fall risk - Gruppo Nazionale di Bioingegneria

La Bioingegneria per il benessere e l’invecchiamento attivo
Gruppo Nazionale di Bioingegneria
Bressanone 2016
Tecnologie indossabili per il monitoraggio
e la prevenzione delle cadute nell’anziano
Rita Stagni, Sabato Mellone, Pierpaolo Palumbo,
Lorenzo Chiari, Angelo Cappello
DEI, Università di Bologna
Schema della lezione
• La caduta: definizione, incidenza e costi
• Fattori di rischio intrinseci, estrinseci e
legati all’attività
• L’equilibrio e il suo controllo: il ruolo
della modellistica
• Stabilità motoria e rischio di caduta
• Rilevazione in tempo reale delle cadute
mediante tecnologie indossabili: misura,
allarmi e feedback
• Predizione della caduta
• Prevenzione della caduta: interventi
mirati alla riduzione del rischio
• Conclusioni
FALL: Definition and Incidence
• Fall: «an unexpected event in which the participant
comes to rest on the ground, floor, or lower level»
[1].
• Incidence rate of falls is quite high: approximately 30% of persons
over 65 fall at least once per year [2].
• 65 % of women and 44% of men fall inside their
usual place of residence [3].
• Up to 30% of these falls leads to injuries.
[1] Axer et al. Falls and gait disorders in geriatric neurology. Clinical Neurology and Neurosurgery, 112: 265-274, 2010.
[2] Masud et al. Epidemiology of falls. Age Ageing 30(S4):3-78, 2001.
[3] Hauer et al. Systematic review of definitions and methods of measuring falls in randomized controlled fall prevention trials. Age Ageing 35: 5-10, 2006.
COSTI
1. Nel solo 2009, le cadute hanno determinato costi che variano tra lo
0,85 e lo 1,5 per cento delle spese sanitarie totali negli Stati Uniti,
Australia, UE e Regno Unito (Heinrich et al., 2009).
2. Le cadute hanno anche un impatto notevole sulle condizioni di salute
generali di uno stato, dato che lo 81-98% delle fratture sono causate
da cadute (Tinetti, 2003), e queste sono la principale causa di
accessi al pronto soccorso in USA (Fuller, 2000).
3. Il rischio di una caduta aumenta con l'età (Mathers e Weiss, 1998);
le cadute rappresentano l'eziologia primaria di morte accidentale in
soggetti con più di 65 anni, e anche il tasso di mortalità associato
aumenta notevolmente con l'età, con picchi pari al 70% delle morti
accidentali nelle persone di 75 anni di età (Fuller, 2000).
4. I principali costi associati tendono quindi a verificarsi in gruppi di età
più avanzata e a seguito di fratture, un problema che si aggrava
ulteriormente con l’invecchiamento della popolazione (Hamacher et
al., 2011).
There are currently over 400 known risk
factors for falls5
EXTRINSIC
•
•
•
•
•
•
•
•
poor lighting
surface elevation
surface roughness
obstacles
clothing/footwear
lack of equipment
perturbations
…
INTRINSIC
•
•
•
•
•
•
•
•
•
•
•
•
…
age and gender
muscular strength
reaction time
visual impairment
ethnicity
use of drugs
living alone
sedentary behavior
psychological status
impaired cognition
cardiovascular issues
foot problems
TASK-RELATED
•
•
•
•
task complexity and
speed
fatigue
load handling
…
[5] Masud & Morris, Age and Ageing, 2001
Fear of falling
The resultant fear of
falling and selfimposed restrictions
in mobility and
function may be
contributing factors
for nursing home
admission.
BALANCE & FALLS: A
CONCEPTUAL MODEL
COM AND COP AS VARIABLES LINKED TO STABILITY
• CoM variations are related to the whole body mass movement
• CoP variations are proportional to the ankle torque (they do not represent any
movement!)
QUIET STANDING
The inverted pendulum model predicts a high correlation
between CoP-CoM and the horizontal acceleration of
CoM in the A/P direction
A subject swaying back and forth while standing quietly
on a force platform. The centre of gravity and the centre
of pressure (p) locations along with the associated
angular accelerations (α) and angular velocities (ω)
Mechanism of balance as diagnostic tool to pinpoint deficits
of the system (vision, vestibular, somatosensory)
Sistema Muscolo-Scheletrico
Effettori muscolari e
osteoarticolari
eseguono i comandi
centrali ed
assecondano i riflessi
mettendo in atto quei
movimenti
compensatori e/o
anticipatori che
consentono di restare
in equilibrio
Le strategie di compenso più
comuni
Strategia di
caviglia
Strategia
d’anca
Stepping
Horak, 1989
I diversi meccanismi che partecipano al
controllo
- un meccanismo puramente biomeccanico ad azione
istantanea, dovuto alla rigidezza (stiffness) muscolare;
- un meccanismo reattivo, che opera in catena chiusa ed
è determinato dai feedback sensoriali che, in un tempo finito,
informano il SNC dello stato attuale del sistema controllato;
- un meccanismo anticipativo, che opera in catena aperta
ed è basato su un modello interno di fusione sensoriale e di
predizione della dinamica del sistema
Winter et al., 1998; Morasso e Schieppati, 1999; Loram e Lakie 2002
GAIT
Dynamics of the balance
of HAT during stance
Center of gravity (COG) and Center
of Pressure (COP) trajectories
during gait
AN INDEX OF STABILITY: THE COM-COP ANGLE
By using wearable sensors we can estimate Centre of Mass and Centre of Pressure by using
a model-based approach.
ϴ
(CoMX , CoMY)
𝐶𝑜𝑀𝑋 − 𝐶𝑜𝑃𝑋
𝜃 = 𝑎𝑠𝑖𝑛
𝐶𝑜𝑀𝑌
[1,2]
Y
X
CoPX
How many sensors do we need to estimate these two variables?
[1] Lee et al., “Detection of gait instability using the centre of mass and centre of pressure inclination angles”, Arch Phys Med Rehab, vol. 87, pp. 569-575,
2006.
[2] Chiari et al., “Stabilometric parameters are affected by anthropometry and foot placement”, Clin Biomech, vol. 17, pp. 666-677, 2002.
Some questions…
1.How many sensors?
2.Best position?
3.Inertial sensors … and nothing else?
4.What about the accuracy of a fall
detection algorithms?
Model  Translational Research
In this respect, wearable inertial sensors (WIS) may be one of the
answers.
There is a growing body of literature suggesting that it is possible to link,
theoretically and/or empirically, specific parameters extracted by WIS to
basic physiological components of postural control.
Furthermore, WIS may allow the measurement of these components
during participation (e.g. in ordinary living conditions), whereas clinical
scales and tools require a standardized setting (that is not so
ecological).
WIS are of course natural candidates also for measuring physical
activity.
Out of the Lab: how is it possible to measure?
Gyroscopes
Accelerometers
They sense angular motion
about one or several axes
They sense linear acceleration along one or
several directions
Inertial Measurement Unit & Smartphones
Smartphones
 High accuracy
 Low cost, dimensions, power consumption
 High portability
WEARABLE DEVICES
SENSOR CONFIGURATIONS
3 Single-axis
accelerometers
Configuration
Shank
Thigh
HAT
1
2
2 IMUs
3
4
1 IMU
BLOCK DIAGRAM
DIRECT
DYNAMICS
FFORCE
PLATFORM
Fx,Fy,Mz
Subject-specific
anthropometric
parameters
CoP
𝐹𝑥, Fy, Mz
DIRECT DYNAMICS
Inertial sensors
Kinematics
Θ,Θ, Θ
Validation
Anthropometric
parameters
CoP, 𝑪𝒐𝑴
𝐹𝑥, Fy, Mz
RESULTS(1): CENTRE OF PRESSURE (COP)
0.3
0.2
CoP X[m]
0.1
1 IMU
0
-0.1
-0.2
-0.3
-0.4
FP
TRUNK SENSOR
0
1
2
3
4
5
6
7
Time[s]
CoP RMSE
2.5 cm
2.6 cm
2.7 cm
2.9 cm
4.5 cm (DeLeva parameters)
RESULTS(2):THE COM-COP ANGLE
Sensor on the trunk
20
CoM-CoP angle[deg]
15
10
5
0
-5
𝐶𝑜𝑀𝑋 − 𝐶𝑜𝑃𝑋
𝜃 = 𝑎𝑠𝑖𝑛
𝐶𝑜𝑀𝑌
-10
-15
0
1
2
3
4
Time[s]
5
6
7
Balance & Falls: Conceptual Model
Fall: the unwanted outcome of a
motor activity over-challenging the
postural control mechanisms
responsible of maintaining balance,
making them surrender to force of
gravity.
Fall
Over-challenging motor activity
Physiology
+
Environment
(gravity)
Under-challenging motor activity
Balance
Balance & Falls: Conceptual Model
Subject j
Which measures?
Balance & Falls: Conceptual Model
• Physiological Measures of the Postural Control System
• Activity Measures
• Environmental Measures
Balance & Falls: Conceptual Model
Postural Control is not a unidimensional variable
De Oliveira et al, JRRD, 2008
He et al, Proc. IEEE, 2008
Long-term recordings
Methods for the quantification of motor stability
for the assessment of fall risk
[1] Tinetti et al., The American Journal of Medicine,, 1986
[2] Fuller, American Family Physician, 2000
[3] Heinrich et al., Osteoporosis International, 2010
[4] Hausdorff et al., Archives of Physical Medicine and Rehabilitation, 2001
Methods for the quantification of motor
stability for the assessment of fall risk
Quantification of fall risk is therefore essential
How can fall risk be
quantified?
QUESTIONNAIRES
MOTOR
FUNCTION
TESTS
BIOMECHANICAL
LABORATORYBASED
MEASUREMENTS
• Questionnaires about risk factors for falling and motor
function tests are commonly used for fall risk assessment
in the clinic, but their predictive value is limited6,7
• Objective methods, suitable for clinical application, are
hence needed to obtain a subject specific quantitative
assessment of fall risk
[6] Skelton et al., Age and ageing, 2007
[7] Hamacher et al., Journal of the Royal Society Interface, 2011
Methods for the quantification of motor stability for
the assessment of fall risk
Falls often occur during gait4,8
GAIT STABILITY
INDIRECT
ASSESSMENT
•
•
•
•
•
•
Standard Deviation
Coefficient of
Variation
Inconsistency of
Variance
Nonstationary Index
Poincaré Plots
…
STABILITYRELATED
MEASURES
DIRECT
ASSESSMENT
•
•
•
Local Dynamic Stability
Orbital Dynamic Stability
…
•
•
•
•
•
Harmonic Ratio
Index of Harmonicity
Recurrence Quantification
Analysis
Multiscale Entropy
…
[4] Hausdorff et al., Archives of Physical Medicine and Rehabilitation, 2001
[8] Berg et al., Age and Ageing, 1997
Methods for the quantification of motor stability for the
assessment of fall risk
STANDARD DEVIATION
Standard deviation (SD) of stride time was simply calculated as the
standard deviation of the stride times in the analyzed time-window
COEFFICIENT OF VARIATION
INCONSISTENCY OF VARIANCE /
NONSTATIONARY INDEX
Each time series was first normalized with respect to its mean and
SD, yielding new time series each with mean = 0 and SD = 1. This
normalized time series was then divided into blocks of five strides
each, and in each segment the (local) average and (local) SD were
computed. The inconsistency of the variance (IV) is the SD of the
local SD. The nonstationary index (NI) is the SD of the local
averages.
Methods for the quantification of motor stability for the
assessment of fall risk
POINCARÈ PLOTS
The plots displays the correlation
between consecutive stride times
data in a graphical manner.
Length (PSD1) and width (PSD2) of
the long and short axes describe
the elliptical nature of the Poincaré
plot images.
The dispersion of points
perpendicular to the line-of-identity
reflects the level of short-term
variability. The dispersion of points
along the line-of-identity is shown to
indicate the level of long-term
variability.
Khandoker et al.,
2008
Methods for the quantification of motor stability for the
assessment of fall risk
ORBITAL STABILITY ANALYSIS
•Introductio
n
•Layout
•Systematic
review
•Orbital
stability of
gait
•Implement
ation
•Fall history
•Directional
changes
•Short term
stability
•Conclusion
Fundamental indicators
of orbital stability are
maximum Floquet
multipliers
Sn+1 = F(Sn )
max FM
<1
STABILITY
>1
INSTABILITY
Poincaré map
linearization
[S
*
*
*
S
»
J(S
)
S
S
]
[
]
n+1
n
Jacobian
eigenvalues
=
Floquet
multipliers
Dingwell et al., 2007
Methods for the quantification of motor stability for the
assessment of fall risk
LOCAL STABILITY ANALYSIS
•Introductio
n
•Layout
•Systematic
review
•Orbital
stability of
gait
•Implement
ation
•Fall history
•Directional
changes
•Short term
stability
•Conclusion
Euclidean distances
between neighboring
trajectories in state space
were computed as a
function of time and
averaged over all original
pairs of initially nearest
neighbors. Local
divergence exponents
were estimated from the
slopes of linear fits to
these exponential
divergence curves:
Dingwell et al., 2007
Methods for the quantification of motor stability for the
assessment of fall risk
HARMONIC RATIO
The HR was calculated by decomposing acceleration signals
into harmonics using a discrete Fourier transform; the
summed amplitudes of the first 10 even harmonics were then
divided by the summed amplitudes of the first 10 odd
harmonics for the AP and V accelerations, and vice-versa for
the ML accelerations.
INDEX OF HARMONICITY
P0 is the power spectral density of the first harmonic and Pi
the cumulative sum of power spectral density of the
fundamental frequency and the first five super-harmonics.
Values close to 1 indicate high harmonicity.
Methods for the quantification of motor stability for the
assessment of fall risk
MULTISCALE ENTROPY
•Introductio
SAMPLE ENTROPY
n
•Layout
•Systematic
review
•Orbital
stability of
Costa et al., 2003
gait
•Implement
ation
•Fall history
•Directional
changes
•Short term
MSE was calculated for values of  ranging from 1 to 6, m = 2 and r
stability= 0.2
•Conclusion
Methods for the quantification of motor stability for the
assessment of fall risk
RECURRENCE QUANTIFICATION ANALYSIS
•Introductio
n
•Layout
•Systematic
review
•Orbital
stability of
gait
•Implement
ation
•Fall history
•Directional
changes
•Short term
stability
•Conclusion
Recurrence rate
Determinism
Average length of diagonal
lines
Maximum length of
diagonal lines
Coordination
Statistical measures of
variability of locomotor
parameters
Local dynamic
stability
Orbital dynamic
stability
Temporal
evolution of
the motor
task
Response to
perturbations
Fall history: is a minimum quantification
setup possible?
Riva F, Toebes MJP, Pijnappels M, Stagni R, van Dieën JH, “Estimating fall risk with inertial sensors using gait stability
measures that do not require step detection”, winner 2012 SIAMOC methodological prize, Gait & Posture, in press.
ASSUMPTIONS
• Falls often occur during walking
• Gait stability is crucial in falls
• The trunk plays a critical role in regulating gait-related oscillations
Analysis of trunk kinematics during gait
• Stride detection can be hazardous in fallprone subjects
• Stability-related, stride-detection
independent nonlinear measures of trunk
accelerations were calculated
AIM: To investigate the association between stability
estimates and fall history in a large sample of
subjects (fallers and non-fallers)
Fall history: is a minimum quantification setup
possible?
Riva F, Toebes MJP, Pijnappels M, Stagni R, van Dieën JH, “Estimating fall risk with inertial sensors using gait stability
measures that do not require step detection”, winner 2012 SIAMOC methodological prize, Gait & Posture, in press.
PROTOCOL
• 131 subjects (age 62.4 + 6.1 years) participated in the study*
• Fall history obtained by self-report
• Subjects walked on a treadmill at 4 km/h
• Accelerometer placed on the trunk
DATA ANALYSIS
• HR, IH, MSE, RQA were calculated
• Accelerations of the trunk in the anterior-posterior (AP) and
medio-lateral (ML) directions
• 3 minutes of walking
• Log transformed parameters were used as inputs for
univariate logistic regression models
• Self-report was considered as the gold standard
*Data acquired by Toebes and colleagues at MOVE Research Institute Amsterdam, Faculty of Human
Movement Sciences, VU University, Amsterdam, Netherlands.For more details, please see Toebes et al.,
Fall history: is a minimum quantification setup
possible?
Riva F, Toebes MJP, Pijnappels M, Stagni R, van Dieën JH, “Estimating fall risk with inertial sensors using gait stability
measures that do not require step detection”, winner 2012 SIAMOC methodological prize, Gait & Posture, in press.
RESULTS
• HR and IH showed poor o no correlation with fall history
• Univariate associations with fall history were found for MSE and
RQA parameters in the AP direction
• The best classification results were obtained for MSE for scale
factor 3 and RQA max_length
• With MSE the model correctly classified 71.8% of the subjects
(sensitivity 21.4%, specificity 95.5%)
• With RQA the model correctly classified 71% of the subjects
(sensitivity 16.7%, specificity 96.6%)
PRIN 2010-2011
Un approccio quantitativo e multifattoriale
per la stima e la prevenzione
del rischio di caduta nell'anziano
(2013-2015)
Rilevazione in tempo reale delle cadute
mediante tecnologie indossabili:
misura, allarmi e feedback
The FARSEEING Consensus
Minimum requirements
- Duration: 24 h (better 72h
or 1 week)
- Sensor type: ACC (gyro,
mag, barometer)
- Resolution: minimum 100
Hz
- Sensor location: L5 or
trunk
- Add semistructured
interviews (context)
- Add baseline medical
information
©Lorenzo Chiari
University of Bologna
24 October 2014
Fall detection: a systematic review
- Technical
specifications of the
devices varied
considerably e.g.
frequency, range
- Real-world fall data
were presented in only
six publications (less
than 10 events in all
studies!)
©Lorenzo Chiari
University of Bologna
24 October 2014
AD vs CADUTE REALI
©Lorenzo Chiari
University of Bologna
24 October 2014
Fall detection: Example of real-world fall
2
Ax
Ay
Az
Orientation
1
0.5
0
SV  A 2x  A 2y  A 2z
-0.5
5
-1
4.5
-1.5
0
10
20
30
Time [s]
40
50
60
4
3.5
Fall Sum Vector [g]
Accelerometer outputs [g]
1.5
Threshold 3 g
3
2.5
2
1.5
Impact
1
0.5
0
0
10
20
30
Time [s]
40
50
60
Esempio di caduta reale
©Lorenzo Chiari
University of Bologna
24 October 2014
What we’ve learnt so far
ADL
©Lorenzo Chiari
FALL
University of Bologna
ADL
24 October 2014
A multiphase fall model
Impact
Pre-fall Phase
t0
Resting Phase
Falling Phase
t1
t2
t3
Recovery Phase
t5 time
t4
Becker et al, Z Gerontol Geriatr, 2012
©Lorenzo Chiari
University of Bologna
24 October 2014
Falls are diverse
©Lorenzo Chiari
University of Bologna
24 October 2014
Can show a clear pattern…
IMPACT
RESPONSE
Vertical
AP
ML
FALLING
PHASE
IMPACT
“In the cellar, wanted to pick an
object from the ground and fell
forward” – (FW; L5; UF; NoInj;
H=178cm; W=92Kg)
Hybrid Klin-1-1
Geriatric Rehab
©Lorenzo Chiari
University of Bologna
24 October 2014
…or very intricate ones
MULTIPLE
IMPACTS,
«BROKEN» FALL
Vertical
AP
ML
“While with the wheeled walker
wanted to pick an object from the
ground and fell forward, laterally
to the left” – (FW; L5; UF; Inj-BRibs; H=154cm; W=56Kg)
MiniMod PSP-1025-3
Community Dwelling
©Lorenzo Chiari
University of Bologna
24 October 2014
Simulated vs Real
* P < 0.05 (exact Wilcoxon test)
Klenk et al, Med Eng & Phys, 2011
©Lorenzo Chiari
University of Bologna
24 October 2014
Sensibilità e specificità
©Lorenzo Chiari
University of Bologna
24 October 2014
Fall detection
Palmerini et al. 2014, submitted
©Lorenzo Chiari
University of Bologna
24 October 2014
Another timescale (falls are rare!)
©Lorenzo Chiari
University of Bologna
24 October 2014
Falls may occur in a number of ADL
PreActivity: Walking
(with an assistive
device)
Vertical
AP
ML
“When trying to go to the side of
the wheeled walker fell forward,
directly on the ground” – (FW; L5;
UF; NoInj; H=178cm; W=92Kg)
Hybrid Klin-1-3a
©Lorenzo Chiari
University of Bologna
24 October 2014
Falls may occur in a number of ADL
PreActivity: Transfer
Vertical
AP
ML
“In the living room, while getting
up from the chair, fell directly on
the ground” – (FW; L5; UF; NoInj;
H=178cm; W=92Kg)
Hybrid Klin-1-2
Geriatric Rehab
©Lorenzo Chiari
University of Bologna
24 October 2014
Predictive tools for falls
input
Subject under
assessment (e.g.
community-dweller
older adult)
Contextual
information (e.g.
weather
conditions)
output
Prediction
(statement about
falls that the subject
under assessment will
experience)
User:
-
physician
subject under
assessment
researcher leading
a RCT
…
Literature
Traditional tools
(1st generation)
2nd generation tools
-
-
-
Often based on subjective
evaluations
No use of statistics, no probabilistic
meaning
(e.g. PPA)
Sensor-based tools
(3rd generation)
-Proof of concept
Validation of traditional tools
1986 POMA
1991 TUG
1986 Get-Up
and Go Test
2008 Lamb’s
screening tree
2003 PPA
(physiological
profile
assessment)
Consciousness in statistics
- development and
validation
2013 Howcroft’s review on
sensor-based tools
2010 Deandrea’s
review on fall risk
factors
Advent and diffusion of highthroughput technology:
-inertial sensors
time
Advance in statistical learning
for high-dimensional
problems
Targets: community‐dwelling elders and
high‐risk groups of fallers
The inclusion of a
longstanding cohort
study (InChianti)
ensures a
representative
population sample,
which is urgently
needed to translate
technological
advance into real
world service
provision.
The system
components will
interact with an open
home automation
infrastructure, able to
support elderly users in
their daily activities.
The architecture of the
database will facilitate
collection, analysis and
processing of data
related to falls, daily
activity and
physiological factors.
Screening the fall risk
 Viste le dimensioni del problema (crescita della popolazione
anziana; prevalenza delle cadute in soggetti 65+; attuale
capacità delle strutture sanitarie) si pone un problema di
fattibilità e di sostenibilità non indifferente
 Quale strategia?
Strumenti di screening di massa?
Analisi per sotto-popolazioni ?
 Come definire adeguatamente il ‘rischio di cadere’?
 Come misurarlo?
63
Un possibile scenario 1/2
Valutazione multidimensionale
‘Modello’ del rischio di caduta
Test strumentati
EMG
Scale
Forza
ROM
Attività
Cognitivo
…
‘Modello’ del rischio di frattura
Un possibile scenario 2/2
Valutazione multidimensionale
‘Modello’ del rischio di caduta
Test strumentati
EMG
Scale
Forza
ROM
Attività
Cognitivo
…
Modello
numerico/probabilistico/
neurobiomeccanico/robotico
Strategie/interventi per modificare
il rischio di caduta
BF; WBV;
FES; esercizio;…
Real-data tools
FRAT-up
Lasso
Predizione in termini puramente
statistici a partire dai fattori di rischio
del soggetto
Sensorbased
Literature
Real Data Tools: FRAT-up
AUC
TPR
FPR
ActiFE
0.57 (95% CI 0.53-0.60)
InCHIANTI
0.64 (95% CI 0.60-0.69)
ELSA
0.70 (95% CI 0.68-0.72)
Real Data Tools: Lasso
Fall prediction from wearable inertial
sensors
~ 8 studies based on prospective falls
small sample size
Fall prediction from wearable inertial sensors
Older adults from the InCHIANTI study
No dementia, no pacemaker
FARSEEINGInCHIANTI
Functional tests:
• Timed-Up-And-Go Test
• Romberg
• 5-time Repeated Chair Stands
• 400 m
Smartphone in L5
6-month follow-up (monthly telephone interview)
• Filter on reliability of features
• Lasso logistic regression
• L/Q discriminant analysis, wrapper feature
selection
• 5-fold cross-validation
Fall prediction from wearable inertial
sensors
FARSEEINGInCHIANTI
257 older adults
25 fallers
90 reliable features
AUC (sd)
Lasso logistic regression 0.58 (0.16)
Linear discriminant
analysis
Quadratic discriminant
analysis
0.61 (0.12)
0.61 (0.15)
Features
TUG, Duration StW
5RCS, RMS V accel
Romberg, Mean velocity AP displ
5RCS, RMS accel AP stand
5RCS, Total duration
400m, Variation coefficient cadence
straight path
Prevenzione della caduta:
interventi mirati alla riduzione
del rischio
PreventIT kick-off meeting, January 25-26, 2016, Trondheim, Norway
Where are most of these results coming from?
FARSEEING (EU FP7, http://farseeingresearch.eu/)
aims to promote better prediction, prevention and
support of older persons, by long-term analysis of
behavioural and physiological data collected using
Smartphones, wearable and environmental sensors:
leading to self-adaptive responses. The FARSEEING
repository is the world’s largest fall repository (inertial
sensor-based)
Where are most of these results coming from?
CuPiD (EU FP7, http://www.cupid-project.eu/ ) project
has a focus on people with Parkinson’s Disease (PD)
specifically on motor learning and rehabilitation
principles. ICT-enabled systems have been developed
to provide, in the home setting, personalized
treatment. Motor rehabilitation programs include: i)
exergaming, ii) training of walking and iii) training for
preventing Freezing of Gait episodes
Smart Environments
Possible ingredients for such an original recipe:
 Home control and automation systems
 Wearable systems
 Portable/mobile devices
Home Control and Automation
 HCA systems can integrate a
variety of environmental
sensors that allow early
detection and warning of
equipment failures or
conditions that exceed userdefined limits.
Wearable/Portable systems
 Wearable sensor systems for health
monitoring are an emerging trend and
are expected to enable proactive
personal health management and
better treatment of various medical
conditions.
Smarthome?
Contrary to visions that consider home automation and
personal health systems as a mean to replace or to simplify
the subject control and actions, in the FARSEEING and
CUPID approach smartphones, wearable sensors, and
home based technology are used to stimulate the user by
making life mentally and physically more challenging
but without losing comfort.
The Smarthome
• A smarthome system in FARSEEING is equipped
with a “Scenario Programmer”. It is possible to
define and compose a set of conditional rules
defining “what”, “when”, and “if” perform
specific actions
• The execution of a scenario can be triggered by
the user but also by external events like the
opening of a door, a detected movement, a
temperature change, or a detected fall.
Wall-mounted
Touchscreen
Power sockets
and Switches
Movement Sensor
plus a power socket
The Smartphone
 Smartphones in FARSEEING are
used for:
1. continuously monitor
physical activity
the
user’s
2. real-time fall detection which is
synchronized with a remote server
for alarm management
3. interacting with the home automation
system
Waist case belt used for
wearning the smartphone
The Smartphone User Interface
Scenarios: Exercise
Scenarios: Exergame-based Exercise
Scenarios: Walking
Scenarios: Fall Detection
 Fall detection: when a fall is detected by the smartphone a message is
sent to the home automation system which also activates the wall
mounted touchscreen.
Personal Goals
 The wall mounted touchscreen also shows the status/progress of the
personal goals of the user. Feedback messages are delivered in the
form of a growing garden with specific elements being awarded on the
basis of the task completed on a daily basis.
Scenarios: rehabilitation at home
 Biofeedback's working principle is based on information from the body
coded into an appropriate signal and provided back to the user in real
time.
Challenge
keep
upright!
“Virtual clinician” continuously
assessing and vocally correcting
patients’ ineffective or unsafe gait
patterns
Solution: CuPiD app
Hardware:
• inertial sensors transmitting Bluetooth 3D accelerations and 3D
angular velocities (@100Hz) to smartphone
Solution: CuPiD app
App
• able to perform in real-time an accurate gait analysis and to act as
an intelligent tutoring system feeding back the voice instructions
usually provided by physiotherapists
• to be used at home independently by patients
Gait analysis algorithms
1. Step detection  temporal parameters
2. Dead-reckoning (Kalman Filter)  ubiquitous localization
3. Automatic increase/decrease difficulty
Temporal gait params
Spatial gait params
Angular velocity
Inertial
sensor
Mechanization eq.
↷ ⊕ ò dt
Accelerometer
Initial contact
Gyroscope
ò dt
position
velocity
-g
ò dt
orientation
Kalman filter
correction
Foot off
ZUPT
K
Step length errors < 4%
Telemedicine
Background service
• remote configuration of settings
• automatic data synchronization (via WiFi)
Validation study
Pre-test
Week1
Week2
Week3
Week4
Week5
Patients
• 20 PD patients with Cupid app
Methods
Training for 6-week ~ 30min – 3 times / week
Pre – Post – 1 month follow up: Gait analysis
Week6
Post test
1motnh
follow-up
Validation study – Preliminarily results
Pre-test
Week1
Week2
Week3
Week4
Week5
Week6
Post test
1motnh
follow-up
Results
14 patients
Self-report: high satisfaction
• Comfortable walking speed improved by 9% ± 3% (p=0.01)
• step length improved by 7% ± 2% (p=0.03)
• 2-minute walk test improved by 12% ± 4% (p=0.03)
• Patients were able to autonomously use system
CuPiD app so far demonstrated its ability in training and treating gait
impairments in patients with Parkinson Disease
Complete Architecture
Scenarios
Scenarios
Scenarios
Scenarios
What we can get out of this architecture?
What we can get out of this architecture? Our proposed
solution for fall detection
Wearable sensors
x
y
z
Acceleration
x
y
z
“…wanted to pick an object from the
ground and fell forward” – (F; L5;
H=178cm; W=92Kg)
Wavelet-based fall detection
Palmerini L, Bagala F, Zanetti A, Klenk J, Becker C, Cappello A. A waveletbased approach to fall detection. Sensors 2015
Performance – ROC Curve
95% sensitivity, 97%
specificity
Sensitivity:
AUC = 0.98
(0.96-0.99)
Specificity:
50% sensitivity,
99.99% specificity
percentage of falls
correctly detected.
percentage of
ADLs correctly
detected.
What do we do now?
Data Mining
The nontrivial extraction of implicit, previously
unknown and potentially useful information from data
[Frawley, 1992]
Data mining is the process of selecting, exploring and
modeling large amount of data in order to discover
unknown patterns or relationships which provide a clear
and useful result to the data analyst [Giudici, 2003]
Data Mining