Available online at www.sciencedirect.com [622624]

ScienceDirect
Available online at www.sciencedirect.com
Procedia Computer Science 110 (2017) 86–93
1877-0509 © 2017 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the Conference Program Chairs.
10.1016/j.procs.2017.06.121
10.1016/j.procs.2017.06.121© 2017 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.
1877-0509Available online at www.sciencedirect.com
00 (2016) 000–000
www.elsevier.com/locate/procedia
The 14th International Conference on Mobile Systems and Pervasive Computing
(MobiSPC 2017)
Activity Recognition and Abnormal Behaviour Detection with
Recurrent Neural Networks
Damla Arifoglu∗, Abdelhamid Bouchachia
Department of Computing and Informatics
Bournemouth University, UK
Abstract
In this paper, we study the problem of activity recognition and abnormal behaviour detection for elderly people with dementia.
Very few studies have attempted to address this problem presumably because of the lack of experimental data in the context of
dementia care. In particular, the paper investigates three variants of Recurrent Neural Networks (RNNs): Vanilla RNNs (VRNN),
Long Short Term RNNs (LSTM) and Gated Recurrent Unit RNNs (GRU). Here activity recognition is considered as a sequence
labelling problem, while abnormal behaviour is flagged based on the deviation from normal patterns. To provide an adequate
discussion of the performance of RNNs in this context, we compare them against the state-of-art methods such as Support Vector
Machines (SVMs), Na ¨ıve Bayes (NB), Hidden Markov Models (HMMs), Hidden Semi-Markov Models (HSMM) and Conditional
Random Fields (CRFs). The results obtained indicate that RNNs are competitive with those state-of-art methods. Moreover,
the paper presents a methodology for generating synthetic data reflecting on some behaviours of people with dementia given the
difficulty of obtaining real-world data.
c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.
Keywords: Smart Homes; Sensor based Activity Recognition; Recurrent Neural Networks; Dementia; Abnormal Behaviour Detection
1. Introduction
Studies indicate that by year 2030, 19% of people will be aged 74 to 84 and nearly half of people who are older
than 84 will have dementia1. Elderly people may suffer from the consequences of dementia, which is a condition that
causes problems with mobility, physical and mental abilities such as memory and thinking2. It also may cause de-
crease in the ability of speaking, writing, distinguishing objects, performing motor activities and performing complex
functional tasks (paying bills, preparing a meal, shopping, managing medication, etc.)3. An elderly person having
such cognitive decline loses independence in daily life and requires care and support from caregivers.
Cognitive diseases like dementia need to be detected at an early stage so that early treatment will be possible.
However, research shows that 75% of dementia and early dementia cases go unnoticed4and many such cases are only
∗Corresponding author. Tel.: +44 (0)1202 524111 ; fax: +44 (0)1202 962736.
E-mail address: [anonimizat]
1877-0509 c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.Available online at www.sciencedirect.com
00 (2016) 000–000
www.elsevier.com/locate/procedia
The 14th International Conference on Mobile Systems and Pervasive Computing
(MobiSPC 2017)
Activity Recognition and Abnormal Behaviour Detection with
Recurrent Neural Networks
Damla Arifoglu∗, Abdelhamid Bouchachia
Department of Computing and Informatics
Bournemouth University, UK
Abstract
In this paper, we study the problem of activity recognition and abnormal behaviour detection for elderly people with dementia.
Very few studies have attempted to address this problem presumably because of the lack of experimental data in the context of
dementia care. In particular, the paper investigates three variants of Recurrent Neural Networks (RNNs): Vanilla RNNs (VRNN),
Long Short Term RNNs (LSTM) and Gated Recurrent Unit RNNs (GRU). Here activity recognition is considered as a sequence
labelling problem, while abnormal behaviour is flagged based on the deviation from normal patterns. To provide an adequate
discussion of the performance of RNNs in this context, we compare them against the state-of-art methods such as Support Vector
Machines (SVMs), Na ¨ıve Bayes (NB), Hidden Markov Models (HMMs), Hidden Semi-Markov Models (HSMM) and Conditional
Random Fields (CRFs). The results obtained indicate that RNNs are competitive with those state-of-art methods. Moreover,
the paper presents a methodology for generating synthetic data reflecting on some behaviours of people with dementia given the
difficulty of obtaining real-world data.
c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.
Keywords: Smart Homes; Sensor based Activity Recognition; Recurrent Neural Networks; Dementia; Abnormal Behaviour Detection
1. Introduction
Studies indicate that by year 2030, 19% of people will be aged 74 to 84 and nearly half of people who are older
than 84 will have dementia1. Elderly people may suffer from the consequences of dementia, which is a condition that
causes problems with mobility, physical and mental abilities such as memory and thinking2. It also may cause de-
crease in the ability of speaking, writing, distinguishing objects, performing motor activities and performing complex
functional tasks (paying bills, preparing a meal, shopping, managing medication, etc.)3. An elderly person having
such cognitive decline loses independence in daily life and requires care and support from caregivers.
Cognitive diseases like dementia need to be detected at an early stage so that early treatment will be possible.
However, research shows that 75% of dementia and early dementia cases go unnoticed4and many such cases are only
∗Corresponding author. Tel.: +44 (0)1202 524111 ; fax: +44 (0)1202 962736.
E-mail address: darifoglu@bournemouth.ac.uk
1877-0509 c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.2 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
diagnosed when such impairment reaches moderate or advanced stage. The detection of early signs of motion and
cognitive impairment (MCI) via activity recognition will be useful to track motion and cognitive capabilities of theelderly, thus improving their life quality and financial saving. Unfortunately, currently there are no dementia friendlysmart homes addressing these people’s special needs.
Most common types of dementia (Alzheimer, Parkinsons disease) can be identified by behavioural changes like
sleep disturbances, difficulty of walking and inability to complete tasks. Such changes can provide key information
about memory, mobility and cognition of a person. For instance, an inhabitant suffering from Alzheimer may forget
his lunch, take multiple lunches instead, wake up in the middle of the night, go to the toilet frequently, or havedehydration problems because of forgetting to drink daily amount of water.
Recent studies suggest that changes in complex daily life tasks can be indicators of early decline
5. The best
markers of cognitive decline may not necessarily be detected based on a person’s performance at any single point intime, but rather by monitoring the trend over time and the variability of change in a duration
5. Thus, tracking an
elderly person’s life over time in a specially designed smart home, doing in-home health assessment and detecting theindicators of dementia at an early step would be beneficial.
The identification of early onsets of dementia using non-medical diagnosis methods requires the development of
new diagnostic tools. Although a few promising methods have been experimentally validated
6,7,8,9,10, the translation
of the current knowledge into smart homes still requires more dedication and work. Current assessment methods
mostly rely on queries from questionnaires or in-person examinations, which depend on recall of events or brief snap-
shots of function that may poorly represent a person’s typical state of function. Moreover, these studies include somepre-defined tasks given to the patients in order to do automatic assessment of cognitive decline by trained experts.
The main motivation for our work is that cognitive decline can be observed in daily activities and routines of
an elderly. Real-time monitoring of activities performed by elderly in a smart home would be beneficial for theearly detection of such decline. In this study, we firstly recognise activities by variants of RNNs, namely VRNNs,
LSTMs and GRUs and model the daily behaviour routines of a person. Whenever a new sequence is introduced, any
abnormality deviating from these regular behaviours are detected and could be used for further investigation by formal
or informal carer.
Unfortunately, there exists no publicly available dataset on abnormal behaviour of people with dementia. Producing
such a dataset require time and adequate experimental environment. Thus we propose in this paper, a way to artificiallyproduce data on abnormal activities reflecting on typical behaviour of elderly people with dementia. We believe that
this an important contribution.
The rest of the paper is organised as follows. Section 2 provides a brief overview of the related research to both
activity recognition and abnormal behaviour detection. Section 3 presents the details of the proposed methodology
together with the datasets and models used. Section 4 describes the experimental set-up and results of the experiments
followed by a discussion. Finally, Section 5 concludes the paper.
2. Literature Review
Activity recognition has been addressed using methods such as decision trees, Bayesian methods (Na ¨ıve Bayes
and Bayesian Networks), k-Nearest Neighbours, Neural Networks (Multilayer perceptron), SVMs, Fuzzy logic, Re-
gression models, Markov models (Hidden Markov Models, Conditional Random Fields) and classifier ensembles
(Boosting and bagging)
11. Recently, there has been growing interest in deep convolutional neural networks12,13,14,15,
Deep Belief Networks16, Restricked Boltzman Machines (RBMs)17,18,16,19and RNNs14,15,20. Previous work shows
that RNNs are useful, but leaves a lot of room for improvement. It is worthwhile to stress that to the best of our
knowledge, this study is the first applying RNNs to detect abnormalities related to dementia in the daily life routines
of an elderly person.
In17, RBMs are used for feature extraction and selection from sequential data. In14, the authors use a combination
of deep convolutional networks and LSTM to do multi-modal wearable activity recognition by showing that their
approach outperforms some of the previously reported results by up to 9% on OPPORTUNITY dataset. In21, the
authors utilised convolutional networks to classify activities using time-series data collected from smart phone sensors.Experiments show that increasing the number of convolutional layers increases the performance, but the complexity of
the derived features decreases with every additional layer. In
15, the authors explore deep, convolutional and recurrent

Damla Arifoglu et al. / Procedia Computer Science 110 (2017) 86–93 87 Available online at www.sciencedirect.com
00 (2016) 000–000
www.elsevier.com/locate/procedia
The 14th International Conference on Mobile Systems and Pervasive Computing
(MobiSPC 2017)
Activity Recognition and Abnormal Behaviour Detection with
Recurrent Neural Networks
Damla Arifoglu∗, Abdelhamid Bouchachia
Department of Computing and Informatics
Bournemouth University, UK
Abstract
In this paper, we study the problem of activity recognition and abnormal behaviour detection for elderly people with dementia.
Very few studies have attempted to address this problem presumably because of the lack of experimental data in the context of
dementia care. In particular, the paper investigates three variants of Recurrent Neural Networks (RNNs): Vanilla RNNs (VRNN),
Long Short Term RNNs (LSTM) and Gated Recurrent Unit RNNs (GRU). Here activity recognition is considered as a sequence
labelling problem, while abnormal behaviour is flagged based on the deviation from normal patterns. To provide an adequate
discussion of the performance of RNNs in this context, we compare them against the state-of-art methods such as Support Vector
Machines (SVMs), Na ¨ıve Bayes (NB), Hidden Markov Models (HMMs), Hidden Semi-Markov Models (HSMM) and Conditional
Random Fields (CRFs). The results obtained indicate that RNNs are competitive with those state-of-art methods. Moreover,
the paper presents a methodology for generating synthetic data reflecting on some behaviours of people with dementia given the
difficulty of obtaining real-world data.
c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.
Keywords: Smart Homes; Sensor based Activity Recognition; Recurrent Neural Networks; Dementia; Abnormal Behaviour Detection
1. Introduction
Studies indicate that by year 2030, 19% of people will be aged 74 to 84 and nearly half of people who are older
than 84 will have dementia1. Elderly people may suffer from the consequences of dementia, which is a condition that
causes problems with mobility, physical and mental abilities such as memory and thinking2. It also may cause de-
crease in the ability of speaking, writing, distinguishing objects, performing motor activities and performing complex
functional tasks (paying bills, preparing a meal, shopping, managing medication, etc.)3. An elderly person having
such cognitive decline loses independence in daily life and requires care and support from caregivers.
Cognitive diseases like dementia need to be detected at an early stage so that early treatment will be possible.
However, research shows that 75% of dementia and early dementia cases go unnoticed4and many such cases are only
∗Corresponding author. Tel.: +44 (0)1202 524111 ; fax: +44 (0)1202 962736.
E-mail address: darifoglu@bournemouth.ac.uk
1877-0509 c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.Available online at www.sciencedirect.com
00 (2016) 000–000
www.elsevier.com/locate/procedia
The 14th International Conference on Mobile Systems and Pervasive Computing
(MobiSPC 2017)
Activity Recognition and Abnormal Behaviour Detection with
Recurrent Neural Networks
Damla Arifoglu∗, Abdelhamid Bouchachia
Department of Computing and Informatics
Bournemouth University, UK
Abstract
In this paper, we study the problem of activity recognition and abnormal behaviour detection for elderly people with dementia.
Very few studies have attempted to address this problem presumably because of the lack of experimental data in the context of
dementia care. In particular, the paper investigates three variants of Recurrent Neural Networks (RNNs): Vanilla RNNs (VRNN),
Long Short Term RNNs (LSTM) and Gated Recurrent Unit RNNs (GRU). Here activity recognition is considered as a sequence
labelling problem, while abnormal behaviour is flagged based on the deviation from normal patterns. To provide an adequate
discussion of the performance of RNNs in this context, we compare them against the state-of-art methods such as Support Vector
Machines (SVMs), Na ¨ıve Bayes (NB), Hidden Markov Models (HMMs), Hidden Semi-Markov Models (HSMM) and Conditional
Random Fields (CRFs). The results obtained indicate that RNNs are competitive with those state-of-art methods. Moreover,
the paper presents a methodology for generating synthetic data reflecting on some behaviours of people with dementia given the
difficulty of obtaining real-world data.
c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.
Keywords: Smart Homes; Sensor based Activity Recognition; Recurrent Neural Networks; Dementia; Abnormal Behaviour Detection
1. Introduction
Studies indicate that by year 2030, 19% of people will be aged 74 to 84 and nearly half of people who are older
than 84 will have dementia1. Elderly people may suffer from the consequences of dementia, which is a condition that
causes problems with mobility, physical and mental abilities such as memory and thinking2. It also may cause de-
crease in the ability of speaking, writing, distinguishing objects, performing motor activities and performing complex
functional tasks (paying bills, preparing a meal, shopping, managing medication, etc.)3. An elderly person having
such cognitive decline loses independence in daily life and requires care and support from caregivers.
Cognitive diseases like dementia need to be detected at an early stage so that early treatment will be possible.
However, research shows that 75% of dementia and early dementia cases go unnoticed4and many such cases are only
∗Corresponding author. Tel.: +44 (0)1202 524111 ; fax: +44 (0)1202 962736.
E-mail address: darifoglu@bournemouth.ac.uk
1877-0509 c/circlecopyrt2016 The Authors. Published by Elsevier B.V .
Peer-review under responsibility of the Conference Program Chairs.2 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
diagnosed when such impairment reaches moderate or advanced stage. The detection of early signs of motion and
cognitive impairment (MCI) via activity recognition will be useful to track motion and cognitive capabilities of theelderly, thus improving their life quality and financial saving. Unfortunately, currently there are no dementia friendlysmart homes addressing these people’s special needs.
Most common types of dementia (Alzheimer, Parkinsons disease) can be identified by behavioural changes like
sleep disturbances, difficulty of walking and inability to complete tasks. Such changes can provide key information
about memory, mobility and cognition of a person. For instance, an inhabitant suffering from Alzheimer may forget
his lunch, take multiple lunches instead, wake up in the middle of the night, go to the toilet frequently, or havedehydration problems because of forgetting to drink daily amount of water.
Recent studies suggest that changes in complex daily life tasks can be indicators of early decline
5. The best
markers of cognitive decline may not necessarily be detected based on a person’s performance at any single point intime, but rather by monitoring the trend over time and the variability of change in a duration
5. Thus, tracking an
elderly person’s life over time in a specially designed smart home, doing in-home health assessment and detecting theindicators of dementia at an early step would be beneficial.
The identification of early onsets of dementia using non-medical diagnosis methods requires the development of
new diagnostic tools. Although a few promising methods have been experimentally validated
6,7,8,9,10, the translation
of the current knowledge into smart homes still requires more dedication and work. Current assessment methods
mostly rely on queries from questionnaires or in-person examinations, which depend on recall of events or brief snap-
shots of function that may poorly represent a person’s typical state of function. Moreover, these studies include somepre-defined tasks given to the patients in order to do automatic assessment of cognitive decline by trained experts.
The main motivation for our work is that cognitive decline can be observed in daily activities and routines of
an elderly. Real-time monitoring of activities performed by elderly in a smart home would be beneficial for theearly detection of such decline. In this study, we firstly recognise activities by variants of RNNs, namely VRNNs,
LSTMs and GRUs and model the daily behaviour routines of a person. Whenever a new sequence is introduced, any
abnormality deviating from these regular behaviours are detected and could be used for further investigation by formal
or informal carer.
Unfortunately, there exists no publicly available dataset on abnormal behaviour of people with dementia. Producing
such a dataset require time and adequate experimental environment. Thus we propose in this paper, a way to artificiallyproduce data on abnormal activities reflecting on typical behaviour of elderly people with dementia. We believe that
this an important contribution.
The rest of the paper is organised as follows. Section 2 provides a brief overview of the related research to both
activity recognition and abnormal behaviour detection. Section 3 presents the details of the proposed methodology
together with the datasets and models used. Section 4 describes the experimental set-up and results of the experiments
followed by a discussion. Finally, Section 5 concludes the paper.
2. Literature Review
Activity recognition has been addressed using methods such as decision trees, Bayesian methods (Na ¨ıve Bayes
and Bayesian Networks), k-Nearest Neighbours, Neural Networks (Multilayer perceptron), SVMs, Fuzzy logic, Re-
gression models, Markov models (Hidden Markov Models, Conditional Random Fields) and classifier ensembles
(Boosting and bagging)
11. Recently, there has been growing interest in deep convolutional neural networks12,13,14,15,
Deep Belief Networks16, Restricked Boltzman Machines (RBMs)17,18,16,19and RNNs14,15,20. Previous work shows
that RNNs are useful, but leaves a lot of room for improvement. It is worthwhile to stress that to the best of our
knowledge, this study is the first applying RNNs to detect abnormalities related to dementia in the daily life routines
of an elderly person.
In17, RBMs are used for feature extraction and selection from sequential data. In14, the authors use a combination
of deep convolutional networks and LSTM to do multi-modal wearable activity recognition by showing that their
approach outperforms some of the previously reported results by up to 9% on OPPORTUNITY dataset. In21, the
authors utilised convolutional networks to classify activities using time-series data collected from smart phone sensors.Experiments show that increasing the number of convolutional layers increases the performance, but the complexity of
the derived features decreases with every additional layer. In
15, the authors explore deep, convolutional and recurrent

88 Damla Arifoglu et al. / Procedia Computer Science 110 (2017) 86–93
Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000 3
approaches across three representative datasets that contain movement data captured with wearable sensors. Moreover,
they describe how to train recurrent approaches in this setting and introduce a novel regularisation approach, showingbetter results over OPPORTUNITY, PAMAP2 and Daphnet Gait datasets. In
19, results with RBM on CASAS dataset
outperformed HMM and Na ¨ıve Bayes Classifier (NBC) in most of the cases. In22, the authors use RNNs to predict
the future values (start time, duration) of the activities.
Most of the aforementioned studies use movement data such as OPPORTUNITY, SKODA17,12,13,14or UCI HAR
smart phone dataset, MIT home dataset21,16, which are obtained through body worn sensors. Except the work by Fang
et al.19,20, none of these studies focus on daily activity datasets collected by sensors placed at home. In this work, we
investigate RNNs on daily activities data obtained by van Kasteren23using various environment sensors (see Sec. 3.1
for more details).
In-home automatic assessment of cognitive decline has been the subject of some studies dedicated24,6,25,18. For
instance, in24, machine learning approaches such as SVMs and Na ¨ıve Bayes are used. In18, Parkinson’s Disease state
assessment in home is explored by means of RBMs using data from body worn sensors. In25, the authors use Markov
Logic Network, which is a probabilistic logic that unifies statistical and symbolic reasoning to detect anomalies.In
24, some instructions to perform some tasks (e.g., sweeping the kitchen, dusting the floor, etc.) are given to the
patients who then receive scores after completing those tasks. These scores are calculated based on the time spent,
the frequency of the sensor triggered, etc. One disadvantage of this scenario is that some pre-selected activities are
performed and instructions are given to the elderly who might not be able to cope with such tasks at all. Moreover,
using rule-based systems, an expert is needed to manually integrate resident-specific rules to the system since every
person has her/his own daily life routines. For example waking up and drinking water in the middle of the night might
be normal for a person, while abnormal for some other person. However, our approach does not require any expertknowledge, since it learns what is normal and abnormal from the training data automatically. Specifically, we aim
in this study to detect anomalies in the natural flow of daily living without giving any instruction and considering
not only some time interval, but everyday living scenario. Continuous assessment of the person is more valid, since
activities are performed in the person’s own home setting.
3. Proposed Method
To assess RNNs in activity recognition and abnormal activity detection, we propose the following steps: Firstly,
raw dataset is segmented into slices by using a sliding window approach. The window size is 60 seconds time of
sensor readings as described in
23. Secondly, sensor-based features are extracted from these slices. These features
arebinary, change-point andlast-fired representations which are used also in23. Thirdly, RNNs (Vanilla, GRU and
LSTM) are trained to recognise daily activities and encode daily-life behaviour routines. Lastly, the trained model is
used to detect anomalies deviating from the normal daily-life sequences.
In the following we describe the dataset as well as the methodology used to generate artificial dataset that reflects
on the typical behaviour of a person with dementia.
3.1. Dataset and Features
We used the popular dataset collected by Van Kasteren23from 3 households which are denoted as dataset A,B
andC. The data captures daily-life activities such as sleeping, cooking, leaving home, etc. using sensors placed at
the homes in less than a month. Please see23for more details. We applied the same sliding window approach as
in23to extract the sensor reading chunks. We also considered three feature representations: binary, change-point and
last-fired which are described as follows:
•Binary: This representation gives 1 when the sensor is triggered and 0 when that sensor is not triggered.
•Change-point : This representation gives information when a sensor changes value. More specifically, it gives
1 when a sensor changes its current state (either from state 1 to state 0 or vice versa) and a 0 when its value
remains the same.
•Last-fired : This representation indicates which sensor is fired last. The sensor that changed state last continues
to give 1 and changes to 0 when another sensor changes state.4 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
3.2. Generation of Abnormal Activities Related to Dementia
Since we do not have any available dataset related to abnormal behaviour of people with dementia, we artificially
create some anomalies in the dataset. In order to show the applicability of the proposed work to detect these anomalies,we focus on two different kinds of anomalies that can be seen in daily-life routines of elderly people with dementia:1) Forgetting or repeating activities 2) Dehydration and disruption in sleep.
1.Forgetting and repeating activities: Elderly people su ffering from dementia may forget whether they performed
a particular daily activity or not, so they may repeat that activity multiple times or they may skip that activity. Forinstance, an elderly person suffering from Alzheimer may forget to have lunch, take multiple lunches instead
26,
to have dinner and start to prepare it in the middle of the night. To reflect on this, we generate this kind ofabnormal activities by manually inserting a specific set of actions within the normal activity sequence. This willresult in multiple occurrences of that activity, which will occur in some inadequate time of the day such as havingdinner in the middle of the night. We inject the instances of the following activities: brushing teeth, preparing
dinner, eating, getting snack into the normal activity sequences to generate abnormal activities related to the
frequency.
2.Dehydration and disruption in sleep: Degeneration of the sleep-waking cycle, sleep disorders and night time
wandering are among the most severe behavioural symptoms of dementia. For example, elderly people may wakeup many times in the night to use the toilet and go back to sleep and may forget to take daily amount of water
26,27.
We simulate these anomalies by inserting some synthetic activities in the normal night-time activity sequences
of a person. More specifically, we inject getting drink, going to toilet into the sleeping activity of normal daily
activity sequences. This will emulate the activities of getting drink and going to the toilet frequently in the middle
of the night.
We generate these abnormal activity instances on dataset Awhich has the following 9 activities: Leave house, use
toilet, take shower, brush teeth, go to bed, prepare breakfast, prepare dinner, get snack, get drink. As a result, we have
multiple instances of those injected instances in order to simulate the anomalies related to dementia. Here, please note
that there is only one subject in the dataset. We take the lifestyle in the training data as a norm and then synthesise the
abnormalities deviating from this norm and introduce these abnormalities in the test data. These activities are totallynormal on their own but they become abnormal when they occur at a wrong time of the day and after or before a
specific activity. Hence, capturing these abnormalities within the context is important. In all, we manually synthesise
135 abnormal activity slices.
3.3. Activity Recognition and Abnormal Behaviour Detection
We believe that the order of activities and their temporal and spatial information is important to encode an elderly
person’s daily life routines. This kind of information can provide important cues to understand the daily patterns and
thus to detect any anomalies in those patterns. Sequence labelling methods such as HMMs and RNNs can capturetemporal and spatial relationship between activities, which some generative methods like SVMs can not do. In thiswork, we investigate the adequacy of RNNs to this task.
In order to recognise daily activities, training instances of the datasets and their corresponding labels are fed into
the RNNs. Then when a new test sequence is introduced, the trained model assigns labels to each activity instancesof that sequence. Each model gives a confidence value about the assigned label for the new sequence. Firstly, wecalculate the mean of confidence values of training instances that are assigned by the model. Then, when a new testsequence is introduced if the model assigns it to a class label with a confidence value which is bigger than the mean,the sequence is considered as a normal activity, otherwise it is abnormal activity.
3.4. RNN Architectures
In the following we give a summary of the RNN architectures used in this work, more specifically Vanilla RNNs,
Long Short Term Memory RNNs, and Gated Recurrent Unit RNNs. Then, we describe how they are used in the
context of daily activity recognition and abnormal activity detection tasks.

Damla Arifoglu et al. / Procedia Computer Science 110 (2017) 86–93 89
Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000 3
approaches across three representative datasets that contain movement data captured with wearable sensors. Moreover,
they describe how to train recurrent approaches in this setting and introduce a novel regularisation approach, showingbetter results over OPPORTUNITY, PAMAP2 and Daphnet Gait datasets. In
19, results with RBM on CASAS dataset
outperformed HMM and Na ¨ıve Bayes Classifier (NBC) in most of the cases. In22, the authors use RNNs to predict
the future values (start time, duration) of the activities.
Most of the aforementioned studies use movement data such as OPPORTUNITY, SKODA17,12,13,14or UCI HAR
smart phone dataset, MIT home dataset21,16, which are obtained through body worn sensors. Except the work by Fang
et al.19,20, none of these studies focus on daily activity datasets collected by sensors placed at home. In this work, we
investigate RNNs on daily activities data obtained by van Kasteren23using various environment sensors (see Sec. 3.1
for more details).
In-home automatic assessment of cognitive decline has been the subject of some studies dedicated24,6,25,18. For
instance, in24, machine learning approaches such as SVMs and Na ¨ıve Bayes are used. In18, Parkinson’s Disease state
assessment in home is explored by means of RBMs using data from body worn sensors. In25, the authors use Markov
Logic Network, which is a probabilistic logic that unifies statistical and symbolic reasoning to detect anomalies.In
24, some instructions to perform some tasks (e.g., sweeping the kitchen, dusting the floor, etc.) are given to the
patients who then receive scores after completing those tasks. These scores are calculated based on the time spent,
the frequency of the sensor triggered, etc. One disadvantage of this scenario is that some pre-selected activities are
performed and instructions are given to the elderly who might not be able to cope with such tasks at all. Moreover,
using rule-based systems, an expert is needed to manually integrate resident-specific rules to the system since every
person has her/his own daily life routines. For example waking up and drinking water in the middle of the night might
be normal for a person, while abnormal for some other person. However, our approach does not require any expertknowledge, since it learns what is normal and abnormal from the training data automatically. Specifically, we aim
in this study to detect anomalies in the natural flow of daily living without giving any instruction and considering
not only some time interval, but everyday living scenario. Continuous assessment of the person is more valid, since
activities are performed in the person’s own home setting.
3. Proposed Method
To assess RNNs in activity recognition and abnormal activity detection, we propose the following steps: Firstly,
raw dataset is segmented into slices by using a sliding window approach. The window size is 60 seconds time of
sensor readings as described in
23. Secondly, sensor-based features are extracted from these slices. These features
arebinary, change-point andlast-fired representations which are used also in23. Thirdly, RNNs (Vanilla, GRU and
LSTM) are trained to recognise daily activities and encode daily-life behaviour routines. Lastly, the trained model is
used to detect anomalies deviating from the normal daily-life sequences.
In the following we describe the dataset as well as the methodology used to generate artificial dataset that reflects
on the typical behaviour of a person with dementia.
3.1. Dataset and Features
We used the popular dataset collected by Van Kasteren23from 3 households which are denoted as dataset A,B
andC. The data captures daily-life activities such as sleeping, cooking, leaving home, etc. using sensors placed at
the homes in less than a month. Please see23for more details. We applied the same sliding window approach as
in23to extract the sensor reading chunks. We also considered three feature representations: binary, change-point and
last-fired which are described as follows:
•Binary: This representation gives 1 when the sensor is triggered and 0 when that sensor is not triggered.
•Change-point : This representation gives information when a sensor changes value. More specifically, it gives
1 when a sensor changes its current state (either from state 1 to state 0 or vice versa) and a 0 when its value
remains the same.
•Last-fired : This representation indicates which sensor is fired last. The sensor that changed state last continues
to give 1 and changes to 0 when another sensor changes state.4 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
3.2. Generation of Abnormal Activities Related to Dementia
Since we do not have any available dataset related to abnormal behaviour of people with dementia, we artificially
create some anomalies in the dataset. In order to show the applicability of the proposed work to detect these anomalies,
we focus on two different kinds of anomalies that can be seen in daily-life routines of elderly people with dementia:1) Forgetting or repeating activities 2) Dehydration and disruption in sleep.
1.Forgetting and repeating activities: Elderly people su ffering from dementia may forget whether they performed
a particular daily activity or not, so they may repeat that activity multiple times or they may skip that activity. Forinstance, an elderly person suffering from Alzheimer may forget to have lunch, take multiple lunches instead
26,
to have dinner and start to prepare it in the middle of the night. To reflect on this, we generate this kind ofabnormal activities by manually inserting a specific set of actions within the normal activity sequence. This willresult in multiple occurrences of that activity, which will occur in some inadequate time of the day such as havingdinner in the middle of the night. We inject the instances of the following activities: brushing teeth, preparing
dinner, eating, getting snack into the normal activity sequences to generate abnormal activities related to the
frequency.
2.Dehydration and disruption in sleep: Degeneration of the sleep-waking cycle, sleep disorders and night time
wandering are among the most severe behavioural symptoms of dementia. For example, elderly people may wakeup many times in the night to use the toilet and go back to sleep and may forget to take daily amount of water
26,27.
We simulate these anomalies by inserting some synthetic activities in the normal night-time activity sequences
of a person. More specifically, we inject getting drink, going to toilet into the sleeping activity of normal daily
activity sequences. This will emulate the activities of getting drink and going to the toilet frequently in the middle
of the night.
We generate these abnormal activity instances on dataset Awhich has the following 9 activities: Leave house, use
toilet, take shower, brush teeth, go to bed, prepare breakfast, prepare dinner, get snack, get drink. As a result, we have
multiple instances of those injected instances in order to simulate the anomalies related to dementia. Here, please note
that there is only one subject in the dataset. We take the lifestyle in the training data as a norm and then synthesise the
abnormalities deviating from this norm and introduce these abnormalities in the test data. These activities are totallynormal on their own but they become abnormal when they occur at a wrong time of the day and after or before a
specific activity. Hence, capturing these abnormalities within the context is important. In all, we manually synthesise
135 abnormal activity slices.
3.3. Activity Recognition and Abnormal Behaviour Detection
We believe that the order of activities and their temporal and spatial information is important to encode an elderly
person’s daily life routines. This kind of information can provide important cues to understand the daily patterns and
thus to detect any anomalies in those patterns. Sequence labelling methods such as HMMs and RNNs can capturetemporal and spatial relationship between activities, which some generative methods like SVMs can not do. In thiswork, we investigate the adequacy of RNNs to this task.
In order to recognise daily activities, training instances of the datasets and their corresponding labels are fed into
the RNNs. Then when a new test sequence is introduced, the trained model assigns labels to each activity instancesof that sequence. Each model gives a confidence value about the assigned label for the new sequence. Firstly, wecalculate the mean of confidence values of training instances that are assigned by the model. Then, when a new testsequence is introduced if the model assigns it to a class label with a confidence value which is bigger than the mean,the sequence is considered as a normal activity, otherwise it is abnormal activity.
3.4. RNN Architectures
In the following we give a summary of the RNN architectures used in this work, more specifically Vanilla RNNs,
Long Short Term Memory RNNs, and Gated Recurrent Unit RNNs. Then, we describe how they are used in the
context of daily activity recognition and abnormal activity detection tasks.

90 Damla Arifoglu et al. / Procedia Computer Science 110 (2017) 86–93
Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000 5
1.Vanilla Recurrent Neural Networks: In feed-forward neural network, it is assumed that all inputs and outputs
are independent of each other, but RNNs have a recurrent hidden state whose activation at each time is dependent
on that of the previous time. This architecture is recurrent as some of the connections within the network form adirected cycle, where the current time-step tconsiders the states of the network in the previous time-step t−1.
They share parameters for different time-steps which enables them to be used in sequential data. RNNs are calledrecurrent because they perform the same task for every element of a sequence, with the output being dependent
on the previous computations. Another way to think about RNNs is that they have a memory which captures
information about what has been calculated so far. However, there is a drawback of Vanilla RNNs, as shownby Bengio et al.
28, Vanilla RNNs are not capable of capturing long term dependencies on sequences because of
the vanishing gradient problem. In theory, RNNs can make use of information in arbitrarily long sequences, butin practice they are limited to looking back only a few steps. Thus, the following two RNN architectures are
exploited to solve this problem.
2.Long Short Term Memory (LSTM) Recurrent Neural Networks: LSTM cells are designed to counter the
effect of diminishing gradients when error derivatives are backpropagated through many layers through time in
recurrent networks
29. Each LSTM unit keeps track of an internal state that represents its memory. Over time
the cells learn to output, overwrite, or null their internal memory based on their current input and the history ofpast internal states, leading to a system capable of retaining information across hundreds of time-steps
29. LSTM
blocks have 3 gates to control the flow of information into or out of their memory. For example, an input gate
controls the extent to which a new value flows into the memory. A forget gate controls the extent to which a
value remains in memory while an output gate is used to compute the output activation of the block (see Figure
1).
Fig. 1. Left: LSTM, Right: GRU. While LSTM can be described as the input signals xtat time t, the output signals yt, the forget gate ft, and
the input gate it, the output gate ot,; GRU, on the other hand, can be described in terms of two internal variables, which retain the previous hand
current hinner states respectively.
3.Gated Recurrent Unit: Cho et al.28recently proposed GRU, which is like LSTM but it has fewer parameters
than LSTM, as GRUs lack an output gate. In GRU, each hidden unit has two gates, which are called update and
reset gates (see Figure 1). GRU also controls the flow of information to prevent vanishing gradient problem, but
without having to use a memory unit.
4. Experiments and Results
We used Keras Deep Learning library’s30and Theano’s31implementations of the RNNs (GRU, LSTM, Vanilla
RNN) in this study. Moreover for the sake of comparison, we also used the One-class SVM from WEKA with
default parameters, Na ¨ıve Bayes (NB), Hidden Markov Models (HMM), Hidden Semi-Markov Models (HSMM) and
Conditional Random Fields (CRF) which are based on the implementation provided in23.
We split the data (see Sec. 3.1) into a test and training set using the leave-one-day-out cross-validation approach.
One full day of sensor readings is used for testing and the remaining days are used for training. Then we cycle over
all days and report the average performance.
We evaluate metrics proposed in23: precision, recall, F-measure and accuracy. We calculate precision and recall
for each class separately and then take the average over all classes. Note that precision and recall measures are used
since these metrics give some idea about how well the models perform on imbalanced datasets like the one in this6 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
Table 1. Activity recognition results on dataset A
Model Feature Precision Recall F-Measure Accuracy
NB Binary 48.3 ±17.7 42.6 ±16.6 45.1 ±16.9 77.1 ±20.8
Change-point 52.7 ±17.5 43.2 ±18.0 47.1 ±17.2 55.9 ±18.8
Last-fired 67.3 ±17.2 64.8 ±14.6 65.8 ±15.5 95.3 ±2.8
HMM Binary 37.9 ±19.8 45.5 ±19.5 41.0 ±19.5 59.1 ±28.7
Change-point 70.3 ±16.0 74.3 ±13.3 72.0 ±14.2 92.3 ±5.8
Last-fired 54.6 ±17.0 69.5 ±12.7 60.8 ±14.9 89.5 ±8.4
HSMM Binary 39.5 ±18.9 48.5 ±19.5 43.2 ±19.1 59.5 ±29.0
Change-point 70.5 ±16.0 75.0 ±12.1 72.4 ±13.7 91.8 ±5.9
Last-fired 60.2 ±15.4 73.8 ±12.5 66.0 ±13.7 91.0 ±7.2
CRF Binary 59.2 ±18.3 56.1 ±17.3 57.2 ±17.3 89.8 ±8.5
Change-point 73.5 ±16.6 68.0 ±16.0 70.4 ±15.9 91.4 ±5.6
Last-fired 66.2 ±15.8 65.8 ±14.0 65.9 ±14.6 96.4 ±2.4
Vanilla Binary 46.5 ±17.7 64.8 ±16.2 53.5 ±16.3 86.8 ±10.6
Change-point 46.3 ±19.5 63.8 ±16.4 53.2 ±17.9 61.4 ±16.4
Last-fired 61.9 ±19.1 74.3 ±12.8 67.2 ±16.4 95.5 ±3.4
LSTM Binary 50.8 ±18.4 63.9 ±16.5 56.2 ±17.1 86.7 ±10.5
Change-point 46.8 ±18.7 63.6 ±14 53.5 ±16.7 61.4 ±16.4
Last-fired 63.7 ±19.9 73.9 ±16.8 68.1 ±18.2 96.7 ±2.6
GRU Binary 47.3 ±18.7 69.1 ±14.9 55.4 ±16.5 86.6 ±10.7
Change-point 42.9 ±19 65.0 ±15.3 51.0 ±17.1 61.4 ±16.4
Last-fired 61.8 ±16.3 80.6 ±11.5 69.5 ±14.0 96.1 ±2.5
SVM Binary 45.6 ±17.9 69.1 ±15.9 54.2 ±15.9 85.4 ±10.4
Change-point 40.3 ±19.1 63.4 ±14.6 48.6 ±17.0 55.9 ±18.7
Last-fired 58.6 ±16.2 77.2 ±14.0 66.3 ±14.9 96.1 ±2.4Table 2. Activity recognition results on dataset B.
Model Feature Precision Recall F-Measure Accuracy
NB Binary 33.6 ±10.9 32.5 ±8.4 32.4 ±8.9 80.4 ±18.9
Change-point 40.9 ±7.2 38.9 ±5.7 39.5 ±5.9 67.8 ±18.6
Last-fired 43.7 ±8.7 44.6 ±7.2 43.3 ±4.8 86.2 ±13.8
HMM Binary 38.8 ±14.7 44.7 ±13.4 40.7 ±12.4 63.2 ±24.7
Change-point 48.2 ±17.2 63.1 ±14.1 53.6 ±16.5 81.0 ±14.2
Last-fired 38.5 ±15.8 46.6 ±19.5 41.8 ±17.1 48.4 ±26.9
HSMM Binary 37.4 ±16.9 44.6 ±14.3 39.9 ±14.3 63.8 ±24.2
Change-point 49.8 ±15.8 65.2 ±13.4 55.7 ±14.6 82.3 ±13.5
Last-fired 40.8 ±11.6 53.3 ±10.9 45.8 ±11.2 67.1 ±24.8
CRF Binary 35.7 ±15.2 40.6 ±12.0 37.5 ±13.7 78.0 ±25.9
Change-point 48.3 ±8.3 51.5 ±8.5 49.7 ±7.9 92.9 ±6.2
Last-fired 46.9 ±12.5 47.8 ±12.1 46.6 ±12.9 89.2 ±13.9
Vanilla Binary 26.7 ±13.5 46.9 ±24.8 32.5 ±17.9 65.2 ±34.7
Change-point 39.6 ±8 62.4 ±15.3 48.3 ±10.2 76.9 ±13.9
Last-fired 41.2 ±12.3 64.4 ±17.8 49.7 ±13.6 87.9 ±13.1
LSTM Binary 29.1 ±12.0 44.0 ±22.0 33.9 ±16.2 63.5 ±32.7
Change-point 40.0 ±11.2 59.0 ±16.4 47.5 ±12.9 76.8 ±14.2
Last-fired 40.8 ±10.7 60.1 ±16.3 48.2 ±12.3 87.2 ±13.2
GRU Binary 28.5 ±15.9 36.3 ±17.2 31.4 ±16.2 64.5 ±32.1
Change-point 37.7 ±7.6 53.5 ±9.2 44.9 ±7.1 76.4 ±14.5
Last-fired 41.7 ±13.2 56.9 ±17.9 47.5 ±14.6 87.0 ±12.9
SVM Binary 39.6 ±10.9 58.5 ±17.4 46.7 ±12.9 81.6 ±18.5
Change-point 32.3 ±6.5 53.6 ±7.5 40.0 ±6.2 67.9 ±28.5
Last-fired 36.4 ±5.4 54.6 ±10.4 43.5 ±6.6 86. 2±14.9
study. On the other hand, the accuracy represents the percentage of correctly classified time slices, therefore morefrequently occurring classes have a larger weight in this measure.
To evaluate the performance of abnormal behaviour detection, we use the following evaluation metrics: True
Positive Rate (TPR) and False Positive Rate (FPR). TPR is the percentage of correctly detected abnormal activities
out of total abnormal activities, FPR is the percentage of normal activities that are detected falsely as abnormal
activities by the algorithm (out of total number of normal activities).
To run experiments on RNNs, we left out 10% of the training data for validation and we used drop-out with a value
of 0.2. We also set the batch size to 10 instances and the epoch to 500 iterations. The internal architecture of RNNs (2layers consisting of 30 and 50 nodes respectively) and time step of the sequences (25 activity slices) were empirically
set.
Note that the results obtained by the models HMM, HSMM, CRF and NB (see Tab. 1 – 4) are taken from the study
by Kasteren et al.
23.
Table 1 refers to the results obtained on dataset Aand shows that there is no clear winner among the three different
feature representations. Considering the accuracy, the results indicate that LSTM is the best method (with the accu-racy of 96 .7%) when last-fired feature is used, while HMM performs the worst. Using change-point feature, HMM
outperforms all other methods. Using binary feature on the other hand shows that CRF (accuracy of 89.8%) is the
best. Also all RNNs, NB and SVM do not perform well when adopting change-point feature. HMM and HSMM
are not good when using binary feature representation. In a nutshell, for the majority of the methods, except HMM
and HSMM, last-fired representation is the best one. In terms of recall which reflects better on performance in the
presence of imbalanced data, the highest value is obtained by GRU (80.6%). This potentially indicate that RNNs aregood to detect relevant class instances. CRF, for instance, score higher on precision, because the most frequent-class
instances are favoured, but then it is not so good at when it comes to the infrequent classes. Overall, there is a clear
hint that that recurrent architectures perform better than HM, NB and HSMM for most of the cases, while CRF is
slightly better than these recurrent architectures on dataset A.
Table 2 refers to the results obtained on dataset Band shows that SVM is the best method when adopting binary
representation achieving the accuracy of 81.6%. On the other hand, CRF is the best when using the change-point
feature and last-fired representations with accuracy 92.9% and 89.2% respectively. It can be noted that HMM is not
as good as the other methods achieving in the best case only 81.0% with the change-point representation. The closest
successful model to CRF is Vanilla RNN and again overall RNNs deliver high recall rates compared to the othermethods. Change-point andlast-fired representations give the highest recall results except for CRF.
Table 4 reports the results on dataset Cshowing that CRF performs best for change-point and binary represen-
tations obtaining 82.2% and 89.7% respectively. Overall, none of the methods performs well when adopting binary

Damla Arifoglu et al. / Procedia Computer Science 110 (2017) 86–93 91
6 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
Table 1. Activity recognition results on dataset A
Model Feature Precision Recall F-Measure Accuracy
NB Binary 48.3 ±17.7 42.6 ±16.6 45.1 ±16.9 77.1 ±20.8
Change-point 52.7 ±17.5 43.2 ±18.0 47.1 ±17.2 55.9 ±18.8
Last-fired 67.3 ±17.2 64.8 ±14.6 65.8 ±15.5 95.3 ±2.8
HMM Binary 37.9 ±19.8 45.5 ±19.5 41.0 ±19.5 59.1 ±28.7
Change-point 70.3 ±16.0 74.3 ±13.3 72.0 ±14.2 92.3 ±5.8
Last-fired 54.6 ±17.0 69.5 ±12.7 60.8 ±14.9 89.5 ±8.4
HSMM Binary 39.5 ±18.9 48.5 ±19.5 43.2 ±19.1 59.5 ±29.0
Change-point 70.5 ±16.0 75.0 ±12.1 72.4 ±13.7 91.8 ±5.9
Last-fired 60.2 ±15.4 73.8 ±12.5 66.0 ±13.7 91.0 ±7.2
CRF Binary 59.2 ±18.3 56.1 ±17.3 57.2 ±17.3 89.8 ±8.5
Change-point 73.5 ±16.6 68.0 ±16.0 70.4 ±15.9 91.4 ±5.6
Last-fired 66.2 ±15.8 65.8 ±14.0 65.9 ±14.6 96.4 ±2.4
Vanilla Binary 46.5 ±17.7 64.8 ±16.2 53.5 ±16.3 86.8 ±10.6
Change-point 46.3 ±19.5 63.8 ±16.4 53.2 ±17.9 61.4 ±16.4
Last-fired 61.9 ±19.1 74.3 ±12.8 67.2 ±16.4 95.5 ±3.4
LSTM Binary 50.8 ±18.4 63.9 ±16.5 56.2 ±17.1 86.7 ±10.5
Change-point 46.8 ±18.7 63.6 ±14 53.5 ±16.7 61.4 ±16.4
Last-fired 63.7 ±19.9 73.9 ±16.8 68.1 ±18.2 96.7 ±2.6
GRU Binary 47.3 ±18.7 69.1 ±14.9 55.4 ±16.5 86.6 ±10.7
Change-point 42.9 ±19 65.0 ±15.3 51.0 ±17.1 61.4 ±16.4
Last-fired 61.8 ±16.3 80.6 ±11.5 69.5 ±14.0 96.1 ±2.5
SVM Binary 45 .6±17.9 69.1 ±15.9 54.2 ±15.9 85.4 ±10.4
Change-point 40.3 ±19.1 63.4 ±14.6 48.6 ±17.0 55.9 ±18.7
Last-fired 58.6 ±16.2 77.2 ±14.0 66.3 ±14.9 96.1 ±2.4Table 2. Activity recognition results on dataset B.
Model Feature Precision Recall F-Measure Accuracy
NB Binary 33.6 ±10.9 32.5 ±8.4 32.4 ±8.9 80.4 ±18.9
Change-point 40.9 ±7.2 38.9 ±5.7 39.5 ±5.9 67.8 ±18.6
Last-fired 43.7 ±8.7 44.6 ±7.2 43.3 ±4.8 86.2 ±13.8
HMM Binary 38.8 ±14.7 44.7 ±13.4 40.7 ±12.4 63.2 ±24.7
Change-point 48.2 ±17.2 63.1 ±14.1 53.6 ±16.5 81.0 ±14.2
Last-fired 38.5 ±15.8 46.6 ±19.5 41.8 ±17.1 48.4 ±26.9
HSMM Binary 37.4 ±16.9 44.6 ±14.3 39.9 ±14.3 63.8 ±24.2
Change-point 49.8 ±15.8 65.2 ±13.4 55.7 ±14.6 82.3 ±13.5
Last-fired 40.8 ±11.6 53.3 ±10.9 45.8 ±11.2 67.1 ±24.8
CRF Binary 35.7 ±15.2 40.6 ±12.0 37.5 ±13.7 78.0 ±25.9
Change-point 48.3 ±8.3 51.5 ±8.5 49.7 ±7.9 92.9 ±6.2
Last-fired 46.9 ±12.5 47.8 ±12.1 46.6 ±12.9 89.2 ±13.9
Vanilla Binary 26.7 ±13.5 46.9 ±24.8 32.5 ±17.9 65.2 ±34.7
Change-point 39.6 ±8 62.4 ±15.3 48.3 ±10.2 76.9 ±13.9
Last-fired 41.2 ±12.3 64.4 ±17.8 49.7 ±13.6 87.9 ±13.1
LSTM Binary 29.1 ±12.0 44.0 ±22.0 33.9 ±16.2 63.5 ±32.7
Change-point 40.0 ±11.2 59.0 ±16.4 47.5 ±12.9 76.8 ±14.2
Last-fired 40.8 ±10.7 60.1 ±16.3 48.2 ±12.3 87.2 ±13.2
GRU Binary 28.5 ±15.9 36.3 ±17.2 31.4 ±16.2 64.5 ±32.1
Change-point 37.7 ±7.6 53.5 ±9.2 44.9 ±7.1 76.4 ±14.5
Last-fired 41.7 ±13.2 56.9 ±17.9 47.5 ±14.6 87.0 ±12.9
SVM Binary 39.6 ±10.9 58.5 ±17.4 46.7 ±12.9 81.6 ±18.5
Change-point 32.3 ±6.5 53.6 ±7.5 40.0 ±6.2 67.9 ±28.5
Last-fired 36.4 ±5.4 54.6 ±10.4 43.5 ±6.6 86.2 ±14.9
study. On the other hand, the accuracy represents the percentage of correctly classified time slices, therefore more
frequently occurring classes have a larger weight in this measure.
To evaluate the performance of abnormal behaviour detection, we use the following evaluation metrics: True
Positive Rate (TPR) and False Positive Rate (FPR). TPR is the percentage of correctly detected abnormal activities
out of total abnormal activities, FPR is the percentage of normal activities that are detected falsely as abnormal
activities by the algorithm (out of total number of normal activities).
To run experiments on RNNs, we left out 10% of the training data for validation and we used drop-out with a value
of 0.2. We also set the batch size to 10 instances and the epoch to 500 iterations. The internal architecture of RNNs (2layers consisting of 30 and 50 nodes respectively) and time step of the sequences (25 activity slices) were empirically
set.
Note that the results obtained by the models HMM, HSMM, CRF and NB (see Tab. 1 – 4) are taken from the study
by Kasteren et al.
23.
Table 1 refers to the results obtained on dataset Aand shows that there is no clear winner among the three different
feature representations. Considering the accuracy, the results indicate that LSTM is the best method (with the accu-racy of 96 .7%) when last-fired feature is used, while HMM performs the worst. Using change-point feature, HMM
outperforms all other methods. Using binary feature on the other hand shows that CRF (accuracy of 89.8%) is the
best. Also all RNNs, NB and SVM do not perform well when adopting change-point feature. HMM and HSMM
are not good when using binary feature representation. In a nutshell, for the majority of the methods, except HMM
and HSMM, last-fired representation is the best one. In terms of recall which reflects better on performance in the
presence of imbalanced data, the highest value is obtained by GRU (80.6%). This potentially indicate that RNNs aregood to detect relevant class instances. CRF, for instance, score higher on precision, because the most frequent-class
instances are favoured, but then it is not so good at when it comes to the infrequent classes. Overall, there is a clear
hint that that recurrent architectures perform better than HM, NB and HSMM for most of the cases, while CRF is
slightly better than these recurrent architectures on dataset A.
Table 2 refers to the results obtained on dataset Band shows that SVM is the best method when adopting binary
representation achieving the accuracy of 81.6%. On the other hand, CRF is the best when using the change-point
feature and last-fired representations with accuracy 92.9% and 89.2% respectively. It can be noted that HMM is not
as good as the other methods achieving in the best case only 81.0% with the change-point representation. The closest
successful model to CRF is Vanilla RNN and again overall RNNs deliver high recall rates compared to the othermethods. Change-point andlast-fired representations give the highest recall results except for CRF.
Table 4 reports the results on dataset Cshowing that CRF performs best for change-point and binary represen-
tations obtaining 82.2% and 89.7% respectively. Overall, none of the methods performs well when adopting binary

92 Damla Arifoglu et al. / Procedia Computer Science 110 (2017) 86–93
Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000 7
Table 3. Activity recognition results on dataset C
Model Feature Precision Recall F-Measure Accuracy
NB Binary 19 .6±11.4 16.8 ±7.5 17.8 ±9.1 46.5 ±22.6
Change-point 39.9 ±6.9 30.8 ±4.8 34.5 ±4.6 57.6 ±15.4
Last-fired 40.5 ±7.4 46.4 ±14.8 42.3 ±6.8 87.0 ±12.2
HMM Binary 15.2 ±9.2 17.2 ±9.3 15.7 ±8.8 26.5 ±22.7
Change-point 41.4 ±8.8 50.0 ±11.4 44.9 ±8.8 77.2 ±14.6
Last-fired 40.7 ±9.7 53.7 ±16.2 45.9 ±11.2 83.9 ±13.9
HSMM Binary 15.6 ±9.2 20.4 ±10.9 17.3 ±9.6 31.2 ±24.6
Change-point 43.8 ±10.0 52.3 ±12.8 47.4 ±10.5 77.5 ±15.3
Last-fired 42.5 ±10.8 56.0 ±15.4 47.9 ±11.3 84.5 ±13.2
CRF Binary 17.8 ±22.1 21.8 ±20.9 19.0 ±21.8 46.3 ±25.5‘
Change-point 36.7 ±18.0 39.6 ±17.4 38.0 ±17.6 82.2 ±13.9
Last-fired 37.7 ±17.1 40.4 ±16.0 38.9 ±16.5 89 .7±8.4
Vanilla Binary 15 .4±5.3 43.1 ±18.1 22.2 ±7.3 50.2 ±22.4
Change-point 31.3 ±7.1 54.9 ±11.3 39.5 ±8.3 72.2 ±13.0
Last-fired 38.3 ±16.3 59.6 ±15.1 45.8 ±14.8 86.7 ±12.5
LSTM Binary 16.8 ±6.2 34.8 ±12.5 22.1 ±7.4 45.3 ±21.2
Change-point 31.0 ±5.1 53.3 ±6.5 38.9 ±5.0 72.0 ±13.0
Last-fired 41.3 ±17.2 57.3 ±15.9 47.5 ±16.1 87.4 ±12.4
GRU Binary 18.7 ±8.3 33.2 ±12.7 23.9 ±9.6 46.7 ±23.4
Change-point 31.2 ±8.3 47. ±10.9 31.2 ±8.5 71.6 ±12.6
Last-fired 40.4 ±16.5 52.7 ±16.4 45.4 ±16.9 86.6 ±12.3
SVM Binary 19.4 ±9.0 35.2 ±12.7 24.0 ±9.2 37.4 ±19.0
Change-point 25.6 ±6.2 51.4 ±9.5 34.0 ±7.2 57.8 ±15.5
Last-fired 37.0 ±7.9 55.5 ±11.6 44.1 ±8.5 87.5 ±12.1
representation. The results are slightly better with change-point but clearly better when applying the last-fired rep-
resentation. RNNs again give the highest recall values for all representations. Overall, the results show that RNNs
perform better than HMM, NB and HSMM in all cases, while CRF is slightly better than RNNs. But in terms of
recall, these later outperform all methods for all feature representations. The reason behind this is that RNNs perform
better for imbalanced data compared to CRF. RNNs variants generally perform equally well.
For abnormal activity detection, we considered LSTM only and compared against NB, HSMM, HMM, SVM and
CRF. TPR and FPR accuracy percentages are correspondingly; 40.40% and 43.50% for NB, 58.36% and 96.20% for
HMM, 68.85% and 32.2% for HSMM, 66.22% and 40.50% for CRF, 72.11% and 44.0% for One-class SVM and
91.43% and 40.96% for LSTM. We used only last-fired feature in this experiment. The results indicate that LSTM
is the best to prune false negatives compared to the other methods. Methods like NB, One-class SVM which do not
capture the data order performs the worst. The models ignore the frequency of the activity, but apply the temporal and
contextual information to make a decision. Results show that LSTM is capable of encoding the order of activities.
Hence, when an activity is introduced in a different context or in a different order, LSTM can detect such anomalies.
Our current approach may fail to detect abnormalities, when there is gradual deterioration regarding the health of
an elderly. We are planning to deal with this issue in the future while collecting real-world data in which gradualdeterioration can be observed.
5. Conclusion
In this paper, we showed that RNNs perform well on the problem of activity recognition. They are also able
to cope quite well with imbalanced data as well as anomaly detection which is very important in the context of
dementia. Compared to a number of traditional and popular techniques used for activity recognition such as SVM,
NB, HMM and HSMM, they perform much better, while remained very competitive with CRF. Furthermore, theempirical experiments showed that the three variants of RNNs generally perform equally well, but LSTM seems to beslightly better across all datasets used in this study. Moreover, in terms of representation, there is no clear preference,
butlast-fired feature seems to be better, at least on the datasets AandC, compared to the change-point andbinary
representations. Overall the study allowed to confirm that RNNs are very appropriate for activity recognition and
abnormal activity detection. In our future investigations, we will extend RNNs to deep neural networks. We will also
aim at collecting a dataset from a smart home dedicated to elderly people with dementia to further study behaviouranomalies related to dementia.8 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
Acknowledgments
We thank to Tim Van Kasteren for providing the code and datasets.
References
1.D. A. Umphred, R. T. Lazaro, M. Roller, and G. Burton. Neurological rehabilitation. In Elsevier Health Sciences, vol. 27, no. 5, 2013.
2.M. S. Albert et al. The diagnosis of mild cognitive impairment due to Alzheimer’s disease: Recommendations from the national institute
on aging-alzheimer’s association workgroups on diagnostic guidelines for alzheimer’s disease. Alzheimer’s &Dementia, Volume 7-Issue
3:270–279, 2011.
3.W. Thies and L. Bleiler. 2013 Alzheimer’s disease facts and figures. Alzheimer’s &dementia: the journal of the Alzheimer’s Association vol.
9, no. 2, page 208245, 2013.
4.M. R. Hodges, K. Kirsch, M. W. Newman, and M. E. Pollack. Automatic assessment of cognitive impairment through electronic observation
of object usage. In Pervasive, pages 192–209, 2010.
5.K. Wild. Aging changes. In Geraotechnology, Vol. 9 No 2, pages 121–125, 2010.
6.P. Dawadi, D. Cook, and M. Schmitter-Edgecombe. Smart home-based longitudinal functional assessment. In ACM UbiComp Workshop on
Smart Health Systems and Applications, 2014.
7.D. Riboni, C. Bettini, G. Civitarese, Z. Haider Janjua, and R. Helaoui. Fine-grained recognition of abnormal behaviors for early detection of
mild cognitive impairment. IEEE International Conference on Pervasive Computing and Communications, 2015.
8.O. D. Lara and M. A. Labrador. A mobile platform for real time human activity recognition. IEEE Conference on Consumer Communications
and Networks, 2012.
9.T. Kirste, A. Hoffmeyer, P. Koldrack, A. Bauer, S. Schubert, S. Schrder, and S. Teipel. Detecting the effect of Alzheimer’s disease on everyday
motion behaviour. Journal of Alzheimer’s Disease, Journal of Alzheimer’s Disease:121–132, 2014.
10.M. Ermes, J. Parkka, and L. Cluitmans. Advancing from offline to online activity recognition with wearable sensors. 30thAnnual International
Conference of the IEEE Engineering in Medicine and Biology Society, page 44514454, 2008.
11.O. D. Lara and M. A. Labrador. A survey on human activity recognition using wearable sensors. IEEE Communications Surveys Tutorials,
15(3):1192–1209, 2013.
12.J. Yang, M. Nguyen, P. San, X. Li Li, and S. Krishnaswamy. Deep convolutional neural networks on multichannel time series for human
activity recognition. pages 3995–4001, 2015.
13.M. Zeng, L. T. Nguyen, B. Yu, O. J. Mengshoel, J. Zhu, P. Wu, and J. Zhang. Convolutional neural networks for human activity recognition
using mobile sensors. In 6thInternational Conference on Mobile Computing, Applications and Services, pages 197–205, 2014.
14.F. J. Ordonez and D. Roggen. Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition.
Sensors, 16(1):115, 2016.
15.N. Hammerla, S. Halloran, and T. Pltz. Deep, convolutional, and recurrent models for human activity recognition using wearables. Proceed-
ings of the 25thInternational Joint Conference on Artificial Intelligence, 2016.
16.S. Choi, E. Kim, and S. Oh. Human behavior prediction for smart homes using deep learning. In 2013 IEEE RO-MAN, pages 173–179, 2013.
17.T. Pl ¨otz, N. Hammerla, and P. Olivier. Feature learning for activity recognition in ubiquitous computing. volume 2, pages 1729–1734, 2011.
18.N. Hammerla, J. Fisher, P. Andras, L. Rochester, R. Walker, and T. Plotz. PD disease state assessment in naturalistic environments using deep
learning. pages 1742–1748, 2015.
19.H. Fang and C. Hu. Recognizing human activity in smart home using deep learning algorithm. In 33rdChinese Control Conference, pages
4716–4720, 2014.
20.H. Fang, H. Si, and L. Chen. Recurrent neural network for human activity recognition in smart home. In Proceedings of 2013 Chinese
Intelligent Automation Conference, pages 341–348, 2013.
21.Charissa A. and Sung-Bae C. Evaluation of deep convolutional neural network architectures for human activity recognition with smartphone
sensors. In Proceedings of the KIISE Korea Computer Congress, page 858860, 2015.
22.A. Lotfi, Caroline Langensiepen, Sawsan M Mahmoud, and M Javad Akhlaghinia. Smart homes for the elderly dementia sufferers: identifi-
cation and prediction of abnormal behaviour. Journal of ambient intelligence and humanized computing, 3(3):205–218, 2012.
23.T. Van Kasteren, G. Englebienne, and B. J. A. Kr ¨ose. Human activity recognition from wireless sensor network data: Benchmark and
software. Activity Recognition in Pervasive Intelligent Environments, pages 165–186, 2011.
24.P. Dawadi, D. Cook, C. Parsey, M. Schmitter-Edgecombe, and M. Schneider. An approach to cognitive assessment in smart homes. In
Knowledge Discovery and Data Mining Workshop on Medicine and Healthcare, 2011.
25.D. Riboni, C. Bettini, G. Civitarese, Z. H. Janjua, and V. Bulgari. From lab to life: Fine-grained behavior monitoring in the elderly’s home.
In Proceedings of the 2015 IEEE International Conference on Pervasive Computing and Communications Workshops, pages 344–349, 2015.
26.J. Saives, C. Pianon, and G. Faraut. Activity discovery and detection of behavioral deviations of an inhabitant from binary sensors. IEEE
Transactions on Automation Science and Engineering, 12(4):1211–1224, 2015.
27.M. Amiribesheli and A. Bouchachia. Smart homes design for people with dementia. In 2015 International Conference on Intelligent
Environments, pages 156–159, 2015.
28.K. Cho, B. Merrienboer, D. Bahdanau, and Y. Bengio. On the properties of neural machine translation: Encoder-decoder approaches. 2014.
29.S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, Nov 1997.
30.C. Franc ¸ois. Keras. https://github.com/fchollet/keras, 2015.
31.Theano Development Team. Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints, abs/1605.02688,
may 2016.

Damla Arifoglu et al. / Procedia Computer Science 110 (2017) 86–93 93
Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000 7
Table 3. Activity recognition results on dataset C
Model Feature Precision Recall F-Measure Accuracy
NB Binary 19 .6±11.4 16.8 ±7.5 17.8 ±9.1 46.5 ±22.6
Change-point 39.9 ±6.9 30.8 ±4.8 34.5 ±4.6 57.6 ±15.4
Last-fired 40.5 ±7.4 46.4 ±14.8 42.3 ±6.8 87.0 ±12.2
HMM Binary 15.2 ±9.2 17.2 ±9.3 15.7 ±8.8 26.5 ±22.7
Change-point 41.4 ±8.8 50.0 ±11.4 44.9 ±8.8 77.2 ±14.6
Last-fired 40.7 ±9.7 53.7 ±16.2 45.9 ±11.2 83.9 ±13.9
HSMM Binary 15.6 ±9.2 20.4 ±10.9 17.3 ±9.6 31.2 ±24.6
Change-point 43.8 ±10.0 52.3 ±12.8 47.4 ±10.5 77.5 ±15.3
Last-fired 42.5 ±10.8 56.0 ±15.4 47.9 ±11.3 84.5 ±13.2
CRF Binary 17.8 ±22.1 21.8 ±20.9 19.0 ±21.8 46.3 ±25.5‘
Change-point 36.7 ±18.0 39.6 ±17.4 38.0 ±17.6 82.2 ±13.9
Last-fired 37.7 ±17.1 40.4 ±16.0 38.9 ±16.5 89 .7±8.4
Vanilla Binary 15 .4±5.3 43.1 ±18.1 22.2 ±7.3 50.2 ±22.4
Change-point 31.3 ±7.1 54.9 ±11.3 39.5 ±8.3 72.2 ±13.0
Last-fired 38.3 ±16.3 59.6 ±15.1 45.8 ±14.8 86.7 ±12.5
LSTM Binary 16.8 ±6.2 34.8 ±12.5 22.1 ±7.4 45.3 ±21.2
Change-point 31.0 ±5.1 53.3 ±6.5 38.9 ±5.0 72.0 ±13.0
Last-fired 41.3 ±17.2 57.3 ±15.9 47.5 ±16.1 87.4 ±12.4
GRU Binary 18.7 ±8.3 33.2 ±12.7 23.9 ±9.6 46.7 ±23.4
Change-point 31.2 ±8.3 47. ±10.9 31.2 ±8.5 71.6 ±12.6
Last-fired 40.4 ±16.5 52.7 ±16.4 45.4 ±16.9 86.6 ±12.3
SVM Binary 19.4 ±9.0 35.2 ±12.7 24.0 ±9.2 37.4 ±19.0
Change-point 25.6 ±6.2 51.4 ±9.5 34.0 ±7.2 57.8 ±15.5
Last-fired 37.0 ±7.9 55.5 ±11.6 44.1 ±8.5 87.5 ±12.1
representation. The results are slightly better with change-point but clearly better when applying the last-fired rep-
resentation. RNNs again give the highest recall values for all representations. Overall, the results show that RNNs
perform better than HMM, NB and HSMM in all cases, while CRF is slightly better than RNNs. But in terms of
recall, these later outperform all methods for all feature representations. The reason behind this is that RNNs perform
better for imbalanced data compared to CRF. RNNs variants generally perform equally well.
For abnormal activity detection, we considered LSTM only and compared against NB, HSMM, HMM, SVM and
CRF. TPR and FPR accuracy percentages are correspondingly; 40.40% and 43.50% for NB, 58.36% and 96.20% for
HMM, 68.85% and 32.2% for HSMM, 66.22% and 40.50% for CRF, 72.11% and 44.0% for One-class SVM and
91.43% and 40.96% for LSTM. We used only last-fired feature in this experiment. The results indicate that LSTM
is the best to prune false negatives compared to the other methods. Methods like NB, One-class SVM which do not
capture the data order performs the worst. The models ignore the frequency of the activity, but apply the temporal and
contextual information to make a decision. Results show that LSTM is capable of encoding the order of activities.
Hence, when an activity is introduced in a different context or in a different order, LSTM can detect such anomalies.
Our current approach may fail to detect abnormalities, when there is gradual deterioration regarding the health of
an elderly. We are planning to deal with this issue in the future while collecting real-world data in which gradualdeterioration can be observed.
5. Conclusion
In this paper, we showed that RNNs perform well on the problem of activity recognition. They are also able
to cope quite well with imbalanced data as well as anomaly detection which is very important in the context of
dementia. Compared to a number of traditional and popular techniques used for activity recognition such as SVM,
NB, HMM and HSMM, they perform much better, while remained very competitive with CRF. Furthermore, theempirical experiments showed that the three variants of RNNs generally perform equally well, but LSTM seems to beslightly better across all datasets used in this study. Moreover, in terms of representation, there is no clear preference,
butlast-fired feature seems to be better, at least on the datasets AandC, compared to the change-point andbinary
representations. Overall the study allowed to confirm that RNNs are very appropriate for activity recognition and
abnormal activity detection. In our future investigations, we will extend RNNs to deep neural networks. We will also
aim at collecting a dataset from a smart home dedicated to elderly people with dementia to further study behaviouranomalies related to dementia.8 Damla Arifoglu and Abdelhamid Bouchachia /00 (2016) 000–000
Acknowledgments
We thank to Tim Van Kasteren for providing the code and datasets.
References
1.D. A. Umphred, R. T. Lazaro, M. Roller, and G. Burton. Neurological rehabilitation. In Elsevier Health Sciences, vol. 27, no. 5, 2013.
2.M. S. Albert et al. The diagnosis of mild cognitive impairment due to Alzheimer’s disease: Recommendations from the national institute
on aging-alzheimer’s association workgroups on diagnostic guidelines for alzheimer’s disease. Alzheimer’s &Dementia, Volume 7-Issue
3:270–279, 2011.
3.W. Thies and L. Bleiler. 2013 Alzheimer’s disease facts and figures. Alzheimer’s &dementia: the journal of the Alzheimer’s Association vol.
9, no. 2, page 208245, 2013.
4.M. R. Hodges, K. Kirsch, M. W. Newman, and M. E. Pollack. Automatic assessment of cognitive impairment through electronic observation
of object usage. In Pervasive, pages 192–209, 2010.
5.K. Wild. Aging changes. In Geraotechnology, Vol. 9 No 2, pages 121–125, 2010.
6.P. Dawadi, D. Cook, and M. Schmitter-Edgecombe. Smart home-based longitudinal functional assessment. In ACM UbiComp Workshop on
Smart Health Systems and Applications, 2014.
7.D. Riboni, C. Bettini, G. Civitarese, Z. Haider Janjua, and R. Helaoui. Fine-grained recognition of abnormal behaviors for early detection of
mild cognitive impairment. IEEE International Conference on Pervasive Computing and Communications, 2015.
8.O. D. Lara and M. A. Labrador. A mobile platform for real time human activity recognition. IEEE Conference on Consumer Communications
and Networks, 2012.
9.T. Kirste, A. Hoffmeyer, P. Koldrack, A. Bauer, S. Schubert, S. Schrder, and S. Teipel. Detecting the effect of Alzheimer’s disease on everyday
motion behaviour. Journal of Alzheimer’s Disease, Journal of Alzheimer’s Disease:121–132, 2014.
10.M. Ermes, J. Parkka, and L. Cluitmans. Advancing from offline to online activity recognition with wearable sensors. 30thAnnual International
Conference of the IEEE Engineering in Medicine and Biology Society, page 44514454, 2008.
11.O. D. Lara and M. A. Labrador. A survey on human activity recognition using wearable sensors. IEEE Communications Surveys Tutorials,
15(3):1192–1209, 2013.
12.J. Yang, M. Nguyen, P. San, X. Li Li, and S. Krishnaswamy. Deep convolutional neural networks on multichannel time series for human
activity recognition. pages 3995–4001, 2015.
13.M. Zeng, L. T. Nguyen, B. Yu, O. J. Mengshoel, J. Zhu, P. Wu, and J. Zhang. Convolutional neural networks for human activity recognition
using mobile sensors. In 6thInternational Conference on Mobile Computing, Applications and Services, pages 197–205, 2014.
14.F. J. Ordonez and D. Roggen. Deep convolutional and LSTM recurrent neural networks for multimodal wearable activity recognition.
Sensors, 16(1):115, 2016.
15.N. Hammerla, S. Halloran, and T. Pltz. Deep, convolutional, and recurrent models for human activity recognition using wearables. Proceed-
ings of the 25thInternational Joint Conference on Artificial Intelligence, 2016.
16.S. Choi, E. Kim, and S. Oh. Human behavior prediction for smart homes using deep learning. In 2013 IEEE RO-MAN, pages 173–179, 2013.
17.T. Pl ¨otz, N. Hammerla, and P. Olivier. Feature learning for activity recognition in ubiquitous computing. volume 2, pages 1729–1734, 2011.
18.N. Hammerla, J. Fisher, P. Andras, L. Rochester, R. Walker, and T. Plotz. PD disease state assessment in naturalistic environments using deep
learning. pages 1742–1748, 2015.
19.H. Fang and C. Hu. Recognizing human activity in smart home using deep learning algorithm. In 33rdChinese Control Conference , pages
4716–4720, 2014.
20.H. Fang, H. Si, and L. Chen. Recurrent neural network for human activity recognition in smart home. In Proceedings of 2013 Chinese
Intelligent Automation Conference, pages 341–348, 2013.
21.Charissa A. and Sung-Bae C. Evaluation of deep convolutional neural network architectures for human activity recognition with smartphone
sensors. In Proceedings of the KIISE Korea Computer Congress, page 858860, 2015.
22.A. Lotfi, Caroline Langensiepen, Sawsan M Mahmoud, and M Javad Akhlaghinia. Smart homes for the elderly dementia sufferers: identifi-
cation and prediction of abnormal behaviour. Journal of ambient intelligence and humanized computing, 3(3):205–218, 2012.
23.T. Van Kasteren, G. Englebienne, and B. J. A. Kr ¨ose. Human activity recognition from wireless sensor network data: Benchmark and
software. Activity Recognition in Pervasive Intelligent Environments, pages 165–186, 2011.
24.P. Dawadi, D. Cook, C. Parsey, M. Schmitter-Edgecombe, and M. Schneider. An approach to cognitive assessment in smart homes. In
Knowledge Discovery and Data Mining Workshop on Medicine and Healthcare, 2011.
25.D. Riboni, C. Bettini, G. Civitarese, Z. H. Janjua, and V. Bulgari. From lab to life: Fine-grained behavior monitoring in the elderly’s home.
In Proceedings of the 2015 IEEE International Conference on Pervasive Computing and Communications Workshops, pages 344–349, 2015.
26.J. Saives, C. Pianon, and G. Faraut. Activity discovery and detection of behavioral deviations of an inhabitant from binary sensors. IEEE
Transactions on Automation Science and Engineering, 12(4):1211–1224, 2015.
27.M. Amiribesheli and A. Bouchachia. Smart homes design for people with dementia. In 2015 International Conference on Intelligent
Environments, pages 156–159, 2015.
28.K. Cho, B. Merrienboer, D. Bahdanau, and Y. Bengio. On the properties of neural machine translation: Encoder-decoder approaches. 2014.
29.S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, Nov 1997.
30.C. Franc ¸ois. Keras. https://github.com/fchollet/keras, 2015.
31.Theano Development Team. Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints, abs/1605.02688,
may 2016.