INTRODUCTION
Anterior cervical spine surgery (ACSS) is widely accepted as an effective procedure for treatment of the degenerative condition in cervical spine, with over 200000 annual cases in the United States alone [1]. However, postoperative prevertebral soft tissue swelling (PSTS) remains a significant concern. While some degree of PSTS is a common radiological finding [4], it can lead to serious clinical complications including dysphagia [8,16], dysphonia, and life-threatening airway obstruction [5]. Systematic reviews report substantial incidence of these complications with dysphagia rates as high as 71% in the immediate postoperative period, though persistent symptoms are less common [7,21]. Fatal airway compromise is rare but severe complication with mortality rates approaching 0.3% in high-risk procedures such as multilevel corpectomy [2]. These findings emphasize the importance of identifying high-risk patients for safety concerns.
Previous studies have shown that risk factors such as multilevel surgery, prolonged operative times, and upper cervical involvement [3,13] are associated with higher chance of post-ACSS complications.
However, current literature focuses on clinical outcomes like dysphagia instead of analyzing PSTS at the quantitative level. Moreover, there are conflicting reports regarding patient-specific factors and most studies include multi-level procedures as a significant confounding variable indicating that our understanding of the underlying pathophysiology is limited [15].
Our recent retrospective study addressed this gap by focusing exclusively on 62 reported cases of single-level ACSS [9] using multivariate linear regression. Our analysis identified male sex (β=1.21, p=0.012), lower preoperative serum albumin (β=-1.63, p=0.014), and upper cervical level surgery above C5 (β=1.44, p=0.008) as independent predictors of PSTS. We showed that the interaction between risk factors and individual risk profiles is non-linear and traditional linear predictive models may not suffice. A framework for integration of multiple variables, through artificial intelligence (AI) and machine learning technology offers an opportunity to surpass simple association analysis in favor of more robust and individualized risk prediction [19].
We hypothesized that a machine learning model designed for small datasets could be a useful clinical tool in pre-operative identification of high-risk patients. Our aim was to design and internally validate an interpretable machine learning model that predicts risk factors for clinically significant PSTS following single-level ACSS.
MATERIALS AND METHODS
This retrospective study followed the principles of the Declaration of Helsinki. The protocol was approved by the Institutional Review Board of Hanyang University Guri Hospital (IRB No. GURI IRB 2023-04-011).
Study design and patient population
We retrospectively reviewed medical records of all patients that underwent single-level ACSS at our institution between January 2014 and December 2022. Inclusion criteria were : 1) age 19-80 years with symptomatic cervical degenerative disc disease (radiculopathy/myelopathy) and 2) single-level ACSS procedure. Exclusion criteria included a history of cervical trauma, previous surgery of the cervical spine, or multi-level fusion. After applying these criteria, 62 patients were included in the analysis.
Data collection and variables
Collected variables included demographics (age, sex, body mass index), clinical information (hypertension, diabetes, smoking), laboratory values (preoperative serum albumin), and surgical details (operative time, level, method). Missing data (<5%) were handled using multiple imputations with chained equations (five iterations).
Radiological assessment and outcome definition
PSTS was quantified using lateral cervical radiographs taken preoperatively and daily for 5 postoperative days. Anteroposterior prevertebral soft tissue thickness was measured from C2 to C6. The primary outcome, ΔSwelling, represented the maximum difference between preoperative and postoperative measurements. For model development, significant PSTS was defined as ΔSwelling >7.0 mm, based on our finding that symptomatic patients in our cohort averaged 6.99 mm of swelling and the 75th percentile value was approximately 7.0 mm. This threshold was chosen to balance clinical relevance with statistical considerations for the binary classification task.
Sample size justification
Based on established guidelines for prediction model development, the minimum sample size requirements were calculated [12]. Our sample size of n=62 falls below these recommendations, positioning this study explicitly as a proof-of-concept investigation that requires external validation before any clinical implementation.
Machine learning model development and validation
An elastic net regularized logistic regression was chosen for its advantages with small datasets : it prevents overfitting through L1/L2 regularization, handles multicollinearity, maintains clinical interpretability, and performs automatic feature selection. Model performance was compared with Random Forest and Support Vector Machines, but the elastic net showed superior performance and interpretability.
A nested 5-fold cross-validation was used, with the outer loop for unbiased performance estimation and the inner loop for hyperparameter tuning. Final model performance was validated using 1000 bootstrap iterations with the 0.632 estimator. Class imbalance was managed through threshold optimization using receiver operating characteristic (ROC) curve analysis.
Performance metrics and statistical analysis
Model performance was assessed by discrimination, calibration, and clinical utility. Discrimination, the model's ability to distinguish whether a patient will or will not develop significant PSTS, was reflected in the area under the ROC curve (AUC-ROC). The ROC curve measures the diagnostic performance of a binary classifier system at various thresholds. An AUC of 1.0 represents a perfect model, while 0.5 indicates a random guess. Calibration, which assesses the agreement between predicted probabilities and observed outcomes, was evaluated using the Hosmer-Lemeshow test. Clinical utility was determined using decision curve analysis (DCA) [17].
Confusion matrix was constructed using sample data and sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were measured. Using sensitivity and specificity, likelihood ratio and Youden’s index were calculated to evaluate diagnostic performance. To compensate for imbalance in the dataset, Balanced accuracy score was taken and Matthews correlation coefficient was calculated to measure performance on all four categories of the confusion matrix. Brier score was obtained to evaluate prediction accuracy. In addition, the model went through bootstrap validation to quantify potential overfitting. SHapley Additive exPlanations (SHAP) analysis was used to quantify the contribution of each predictor to the output for individual patients, enhancing model interpretability. Between-group comparisons were done by T-test and chi-square test. Continuous variables were reported as mean±standard deviation, while categorical variables were reported as frequencies (%). All analyses were performed using Python 3.9, with p<0.05 considered significant.
RESULTS
Patient characteristics
Table 1 summarizes baseline characteristics of our sample of 62 patients. Of these, 16 (25.8%) developed significant PSTS. The significant PSTS group was older (62.9±9.2 vs. 58.2±12.1 years, p=0.184), had significantly lower preoperative albumin levels (3.8±0.5 vs. 4.2±0.3 g/dL, p<0.001), a higher proportion of surgeries at the upper cervical levels (81.3% vs. 32.6%, p<0.001), and a predominance of male patients (75.0% vs. 47.8%, p=0.048).
Model performance and interpretability
The elastic net model demonstrated strong predictive performance, achieving a bootstrap-validated AUC of 0.84 (95% confidence interval [CI], 0.73-0.92) (Fig. 1 and Table 2). Sensitivity was 0.75 (95% CI, 0.62-0.88) and specificity 0.78 (95% CI, 0.68-0.87) at the optimal threshold (0.31), and balanced accuracy was 0.77. Positive likelihood ratio was 3.41 and negative likelihood ratio was 0.32, indicating good accuracy at predicting both high-risk and low-risk patients. The model achieved a Brier score of 0.14 with minimal optimism (0.03) despite small sample size, demonstrating good calibration with low level of overfitting.
The final model equation was determined as : Logit(P) = -2.84 + 1.28×male - 1.92×albumin + 1.51×uppercervical + 0.42×age_scaled. SHAP analysis revealed the most influential predictors to be lower preoperative albumin, upper cervical level surgery, and male sex (Fig. 2), and the logistic regression analysis showed good fit (Hosmer-Lemeshow p=0.42) (Fig. 3A) and a high NPV of 0.90 (95% CI, 0.81-0.96). These findings suggest that preoperative albumin level contributes most substantially to individual risk predictions, followed by anatomical factors (cervical level) and patient demographics (sex).
Clinical utility
DCA showed net benefit across probability thresholds of 15-45%, supporting the model’s clinical usefulness for risk-stratified management strategy over “treat all” or “treat none” strategies (Fig. 3B).
Model robustness
Internal validation confirmed model stability. Cross-validation yielded a mean AUC of 0.82 compared to the bootstrap-validated performance of 0.84; consistent performance was achieved across different data subsets. The calibration intercept was near zero (0.08; 95% CI, -0.15 to 0.31) while the calibration slope of 0.92 approached the ideal value of 1.0. Excellent calibration was demonstrated across the whole risk spectrum with minimal systematic over- or under-prediction.
DISCUSSION
In this study, we developed an interpretable machine learning model for prediction of PSTS following single-level ACSS. Despite the small cohort, our model achieved an AUC of 0.84, demonstrating that AI-based risk stratification is feasible in spine surgery even with limited data.
PSTS after ACSS is a known complication due to the anatomical characteristics of the surgical site. Previous studies have reported airway complication rates reaching 6.1% postoperatively, with surgery involving three or more segments and upper cervical procedures (C2-C4) identified as major risk factors [13]. A large longitudinal cohort study confirmed that the risk of dysphagia increases in multi-level surgery [10]. The clinical significance of this issue is underscored by reports of reintubation rates of 0.2-1.2% and associated mortality rates reaching 0.3% in high-risk cases, reaffirming the need for robust predictive tools [2,5].
Comparison with prior work and methodological advancements
This study represents a significant methodological advancement over our previous work on the same patient cohort [9]. Our prior analysis successfully utilized multivariate linear regression to identify key independent predictors of PSTS, and its importance was establishing the linear associations between factors like male sex, low albumin, and upper cervical surgery with the degree of postoperative swelling. However, traditional regression models are, by design, limited in their ability to capture more complex, non-linear relationships and interactions between variables common in biological systems. This study employs an elastic net regularized logistic regression model to overcome such limitations.
Our logistic regression framework is designed with the ability to handle a large number of predictors relative to sample size. Robustness of the model is reinforced by regularization of data to prevent overfitting, automatic feature selection and by advanced validation techniques such as nested cross-validation and bootstrapping.
Our model offers several advantages over linear regression analyses. First, it predicts the probability that a clinically significant event may occur, rather than providing a numerical estimate of how much swelling is to be expected. It also weighs each risk factor’s contribution to overall predicted risk through SHAP analysis, for relevance in clinical decision-making. Our previous study focused on identifying what the risk factors are, whereas this study demonstrates how these factors can be interpreted in practice.
The three key predictors identified by our model are consistent with existing literature. Upper cervical surgery (above C5) emerging as a strong predictor aligns with research showing that PSTS is significantly more severe in surgeries proximal to C5 [18]. This is likely due to the larger potential retropharyngeal space in the upper cervical spine and the need for more extensive dissection of the longus colli and longus capitis muscles, which can lead to venous and lymphatic drainage obstruction [3].
Hypoalbuminemia is also identified as a clinically significant predictor of PSTS. Albumin is known to be a factor in maintaining oncotic pressure and modulation of inflammatory responses; low albumin levels can contribute to a cycle of inflammation and edema [19]. Thus, nutrition may be a potentially modifiable preoperative risk factor, as enhanced recovery after surgery protocols emphasize.
Male sex was a risk factor for PSTS. This result differs from some dysphagia studies that report higher rates in females [6,11], but is consistent with other large-scale analyses [14]. Anatomical factors may lead to these sex-based differences, such as larger soft tissue volume in males.
This study followed current best practice in medical science regarding application of machine learning with small datasets [20]. We applied elastic net regression and built a rigorous bootstrap validation framework to minimize the risk of overfitting while providing a reliable estimate of our model’s performance.
Key metrics of our model included its NPV (0.90), diagnostic ratio (10.66), and Matthews correlation coefficient (0.49). High NPV and diagnostic ratio represent effectiveness at identifying low-risk patients. Matthews correlation coefficient shows moderate to strong correlation between predicted and observed outcomes. F1 score (0.63) at optimal probability threshold of 0.31, derived from Youden’s index (0.53), indicates reasonable precision and recall. Somewhat less effective was its ability to predict outcome for high-risk patients, as implied by PPV of 0.55. Approximately half of those patients identified as high-risk according to our model may not develop significant PSTS.
In summary, this study shows that patients about to receive upper cervical level surgery, male patients, and those with serum albumin level less than 4.0 g/dL are at a high risk of clinically significant PSTS and could benefit from additional safety protocols. Examples would include nutrition regimen, thorough perioperative monitoring, and perhaps even prophylactic corticosteroid administration.
This study has several limitations. Generalizability is limited by small sample size and by our single-center cohort design. Integration of more complex algorithms and incorporation of more predictive variables may be possible with larger datasets. Future research across multiple institutions focused on external validation could reinforce robustness of the model. While the model has minimal optimism (0.03) and good calibration slope (0.92), several metrics are limited by wide CIs, possibly due to limited sample size. While the AUC-precision-recall (0.71) is a respectable value, it is still lower than the AUC-ROC. Given 25.8% event rate, this shows the challenge of building a prediction model from an imbalanced dataset.
External validation through large-scale multicenter studies with independent cohorts are required to improve our model’s predictive accuracy. Furthermore, prospective studies using it could provide opportunities for recalibration and improvements. Refinement and integration with electronic health record systems could facilitate automated risk stratification at point of care.
CONCLUSION
This study successfully developed and internally validated a machine learning model for predicting significant PSTS after single-level ACSS. Despite its limitations, this proof-of-concept study shows that use of AI-based tools may enhance the preoperative patient assessment process and improve healthcare quality. Further validation and refinement would enhance clinical capability of our model.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation 
Print 
