Fall Risk Assessment Using Wearable-Based Turn Detection:

Comparison of Different Algorithms During Real-World Monitoring

Jose Albites-Sanabria

1,2 a

, Pierpaolo Palumbo

, Stefania Bandinelli

, Luca Palmerini

1,4 d

and Lorenzo Chiari

1,4 e

Department of Electrical, Electronic, and Information Engineering – DEI, University of Bologna, Italy

Institute of Advanced Studies, University of Bologna, Italy

Azienda Sanitaria Toscana Centro, Firenze, Piero Palagi Hospital, Firenze, Italy

Health Sciences and Technologies-Interdepartmental Center for Industrial Research, University of Bologna, Italy

Keywords: Turning, Wearable Sensors, Continuous Monitoring, Older Adults, Falls.

Abstract: Turning deficits have been linked to aging and movement disorders and are a common cause of falls and

fractures. Despite previous works on the automatic identification of turns and on its relation to fall risk,

different algorithms for turn identification have been used, but their agreement and differences have not been

investigated. In this study, we compared the two most-used turn-validated algorithms (El-Gohary and Pham)

using a dataset comprising real-world data from 171 community-dwelling older adults monitored for one

week with a single wearable sensor. The quantity and quality of turn parameters were calculated and used as

predictors of future falls. After the analysis, the El-Gohary and Pham algorithms identified 1,063,810 and

942,845 turns, respectively. The agreement of the algorithms showed a very high to moderate correlation for

all turn parameters. We found that prospective fallers take longer to perform a turn, and their movements are

less smooth when compared to non-fallers. A fall risk assessment model built only on turn parameters showed

reasonable performance for both algorithms (AUC = 0.6). Our results show that differences between turn

parameters in the algorithms, when averaged at the single-subject level, are less of a concern when looking

for associations with prospective falls.

1 INTRODUCTION

Turning represents a major component of everyday

walking behavior, as between 35 and 45% of steps

occur within turns (Glaister et al., 2007). However,

up until recent years, studies only focused primarily

on straight-ahead walking. Turning requires a

continuous change of the center of mass and multi-

limb coordination, so it is not surprising that its

deficits are associated with movement disorders and

the risk of falling.

Several studies have noted that turns can

challenge stability maintenance and increase energy

expenditure, and that turning time, steps per turn, and

variability in the number of steps across different

https://orcid.org/0000-0001-7688-6221

https://orcid.org/0000-0002-4438-1787

https://orcid.org/0000-0002-6491-0850

https://orcid.org/0000-0003-4758-662X

https://orcid.org/0000-0002-2318-4370

turns are valuable features for distinguishing fallers

from non-fallers (Mancini et al., 2016). Subtle fall-

risk-related gait-based measures may become highly

effective fall-risk indicators when applied to turns due

to the increased challenge to stability compared to

straight walking. Individuals at high risk of falling

employ different turning methods than healthy

individuals.

Assessment of turning is not trivial. Optical

systems have been widely used in previous studies

but are cumbersome, expensive, and can only be used

in controlled environments (Marín et al., 2020;

Thigpen et al., 2000). Wearable sensors, which can

measure for days or even weeks, are a promising

294

Albites-Sanabria, J., Palumbo, P., Bandinelli, S., Palmerini, L. and Chiari, L.

Fall Risk Assessment Using Wearable-Based Turn Detection: Comparison of Different Algorithms During Real-World Monitoring.

DOI: 10.5220/0011727700003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 4: BIOSIGNALS, pages 294-300

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

alternative. Hence, they are ideal in unconstrained

environments over long periods of time.

Algorithms for analyzing the turning movements

of older adults and Parkinson’ disease patients have

already been published. To the best of our knowledge,

only two algorithms based on one inertial sensor

(accelerometer and gyroscope) worn on the lower

back have been validated against video observation

(gold standard) with reasonable agreement. These

algorithms have been and are currently used by other

studies to extract relevant turning parameters

associated with movement disorders and the risk of

falling (Haertner et al., 2018; Leach et al., 2018;

Roussos et al., 2022; Thierfelder et al., 2022).

However, using different algorithms increases

heterogeneity in remote monitoring studies;

validation and adoption of standardized

digital

mobility biomarkers is an ongoing task being

addressed by different initiatives.

In this study, we tested two algorithms to identify

turns and extract turning characteristics in real-world

conditions. We aim to compare here the performance

of the two algorithms and their impact on assessed

turn quantity and quality during a week of monitoring

relative to prospective falls. To the best of our

knowledge, this is the first study characterizing

different turning biomarkers worn on the lower back

for fall risk assessment in real-world conditions.

2 METHODS

Study Participants and Settings

The present study is based on data from the 4

wave

of the “Invecchiare in Chianti” (InCHIANTI) study.

One hundred and seventy-one community-dwelling

older adults over 65 (79·7±6·6) years, 50·9% female,

were monitored for 5–9 days using a smartphone

(Samsung Galaxy SII), embedded with a tri-axial

accelerometer and gyroscope with a 100 Hz sampling

rate, worn on the midsagittal plane of the lower back

during all waking hours.

Participants brought the device home, used it for

one week, and then returned it to the clinical staff at

the end of the monitoring period. Telephone

interviews were used to collect prospective fall

incidence data between 6 and 12 months after the start

of continuous monitoring. Participants who did not

fall were defined as non-fallers [NFs] and participants

who fell one or more times were defined as fallers

[Fs].

The study protocol was approved by the ethical

committee of the Italian National Institute of

Research and Care of Aging and complies with the

Declaration of Helsinki. All participants received a

detailed description of the study purpose and

procedures and gave their written informed consent.

Turns

Two validated algorithms for turning detection were

implemented in Python 3.8

(El-Gohary et al., 2013) algorithm measures the

angular rotational rate of the pelvis about the vertical

axis (𝑤



). Candidate turns are detected in segments

where the maxima of the low-pass filtered ( 𝑓



1.5 𝐻𝑧) 𝑤



exceed a threshold of 15°/s. The start and

end of turns are found when the filtered signal drops

below 5°/s. The direction of the turn (right or left) was

defined by the sign of 𝑤



(Pham et al., 2017) algorithm estimates the

angular displacement around the vertical axis through

attitude estimation. The start of a right turn is defined

by a change from an increase to a decrease of the

angular displacement, and the end by a change from

a decrease to an increase of the angular displacement.

The opposite operation is applied to the definition of

a left turn.

Both methods rely on a single inertial sensor worn

on the lower back to detect turns. Still, different post-

processing cutoffs are suggested to improve the

performance of the algorithm based on heuristics. The

thresholds were optimized and validated using video

observations according to the information reported by

the authors of the algorithms. Table 1 presents a

description of both algorithms.

Table 1: Characteristics of turn algorithms.

El-Gohar

Pha

Sensor accelerometer

roscope

accelerometer +

roscope

Location Low bac

Low bac

Identification

method

Maxima from

filtered vertical

angular

velocit

Changes in

vertical angular

displacement

Turn

duration

threshold*

0.5 – 5 s 0.1 – 10 s

Turn angle

threshold*

45° 90°

* Thresholds Suggested by Authors

To standardize the comparison of both algorithms,

turns with angles between 50–200° and durations

between 0.5–5 seconds were applied in the

implementation of the algorithms and were analyzed.

Fall Risk Assessment Using Wearable-Based Turn Detection: Comparison of Different Algorithms During Real-World Monitoring

295

Turns were divided into three subsets based on turn

angle (small (50–100°], medium (100–150°], and

large (150–200°]) to account for different motor

planning strategies within our analysis.

Following what was defined in previous studies,

we calculated different quantity and quality turn

parameters. Turn quantity was characterized by the

number of turns per hour (TPH). Turn quality was

characterized by the turn duration (DUR), turn angle

(ANG), mean velocity (MV), and peak turn velocity

(PV)(Caby et al., 2011; Leach et al., 2018), and the

spectral arc length (SPARC)(Figueiredo et al., 2020;

Gulde & Hermsdörfer, 2018).

Statistical Analyses

The degree of agreement for turn detection between

the two algorithms was calculated using a correlation

matrix of quantity and quality parameters of turns.

Univariate and k-fold cross validation logistic

regression analysis was used to evaluate the

association of turn parameters with prospective falls

for both algorithms. The quantity and quality turn

parameters were included as independent variables in

the univariate model. The correlation between

quantity and quality parameters was used to select a

set of possible explanatory variables in the

multivariate model. All analyses were performed

using Python 3.8. All p values were two-tailed, and p

< 0.05 was considered significant.

3 RESULTS

Cohort and Fall Status

Table 2 presents demographic and clinical data about

participants included in the study, labeled as fallers

and non-fallers.

Table 2: Cohort characteristics for 12-month prospective

falls.

Non-Fallers

[NFs]

(N=142)

Fallers [Fs]

(N=29)

Combined

(N=171)

Gender

(

M/F

)

73/69 11/18 84/87

Age

(years)

79.4 ± 6.7 81.1 ± 5.5 79.7 ± 6.5

Height

(cm)

159.8 ± 9.1 159 ± 9.5 159.6 ± 9.1

Weight

(

70.7 ± 13.1 70.2 ± 14.4 70.6 ± 13.3

MMSE 27.3 ± 1.9 27.1 ± 1.8 27.3 ± 1.8

Turns Characterization

A total number of 1,063,810 and 942,845 turns were

detected from the dataset with El-Gohary and Pham

algorithms, respectively (Figure 1).

Figure 1: Turns identified during real-world monitoring.

The average across days was computed for each

participant. “El-Gohary and Pham algorithms showed

very high agreement on TPH ( 𝑅



=0.97), high

agreement on DUR, ANG, and SPARC (𝑅



between

0.74 and 0.82) and moderate agreement on PV and

MV (𝑅



0.51-0.61) (Figures 2-3, Table 3).

Figure 2: Turn quantity correlation identified by turn

algorithms.

Table 3 summarizes descriptive characteristics for

turn quantity and quality parameters. Computed

DUR, ANG, and SPARC revealed high agreement

between quality turn characteristics identified by both

algorithms. MV and PV showed moderate correlation

among the computed parameters.

While not reported in the present manuscript,

outliers were identified in MV and PV, which may be

responsible for the lower agreement.

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

296

Figure 3: Correlation (upper) and Bland-Altman plot

(bottom) for turn duration (s).

Table 3: El-Gohary and Pham correlation and mean

difference.

Overall means + SD

𝑅



Mean

diff

El-Gohary Pham

TPH (/h) 74.87±34.8 66.44±31.71 0.97 8.43

DUR (s) 2.57±0.33 2.61±0.46 0.79 0.04

ANG (°) 86.77±3.88 99.08±4.47 0.74 12.31

MV (°/s) 43.76±7.02 44.8±4.9 0.61 1.04

PV (°/s) 96.04±16.16 92.07±8.4 0.51 3.97

SPARC -2.14±0.09 -2.04±0.07 0.82 0.1

Taking physical properties of body movement

into account, it is expected that some of the quality

parameters extracted from turns will be correlated

(angle, velocity, duration). Therefore, to avoid

collinearity problems in the following multivariate

analysis, we analyzed potential correlations between

parameters. Figure 4 shows the correlation matrix for

all parameters (turn quantity and quality).

To account for different motor planning strategies

individuals take when performing a turn, three

subsets based on turn angle (small (50–100°],

medium (100–150°], and large (150–200°]) were

analyzed. As shown in figure 5, despite a high

agreement in angle estimation of both algorithms, the

subtle differences in the estimation techniques lead to

considerable differences when differentiating turns

based on their angle ranges.

Figure 4: Correlation matrix for turn quantity and quality

parameters.

Figure 5: Turns’ subsets division based on turn angle.

Turns and Prospective Falls

To identify associations between turn parameters and

prospective falls, the fall incidence used in the

analysis was calculated after a 6-month (NFs: 157,

Fs: 14) and a 12-month period (NFs: 142, Fs: 29).

The odds ratios that quantify the univariate

associations between turn quantity and quality

parameters and fall status after 12 months are shown

in Figure 6. Z-scored was applied for better

Fall Risk Assessment Using Wearable-Based Turn Detection: Comparison of Different Algorithms During Real-World Monitoring

297

visualization of the forest plot. The parameters were

grouped by characteristics according to the algorithm

used for turn detection and angle-range subsets (the

prefix 50, 100, or 150 defines the type of subset

analyzed).

Turn measures were associated with prospective

fall status when analyzing all turns for both

algorithms. Despite some differences, both

algorithms identified the same parameters that were

strongly associated with future falls. More TPH,

longer DUR, and less smooth movements (SPARC)

were associated with the risk of falling (Figure 6). PV

and MV demonstrated similar trends to DUR, which

is in agreement with findings in the turns

characterization section (correlation matrix, figure 4).

Finally, specific angle-range subsets (e.g., (150–

200°] TPH) seemed to provide stronger evidence for

turn associations with prospective falls.

Figure 6: Forest plot of univariate analysis for turn

parameters associated with 12-month prospective falls, El-

Gohary (top), Pham (bottom).

For multivariate analysis, we then performed a

selection of parameters based on the univariate

analysis and previous results from the correlation

matrix analysis. Since PV and MV were moderately

and highly correlated with DUR in the El-Gohary and

Pham parameters, respectively, both parameters were

removed from the following multivariate analysis.

The results of the ROC curve analysis using TPH,

DUR, ANG, and SPARC to classify fallers vs. non-

fallers over a 6-month and 12-month period are

shown in Figure 7. A different set of parameters (e.g.,

based on specific angle ranges) was also analyzed and

was found to only marginally improve the

performance of the classifier.

Figure 7: ROC curve for 6-month (top) and 12-month

(bottom) prospective falls.

4 CONCLUSIONS

In this study, we compared two wearable-based turn

detection algorithms and assessed their importance in

real-world fall risk assessment.

Although both algorithms are based on the same

“principle” (e.g., estimating turns based on the

rotation of the pelvis around the vertical axis),

different processing steps to identify turn events lead

to significant differences in the number of detected

turns and angles estimated by both algorithms. The

readings from the gyroscope (i.e., the angular speed)

are generally very accurate; however, drift might

occur when integrating gyroscope readings over

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

298

longer periods, such as in continuous monitoring

experiments. Future studies could apply available

techniques to avoid drifting, such as the integration of

data coming from the orientation sensor (magnetic

plus acceleration) and data coming from the

gyroscope. The use of additional sensors combined

with data fusion techniques could improve accuracy

in the identification of turns while increasing

computational and power costs.

Despite some differences and potential errors in

estimating some quantity and quality parameters,

both algorithms showed a moderately to very high

correlation. We hypothesize that the differences

among turn parameters at the single-subject level are

less of a concern when looking for associations with

prospective falls. In line with this discussion, we

could summarize a pipeline-process: turn detection,

calculation of turn parameters at the single-turn level,

and calculation of the average over turns of each

subject to extract turn parameters at the subject level.

The last two steps downstream (probably, the average

step in particular) attenuate the discrepancies, making

the two algorithms exchangeable. Initial evidence for

this statement is given by the similar performance of

the logistic regression model built on the identified

turning parameters with both algorithms.

All in all, the results and parameters presented

here are in line with previous research studies and

with current clinical standards tests. In fact, turning

ability is a fundamental aspect of several walking

tests, including the Timed Up and Go Test (TUG),

which is used to discriminate fallers from non-fallers.

Other cohorts could also be explored in prospective

longitudinal studies, it should be noted that the

percentage of fallers after 6 and 12 months in this

cohort was significantly lower than the global

statistics for falls in older adults.

Last but not least, a quick review of the literature

shows an exponential increase in reports related to

wearable-based monitoring for fall prevention.

However, despite several efforts to use this

technology for assessments of both healthy and

pathological movement patterns, the high level of

heterogeneity in the use of wearables (e.g., sensor

location and extracted gait parameters) makes it hard

to yield conclusive results. While some ongoing

initiatives aim to establish the clinical validity of

digital mobility biomarkers in different cohorts, some

real-world characteristics, such as turning, deserve

deeper analysis.

ACKNOWLEDGEMENT

This study was partially funded by the Innovative

Medicines Initiative 2 Joint Undertaking under grant

agreement No 820820 (Mobilise-D). This Joint

Undertaking receives support from the European

Union’s Horizon 2020 research and innovation

programme and EFPIA.

REFERENCES

Caby, B., Kieffer, S., de Saint Hubert, M., Cremer, G., &

Macq, B. (2011). Feature extraction and selection for

objective gait analysis and fall risk assessment by

accelerometry. BioMedical Engineering OnLine, 10(1),

1. https://doi.org/10.1186/1475-925X-10-1

El-Gohary, M., Pearson, S., McNames, J., Mancini, M.,

Horak, F., Mellone, S., & Chiari, L. (2013). Continuous

monitoring of turning in patients with movement

disability. Sensors (Basel), 14(1), 356–369.

Figueiredo, A. I., Balbinot, G., & Brauner, F. O. (2020).

SPARC Metrics Provide Mobility Smoothness

Assessment in Oldest-Old With and Without a History

of Falls : A Case Control Study. 11(June), 1–11.

https://doi.org/10.3389/fphys.2020.00540

Glaister, B. C., Bernatz, G. C., Klute, G. K., & Orendurff,

M. S. (2007). Video task analysis of turning during

activities of daily living. 25, 289–294. https://

doi.org/10.1016/j.gaitpost.2006.04.003

Gulde, P., & Hermsdörfer, J. (2018). Smoothness Metrics

in Complex Movement Tasks. 9(September), 1–7.

https://doi.org/10.3389/fneur.2018.00615

Haertner, L., Elshehabi, M., Zaunbrecher, L., Pham, M. H.,

Maetzler, C., van Uem, J. M. T., Hobert, M. A., Hucker,

S., Nussbaum, S., Berg, D., Liepelt-Scarfone, I., &

Maetzler, W. (2018). Effect of fear of falling on turning

performance in Parkinson’s disease in the lab and at

home. Frontiers in Aging Neuroscience, 10(MAR), 1–

8. https://doi.org/10.3389/fnagi.2018.00078

Leach, J. M., Mellone, S., Palumbo, P., Bandinelli, S., &

Chiari, L. (2018). Natural turn measures predict

recurrent falls in community-dwelling older adults: a

longitudinal cohort study. Scientific Reports, 8(1),

4316. https://doi.org/10.1038/s41598-018-22492-6

Mancini, M., Schlueter, H., El-Gohary, M., Mattek, N.,

Duncan, C., Kaye, J., & Horak, F. B. (2016).

Continuous Monitoring of Turning Mobility and Its

Association to Falls and Cognitive Function: A Pilot

Study. The Journals of Gerontology Series A:

Biological Sciences and Medical Sciences, 71(8),

1102–1108. https://doi.org/10.1093/gerona/glw019

Marín, J., Blanco, T., de la Torre, J., & Marín, J. J. (2020).

Gait analysis in a box: A system based on

magnetometer-free IMUs or clusters of optical markers

with automatic event detection. Sensors (Switzerland),

20(12), 1–27. https://doi.org/10.3390/s20123338

Fall Risk Assessment Using Wearable-Based Turn Detection: Comparison of Different Algorithms During Real-World Monitoring

299

Pham, M. H., Elshehabi, M., Haertner, L., Heger, T.,

Hobert, M., Faber, G., & Maetzler, W. (2017).

Algorithm for Turning Detection and Analysis

Validated under Home-Like Conditions in Patients with

Parkinson ’ s Disease and Older Adults using a 6

Degree-of-Freedom Inertial Measurement Unit at the

Lower Back. 8(April), 1–8. https://doi.org/10.3389/

fneur.2017.00135

Roussos, G., Herrero, T. R., Hill, D. L., Dowling, A. V.,

Müller, M. L. T. M., Evers, L. J. W., Burton, J., Derungs,

A., Fisher, K., Kilambi, K. P., Mehrotra, N., Bhatnagar,

R., Sardar, S., Stephenson, D., Adams, J. L., Ray Dorsey,

E., & Cosman, J. (2022). Identifying and characterising

sources of variability in digital outcome measures in

Parkinson’s disease. Npj Digital Medicine, 5(1), 1–10.

https://doi.org/10.1038/s41746-022-00643-4

Thierfelder, A., Seemann, J., John, N., Harmuth, F., Giese,

M., Schüle, R., Schöls, L., Timmann, D., Synofzik, M.,

& Ilg, W. (2022). Real-Life Turning Movements

Capture Subtle Longitudinal and Preataxic Changes in

Cerebellar Ataxia. Movement Disorders, 37(5), 1047–

1058. https://doi.org/10.1002/mds.28930

Thigpen, M. T., Light, K. E., Creel, G. L., & Flynn, S. M.

(2000). Turning difficulty characteristics of adults aged

65 years or older. Physical Therapy, 80(12), 1174–

1187. https://doi.org/10.1093/ptj/80.12.1174

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

300