A Hierarchical Approach to Recognize Purposeful Movements Using Inertial Sensors: Preliminary Experiments and Results

Carme Zambrana1, Sebastian Idelsohn-Zielonka1, Mireia Claramunt-Molet1, Maria Almenara-Masbernat2, Eloy Opisso2, Josep Maria Tormos2,Felip Miralles1 and Eloisa Vargiu1

1 2

A Hierarchical Approach to Recognize Purposeful Movements Using Inertial Sensors:Preliminary Experiments and Results

CONTEXT

POST STROKE REHABILITATION

Stroke Acute careOutpatient

rehabilitation

Inpatient

rehabilitation

24-48h 2-8 weeks > 2 months0

ADALT

METHOD

Hierarchical Approach

Preprocessing (SMV)

Threshold

Tartaglia’s filter

𝑆𝑀𝑉 = 𝐴𝑐𝑐𝑥2 + 𝐴𝑐𝑐𝑦

2 + 𝐴𝑐𝑐𝑧2

Experimentally calculated

Module 1 (M1)

(𝜃)

INPUT

MODEL

OUTPUT

𝑣𝑖 = 𝑞=−2(≠0)

2

𝑙𝑖+𝑞𝑓𝑞

𝑣𝑖 > 𝑣𝑚𝑎𝑥𝑣𝑖 ≤ 𝑣𝑚𝑎𝑥

: Movement: Non-Movement

𝑣𝑚𝑎𝑥 = 𝑞=−2(≠0)

2

1 𝑓𝑞 = 8

Supervised binary classifier

Experimentally selected

Module 2 (M2)

INPUT

OUTPUT

MODEL

PRELIMINARY EXPERIMENTS AND RESULTS

Setting up

Devices• 2 IMUs (one on each wrist)• 3-axial accelerometers• 20 Hz• Bluetooth 2.0

Volunteers• 9 healthy volunteers

• 31.22 ± 4.59 years• 5 female / 4 male• 1 left-handed

Activities• Purposeful

• Non-purposeful

Results

Module 1

Preprocessing (SMV)

Threshold

Tartaglia’s filter

𝐴𝑐𝑐 =𝑇𝑃 + 𝑇𝑁

𝑇𝑃 + 𝑇𝑁 + 𝐹𝑁 + 𝐹𝑃

Grid-search over the range 0 to 1 by 0.005, maximizing the accuracy function.

(𝜃)

Module 2

Supervised binary classifier

Features:• Time domain• Frequency domain

Models:• K-Nearest Neighbor (KNN)• Random Forest (RF)• Support Vector Machine (SVM)

Windows Lengths:• 2.0 seconds• 4.0 seconds

Randomly Split 70-30% Train-Test10-fold cross validation

𝐹1 = 2𝑇𝑃

2𝑇𝑃 + 𝐹𝑁 + 𝐹𝑃

Results

Overall results

Purposeful Other

87.46%

92.38%7.62%

12.54%

Purp

osefu

lO

ther

Predicted

Tru

e

How it works

Non-movement

Movement

M1 output

Non-purposeful

Purposeful

M2 output

Eating Pouring

waterDrinking Brushing

their teethFolding a towel

Walking

FINAL REMARKS AND FUTURE WORK

Final remarks and future work

Future work:• Test the Hierarchical Approach with post-stroke patients• Create new modules to recognize other non-purposeful movements

Final remarks• We have developed a hierarchical approach to recognize purposeful movements• Tested with 9 healthy volunteers, obtaining encouraging results: Acc = 0.90701

THANK YOU

Contact:Carme Zambranacarme.zambrana@eurecat.org

Acknowledgments: The study has been partially funded by ACCIÓ(Pla d’Actuació de Centres Tecnològics 2016)