Predicting falls using electronic health records: a time series approach

Area of Science:

• Healthcare Informatics
• Machine Learning in Medicine
• Patient Safety

Background:

• Traditional fall risk assessment tools like the Morse Fall Scale (MFS) have limitations in accuracy.
• Accurate fall prediction is crucial for patient safety in acute care settings.
• Electronic health records (EHRs) contain rich data for developing predictive models.

Purpose of the Study:

• To develop and validate a more accurate fall prediction model for patients within the Veterans Health Administration (VHA).
• To leverage existing EHR data for automated and efficient fall risk assessment.
• To improve upon the predictive capabilities of existing fall risk assessment tools.

Main Methods:

• A cohort of Veterans admitted to VHA acute care settings between July 2020 and June 2022 was analyzed.
• Demographic and clinical data were extracted from EHRs, with falls identified via clinical progress notes.
• A transformer model was employed for feature extraction, followed by training a Light Gradient-Boosting Machine for classification.

Main Results:

• The study analyzed 242,844 Veterans, with 2.5% experiencing a documented fall.
• The developed model achieved an Area Under the Curve (AUC) of .851 and an Area Under the Precision-Recall Curve (AUPRC) of .285.
• The model demonstrated 76.3% accuracy, with 76.2% specificity and 77.3% sensitivity.

Conclusions:

• The developed machine learning model significantly outperforms traditional methods like the Morse Fall Scale for fall risk prediction.
• This automated model, derived from existing EHR data, offers a more efficient and accurate approach to identifying at-risk patients.
• Further research is needed to address class imbalance and conduct prospective validation for broader implementation.