The study by Taka et al, reviewed here, evaluated risk factors for postoperative hematoma following anterior cervical spine surgery using a machine learning approach. Postoperative hematoma is a rare complication following anterior cervical procedures but can be particularly consequential given the potential for airway compromise and need for urgent reoperation. As anterior cervical procedures are increasingly performed in outpatient and short-stay settings, identifying patients who may be at high risk for hematoma is a clinically relevant goal. The authors applied machine learning techniques to a large national surgical database to explore whether readily available preoperative factors could inform risk stratification.

The authors queried the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) database for adult patients undergoing elective anterior cervical spine procedures between 2012 and 2018. Procedures included anterior cervical discectomy and fusion (ACDF), anterior cervical discectomy without fusion, cervical disc arthroplasty, and cervical corpectomy. The primary outcome was postoperative hematoma resulting in readmission or reoperation within 30 days of surgery. A variety of demographic characteristics, comorbidities, functional status, and preoperative laboratory values were considered as potential predictors.

A total of 54,427 patients met study criteria, of whom 176 developed a postoperative hematoma requiring readmission or reoperation. Thus, the observed incidence of hematoma was only 0.32%, emphasizing both the rarity of the event and the challenge of studying it. To address imbalance between patients with and without hematoma, the authors performed 1:5 propensity score matching based on age and sex, resulting in a matched cohort of 1,056 patients consisting of 176 hematoma cases and 880 controls.

The authors evaluated five supervised machine learning classifiers - random forest, XGBoost, LightGBM, CatBoost, and a multilayer perceptron - alongside logistic regression as a standard statistical comparator. The models were developed using a 70:30 train-test split, with the training data set used for model development, and the test set reserved for final performance evaluation. This approach reduces overfitting and provides a more reliable estimate of how the models perform on unseen data. Model performance was assessed using several complementary measures, including accuracy, sensitivity, specificity, positive and negative predictive values, F1 score, and area under the receiver operating characteristic curve (AUROC). This broad assessment of model performance is important in the setting of a rare outcome, where overall accuracy alone can give an incomplete impression of predictive capability.

Across the evaluated algorithms, mean accuracy was 87.75%, mean specificity was 99.3%, and mean AUROC was 0.824. However, as expected for a rare outcome, mean sensitivity was low at 30.2%. XGBoost demonstrated the strongest overall balance of performance measures, including a sensitivity of 35.9%, AUROC of 0.849, and the highest F1 score among the evaluated models. Notably, the inclusion of a standard logistic regression model provided a useful benchmark, highlighting the improved performance of the more complex machine learning models on some metrics and suggesting that nonlinear interactions among clinical variables may contribute to hematoma risk in ways not fully captured by traditional regression.

Using permutation feature importance (PFI) from the XGBoost model, several preoperative patient and laboratory variables were identified as significantly associated with postoperative hematoma. These included diabetes, smoking history, higher preoperative white blood cell count, lower preoperative sodium, and dependent functional status. The PFI analysis adds an element of explainability to the otherwise black-box machine learning models and was an excellent addition to the manuscript.

Overall, this study represents an important step towards data-driven, multifactorial preoperative risk stratification in spine surgery patients. As granular patient data is increasingly captured and leveraged for clinical decision making, the machine learning techniques used in this study appear to be promising tools for processing such data. However, as highlighted by the authors, these models are not yet ready to be load-bearing components of clinical decision making. Despite favorable accuracy and discrimination metrics, a substantial proportion of patients who ultimately developed a hematoma would not have been identified prospectively. This emphasizes the continued importance of postoperative vigilance regardless of a patient’s predicted preoperative risk.

As the authors acknowledge, there are limitations inherent to this type of investigation. Importantly, the rarity of postoperative hematoma created substantial class imbalance and contributed to low sensitivity of the models. Additionally, the ACS-NSQIP database, while having the advantage of a large, multicenter patient population, only captures a limited set of data variables. Granular surgical-site variables and specific intraoperative factors likely influence hematoma formation but were not able to be assessed in this database. Finally, ACS-NSQIP is limited to 30 days of postoperative follow-up; however, since hematoma is predominantly an early complication, this limitation is likely less consequential.

In conclusion, Taka et al provide an interesting and timely application of machine learning to a clinically important complication of anterior cervical spine surgery. As more anterior cervical procedures transition to outpatient or short-stay pathways, preoperative risk stratification has become increasingly important. While machine learning techniques are promising and may enhance risk awareness, they do not obviate the importance of surgical optimization, postoperative vigilance, and sound clinical judgment. This reviewed paper provides a strong foundation for future studies and a meaningful contribution towards the goal of data-driven risk stratification. It will be exciting to see how these efforts develop in future work.