Evaluation of Wellness Detection Techniques using Complex Activities Association for Smart Home Ambient

Wireless Sensor Network based smart homes have the potential to meet the growing challenges of independent living of elderly people in smart homes. However, wellness detection of elderly people in smart homes is still a challenging research domain. Many researchers have proposed several techniques; however, majority of these techniques does not provide a comprehensive solution because complex activities cannot be determined easily and comprehensive wellness is difficult to diagnose. In this study’s critical review, it has been observed that strong association lies among the vital wellness determination parameters. In this paper, an association rules based model is proposed for the simple and complex (overlapped) activities recognition and comprehensive wellness detection mechanism after analyzing existing techniques. It considers vital wellness detection parameters (temporal association of sub activity location and sub activity, time gaps between two adjacent activities, temporal association of inter and intra activities). Activity recognition and wellness detection will be performed on the basis of extracted temporal association rules and expert knowledgebase. Learning component is an important module of our proposed model to accommodate the changing trends in the frequent pattern behavior of an elderly person and recommend a caregiver/expert to adjust the expert knowledgebase according to the found abnormalities. Keywords—Wellness detection; Elderly people; WSN Smart homes; Activity recognition


INTRODUCTION
World population is increasing and according to the World Health Organization (WHO) elderly people population will also increase drastically in the near future [1].Independent living is a concept that defines a person can live alone without any assistance of any other human being (care giver).Automated mechanisms for the monitoring of an individual are required to ensure the wellness of the elderly person.This monitoring should be accurate, error free and automated (without any human involvement) for an independent livening ambient.
Elderly people have health related issues with the age such as limitations in physical functions, different diseases such as Alzheimer"s disease, diabetes, cardio vascular disease.Smart homes technologies have potential to meet the emerging challenges of elderly people independent livings with enhance quality of life.
Wireless sensor network is a vital component of smart home for monitoring of the elderly persons.Variety of sensors deployed all over the house for monitoring purpose such as in electric appliances, gadgets and objects of daily usage for the monitoring of elderly persons as reported in [2].
Activity recognition is one of the active areas of research domain.The aim is to recognize the actions an individual takes to conduct an act of daily living.Most of the researcher have divided these actions into a set of defined activities and termed them "Activities of daily living".This concept was originally proposed in 1950 [3].There are broadly three categories of activities.These categories are concurrent activities, interleaved and simple atomic activities as reported in [4].However, most of the literature does not consider all type of activities for their activity recognition and wellness detection of elderly people in smart homes ambient.
Frequent pattern of an individual exists while conducting a specific activity, resulting in sequential frequent pattern as reported [1,5,6].There is a strong association among subactivities and locations where the sub activity is being performed, time duration taken to perform a specific sub activity and specific day of time [7][8][9].Wellness detection is determined, when an individual deviates from its sequential frequent pattern among sub activities, individual is not conducting a specific activity on a specific location.
Accurate wellness detection is a challenging research area, several researchers have employed artificial intelligence and machine learning techniques (such as artificial neural networks, support vector machines, Naïve Bayes) as a classifier for the determination of wellness of elderly people in [9,10].Few researchers have also proposed fuzzy rules based classifiers for the activity detection and wellness determination of elderly people in smart home environments [11].In probabilistic approaches, Markov and Hidden Markov models, where each sub activity is considered as a component of main activity for the activity recognition.Human activities pattern rely on temporal sequence as reported in [10,12,13].Similarly data mining techniques also attracted many researchers to explore for the activity recognition and wellness detection, such as classification techniques, association rule mining techniques and clustering techniques, as reported in [1,5,10].However, wellness detection is a complex research problem; many researchers proposed hybrid models to achieve the said www.ijacsa.thesai.orgobjective.There are few hierarchal models proposed for wellness and activity detection [6,14].
In this paper, we have proposed an association rules based model with the following contribution:  Temporal association of sub activity and their location  Time gaps between two adjacent activities consideration  Temporal association of inter and intra activities  Learning module  Recommendations to the caregiver or expert Rest of the paper is organized as follow: Section II summarizes literature review and highlighting strength and weakness of existing technique by critically analyzing them in tabular formats.In section III a new model is proposed to address the limitation of existing techniques.Finally, paper is concluded and future work is presented in section IV.

II. RELATED WORK
Wellness detection is a complex and challenging research domain.Majority of proposed models does not provide a comprehensive solution from different perceptive of wellness and activity detection.J. Wen., et al [15], proposed a model based on weighted frequent pattern (association mining).The objectives of this model is to find the association rules of activities performed by the elderly person, and build a classifier which is based on these extracted association rules.However, time series analysis of activities with the location of activities is not considered.M.T Moutacalli., et al [5], have employed frequent pattern mining for the determination of activity recognition and wellness detection.In proposed model, association mining is performed for extraction of homogenous frequent pattern activity model.The strength of this work is that this model considers event duration and time gap between adjacent activities.However, this model cannot recognize complex activities.S.T.Bourobou., et al [10] proposed a hybrid model based of association mining and AAN.The proposed model consists of two steps of repeated activity discovery by using FP growth algorithm and clustered these patterns into activities using K-pattern algorithm.ANN based classifier is used for the classification of the activities.However, integration of multiple techniques into a single model makes the whole process difficult and complex to implement.L.G. Fahad., et al [6] proposed a hybrid hierarchal model based on frequent patterns and probabilistic approach to find the activity tasks.Frequent pattern mining is being conducted to extract the inhabitant behavior pattern.However, it is a tedious task to create routine tasks lists manually.This model does not follow temporal analysis for wellness determination.S.Nasreen., et al [16], a hybrid model is proposed based on Association rules for inhabitant pattern.These association rules are used for the assignment of activity category when new activity arrives as an input.Binary support vector machine is used to detect correct or incorrect assignment.However, inter and intra relationship of activities has not been considered.L.G. J. Saives., et al [17] have proposed a hybrid model based on association rule mining and extended finite automata for the mapping of activities found from the association rule mining.However, prior knowledge is required and temporal deviations are not considered for wellness detection.Similarly repeated activities might not represent the same sequence of events.A. Forkan., et al [11] proposed a probabilistic model to detect the abnormality in the activity build model, to identify the wellness of an elderly person.Hidden Markov Model (HMM) is used for the change detection in the activities of daily living along with identification of change in daily routine using statistical history and expert knowledgebase for abnormality detection.This model does not follow single wellness detection criteria which is the major contribution of this study.The outcome of each module is then serves as an input to fuzzy rules based classifier for the classification of activities and detection of abnormality.However, this model might have less acceptability for elderly as wearable sensors must be used for expert knowledgebase.Similarly, different modules working together making the process complex and prior knowledge is required as a statistical histories for the prediction and abnormality detection.Similarly N.K. Suryadevara [8], proposed another probabilistic model is for the prediction wellness in the smart home environment based on appliance usage.Two wellness functions been used.Irregular patterns will lead to the abnormality which helps to identify whether an elderly person is well or unwell.However, activities are presented in simple manner and appliance usage monitoring might not be suitable when the person is not using any specific appliance.F.Ordonez., et al [2], proposed a hybrid model which is based on training phase and classification phase.For temporal and sequential activities analysis Hidden Markov model is used.The objective of this study is to validate hybrid model performs better than single technique based model.It used Hidden Markov Model for training and for classification phase SVM reported better results.However, while validating the results for temporal and sequential activities, other sequential and association mining techniques were not considered.L.Kalra., et al [13], proposed two stage probabilistic model based on Markov model.The proposed model process the sensor data in two stages, in the first stage sensor data is analyzed for the patterns of activity events by using Markov Model.However, time gap between adjacent activities are not considered for the determination of wellness of elderly person.K. Gayathri., et al [14] proposed a probabilistic hierarchal model for abnormality detection to determine the wellness.This model combines the data and knowledge driven parameters by employing Markov logic networks for abnormality detection.This model considers location, object usage, time of usage for the abnormality detection.However, manual rules based systems requires effort to create knowledgebase.Sensor and data relationship is ignored.E. Kim., et al [4] have discussed the uncertainty parameters in the wellness determination.This study serves as groundwork to develop diagnostic system.In [7], tree based regression algorithm is being proposed for the classification of activities in the training phase and in the later phase when new activity can be predicted as future activity.However, this model does not consider complex activities.www.ijacsa.thesai.orgE. Nazerfard., et al [9] proposed an activity classification model based on Bayesian network.This proposed model proposes two steps of inference for the prediction of upcoming activity and its label.This model considers time analysis and activity features.However, complex activities are not considered for the prediction.X. Hong., et al [18] proposed model based on time segmentation.The model extracts events that are associated with a complete activity.The first segmentation is location based, second segmentation is model-based, and third is based on modeling of each activity in terms of sensor activation to determine.The third proposed algorithm identifies the most predominate sensor in the activity to identify the event which that sensor exhibits.However in this approach, rules need to be defined manually for the activity discovery.

A. Critical Evaluation
In this sub section of literature review, critical evaluation is being conducted.Fifteen papers are represented in three different tables for the cumulative review of each literature under the set of predefined vital parameters.
Table.I represents the methodology of the proposed techniques for wellness and activity detection for smart home.Majority of the existing techniques aimed to extract elderly persons frequent behavior patterns, whether they have employed Hidden Markov model [2, 9, 10], [7] tree based regression analysis association rules or frequent pattern mining [5,6,15,16] and deviation from this pattern leads to the abnormality.A proven technique should be employed for the extraction of frequent patterns of an elderly person, like frequent pattern or association mining.

Table.II represents the implementation details of existing
techniques which helps to explain the implementation details, environment and validation.Majority of the proposed models have not validated and compared their results with the existing models and methods.
Table.III represents the vital parameters which are necessary for the wellness detection.These parameters have strong impact of elderly person"s activities and wellness detection.In table.III, type of the activities which different proposed models have used in activity detection, similarly the association of intra and inter activities is necessary to understand pattern of elderly person behaviors.Time gap between adjacent activities is important criteria to measure the wellness of elderly person.Association of time with location is a necessary parameter to understand pattern where an individual will perform a specific activity.Temporal analysis is important criteria to understand the elderly person is also one of the most vital criteria to understand elderly person activity pattern.

III. PROPOSED MODEL
In this section, a model is proposed for the wellness detection based on association rules mining and frequent patterns of the behavior of the elderly people.From the critical evaluation section it is evident that association exists among different parameters of activity recognition and wellness detection.Association rules between sub activities and location of sub activities have been extracted from the sensor data, correlation among sub activities for as inter temporal association among sub activities rules are extracted and time gap between adjacent activities have proposed to be found to determine the wellness of elderly people.Model is proposed by considering all those parameter which were mainly ignored in existing literature such as detection of overlapped activities; location based association, intra and inters activities associations.
The proposed model consists of 9 modules, namely, data transformation module, sub activity (location and time) correlation module, time gap between two adjacent activities module, sub activity (sequence of sub activities and time) correlation module, sub activity (location and sequence of activities) correlation w.r.t time module, expert knowledgebase module, contextual database module, learning component and activity recognition module as shown in the Fig. 1.

A. Data Transformation module
In smart home data is generated from the WSN"s network.To make the sensor data suitable for processing, sensor data is transformed with labels (for example, if a person is sleeping, bed sensor is active, the activity will be marked as "sleeping", time duration of this active sensor will be the duration of the activity and location of the sensor will be the location of activity).For example, an elderly person of a smart home wants to make coffee, he/she will achieve the said task in a series of sub activities for example: 1. Staring of stove, 2. Filling the pan with water, 3. Opening coffee jar and 4. Mixing coffee & water etc.These steps are considered as sub activities and while making of coffee are considered as an activity.Similarly it is also important to consider time gap between two adjacent sub activities for the determination of wellness of an elderly person.

B. Temporal Association of sub activity's location and sub activities
For activity recognition and wellness detection it is important to determine and extract the association of each sub activity time duration with the location where that sub activity has been performed along with the day of time.Most frequent association rules are extracted by using FP-Growth.For example, the normal pattern of an individual sleeping routine at around 2200hrs in "bedroom" for approximately 8 hours.If a person is in "dining room" at around 2200hrs and duration of the activity is more than it should spend in the "dining room" so the violation of rule will indicate that person is not well.

C. Sub activity (sequence of sub activities and time) correlation module:
For activity recognition the correlation of sequence of sub activities with their time duration is required to handle complex activities.As discussed earlier an activity consists of sequence of sub activities.An example is discussed earlier of coffee making.The correlation among the sub activities and time in 24hrs is determined by using Generalized Sequential Pattern mining (GSP) [19].The outcome of this process will be association rules (correlation among sub activities and time).www.ijacsa.thesai.org

D. Time gap determination of two adjacent activities over a specific time:
The input of this process is sub activities labels and time duration to perform each sub activity to determine the time gap between two adjacent sub activities over a specific day of time.The benefit of this outcome is the determination of the wellness of an elderly person most of the literature surveyed has ignored this important wellness parameter.For example, a healthy person will perform sub activities in normal routine with a normal time gap between two adjacent activities, while an unhealthy elderly person might not execute the sub activities with the normal time gap.

E. Sub activity (location, time gap and sequence of activities) correlation w.r.t time module:
This module servers as global perceptive of activity, in this module activities rules are formulated from vital parameters of complete activity detection rules.In this module association rules of (location + time duration), association rules of sub activities and time gap between adjacent sub activities with respect to time of day respectively are correlated using association rule mining (using FP-Growth) to extract pattern of these vital parameter over a specific time of day.The outcome of this module will be association rules of the required parameters (location, time gap and sequence of sub activities) over a specific time of day to accurately determine the wellness of an elderly person.For example, a healthy person has a routine to wake up (S1) around (0700 -0715hrs).

(S1, S2, S3) ^ (L1, L2, L3) ^ (G1, G2, G3) ^ T Activity
For overlapped activities, a set of rules have same location and temporal association (location of sub activities and time of day) but different sequence of sub activities and different time duration between adjacent activities.

F. Expert (medical) knowledgebase:
This module will have hard rules for an individual from medical history perspective.This rules are the recommendation for an individual like walk time, sleep time (recommended).

G. Database module
In this module association rules which are extracted from above mentioned model will be saved in this module.Expert (medical) knowledgebase rules will also be saved.This module will serve as a rules repository for the online activity recognition module.

H. Learning module:
The learning module updates the existing set of rules by monitoring the sensor data streams and extracting the contextual spatial-temporal association of the changing trends in the behavior pattern of an elderly person.The learning component also monitors of the outcome of the online activity recognition and wellness detection module, on correct diagnoses the weight of the rule will increase and will have high priority.Similarly to accommodate the changing trends in the behavior pattern, rules will be updated in the database repository.
The second novel function of the learning module is to recommend the caregiver the abnormalities found in the inhabitant behavior to make necessary changes in the expert knowledgebase according to the recommendation.

I. Online activity recognition and wellness detection module:
In this module have the association rules generated and the hard rules of expert knowledge base will be mapped with sensor data.Data from the sensor network is coming to this module and data is being compared with knowledge base and association rules.If the rules are met for the input data the activities are considered normal and the elderly person is considered well.Incase rules are violated the elderly person is considered unwell.www.ijacsa.thesai.orgThis study includes literature survey and critical evaluation of existing techniques for wellness detection of elderly people in WSN based smart homes.There are certain vital parameters such as sub activity, location of sub activity, time gap between adjacent sub activities and their temporal associations are identified for the activity detection and wellness detection.Similarly existing literature used different methods for the identification of frequent patterns of elderly people behaviors and finding correlation and association among different parameters.On the basis of the critical evaluation, a model is proposed which is based on association rule mining.The proposed model is based on temporal association of location and sub activities, time gap and temporal association among sequence of sub activities been extracted for the determination of complex activities such as overlapped activities.Association rules and expert knowledge base classifier is proposed for the wellness detection of online data stream from sensor data of smart homes.Learning component is proposed to accommodate the changing trends in the activities pattern of the behavior of the elderly people.In future proposed model will be implemented and tested for 50 elderly people.Furthermore, result will be compared to existing techniques to evaluate its performance.

TABLE III .
CRITICAL EVALUATION OF RELATED WORK BASED ON WELLNESS AND ACTIVITY RECOGNITION PARAMETERS