinement, these systems could ultimately serve as adjunctive tools to assist with the mechanical aspects of TI. However, substantial challenges remain. Complex procedures performed in unstructured environments require enhanced structural flexibility, real-time adaptive control, and robust human–machine interfaces. In addition, the “black-box” nature of many AI algorithms necessitates greater interpretability to ensure clinical transparency. Ethical and legal considerations, particularly related to data privacy and responsibility allocation, also require thorough examination. Future research should prioritize optimizing device design and validating these systems through multicenter clinical trials. It is essential to emphasize that robotic and AI technologies are adjunctive tools intended to enhance procedural precision and optimize workflow efficiency; however, ultimate responsibility for safety–critical decisions, including the decision to forgo intubation, remains with experienced clinicians.

This review summarizes current research on robotics and AI in TI and discusses potential future directions. However, a key limitation is its exclusive focus on these technologies for intubation, without consideration of other advancements in airway management.

Abbreviations

4IR

Fourth industrial revolution

AI

Artificial intelligence

TI

Tracheal intubation

KIS

Kepler intubation system

DOF

Degree of freedom

VL

Video laryngoscopy

CNNs

Convolutional neural networks

VC

Vocal cords

FPS

Frames per second

EDs

Emergency departments

DL

Deep learning

CC

Corniculate cartilage

DSCs

Dice similarity coefficients

IoU

Intersection over union

CPR

Cardiopulmonary resuscitation

VD

Visual difficulty

Author contributions

W.C, Y.T.: take responsibility for maintaining data integrity and ensuring the accuracy of the analysis. W.C, Y.T., L.F., Y. W.: conceptualization, data curation, formal analysis, methodology, and original draft writing. M.A.: data curation, investigation, and editing of the manuscript. W.L, Y.H., Y.T., S.L.: securing funding, supervision, and review and editing of the manuscript. All authors confirm their consent to the submission of this manuscript.

Funding

This study was supported by grants-in-aid for scientific research from the Perioperative Airway Refinement Management Programs of the Health Human Resources Development Center, National Health Commission (Grant Number: RCLX2315001); China Postdoctoral Science Foundation (Grant Number: 2023M740664); Continuing Education Programs of Shanghai Medical College Fudan University (Grant Number: FDYXYBJ-20222004); National Natural Science Foundation of China (Grant Number: 82201375); National Natural Science Foundation of China (Grant Number: 82271295).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This article does not include any studies involving human participants or animals conducted by the authors; therefore, informed consent or Ethics Committee approval was not required.

Competing interests

The authors declare no competing interests.