Identification of important factors in an inpatient fall risk prediction model to improve the quality of care using EHR and electronic administrative data: A machine-learning approach

BACKGROUND

Inpatient falls, many resulting in injury or death, are a serious problem in hospital settings. Existing falls risk assessment tools, such as the Morse Fall Scale, give a risk score based on a set of factors, but don't necessarily signal which factors are most important for predicting falls. Artificial intelligence (AI) methods provide an opportunity to improve predictive performance while also identifying the most important risk factors associated with hospital-acquired falls. We can glean insight into these risk factors by applying classification tree, bagging, random forest, and adaptive boosting methods applied to Electronic Health Record (EHR) data.

OBJECTIVE

The purpose of this study was to use tree-based machine learning methods to determine the most important predictors of inpatient falls, while also validating each via cross-validation.

MATERIALS AND METHODS

A case-control study was designed using EHR and electronic administrative data collected between January 1, 2013 to October 31, 2013 in 14 medical surgical units. The data contained 38 predictor variables which comprised of patient characteristics, admission information, assessment information, clinical data, and organizational characteristics. Classification tree, bagging, random forest, and adaptive boosting methods were used to identify the most important factors of inpatient fall-risk through variable importance measures. Sensitivity, specificity, and area under the ROC curve were computed via ten-fold cross validation and compared via pairwise t-tests. These methods were also compared to a univariate logistic regression of the Morse Fall Scale total score.

RESULTS

In terms of AUROC, bagging (0.89), random forest (0.90), and boosting (0.89) all outperformed the Morse Fall Scale (0.86) and the classification tree (0.85), but no differences were measured between bagging, random forest, and adaptive boosting, at a p-value of 0.05. History of Falls, Age, Morse Fall Scale total score, quality of gait, unit type, mental status, and number of high fall risk increasing drugs (FRIDs) were considered the most important features for predicting inpatient fall risk.

CONCLUSIONS

Machine learning methods have the potential to identify the most relevant and novel factors for the detection of hospitalized patients at risk of falling, which would improve the quality of patient care, and to more fully support healthcare provider and organizational leadership decision-making. Nurses would be able to enhance their judgement to caring for patients at risk for falls. Our study may also serve as a reference for the development of AI-based prediction models of other iatrogenic conditions. To our knowledge, this is the first study to report the importance of patient, clinical, and organizational features based on the use of AI approaches.

A data-driven and practice-based approach to identify risk factors associated with hospital-acquired falls: Applying manual and semi- and fully-automated methods

BACKGROUND AND PURPOSE

Electronic health record (EHR) data provides opportunities for new approaches to identify risk factors associated with iatrogenic conditions, such as hospital-acquired falls. There is a critical need to validate and translate prediction models that support fall prevention clinical decision-making in hospitals. The purpose of this study was to explore a combined data-driven and practice-based approach to identify risk factors associated with falls.

PROCEDURES

We conducted an observational case-control study of EHR data from January 1, 2013 to October 31, 2013 from 14 medical-surgical units of a tertiary referral teaching hospital. Patients aged 21 or older admitted to medical surgical units were included in the study. Manual and semi- and fully-automated methods were used to identify fall risk factors across four prediction models. Sensitivity, specificity, and the Area under the Receiver Operating Characteristic (AUROC) curve were calculated for all models using 10-fold cross validation.

FINDINGS

We confirmed the significance of a set of valid fall risk factors (i.e., age, gender, fall risk assessment, history of falling, mental status, mobility, and confusion) and identified set of new risk factors (i.e., # of fall risk increasing drugs, hemoglobin level, physical therapy initiation, Charlson Comorbity Index, nurse skill mix, and registered nurse staffing ratio) based on the most precise prediction approach, namely stepwise regression.

CONCLUSIONS

The use of semi- and fully-automated approaches with expert clinical knowledge over expert or data-driven only approaches can significantly improve identifying patient, clinical, and organizational risk factors of iatrogenic conditions, including hospital-acquired falls.