JAAOS - 2026-07-15 - Journal Article
Prediction of Short-Term Postoperative Complications Following Open Reduction Internal Fixation of Ankle Fractures.
Agarwalla A, Gowd AK, Cody EA, Tan EW, Peterson AB, Liu JN
Topics
Key Takeaway
Logistic regression ML outperformed ASA classification for predicting transfusion after ankle ORIF (AUC 93% vs. 82%) across 42,254 NSQIP cases.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study queried ACS-NSQIP (2011–2020) to develop ML models predicting short-term complications after ankle ORIF and benchmarked them against ASA classification, Charlson Comorbidity Index, and modified frailty index. Logistic regression ML outperformed ASA classification across all outcomes: any adverse event (AUC 73% vs. 69%), transfusion (93% vs. 82%), extended LOS >3 days (82% vs. 75%), DVT/PE (55% vs. 53%), SSI (62% vs. 58%), and discharge home (89% vs. 79%). Hematocrit, BMI, age, and surgical time were the highest-importance predictive variables.
Key Limitation
NSQIP's 30-day capture window excludes delayed wound complications and implant-related failures, which represent a substantial portion of morbidity after ankle ORIF and likely underestimate the true complication burden the model should predict.
Original Abstract
BACKGROUND
Minimizing postoperative complications is imperative to improving patient outcomes. The purpose of this investigation is to develop machine learning (ML) models that can predict complications following open reduction and internal fixation of ankle fractures and compare them with legacy indices.
METHODS
The ACS-NSQIP database was queried from 2011 to 2020 for ankle fractures. Training and validation sets were created by randomly assigning 80% and 20% of the data set, respectively. Age, body mass index (BMI), surgical time, smoking status, comorbidities, and preoperative hematocrit and albumin were included. Complications included any adverse event, transfusion, extended length of stay (>3 days), surgical site infection, deep vein thrombosis/pulmonary embolism, and discharge home. ML algorithms were compared with legacy indices, such as the American Society of Anesthesiologists classification, Charlson Comorbidity Index, and modified frailty index. Model strength was evaluated by calculating the area under the receiver operating characteristic.
RESULTS
A total of 42,254 cases were identified. Mean age, BMI, and length of stay were 44.5 ± 18.5 years, 30.6 ± 7.7 kg/m 2 , and 1.6 ± 4.4 days. Percentage hematocrit, BMI, age, and surgical time were among the highest importance in outcome prediction. Logistic regression ML algorithm outperformed American Society of Anesthesiologists classification for predicting any adverse event (73% vs. 69%), transfusion (93% vs. 82%), extended length of stay (82% vs. 75%), deep vein thrombosis/pulmonary embolism (55% vs. 53%), surgical site infection (62% vs. 58%), and discharge home (89% vs. 79%). Logistic regression ML had the highest positive predictive value (91.4%) for discharge home and negative predictive value (99.4%) for blood transfusion.
CONCLUSION
ML algorithms can calculate patient-specific risk for complications following open reduction and internal fixation of ankle fractures. These models have greater utility in predicting adverse events than legacy indices. With continued validation, ML can stratify surgical candidates, identify site of surgery, and allocate resources postoperatively.
LEVEL OF EVIDENCE
IV, Cohort Study.