Prediction of fall events during admission using eXtreme gradient boosting: a comparative validation study

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Geriatric Medicine

Background:

• Current fall risk assessment tools have limitations, posing challenges for clinicians in identifying at-risk patients.
• In-hospital falls are a significant concern, necessitating improved prediction and prevention strategies.

Purpose of the Study:

• To develop and validate an automatic fall risk prediction model using eXtreme gradient boosting (XGB).
• To compare the performance of the XGB model against the standard Morse Fall Scale (MFS).

Main Methods:

• A data-driven approach utilizing standardized medical records from 639 participants.
• Model construction using XGB on a derivation cohort (507 participants).
• Comparative validation against MFS using a prospective cohort (132 participants) and ROC curve analysis.

Main Results:

• The XGB model demonstrated higher sensitivity for fall risk stratification compared to the MFS.
• Key predictors identified include specific hospital departments, medication use, admission source, and bed rest status.
• The model offers novel insights into in-hospital fall event prediction.

Conclusions:

• The XGB-based machine learning model shows promise for accurate and sensitive fall risk prediction.
• Identified predictors offer valuable information for clinical decision-making and targeted interventions.
• This approach enhances the identification of patients requiring closer monitoring to prevent falls.