The FACT-Score: Development and Temporal Validation of a Clinical Risk Stratification Tool for 30-Day Complications After Operative Facial Fracture Repair.

BACKGROUND

No validated, procedure-specific tool exists for predicting 30-day complications after operative facial fracture repair. This study aimed to develop and temporally validate a clinical risk stratification tool for 30-day complications after operative facial fracture repair.

METHODS

Adults undergoing operative facial fracture repair were identified in the American College of Surgeons National Surgical Quality Improvement Program from 2007 to 2022 (n=4805). The primary outcome was a 30-day ACS-NSQIP composite surgical morbidity endpoint comprising surgical site infection, wound dehiscence, and unplanned return to the operating room. Seven algorithms were benchmarked on a random 70/30 train-test split, including logistic regression, LASSO, random forest, XGBoost, support vector machine, naive Bayes, and neural network. The best-performing model was used to develop a 3-tier risk stratification system, resulting in the Facial Fracture Adverse-event Clinical Triage (FACT)-Score. Temporal validation was performed by training on patients who underwent surgery between 2007 and 2018 (n=3299) and testing on those undergoing surgery from 2019 to 2022 (n=1506). A parsimonious 5-variable clinical nomogram and web-based calculator were constructed for clinical deployment.

RESULTS

The overall 30-day complication rate was 6.7% (n=320). Logistic regression achieved the highest discrimination (AUC 0.764, 95% CI 0.713-0.814) and best calibration (Brier score 0.056), significantly outperforming 5 of 6 alternative algorithms (DeLong P<0.05). On temporal validation, the FACT-Score achieved an AUC of 0.797 (95% CI 0.755-0.839) with strong calibration (intercept -0.02, slope 0.98). The 3 risk tiers showed clear separation in observed complication rates: low risk 1.8% (95% CI 0.9%-3.0%), moderate risk 7.1% (4.9%-9.3%), and high risk 22.8% (17.8%-27.8%), with an odds ratio for high versus low risk of 15.88 (95% CI 8.78-31.25).

CONCLUSIONS

In this national cohort, machine learning did not improve prediction beyond logistic regression for facial fracture complications. The 5-variable FACT-Score provides a simple, temporally validated tool for postoperative risk stratification in facial fracture surgery. A web-based calculator is freely available for clinical use at https://factscore.github.io/FACT-SCORE/.