Enhancing multimodal inpatient fall prediction via nursing statement integration within the OMOP common data model

Accurate prediction of inpatient falls is a cornerstone of patient safety, yet conventional models often overlook the rich clinical context embedded in nursing records. This study aimed to develop and validate machine learning (ML) models by integrating structured clinical data from a Common Data Model (CDM) with nursing statements, evaluating whether nursing documentation provides incremental predictive value for fall risk stratification. We conducted a retrospective cohort study using adult inpatient data mapped to the Observational Medical Outcomes Partnership (OMOP) CDM, linked with nursing statements characterizing patients' functional status, mobility, mental status, and care dependency. Four algorithms-Logistic Regression, Random Forest, Gradient Boosting, and XGBoost-were trained and evaluated across three distinct feature sets: CDM-only, nursing-only, and combined. Model performance was quantified using the area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity. Model interpretability and feature contributions were elucidated using Shapley Additive Explanations (SHAP) analysis. Integrated models (CDM + nursing) consistently outperformed single-source models, demonstrating superior discriminative power and a more balanced sensitivity-specificity profile. Notably, models utilizing only nursing statements achieved performance comparable to CDM-only models, indicating that nursing documentation captures critical safety-relevant information often latent in structured electronic health records. Ensemble methods, particularly XGBoost, demonstrated the highest discrimination. SHAP analysis revealed that nursing-derived indicators, such as documented "Reports weakness" and "At risk of falling" were among the most influential predictors, providing context-rich insights beyond standard clinical variables. The integration of coded nursing statements with structured CDM data enhances both the accuracy and interpretability of inpatient fall risk predictions. These findings underscore the importance of leveraging unstructured clinical documentation to develop transparent, AI-driven clinical decision support systems for proactive fall prevention in hospital settings.