Digital assessment of falls risk, frailty, and mobility impairment using wearable sensors

Falls are among the most frequent and costly population health issues, costing $50bn each year in the US. In current clinical practice, falls (and associated fall risk) are often self-reported after the “first fall”, delaying primary prevention of falls and development of targeted fall prevention interventions. Current methods for assessing falls risk can be subjective, inaccurate, have low inter-rater reliability, and do not address factors contributing to falls (poor balance, gait speed, transfers, turning). 8521 participants (72.7 ± 12.0 years, 5392 female) from six countries were assessed using a digital falls risk assessment protocol. Data consisted of wearable sensor data captured during the Timed Up and Go (TUG) test along with self-reported questionnaire data on falls risk factors, applied to previously trained and validated classifier models. We found that 25.8% of patients reported a fall in the previous 12 months, of the 74.6% of participants that had not reported a fall, 21.5% were found to have a high predicted risk of falls. Overall 26.2% of patients were predicted to be at high risk of falls. 29.8% of participants were found to have slow walking speed, while 19.8% had high gait variability and 17.5% had problems with transfers. We report an observational study of results obtained from a novel digital fall risk assessment protocol. This protocol is intended to support the early identification of older adults at risk of falls and inform the creation of appropriate personalized interventions to prevent falls. A population-based approach to management of falls using objective measures of falls risk and mobility impairment, may help reduce unnecessary outpatient and emergency department utilization by improving risk prediction and stratification, driving more patients towards clinical and community-based falls prevention activities.


Wearable sensor data
For each TUG test, participants were fitted with two wearable sensors (wireless inertial measurement units) which were attached by a user, using dedicated Velcro straps (or elasticated bandages if straps did not fit), to the mid-point of the left and right anterior shank (shin) 5 .
Supplementary figure 1 details the setup. Participants were asked to complete the TUG test, 'as fast as safely possible'. A standard chair with armrests was recommended. The timer was started the moment the clinician said 'go', and stopped the moment the participant's back touched the back rest of the chair. It was recommended that each participant be given time to become familiar with the test and that the test be demonstrated to them beforehand.
Wearable inertial sensors were oriented to capture movement about the anatomical medio-lateral axis. Inertial sensor data were simultaneously acquired from the two sensors via Bluetooth (QTUG™, Kinesis Health Technologies; Dublin, Ireland). Each sensor was tri-axial and contained an accelerometer and gyroscope. Sensor data were sampled at 102.4 Hz with a full scale range of 500 °/s and a sensitivity of 2 mV/°/s. Sensors were calibrated using a standard method 6 . The raw tri-axial gyroscope signals were low-pass filtered (zero-phase 2 nd order Butterworth filter, 20 Hz corner frequency). Each test takes approximately five minutes, including application of the sensors and explaining the protocol.

Fall risk estimate
The AGS/BGS offer guidelines aimed to capture the main clinical risk factors linked to falls in older adults 4 . A logistic regression model was created using a number of the self-reported factors discussed in the AGS/BGS guidelines. History of falls in the past 5 years for each participant was obtained through a questionnaire and used as the target variable for each model. A fall was defined as "an event which resulted in a person coming to rest on the lower level regardless of whether an injury was sustained, and not as a result of a major intrinsic event or overwhelming hazard" 7 . Fall outcome data were verified using available hospital records as well as information provided by relatives.
The features included and used to classify participants according to falls history were as follows: gender, height, weight and age on date of assessment, polypharmacy, vision problems and orthostatic hypotension. A logistic regression model is used to produce a fall risk estimate (FREclinical); An estimate of the classifier performance of each model on unseen data was obtained using ten repetitions of ten-fold cross-validation 8 .
The sensor-based fall risk estimate (FREsensor) method uses a subset of the QTUG mobility parameters applied to a regularized discriminant (RD) classifier model 9 , with regularization parameter values set to λ=0.1 and r=0.1 prior to analysis. Ten repetitions of ten-fold cross validation 8 was used to estimate the generalized classifier performance. Using only the training data for each iteration of the crossvalidation routine, a potential feature set was evaluated using a second inner cross-validation loop.
Once a feature set is identified using the training data, it is tested using the withheld data for this iteration of the outer cross-validation loop 10 , a process known as 'nested' cross-validation. Training and testing sets were randomly selected for each repetition.

Frailty estimate
A logistic regression model with interaction terms included was used to calculate FEsensor, using QTUG sensor features, combined with gender, age, height and weight, with a separate classifier model per gender. The features included in each model were selected using a cross-validated sequential forward feature selection procedure. Models were then evaluated using a separate repeated crossfold validation, with 10 folds and 10 repetitions. Frailty was considered as a binary classification problem; grouping participants identified as frail and pre-frail (by the Fried phenotype criteria) together into one frail class and comparing this to a non-frail (robust under the Fried phenotype criteria) class. The output of this model was an estimate of the frailty category (frail/non-frail). This estimated frailty category was then compared to the true frailty category (as defined using modified Fried criteria 11 ) to yield an estimate of the accuracy in classifying each participant according to frailty category. FEclinical and FEcombined were calculated using the methodology outlined above for FREclinical and FREcombined.

Mobility impairment scores
The sensor data for each participant was processed using a previously reported algorithm 5,12 for assessment of gait and mobility. 59 features were calculated from the sensor data for each participant (known as the QTUG parameters), in order to characterize mobility.
Mobility issues are identified by grouping the 59 calculated mobility parameters into five functional categories: Speed, Variability, Symmetry, Transfers, Turning.
Mobility issues per functional category are identified by calculating a z-score calculated for each QTUG parameter per group; = −µ where µ is population mean for a given parameter x, and σ is the population standard deviation. The population reference data are obtained from a previously reported independent sample 13 . The population data are stratified by gender to produce genderspecific values for population mean and standard deviation. The mean z-score per group is then calculated; if |zµ| ≥ 2, group is determined to be out of normal range. An estimate of the percentile is calculated by applying the normal cumulative distribution function = Positive parameters are defined as those for which a large value is considered to be a clinical indicator of good mobility (e.g. gait velocity), whereas a negative parameter is one where a large value is one where a large value is considered to be a clinical indicator of poor mobility (e.g. TUG time). A neutral parameter is then defined as one that does not fit into either category. If the mean