Text-based fall prediction in hospital: Development and internal validation of a model to predict in-hospital falls in older patients using free text from daily nursing records

Abstract

Aim

The aim of this study was to evelop a predictive model, estimating the probability of an in-hospital fall using previously identified associated words, and word combinations in daily nursing records. To assess the difference in discriminatory ability between the predictive model and currently used screening questions.

Background

Hospital falls are a persistent challenge. Identifying patients at high risk before fall incidents occur is essential to optimize preventive measures and reduce the burden on nursing staff.

Method

Words from daily nursing records were used as predictive variables to construct and validate the model. The DeLong's test was used to determine statistical differences between the developed model and the current screening questions.

Results

A total of 3255 consecutive admissions of patients aged 70 and over were included, of whom 110 experiences a fall. Upon internal validation, the predictive text model demonstrated moderate discriminatory ability (AUC-ROC 0. 737 (CI 95 % 0. 683–0.791)) and good calibration across a range of the risk groups. Compared to the screening questions (AUC-ROC 0.603 (CI 95 % 0.555–0.652)) the text model (AUC-ROC 0.734 (CI 95 % 0.679–0.788)) showed significantly better discriminatory ability (DeLong's − 3.93, p ≤0.001).

Conclusion

Daily nursing records can be used to estimate the probability of in-hospital falls. A text-based predictive model outperforms the currently employed screening questions and provides insights for the efficient use of fall prevention interventions. Further research should focus on improving the accuracy and external validation of the model and implementation strategies.