Sensor-derived physical activity parameters can predict future falls in people with dementia

There is a need for simple clinical tools that can objectively assess the fall risk in people with dementia. Wearable sensors seem to have the potential for fall prediction; however, there has been limited work performed in this important area.

To explore the validity of sensor-derived physical activity (PA) parameters for predicting future falls in people with dementia. To compare sensor-based fall risk assessment with conventional fall risk measures.

This was a cohort study of people with confirmed dementia discharged from a geriatric rehabilitation ward. PA was quantified using 24-hour motion-sensor monitoring at the beginning of the study. PA parameters (percentage of walking, standing, sitting, and lying; duration of single walking, standing, and sitting bouts) were extracted using specific algorithms. Conventional assessment included performance-based tests (Timed Up and Go Test, Performance-Oriented Mobility Assessment, 5-chair stand) and questionnaires (cognition, ADL status, fear of falling, depression, previous faller). Outcome measures were fallers (at least one fall in the 3-month follow-up period) versus non-fallers.

77 people were included in the study (age 81.8 ± 6.3; community-dwelling 88%, institutionalized 12%). Surprisingly, fallers and non-fallers did not differ on any conventional assessment (p = 0.069-0.991), except for 'previous faller' (p = 0.006). Interestingly, several PA parameters discriminated between the groups. The 'walking bout average duration', 'longest walking bout duration' and 'walking bout duration variability' were lower in fallers, compared to non-fallers (p = 0.008-0.027). The 'standing bout average duration' was higher in fallers (p = 0.050). Two variables, 'walking bout average duration' [odds ratio (OR) 0.79, p = 0.012] and 'previous faller' (OR 4.44, p = 0.007) were identified as independent predictors for falls. The OR for a 'walking bout average duration' <15 s for predicting fallers was 6.30 (p = 0.020). Combining 'walking bout average duration' and 'previous faller' improved fall prediction (OR 7.71, p < 0.001, sensitivity/specificity 72%/76%).

RESULTS demonstrate that sensor-derived PA parameters are independent predictors of the fall risk and may have higher diagnostic accuracy in persons with dementia compared to conventional fall risk measures. Our findings highlight the potential of telemonitoring technology for estimating the fall risk. RESULTS should be confirmed in a larger study and by measuring PA over a longer period of time.