S. Chakraborty, M. O'Brien
Patient falls during the early recovery period after total joint arthroplasty can delay rehabilitation and lead to preventable injury. Existing bedside screening tools are practical but rely on a small set of categorical risk factors and may not distinguish risk well when the event rate is low. We evaluated whether supervised machine learning models trained on routinely collected electronic health record data could improve the prediction of documented in-hospital and 30-day post-discharge falls after elective primary hip or knee arthroplasty. The retrospective cohort included 16,408 adult patients treated at a single academic medical center between July 2013 and December 2019. Four models were compared with the Hendrich~II Fall Risk Model. They were logistic regression, decision tree, gradient-boosted ensemble, and neural network. Falls were documented for 213 patients (1.30%). To address class imbalance, SMOTE was applied only within the training folds, followed by cross-validated hyperparameter tuning and evaluation on a temporally distinct hold-out set. The gradient-boosted ensemble had the highest test-set area under the receiver operating characteristic curve (AUC~=~0.831), followed by the neural network (0.812), logistic regression (0.794), decision tree (0.762), and Hendrich~II baseline (0.641). The most influential predictors were comorbidity burden, perioperative medication count, age, body mass index, and benzodiazepine exposure. These results support further external validation of EHR-based fall-risk scores as an adjunct to, rather than a replacement for, clinical assessment in arthroplasty recovery.
Pearl Academic Publishing. All rights reserved.
Content is licensed under a Creative Commons Attribution 4.0 License (CC-BY).
Privacy Policy | Terms of Service