JOA - 2026-04-09 - Journal Article
Utility of Machine Learning in Predicting Catastrophic Cardiac Complications Following Hip and Knee Periprosthetic Fracture Surgery Using Preoperative Parameters.
Sampson WT, Freeman IA, Shimizu MR, Kwon OJ, Xiao P, Kwon YM
Topics
Key Takeaway
Four machine learning models predicted MI (AUC 0.82–0.91) and cardiac arrest (AUC 0.76–0.94) following periprosthetic fracture surgery using preoperative labs, with elevated creatinine, WBC, and sodium and lower BMI as the dominant predictors.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study asked whether machine learning models could predict MI and cardiac arrest following PPF surgery (revision arthroplasty or ORIF) using preoperative variables extracted from a national quality improvement database. Four models—artificial neural network, random forest, histogram-based gradient boosting, and k-nearest neighbor—were trained on 3,805 patients with MI rates of 1.97% (revision) and 1.81% (ORIF) and CA rates of 0.92% and 0.67%, respectively. All MI models achieved AUC 0.82–0.91 with Brier scores 0.031–0.044; CA models achieved AUC 0.76–0.94 with Brier scores 0.013–0.037, with preoperative creatinine, WBC, sodium, and BMI as top predictors.
Key Limitation
Event rates for CA were extremely low (0.67–0.92%), creating class imbalance that may inflate AUC metrics and limit generalizability of the CA models despite reported Brier score performance.
Original Abstract
BACKGROUND
With the rising prevalence of primary total hip (THA) and knee arthroplasties (TKA), the number of periprosthetic fractures (PPF) is expected to increase. Myocardial infarction (MI) and cardiac arrest (CA) are devastating complications following PPF surgery. There is a paucity of literature demonstrating the utility of machine learning (ML) in predicting rare clinical outcomes in revision arthroplasty. This study aimed to develop ML models to predict cardiac complications following PPF surgery.
METHODS
Patients who underwent revision arthroplasty or open reduction internal fixation (ORIF) due to PPF were extracted from a national quality improvement database. There were four ML models that were developed to predict MI and CA: artificial neural network, random forest, histogram-based gradient boosting, and k-nearest neighbor. Area under the curve (AUC) was used to assess discriminatory ability. Validation was performed with calibration slopes, intercepts, and Brier scores.
RESULTS
The final population included 3,805 PPF surgery patients, consisting of 1,869 revision arthroplasty (1,378 THA, 491 TKA) and 1,936 ORIF (811 THA, 1,125 TKA) procedures. Among revisions, 1.97% suffered MI (27 THA, 10 TKA) and 0.92% suffered CA (13 THA, 5 TKA). In ORIF patients, 1.81 and 0.67% experienced MI (14 THA, 21 TKA) and CA (1 THA, 12 TKA), respectively. All MI models yielded AUCs and Brier scores between 0.82 and 0.91 and 0.031 and 0.044, respectively. The CA analyses were similar, with AUCs between 0.76 and 0.94 and Brier scores between 0.013 and 0.037. Elevated preoperative serum creatinine, white blood cell count, and sodium, as well as lower body mass index were the most critical predictors.
CONCLUSION
All models demonstrated excellent predictive power and accuracy in assessing postoperative cardiac complications. This highlights the effectiveness of ML in predicting rare, catastrophic outcomes and offers potential areas of intervention to mitigate the risk of adverse cardiac events.