Introducing Knowledge in the Process of Supervised Classification of Activities of Daily Living in Health Smart Homes A. Fleury2 , N. Noury1,3 and M. Vacher4 1 TIMC-IMAG Lab, Team AFIRM, UMR CNRS/UJF 5525, Grenoble 2 École des Mines de Douai, Computer Sciences and Control Dpt., Douai, France 3 INL Lab, Team MMB, UMR CNRS/ECL/INSA/UCBL 5270, Lyon, France 4 LIG Lab, Team GETALP, UMR UJF/CNRS/INPG 5217, Grenoble, France

HealthCom 2010 – 12th International Conference on E-Health Networking, Application & Services

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Outline of the presentation

1

Introduction and Objectives

2

Data Acquisition

3

Classification of ADLs

4

Prior Knowledge Introduction

5

Conclusion and Discussion

A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

2 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Context of the study Related Works Objectives of the Work

Context of the study Facts Ï Ageing population is growing faster than institution equipped to welcome them in developed countries. Ï In France, population over 85 is nowadays 1.3 million and previsions for 2015 are more than 2 millions. Ï Same previsions in the whole world =⇒ Loss of autonomy is an important problem to care about. Possible solution ? Smart sensors and smart homes to: Monitor the persons at home and detect distress situations Evaluate and observe his/her activity continuously (to detect early symptoms or monitor autonomy) A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

3 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Context of the study Related Works Objectives of the Work

Related Works Several works for automatic classification of activities in different environments: Philipose et al., 2004 =⇒ RFID tags on hundreds of objects, 14 activities, Dynamic Bayesian Networks Hong et al., 2008 =⇒ RFID tags on foods etc., Hygiene vs. preparing a drink, Dempster-Shafer. Nugent et al. tested also the impact of sensor failure. Kröse et al., 2008 =⇒ Lots of sensors (environmental, switches...), Going to the toilets vs exit from the flat, HMM. Berenguer et al., 2009 =⇒ Sensor: electrical powerline, activity of taking a meal. A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

4 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Context of the study Related Works Objectives of the Work

Objectives of the Work Objectives of the project AILISA and HIS projects (Health Smart Home) of the TIMC-IMAG Laboratory, faculty of Medicine of Grenoble. Objectives: Find the optimal set-up of sensors for analyzing the activities of daily living at home, Identifying the higher number of activities of daily living, Linking it to autonomy-related scales used by geriatricians.

Objectives of this work Improving the results of the classification of ADLs obtained with SVM For this, testing different priors to improve the classification A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

5 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Health Smart Home of Grenoble Experimental protocol Data Processing and Indexing

Health Smart Home - TIMC-IMAG Lab, Grenoble

Kitchen

Infra−Red Presence sensors

Wide Angle Webcamera

Living Room Bedroom

Bathroom WC

Microphones

Temperature and Hygrometry sensors

Midsagittal

Door contacts

Technical Room Hall

Transverse

Wearable kinematic sensor (accelerometers and magnetometers) Frontal

Description & equipment of the smart home of TIMC-IMAG. A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

6 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Health Smart Home of Grenoble Experimental protocol Data Processing and Indexing

Experimental Protocol Population and experimentation Data on 13 young and healthy subjects (6 women et 7 men) Mean age: 30.4 yo (24 – 43 yo, min–max) Mean execution time: 51min 40s (23min 11s – 1h 35min 44s, min–max) Seven activities, to perform at least once by every person 3 minutes time frames, with following repartitions: Class

Name

Sleeping Resting Dressing/undressing Having meal Elimination Hygiene Communication Total

A. Fleury, N. Noury and M. Vacher

C1 C2 C3 C4 C5 C6 C7

Repartition 49 73 16 45 16 14 19

21.1% 31.5% 6.9% 19.4% 6.9% 6.0% 8.2%

252

100%

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

7 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Health Smart Home of Grenoble Experimental protocol Data Processing and Indexing

Data Processing Modality

Features

Actimeter

Percentage of time spent in various postures and walking

Microphones

Number of events per kind of sounds and for each microphone

IPR

Percentage of time in each room, number of events for each PIR

Door contacts

Percentage of time “open” and predominant position (open or close)

Environmental

Differential measure for the last 15 minutes for both temperature and hygrometry

A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

8 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Health Smart Home of Grenoble Experimental protocol Data Processing and Indexing

Time-stamping Manual indexation using video recordings:

A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

9 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

SVM for ADLs Classification First Results

Multimodality for ADLs Classification

Location:  LIVING ROOM

Location :  BEDROOM

Possible activities: Dressing/undressing or sleeping

Possible Activities: Resting, communicating, dressing

Données de posture Postural Data 

Données de posture Postural data 

(ACTIM6D) (ACTIM6D)

Sound data Contacteur de porte ((speech, p la commode p phone de ringing…)

Classified f activityy (communicating OR resting OR  dressing)

A. Fleury, N. Noury and M. Vacher

(ACTIM6D) (ACTIM6D)

Door contact on  Contacteur de porte de convenience la commode the the convenience

Knowledge for supervised classification of ADL

Classified activity  f y (Sleeping OR  Dressing/undres‐ sing)

Friday July 2nd, 2010

10 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

SVM for ADLs Classification First Results

SVM for ADLs Classification SVM (Boser et Vapnik, 1992): Binary classification (linear or non-linear using kernel and feature space) Method multiclass classification: one-against-one

C2

C2

C1

000 111 000 111 000 111 000 111 00 11 00 11 000 111 000 111 000 111 00 11 11 00 00 000 000 11 111 00111 11 00 11 11 00 00 000 11 111 00 11 000 111 00 11 000 111 000 111 00 111 11 00 11 00 11 11 000 111 000 00 11 00 00 11 00 11 11 00 11 11 00 11 00 00 11 00 11 00 00 11 00 11 11 00 11 00 11 00

11 00 00 11 00 11 00 11

00 11 00 11 00 11 11 00 00 11 11 00 11 00 00 11 00 11 00 11 00 11 00 11 00 11 00 11 000 111 000 111 000 111 000 111

C2 C3 111 111 000 000 00 11 000 111 000 111 00 11 00 000 11 111 000 00 11 00 11 00 11 00111 11 00111 11 00 11 000 111 000 00 11 000 111 000 111 000 111 000 111 00 11 000 111 000 000 111 00111 11 000 111 000 111 00 11 00 11

C1 00 11 11 00 00 11 00 11 00 11 00 11 000 111 00 11 00 11 11 000 111 00 00 11 000 111 00111 11 000 00 11 000 00111 11 000 11 111 00 00 000 111 00 11 11 00 11 000 111 00 00 11 000 11 111 00 11 00 000 111 00 11 11 00 11

C1 C3

N ·(N −1) 2

SVM (for N classes)

Differentiate the classes Ci and Cj , 0 < i ≤ N et 0 < j < i Majority voting: C = max Card({yi ,j } ∩ {k }) k =1..N

C3

A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

11 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

SVM for ADLs Classification First Results

First Results Global Error Rate: 13.79%. Classification Results

Activities

C1 C2 C3 C4 C5 C6 C7 Sleeping Resting Dressing Eating Elimination Hygiene Communication C1 C2 C3 C4 C5 C6 C7

98% 2% 16.4% 78.1% 13.3% 6.7% 0% 0% 0% 6.2% 7.1% 0% 5% 10%

0% 0% 80% 2.2% 0% 0% 5%

0% 1.4% 0% 97.8% 6.3% 7.1% 0%

0% 4.1% 0% 0% 81.2% 14.3% 0%

0% 0% 0% 0% 6.3% 71.5% 0%

0% 0% 0% 0% 0% 0% 80%

Table: Confusion Matrix for the leave-one-out validation protocol with Generic Models. A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

12 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Temporal Knowledge Spatial Knowledge Hybridization of Spatial and Temporal Knowledge

Introducing Temporal Knowledge Slots of activities Each activity is affected to a restricted number of time slots:

Time slots Seven time slots: T1: breakfast (7–9 AM) T2: morning (9–12 AM) T3: lunch (12 AM–2 PM) T4: afternoon (2–7 PM) T5: diner (7 PM–9 PM) T6: evening (9–11 PM) T7: night (11 PM–7 AM)

Class Sleeping Resting Dress/undress Having meal Elimination Hygiene Communication

T1 T2 T3 T4 T5 T6 T7 p p p p p p p

p p

p p p p p

p p p p

p p

p p p p p p

p p p

p

p

p

p

p

Interest of this introduction Reducing the possibilities for the frame considering the slot A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

13 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Temporal Knowledge Spatial Knowledge Hybridization of Spatial and Temporal Knowledge

Results of Introduction of Temporal Knowledge Global Error Rate: 9.91%. Classification Results

Activities

C1 C2 C3 C4 C5 C6 C7 Sleeping Resting Dressing Eating Elimination Hygiene Communication C1 C2 C3 C4 C5 C6 C7

100% 0% 0% 0% 11% 83.6% 0% 0% 0% 13.3% 86.7% 0% 0% 0% 2.2% 97.8% 0% 12.5% 0% 0% 0% 0% 0% 7.1% 0% 5% 5% 5%

0% 2.7% 0% 0% 87.5% 14.3% 0%

0% 1.3% 0% 0% 0% 78.6% 0%

0% 1.4% 0% 0% 0% 0% 85%

Table: Confusion Matrix for the leave-one-out validation protocol with temporal knowledge. A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

14 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Temporal Knowledge Spatial Knowledge Hybridization of Spatial and Temporal Knowledge

Introducing Spatial Knowledge

Kitchen Living Room

Resting Communication Sleeping Dressing

Eating

Bedroom

}

Bathroom WC

Elimination Hygiene

Technical Room Hall

A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

15 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Temporal Knowledge Spatial Knowledge Hybridization of Spatial and Temporal Knowledge

Results of Introduction of spatial Knowledge Global Error Rate: 21.12%. Classification Results

Activities

C1 C2 C3 C4 C5 C6 C7 Sleeping Resting Dressing Eating Elimination Hygiene Communication C1 C2 C3 C4 C5 C6 C7

69.4% 11% 13.3% 0% 0% 0% 5%

2% 78% 6.7% 0% 0% 0% 10%

0% 28.6% 0% 6.9% 73.3% 6.7% 2.2% 97.8% 0% 6.2% 0% 14.3% 10% 10%

0% 4.1% 0% 0% 87.5% 14.3% 0%

0% 0% 0% 0% 6.3% 71.4% 0%

0% 0% 0% 0% 0% 0% 65%

Table: Confusion Matrix for the leave-one-out validation protocol with spatial knowledge. A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

16 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Temporal Knowledge Spatial Knowledge Hybridization of Spatial and Temporal Knowledge

Introducing Hybridized (Spatial/Temporal) Knowledge

Construction First consider the decision taken by the classification using Temporal Knowledge Then consider the decision taken with Spatial Knowledge. Two cases then: Decisions are coherent: this decision is kept as the result Decisions are not: Generic estimation is performed.

A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

17 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Temporal Knowledge Spatial Knowledge Hybridization of Spatial and Temporal Knowledge

Results of Introduction of Hybridized Knowledge Global Error Rate: 13.2%. Classification Results

Activities

C1 C2 C3 C4 C5 C6 C7 Sleeping Resting Dressing Eating Elimination Hygiene Communication C1 C2 C3 C4 C5 C6 C7

98% 2% 16.4% 78.1% 13.3% 6.7% 0% 0% 0% 0% 7.1% 0% 5% 10%

0% 0% 80% 2.2% 0% 0% 5%

0% 1.4% 0% 97.8% 6.2% 7.1% 0%

0% 4.1% 0% 0% 87.5% 14.3% 0%

0% 0% 0% 0% 6.3% 71.5% 0%

0% 0% 0% 0% 0% 0% 80%

Table: Confusion Matrix for the leave-one-out validation protocol with hybridized knowledge. A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

18 / 20

Introduction and Objectives Data Acquisition Classification of ADLs Prior Knowledge Introduction Conclusion and Discussion

Discussion and Conclusion Conclusion Time of the day is the best indicator in this case =⇒ Effect of Perfection of the sensor Spatial knowledge have lower results =⇒ lots of mis-detection for the PIR sensors =⇒ Hybridization does not significantly improve the results

Discussion Generic models: more complicated but best results with imperfect sensors Knowledge do not improve significantly the results but degrade the generality of the models As a consequence, generic models should represent the best solution More experimentations and improvement of the quality of the sensors should confirm these results A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

19 / 20

Introducing Knowledge in the Process of Supervised Classification of Activities of Daily Living in Health Smart Homes

Thank you for your attention. Questions ?

A. Fleury, N. Noury and M. Vacher

Knowledge for supervised classification of ADL

Friday July 2nd, 2010

20 / 20

Introducing Knowledge in the Process of Supervised ...

Jul 2, 2010 - Data Acquisition. Classification of ADLs. Prior Knowledge Introduction. Conclusion and Discussion. Health Smart Home of Grenoble.

2MB Sizes 1 Downloads 151 Views

Recommend Documents

How to Manage Knowledge in the Software Maintenance Process
Process", Lecture Notes in Computer Science, Springer, Vol. 3096: p. 78-87. ... Keywords: Knowledge management, software maintenance, agents. 1. Introduction ... year 2004 the cost of knowledge management is expected to reach USD ..... makes the work

How to Manage Knowledge in the Software Maintenance Process
very few software companies use knowledge management techniques to ... identifies which parts of the system should be modified, which modules could be.

[PDF BOOK] Integration of Process Knowledge into Design Support ...
Online PDF Integration of Process Knowledge into Design Support Systems: Proceedings of the 1999 CIRP International Design Seminar, University of Twente, ...

The Relevance of Introducing Design Education in ...
the development and education of children will play a very important role in .... can opt to offer design and technology course to students who would study it as ...

Blue Prism - Introducing your Process to Live Data.pdf
requirement (or scenario) is not documented, it will not be tested. ... transition of workload from the human to the digital workforce. ... the defect or scenario in.

Blue Prism - Introducing your Process to Live Data.pdf
Cases are worked end-to-end with no one. watching the screen. Straight Through Processing Rate. No matter how well the Subject Matter Experts know the ...

Auto-supervised learning in the Bayesian Programming Framework∗
tion [2]. Finding regularities means finding compact data ... In this framework, a Bayesian programmer starts to ... Choose prior distributions on those variables.

Exploring the Role of Traditional Ecological Knowledge in Climate ...
Exploring the Role of Traditional Ecological Knowledge in Climate Change Initiatives.pdf. Exploring the Role of Traditional Ecological Knowledge in Climate ...

Depth of Knowledge in Music
Jan 30, 2009 - Measures the degree to which the knowledge ... for Science Education. Several .... balance, and expression without the aid of a teacher. 19 ...

THE IMPACT OF THE FLOW ARRIVAL PROCESS IN ...
ARC Special Research Center ... E-mail: {n.hohn, d.veitch}@ee.mu.oz.au. P. Abry∗. CNRS .... ets carrying web (HTTP), or e-mail (SMTP) data), packet volume.

The Influence of Tax Knowledge on Tax Compliance Attitude in ...
The Influence of Tax Knowledge on Tax Compliance Attitude in Rwanda.pdf. The Influence of Tax Knowledge on Tax Compliance Attitude in Rwanda.pdf. Open.

The Role of IT in Successful Knowledge Management ...
knowledge management initiatives. By Atreyi ... software life cycle is short, an up-to-date expert direc- .... Technologies makes use of telephone, email, and.

The role of common knowledge in achieving collaboration across ...
There was a problem previewing this document. Retrying... Download. Connect more apps... Try one of the apps below to open or edit this item. The role of ...

Use of Patterns for Knowledge Management in the ... - Springer Link
Data Management), cPDm (collaborative Product Definition management) and PLM. (Product ... internal knowledge of the enterprise, and also to customers and stakeholders. In general ... and the patterns created in the chosen software.

On the Semi-Supervised Learning of Multi-Layered ...
several phone classification tasks over baseline MLPs. Index Terms: ... Foundation under grant IIS-0326382 and by the Office of Naval Re- search under MURI ...

Supervised fuzzy clustering for the identification of fuzzy ...
A supervised clustering algorithm has been worked out for the identification of this fuzzy model. ..... The original database contains 699 instances however 16 of ...

Knowledge Continuous Integration Process (K-CIP)
Every time he/she commits a new set of modifications, the dedicated server automatically launches an integration build to validate the changes and to test the ...

Semi-supervised learning of the hidden vector state model for ...
capture hierarchical structure but which can be ... sibly exhibit the similar syntactic structures which ..... on protein pairs to gauge the relatedness of the abstracts ...

Verifying Performance of Supervised Plants
Email: {j.markovski, m.a.reniers}@tue.nl. Abstract—Supervisory control ... Supervisory control the- ory tackles a part of this problem by investigating automated.