Text

Objectives and questions: Inpatient falls are a relevant adverse event during post-acute rehabilitation after total hip arthroplasty (THA) and total knee arthroplasty (TKA) and can lead to dislocation, periprosthetic fractures, delayed recovery, and prolonged length of stay. Falls in patient rooms or bathrooms are particularly concerning, as patients may remain undiscovered. Descriptive analyses suggest that falls tend to cluster in specific environments and may be preceded by potentially modifiable behaviors such as not using the nurse call system before mobilizing. However, tools to identify characteristics of high-risk fall situations at scale remain limited. We aimed to identify factors associated with room and bathroom falls specifically, using machine learning (ML) models trained on routinely collected fall reports from inpatient rehabilitation settings, with external validation on held-out clinical sites.

Material and methods: We performed a multicenter retrospective study using standardized fall reports from 11 rehabilitation clinics (2016–2026) for THA/TKA patients. Predictors included temporal factors (hour, weekday/weekend), lighting conditions, patient orientation (familiarity with call system and ward), mobility aids (crutches, rollator, walker, none), and footwear (barefoot, socks, slippers, proper shoes). The outcome was falls in patient rooms/bathrooms versus other locations. Nine clinics (n=1,116) served as the internal development set; two clinics (n=338, n=341) were held out for external validation. We trained Logistic Regression, Random Forest, XGBoost and CatBoost models using stratified 5-fold cross-validation. Performance was assessed using AUROC, F1, precision, sensitivity, and specificity. Interpretability was assessed using SHAP values.

Results: Among 1,795 fall events, 58% occurred in rooms/bathrooms. At the time of the event, most patients were using crutches (53.4%) or a rollator (22.2%), while mobilization without an assistive device was uncommon (0.3%). 70.9% had not used the call system beforehand. CatBoost achieved internal cross-validation AUC of 0.832 (±0.032), F1 0.763, sensitivity 0.702, and specificity 0.802. External validation demonstrated robust generalizability: AUC 0.856 and 0.843 at the two held-out sites respectively. SHAP analysis identified nighttime hours (midnight–6 AM) as the strongest predictor, followed by inadequate footwear (barefoot/socks). Proper shoes were protective.

Discussion and conclusions: Using fall reports from 11 rehabilitation clinics over 10 years, ML models identified reproducible patterns associated with room/bathroom falls after THA/TKA, with robust external validation (AUC 0.856 and 0.843 at two held-out sites). Nighttime hours (midnight–6 AM) and inadequate footwear (barefoot/socks) were the strongest predictors. These findings support prevention strategies that combine behavioral advice (“wear proper shoes”, “call for assistance before getting up in the morning”) with improved accessibility of the call system and targeted monitoring during high-risk hours.