Mining geriatric assessment data for in-patient fall prediction models and high-risk subgroups

Area of Science:

• Geriatric Medicine
• Data Science
• Patient Safety

Background:

• Hospital in-patient falls are a significant issue in geriatric care.
• Existing screening tools lack consistent predictive accuracy for falls.
• This study aimed to develop and evaluate fall risk classification models for geriatric in-patients.

Purpose of the Study:

• To derive comprehensible fall risk classification models from geriatric in-patient assessment data.
• To evaluate the predictive performance of these models.
• To identify high-risk patient subgroups within geriatric populations.

Main Methods:

• Utilized a dataset of 5,176 in-patient episodes from a geriatric hospital over 1.5 years.
• Developed classification tree (C4.5 algorithm) and logistic regression models.
• Matched admissions data with 493 fall incident reports to assess model performance.

Main Results:

• The classification tree model achieved 66% overall accuracy, with 55.4% sensitivity and 93.5% specificity.
• Identified five high-risk subgroups characterized by advanced age, low Barthel index, cognitive impairment, polypharmacy, and co-morbidity.
• Positive predictive value was 15%, indicating limitations in predicting actual fallers.

Conclusions:

• The developed models can identify non-fallers effectively but have limitations in predicting fallers due to low positive predictive value.
• Identified risk factors align with existing knowledge and can aid in screening high-risk geriatric patients.
• Data mining-derived models show potential comparable to current tools but require further validation in prospective settings.