Introduction
Acute postoperative airway obstruction following anterior cervical spine surgery is a rare but potentially life-threatening complication. If not promptly recognized and managed, it can lead to severe outcomes, such as prolonged hypoxemia, anoxic brain injury, and the need for emergency endotracheal intubation. The reported incidence of this complication ranges from 0.7% to 2.4% of anterior spinal fusion procedures [1,2]. The underlying causes include retropharyngeal hematoma with or without pharyngeal edema, often exacerbated by vocal cord paralysis secondary to recurrent laryngeal nerve injury [3–6]. However, the risk factors contributing to the development of this complication are not well characterized. Previously identified risk factors include a greater number of surgical levels, anticoagulant use, prolonged operative time, and preexisting conditions such as diffuse idiopathic skeletal hyperostosis and ossification of the posterior longitudinal ligament (OPLL) [2,7]. Early detection of radiographic warning signs is critical for timely intervention and prevention of irreversible damage. Prompt wound reopening, hematoma evacuation (if present), and emergent intubation can help preserve airway access; if these measures are not feasible, cricothyroidotomy may be required [4,5,8]. Differentiating benign postoperative imaging changes from early indicators of potentially adverse outcomes is challenging due to substantial variability in the normal appearance and thickness of prevertebral soft tissue during the early postoperative period [9]. The aim of this study was to develop novel artificial intelligence (AI) software to predict and identify postoperative airway obstruction at an early stage based on routine postoperative imaging.
Materials and Methods
Following institutional review board approval, all adult patients who underwent anterior cervical fusion between 2012 and 2019 at a single tertiary care medical center for any indication were identified using keyword-based queries of the electronic medical record. Patients who experienced respiratory adverse events in the immediate or intermediate postoperative period were identified through manual review of all hospitalization and outpatient follow-up notes. Only patients with respiratory distress attributable to airway obstruction or narrowing were included (Fig. 1). The subsequent clinical course, including emergent wound reopening, intubation, and surgical exploration, was documented. Mildly distressed patients with a narrow, nonobstructed airway on imaging were conservatively managed and observed in the neurosurgical intensive care unit (ICU).
Exclusion criteria were high cervical spinal cord injuries causing severe respiratory muscle dysfunction (American Spinal Injury Association [ASIA] grade ≤C) and absence of postoperative imaging prior to airway obstruction (e.g., emergent surgical revisions performed without imaging, n=2).
Surgical technique and postoperative assessment
All surgeries were performed under general anesthesia with endotracheal intubation, nasogastric tube insertion, and conventional prevertebral dissection. Bilateral fixation of sharp retractors to the longus coli muscles was performed using Caspar devices (Aesculap Corp., Tuttlingen, Germany). Multilevel stenosis was managed with either multiple discectomies or hybrid constructs incorporating corpectomy and discectomy. The posterior longitudinal ligament was routinely resected to achieve complete dural decompression. Standard rotational and translational plates were used for fusion (Atlantis [Medtronic Corp., Minneapolis, MN, USA] and Helix-T [Nuvasive Corp., San Diego, CA, USA]). A prevertebral drain was placed in all cases. Postoperatively, patients were monitored for 2 hours in the postanesthesia care unit and then transferred to the ward. Routine cervical imaging, either computed tomography (CT) or plain radiography, was performed on the first postoperative day, typically 12–18 hours after surgery. The choice of imaging modality was determined by the most caudal operated level and the patient’s body habitus to ensure adequate visualization of the construct.
Image processing and optimization
Digital Imaging and Communications in Medicine files from postoperative imaging studies were anonymized and prepared for segmentation. The aim was to identify potential airway obstructions by segmenting osseous vertebral structures, implanted hardware, the retropharyngeal space, and the airway, using PyCharm’s Python Integrated Development Environment (JetBrains, Prague, Czech Republic). For CT scans, three-dimensional reconstructions of the airway and retropharyngeal space were generated to facilitate comprehensive volumetric assessment (in voxels). For radiographs, measurements were performed in two dimensions (in pixels). Data augmentation techniques were applied to enhance accuracy and reproducibility (Medical Image Processing Toolbox; MATLAB 2022b; The MathWorks Inc., Natick, MA, USA). A detailed description of the augmentation strategy is provided in the Supplement 1. The resultant data was divided into training and validation sets in a 4:1 ratio to mitigate overfitting bias. Augmentation of the training set was further achieved by randomizing data allocation between training and validation groups 5 times (Fig. 2).
Gradient boosting algorithm
CatBoost (Yandex Inc., Moscow, Russian Federation), a gradient boosting algorithm, was employed to process the automatically segmented voxel data from CT scans and pixel data from radiographic scans. The decision trees within the model sequentially differentiated between radiographs and CT scans, enabling adaptation to diverse patient populations in which postoperative imaging modality varies according to institutional practice. Multiple decision trees were generated and permutated by the model to account for data variability. In instances where data were missing, such as the absence of radiographs for specific patients, the algorithm treated the missing information as a distinct learning category rather than generating synthetic data. This approach allowed the model to preserve data integrity while maintaining predictive performance across heterogeneous imaging inputs.
Statistical analysis
Descriptive statistics were used to summarize patient demographics, surgical characteristics, operative duration, and imaging modalities. Continuous variables were presented as mean±standard deviation and median with interquartile range, while categorical variables were expressed as frequencies and percentages. Model performance was evaluated using key classification metrics, such as positive predictive value (PPV), negative predictive value (NPV), F1 score, and overall accuracy.
Results
Clinical outcomes
A total of 815 anterior cervical fusion procedures were identified. After excluding 20 patients with high cervical spinal cord injuries graded as ASIA ≤C, the study cohort comprised 795 patients (Fig. 1). The mean follow-up period was 7.65±0.6 months. Respiratory complications occurred in 35 patients (4.4%), of which 11 (1.3%) were attributed to airway obstruction. Postoperative cervical CT or radiographs were available for nine patients (11.3%). Patient characteristics are summarized in Table 1. Two patients developed prolonged hypoxemia leading to anoxic brain injury. The remaining nine patients were successfully intubated without permanent hypoxic damage. Urgent wound reopening was performed in six patients, whereas cricothyroidotomy was not required in any case. Following endotracheal extubation, six patients underwent surgical exploration; intraoperative examination revealed edema in three and a frank hematoma in six patients. Patients who required intubation but did not undergo surgical exploration were managed conservatively with serial clinical and imaging assessments until sufficient laryngeal space reexpansion permitted safe extubation.
AI algorithm performance
A total of 336 patients were included in the training set for the model. Data augmentation was applied to achieve an effective 1:4 ratio of affected to unaffected cases. Because airway obstruction was rare (nine cases), nonobstructed controls were intentionally limited to achieve a 1:46 ratio (9:411), with control cases randomly selected. This design ensured that each cross-validation fold contained at least one positive case, reducing instability and the risk of overfitting. Within each training fold, augmentation was again applied to maintain an effective 1:4 ratio. The AI model’s confusion matrix is displayed in Fig. 3. The model achieved a PPV of 0.98, NPV of 0.90, sensitivity of 0.91, and specificity of 0.99. The decision boundary plot incorporating airway volume, retropharyngeal space volume, and weighted radiograph-based features is shown in Fig. 4.
Discussion
This study developed an AI model capable of predicting and detecting early postoperative airway obstruction following anterior cervical discectomy and fusion (ACDF). The model successfully distinguished benign postoperative soft tissue changes from potentially life-threatening airway compromise, achieving a PPV, NPV, sensitivity, and specificity of 0.98, 0.9, 0.91, and 0.99, respectively.
Comparison with previous research
Although several studies have explored AI-based prediction of adverse outcomes following spinal surgery [10], this study represents a novel effort specifically aimed at predicting postoperative airway obstruction. Chen et al. [11] used a K-Nearest Neighbor model to predict the need for tracheostomy in patients with deep neck infections based on manually derived CT features, achieving a sensitivity of 62.5% and a specificity of 80.60%.
Model characteristics and rationale
The present study employed CatBoost, a gradient boosting algorithm. Consistent with definitions established at the recent American Orthopedic Association Symposium on AI in Orthopedic Surgery [12], CatBoost integrates both supervised and unsupervised learning approaches, two crucial subsets of machine learning [13–15]. Supervised learning was applied by labeling and optimizing decision trees, while unsupervised learning was achieved through augmentation of semisynthetic data when imaging data were unavailable. CatBoost was chosen for its robustness to overfitting, tolerance for missing data, and superior handling of categorical features, all of which are critical advantages given the diversity and inconsistency of postoperative imaging [3,9]. The model’s ability to predict airway obstruction based on routinely obtained postoperative imaging may enable earlier intervention, such as delayed discharge, ICU monitoring, or prophylactic intubation before critical airway compromise.
Study limitations
Some limitations of this study should be acknowledged. First, despite CatBoost’s inherent resistance to overfitting, the small number of airway obstruction cases can potentially cause overfitting. To mitigate this, we utilized a k-fold cross-validation approach [16], wherein the data was split into k segments, and the model was subsequently trained on each, bolstering its dependability and minimizing overfitting. Second, the study cohort had a high proportion of corpectomy cases, which may limit generalizability to centers where cervical discectomy is more commonly performed. Third, since this model relies on postoperative imaging, it cannot be applied to patients who deteriorate before imaging is obtained.
Future directions: toward a decision support system
The development of a clinically available and user-friendly graphical user interface is envisioned as an essential next step for translating this tool into a usable decision support system. Given the time-critical nature of postoperative airway obstruction, the integration of an automated alarm mechanism will be vital. Once fully operational, this system could support less experienced clinicians or those managing heavy caseloads by providing reliable, data-driven guidance in patient monitoring and intervention. Furthermore, the model’s high NPV may help avoid unnecessary and potentially harmful actions such as wound reopening and unnecessary ICU admissions.
Conclusions
The AI model developed in this study demonstrated strong potential for predicting post-ACDF airway compromise, despite variability between radiographic environments using plain radiographs and CT scans. This tool may enable early identification of patients at risk of postoperative airway obstruction and help differentiate benign postoperative changes from complications such as edema or hematoma. Further studies are necessary to validate and confirm these initial findings. To our knowledge, this is the first study to employ an AI-based approach for predicting airway compromise following anterior cervical spine surgery.
Key Points
Acute postoperative airway obstruction after anterior cervical spine surgery due to edema or hematoma is rare but potentially life-threatening.
The AI model developed in this study shows potential for predicting post-anterior cervical discectomy and fusion airway compromise, despite the variability across radiographic and computed tomography environments.
Early detection of complications by distinguishing benign postoperative changes from edema or hematoma may help prevent life-threatening airway obstruction.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement
