ݺߣ

ݺߣShare a Scribd company logo

Improving Activity Monitoring through a Hierarchical Approach

•
•
Xavier Rafael Palou
Xavier Rafael Palou

This document describes a hierarchical approach to improve activity monitoring through home sensor data. It presents a two-level classifier system: an upper level to detect if a user is home or away, and a lower level to detect if the user is alone or has visitors. Machine learning methods are able to outperform rule-based approaches for both classification tasks. When combined in a hierarchical model and tested on one month of sensor data, it achieves 94% accuracy in detecting user location and presence of visitors.

1 of 24
Download to read offline
ICT4AgeingWell 2015 International Conference on Information and Communication Technologies for Ageing Well and e-Health Improving Activity Monitoring through a Hierarchical Approach Xavier Rafael-Palou, Eloisa Vargiu, Guillem Serra, Felip Miralles
- Motivation - Sensor-based Home telemonitoring & support - Hierarchical approach - Upper lever - Lower lever - System Integration - Conclusions - Future work 2 Contents Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 BackHome FP-7 project which aims to assist people with disabilities back home after a discharge Goals - To study the transition from the hospital to the home - To learn how different BNCIs and other assistive technologies work together - To learn how different BNCIs and other assistive technologies can help in the transition from the hospital to the home - To reduce the cost and hassle of the transition from the hospital to the home Motivation 3
Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 BackHome FP-7 project which aims to assist people with disabilities back home after a discharge Main outcomes - New and better integrated practical electrode systems - Friendlier and more flexible BNCI software - Better telemonitoring and home support tools Motivation 4
The focus of this study is to improve performance and limitations of sensor based home telemonitoring systems Benefits: - Quantify physical activity in daily life which is an important predictor of risk of hospital readmission and mortality in patients with chronic diseases - Enable therapists, caregivers, and relatives to become aware of user context, infer user habits and behaviour to provide an enhanced and personalised assistance and support - Allow to anticipate or identify if the person requires a form of assistance since an unusual activity has been recognized 5 Motivation Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
At home - A mesh network of wireless sensors (e.g. presence, door, mattress) - A data collector to parse and transmit data On the cloud - A middleware to store data securely - An intelligent monitoring system to continuously and concurrently analyse new sensor data according to a sliding window approach On a Healthcare center - Therapist web application consumes crunched data 6 Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS - Sensor Based Home Telemonitoring and support Sensor-based Telemonitoring & Home Support Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 In Backhome we developed a sensor based home telemonitoring and support system:
Limitations: - By requirements, sensor system should be unobtrusive and cheap (e.g. open source, wireless, minimum set of sensors, avoid cameras) - System is prone to sensor errors, sometimes they lose or add noisy events (e.g. when sensors remain with a low battery charge) These conditions reduce the activities that the system can recognize and makes more complex to produce an accurate detection of user activities Also makes crucial understanding the user location, i.e. away, at home, alone or with visits, before to proceed with the recognition of more specific activities 7 Sensor-based Telemonitoring & Home Support Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS - Sensor Based Home Telemonitoring and support
Examples of the importance of detecting the user's location: - The system may register a status door change but the door was not opened, so the system believes the user is away but sensor events are being detected at user’s home - On the contrary, the system may not register that a door was opened so the system still believes the user is at home - Also, the system may fail to detect if a user is away after a door change if multiple users were at home or if a false positive door event happened 8 Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 Sensor-based Telemonitoring & Home Support Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS - Sensor Based Home Telemonitoring and support
- Large literature on recognition of activities at home - Great variability in the settings of the experiments either in the number of sensors and their type, individuals involved or the duration - Large amount of techniques for daily life activities (e.g. supervised, either generative or discriminative; and unsupervised) - This diversity makes it extremely difficult to compare performances and draw conclusive findings from those studies - Despite this we have not found extensive studies that analyze altogether the detection of visits, presence or absence of users at home using wireless binary sensors 9 Alternatives Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
To solve, or improve the user location ability in noisy and limited sensorized environments: - We partitioned the problem in two (upper and lower classifiers) and aggregated their results - The classifier at the upper level aimed at augmenting performance of the door sensor, improving its capability of recognizing if the user is at home, or not - The classifier at the lower level is aimed at recognizing if the user is really alone or if s/he received some visits 10 Hierarchical approach Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
Data preparation - 4 month training & evaluation / 1 month test - Extract relevant periods (door events) - Compute features by counting motions before, after and during door periods (total, avg, max, min, ratios) - Label data (0 home/ 1 away) contrasting with a mobile activity tracker + user Rule based approaches explored: - Naive rule algorithm outperformed 77% (accuracy score) Can we improve this result? 11 Hierarchical approach: Upper level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 PCA explained variance ratio for 5 features: 0.32 0.19 0.14 0.10 0.09
Methods - Supervised algorithms: SVM, LogisticRegression, RandomForest, AdaBoost - Search best learning parameters - Use of 10-fold cv over evaluation set - Study of the trade off between bias vs variance 12 Hierarchical approach: Upper level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
Best results for high-level classifier during training (T) and evaluation (E) phase Selected classifier SVM (gamma=1, C=0.452) On testing phase: f1-score 0.97 and accuracy 0.968. Rule based Best ML Improvement correct 48 60 +12 errors 14 2 -12 accuracy 0.77 0.968 +0.19 13 Hierarchical approach: Upper level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
Data preparation - 4 month training & evaluation / 1 month test - Extract relevant data periods (door events) - Filter those when user is away - Compute features by counting motions before, after and during door periods (total, avg, max, min, ratios) - Label data (0 alone / 1 multi-user) by contrasting them with a mobile activity tracker + user Rule based approach was explored unsatisfactorily: ● Num movements every 60 secs g.t. 6: 23 anomalies ● Num of movements every 60 secs g.t. 10: 1 anomaly ● Simultaneous movements: 20 anomalies ● Simultaneous movements in less 2 secs: 23 anomalies ● Max num of movements in 60 secs g.t. 6 or simultaneous moves in less 2 secs: 28 anomalies ● Max num of movements in 60 secs g.t. 6 or simultaneous moves: 21 anomalies 14 Hierarchical approach: Lower level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
Eventually, we relied on a novelty detection approach, due to: - Allows the identification of new or unknown data that a machine learning system has not been trained with and was not previously aware of - Indicated when data is skewed i.e. has few positive cases (i.e., anomalies) compared with the negatives (i.e., regular cases) Thus, we estimate a function f that is positive on the dataset and negative on the complement. The functional form of f is given by a kernel expansion in terms of a potentially small subset of the training data; f is regularized by controlling the length of the weight vector in an associated feature space. 15 Hierarchical approach: Lower level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
The Novelty approach was implemented by using a one-class SVM with RBF classifier - The classifier was trained considering only the normal instances (person is alone) and then evaluated introducing the anomalies (person receives a visit) - According to the best evaluation results, we selected the regularization parameters (n) =0.01 and g = 0.1 - The classifier was tested with the data coming from the 1-month window of monitored events and results showed an accuracy of 0.94 16 Hierarchical approach: Lower level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
Once both classifiers have been trained and chained we tested its performance with the testing dataset corresponding to a window of 1 month We compared the overall results with those using the rule-based approach in both levels of the hierarchy 17 Hierarchical approach: Overall evaluation Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
The intelligent monitoring system is composed of 5 modules: - PP, pre-processing module to encode the data for the analysis - ED, emergency detection module to notify, for instance, in case of smoke and gas leakage - AR, the activity recognition module to identify the location, position, activity and sleeping-status of the user - EN, the event notification module to inform when a new event has been detected - ST, summary computation module to perform summaries from the data 18 System integration: Intelligent monitoring system Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
The hierarchical approach was integrated in AR - Once the current window recognizes a door event at time tb it looks for the previous one in the window or before (in the example ta) - Then, the period from that door events (i.e, tb-ta) is classified by the hierarchical classifier. Seemly, when the event tc has been recognized, the period from tb and tc is classified - Finally, the period from tc to the end of the window is classified - In case of no door events have been recognized, the period from ta to the end of the window is classified 19 System integration: Intelligent monitoring system Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
SB-TMHSS is currently running in 5 end user homes, 2 in Barcelona and 3 in Belfast 20 System integration: Therapist web application Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
21 System integration: Therapist web application Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS is currently running in 5 end user homes, 2 in Barcelona and 3 in Belfast
- Performance of sensor-based telemonitoring and home support systems depends, among other issues, on the reliability of the sensors - Binary sensors are quite adopted in the activity recognition literature and also in commercial solutions, they are prone to noise and errors - We presented a solution, based on machine learning techniques, aimed at reducing error from the sensors and helping to discriminate when user is at home, away or with a visit - The proposed approach can be used as a preprocessing phase in every activity recognition system, being completely independent from that - Results showed an overall improvement of 15% in accuracy with respect to a rule- based approach - The system is part of the BackHome project and is currently running in 2-healthy- users’ home and in 3-end-users’ home 22 Conclusions Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
We are currently setting-up new experiments aimed at comparing the hierarchical approach with a multi-class classifier and with an ensemble (not hierarchical) of classifiers Moreover, we are improving the approach in order to be totally automatic, by using data from Moves as feature instead of to validate the initial dataset We are also interested in studying if we can generalize the proposed approach to use it for all the users of the system, or if we have to use a personalized approach for each different user 23 Future work Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
www.Backhome-FP7.eu BackHome-FP7-Research-Innovation @BackHomeFP7 BackHomeFP7

More Related Content

Improving Activity Monitoring through a Hierarchical Approach

  • 1. ICT4AgeingWell 2015 International Conference on Information and Communication Technologies for Ageing Well and e-Health Improving Activity Monitoring through a Hierarchical Approach Xavier Rafael-Palou, Eloisa Vargiu, Guillem Serra, Felip Miralles
  • 2. - Motivation - Sensor-based Home telemonitoring & support - Hierarchical approach - Upper lever - Lower lever - System Integration - Conclusions - Future work 2 Contents Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 3. Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 BackHome FP-7 project which aims to assist people with disabilities back home after a discharge Goals - To study the transition from the hospital to the home - To learn how different BNCIs and other assistive technologies work together - To learn how different BNCIs and other assistive technologies can help in the transition from the hospital to the home - To reduce the cost and hassle of the transition from the hospital to the home Motivation 3
  • 4. Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 BackHome FP-7 project which aims to assist people with disabilities back home after a discharge Main outcomes - New and better integrated practical electrode systems - Friendlier and more flexible BNCI software - Better telemonitoring and home support tools Motivation 4
  • 5. The focus of this study is to improve performance and limitations of sensor based home telemonitoring systems Benefits: - Quantify physical activity in daily life which is an important predictor of risk of hospital readmission and mortality in patients with chronic diseases - Enable therapists, caregivers, and relatives to become aware of user context, infer user habits and behaviour to provide an enhanced and personalised assistance and support - Allow to anticipate or identify if the person requires a form of assistance since an unusual activity has been recognized 5 Motivation Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 6. At home - A mesh network of wireless sensors (e.g. presence, door, mattress) - A data collector to parse and transmit data On the cloud - A middleware to store data securely - An intelligent monitoring system to continuously and concurrently analyse new sensor data according to a sliding window approach On a Healthcare center - Therapist web application consumes crunched data 6 Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS - Sensor Based Home Telemonitoring and support Sensor-based Telemonitoring & Home Support Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 In Backhome we developed a sensor based home telemonitoring and support system:
  • 7. Limitations: - By requirements, sensor system should be unobtrusive and cheap (e.g. open source, wireless, minimum set of sensors, avoid cameras) - System is prone to sensor errors, sometimes they lose or add noisy events (e.g. when sensors remain with a low battery charge) These conditions reduce the activities that the system can recognize and makes more complex to produce an accurate detection of user activities Also makes crucial understanding the user location, i.e. away, at home, alone or with visits, before to proceed with the recognition of more specific activities 7 Sensor-based Telemonitoring & Home Support Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS - Sensor Based Home Telemonitoring and support
  • 8. Examples of the importance of detecting the user's location: - The system may register a status door change but the door was not opened, so the system believes the user is away but sensor events are being detected at user’s home - On the contrary, the system may not register that a door was opened so the system still believes the user is at home - Also, the system may fail to detect if a user is away after a door change if multiple users were at home or if a false positive door event happened 8 Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 Sensor-based Telemonitoring & Home Support Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS - Sensor Based Home Telemonitoring and support
  • 9. - Large literature on recognition of activities at home - Great variability in the settings of the experiments either in the number of sensors and their type, individuals involved or the duration - Large amount of techniques for daily life activities (e.g. supervised, either generative or discriminative; and unsupervised) - This diversity makes it extremely difficult to compare performances and draw conclusive findings from those studies - Despite this we have not found extensive studies that analyze altogether the detection of visits, presence or absence of users at home using wireless binary sensors 9 Alternatives Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 10. To solve, or improve the user location ability in noisy and limited sensorized environments: - We partitioned the problem in two (upper and lower classifiers) and aggregated their results - The classifier at the upper level aimed at augmenting performance of the door sensor, improving its capability of recognizing if the user is at home, or not - The classifier at the lower level is aimed at recognizing if the user is really alone or if s/he received some visits 10 Hierarchical approach Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 11. Data preparation - 4 month training & evaluation / 1 month test - Extract relevant periods (door events) - Compute features by counting motions before, after and during door periods (total, avg, max, min, ratios) - Label data (0 home/ 1 away) contrasting with a mobile activity tracker + user Rule based approaches explored: - Naive rule algorithm outperformed 77% (accuracy score) Can we improve this result? 11 Hierarchical approach: Upper level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 PCA explained variance ratio for 5 features: 0.32 0.19 0.14 0.10 0.09
  • 12. Methods - Supervised algorithms: SVM, LogisticRegression, RandomForest, AdaBoost - Search best learning parameters - Use of 10-fold cv over evaluation set - Study of the trade off between bias vs variance 12 Hierarchical approach: Upper level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 13. Best results for high-level classifier during training (T) and evaluation (E) phase Selected classifier SVM (gamma=1, C=0.452) On testing phase: f1-score 0.97 and accuracy 0.968. Rule based Best ML Improvement correct 48 60 +12 errors 14 2 -12 accuracy 0.77 0.968 +0.19 13 Hierarchical approach: Upper level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 14. Data preparation - 4 month training & evaluation / 1 month test - Extract relevant data periods (door events) - Filter those when user is away - Compute features by counting motions before, after and during door periods (total, avg, max, min, ratios) - Label data (0 alone / 1 multi-user) by contrasting them with a mobile activity tracker + user Rule based approach was explored unsatisfactorily: ● Num movements every 60 secs g.t. 6: 23 anomalies ● Num of movements every 60 secs g.t. 10: 1 anomaly ● Simultaneous movements: 20 anomalies ● Simultaneous movements in less 2 secs: 23 anomalies ● Max num of movements in 60 secs g.t. 6 or simultaneous moves in less 2 secs: 28 anomalies ● Max num of movements in 60 secs g.t. 6 or simultaneous moves: 21 anomalies 14 Hierarchical approach: Lower level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 15. Eventually, we relied on a novelty detection approach, due to: - Allows the identification of new or unknown data that a machine learning system has not been trained with and was not previously aware of - Indicated when data is skewed i.e. has few positive cases (i.e., anomalies) compared with the negatives (i.e., regular cases) Thus, we estimate a function f that is positive on the dataset and negative on the complement. The functional form of f is given by a kernel expansion in terms of a potentially small subset of the training data; f is regularized by controlling the length of the weight vector in an associated feature space. 15 Hierarchical approach: Lower level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 16. The Novelty approach was implemented by using a one-class SVM with RBF classifier - The classifier was trained considering only the normal instances (person is alone) and then evaluated introducing the anomalies (person receives a visit) - According to the best evaluation results, we selected the regularization parameters (n) =0.01 and g = 0.1 - The classifier was tested with the data coming from the 1-month window of monitored events and results showed an accuracy of 0.94 16 Hierarchical approach: Lower level Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 17. Once both classifiers have been trained and chained we tested its performance with the testing dataset corresponding to a window of 1 month We compared the overall results with those using the rule-based approach in both levels of the hierarchy 17 Hierarchical approach: Overall evaluation Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 18. The intelligent monitoring system is composed of 5 modules: - PP, pre-processing module to encode the data for the analysis - ED, emergency detection module to notify, for instance, in case of smoke and gas leakage - AR, the activity recognition module to identify the location, position, activity and sleeping-status of the user - EN, the event notification module to inform when a new event has been detected - ST, summary computation module to perform summaries from the data 18 System integration: Intelligent monitoring system Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 19. The hierarchical approach was integrated in AR - Once the current window recognizes a door event at time tb it looks for the previous one in the window or before (in the example ta) - Then, the period from that door events (i.e, tb-ta) is classified by the hierarchical classifier. Seemly, when the event tc has been recognized, the period from tb and tc is classified - Finally, the period from tc to the end of the window is classified - In case of no door events have been recognized, the period from ta to the end of the window is classified 19 System integration: Intelligent monitoring system Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 20. SB-TMHSS is currently running in 5 end user homes, 2 in Barcelona and 3 in Belfast 20 System integration: Therapist web application Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 21. 21 System integration: Therapist web application Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015 SB-TMHSS is currently running in 5 end user homes, 2 in Barcelona and 3 in Belfast
  • 22. - Performance of sensor-based telemonitoring and home support systems depends, among other issues, on the reliability of the sensors - Binary sensors are quite adopted in the activity recognition literature and also in commercial solutions, they are prone to noise and errors - We presented a solution, based on machine learning techniques, aimed at reducing error from the sensors and helping to discriminate when user is at home, away or with a visit - The proposed approach can be used as a preprocessing phase in every activity recognition system, being completely independent from that - Results showed an overall improvement of 15% in accuracy with respect to a rule- based approach - The system is part of the BackHome project and is currently running in 2-healthy- users’ home and in 3-end-users’ home 22 Conclusions Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015
  • 23. We are currently setting-up new experiments aimed at comparing the hierarchical approach with a multi-class classifier and with an ensemble (not hierarchical) of classifiers Moreover, we are improving the approach in order to be totally automatic, by using data from Moves as feature instead of to validate the initial dataset We are also interested in studying if we can generalize the proposed approach to use it for all the users of the system, or if we have to use a personalized approach for each different user 23 Future work Improving Activity Monitoring through a Hierarchical Approach. ICT4AgeingWell, 2015