Abstract
Background
Anterior approaches to the cervical spine have consistently increased annually, with commonly performed procedures demonstrating low morbidity and mortality rates. However, rare complications of postoperative hematoma requiring readmission or reoperation poses risks such as respiratory compromise and reintubation. This study sought to utilize machine learning algorithms (MLA) on American College of Surgeons National Surgical Quality Improvement Program data to characterize clinical risk profile for readmission and reoperation secondary to postoperative hematoma in patients following anterior cervical procedures.Methods
A query of the American College of Surgeons National Surgical Quality Improvement Program database identified adult patients undergoing elective anterior cervical spine procedures from 2012 to 2018 and who developed a postoperative hematoma requiring readmission or reoperation within 30 days. 1:5 Propensity score matching was employed to ensure comparable groups and reduce data bias. Six MLAs were constructed to evaluate the association of preoperative variables with postoperative hematoma development within the matched cohort. Permutation feature importance (PFI) was derived from the top-performing MLA to identify and quantify the relative contribution of individual clinical factors to overall risk variance.Results
Of 54,427 patients, following the 1:5 Propensity score matching, 1,056 patients remained, with 176 (16.67%) developing a postoperative hematoma that required either readmission and/or reoperation. The 6 MLAs generated predictions with an average AUC of 0.824 and average accuracy of 87.75%. However, reflecting the class imbalance of this exceedingly rare complication, the models demonstrated a low average sensitivity of 30.2%. Analysis of PFIs from the top performing algorithm identified diabetes (PFI = 0.020, p = .020), history of smoking (PFI = 0.026, p = .026), preoperative WBC (PFI = 0.028, p = .028), preoperative sodium (PFI = 0.001, p = .001), and dependent function status (PFI = 0.001, p = .001) as statistically significant preoperative factors in the development of postoperative hematomas.Conclusions
MLAs identified several variables associated with clinically significant postoperative hematoma formation following anterior cervical spine surgery. However, given the low sensitivity driven by the rarity of this event, these algorithms cannot reliably predict individual patient outcomes or serve as independent screening tools. Instead, they function as adjunctive assets to enhance perioperative risk awareness and guide risk stratification.