Fall Risk is an ML-derived value assigned to an individual based on the results of their most recent Stability Scan
Fall Risk is a representation of an individual’s risk of suffering a fall relative to their peers. It is derived utilizing cohort-specific modeling and is applicable to older adults (65+).
How does the scale work?
Fall Risk is a categorical metric with three unique values of risk on a scale: Low, Average, and High.
The model is developed utilizing data from an older adult population (65+), with the resulting risk values relative to this population. While most older adults are classified as Average based on the model, these patients are still candidates for treatment/intervention as well as additional balance, mobility, and fall risk factor evaluations.
Historically, traditional fall risk assessment has utilized binary classification (fall risk/no fall risk) to simplify interpretation. This ternary (three elements) classification provides further granularity to better engage patients and more accurately assess and address fall risk.
How does the model know this?
The Fall Risk model is iteratively developed, leveraging accumulating data of relevant balance force plate assessments and securely shared clinical data collected through research partners in senior healthcare. The software analyzes the results derived from these balance assessments and utilizes supervised machine learning models to stratify risk.
Read more about risk modeling in our white paper: Injury Risk: Is it predictable?
The current model iteration was developed as part of a research study. As part of model development, various data sources were collected and analyzed, along with the balance force plate assessment data (Stability Scan). This includes:
- General health questionnaire data (PROMIS Global Health)
- Relevant CDC health and fall-risk-related questionnaire data (Stay Independent, STEADI)
- Timed Up & Go (TUG) assessment data
The development process primarily focused on enhancing the classification model's performance (accuracy) and utility (real-world usefulness). Exploration included different modeling techniques, data source combinations, and TUG assessment cutoff thresholds. The present iteration is trained on TUG assessment data which employed a cutoff of 13.5 seconds. In the context of model development, the model achieves an AUC of greater than 0.70, signifying acceptable performance in accurately identifying individuals at risk of falling.
The current features that contribute to the model are listed below:
- Timed Up & Go (TUG) assessment data
- Force-plate derived balance data (posturography)
- Assigned Sex at Birth
- Mass
- Age
Read more: