IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO.

1, JANUARY 2014

375

A Predictive Model for Assistive Technology Adoption for People With Dementia
Shuai Zhang, Sally I. McClean, Member, IEEE, Chris D. Nugent, Member, IEEE, Mark P. Donnelly, Leo Galway, Bryan W. Scotney, and Ian Cleland

AbstractAssistive technology has the potential to enhance the level of independence of people with dementia, thereby increasing the possibility of supporting home-based care. In general, people with dementia are reluctant to change; therefore, it is important that suitable assistive technologies are selected for them. Consequently, the development of predictive models that are able to determine a persons potential to adopt a particular technology is desirable. In this paper, a predictive adoption model for a mobile phone-based video streaming system, developed for people with dementia, is presented. Taking into consideration characteristics related to a persons ability, living arrangements, and preferences, this paper discusses the development of predictive models, which were based on a number of carefully selected data mining algorithms for classication. For each, the learning on different relevant features for technology adoption has been tested, in conjunction with handling the imbalance of available data for output classes. Given our focus on providing predictive tools that could be used and interpreted by healthcare professionals, models with ease-of-use, intuitive understanding, and clear decision making processes are preferred. Predictive models have, therefore, been evaluated on a multi-criterion basis: in terms of their prediction performance, robustness, bias with regard to two types of errors and usability. Overall, the model derived from incorporating a k Nearest-Neighbour algorithm using seven features was found to be the optimal classier of assistive technology adoption for people with dementia (prediction accuracy 0.84 0.0242). Index TermsAssistive technology, classication, dementia, prediction models, technology adoption.

I. INTRODUCTION

SSISTIVE technology has the potential to enhance the level of independence for people with dementia (PwD) by providing targeted support with their activities of daily living (ADLs). A popular approach is the use of reminding technologies, which issue messages to PwD in order to remind them to

Manuscript received February 4, 2013; revised April 10, 2013; accepted May 31, 2013. Date of publication June 10, 2013; date of current version December 31, 2013. This work was supported in part by f the Engineering and Physical Sciences Research Council through the MATCH programme under Grant EP/F063822/1 and Grant EP/G012393/1, and also in part supported by the Alzheimers Association under Grant ETAC-12242841. S. Zhang, C. D. Nugent, M. P. Donnelly, L. Galway, and I. Cleland are with the School of Computing and Mathematics, University of Ulster at Jordanstown, Antrim BT37 0QB, U.K. (e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]). S. I. McClean, and B. W. Scotney are with the School of Computing and Information Engineering, University of Ulster at Coleraine, Londonderry BT52 1SA, U.K. (e-mail: [email protected]; [email protected]). Digital Object Identier 10.1109/JBHI.2013.2267549

perform ADLs at appropriate times. Acceptance and successful use of such assistive technologies increases the possibility for PwD to remain in their own homes for longer, resulting in reduced caregiver burden, lower care costs, andpossibly improved the quality of life for the PwD. Nevertheless, a change of routine is generally a challenge to PwD and their caregivers [1]. In addition, PwD are often afraid of making mistakes, including misuse of technology, or are simply unable to use it given their limited technological experience. It is, therefore, important to identify the most suitable technology for the PwDcaregiver dyad. Two key elements in achieving this are 1) being able to identify the relevant features and 2) discovering the relationship between those features and the likelihood of adoption of a particular technology. The mobile phone-based video streaming (MPVS) system has been developed to provide video-based reminders for everyday tasks for PwD. Videos, delivered through a mobile phone, are recorded and scheduled by the carers or family members [2]. The MPVS system has been evaluated using a cohort of PwD, where a wide range of features were collected both automatically through the use of the system and through the usage of preand post-trial questionnaires, ranging from categories of participants ability to perform ADLs, technology experience, home technology infrastructure to care burden, and caregiver technology experience [3][5]. Based on these previous evaluations, the aim of the current study is to identify those features that may inuence technology adoption. Using such features, a predictive model can be developed to determine whether PwD may or may not adopt assistive technology, through investigation of a combination of key aspects, including a range of popular classication algorithms, feature sets, and data imbalance handling. It is envisaged that such a model could become a part of a screening process to the benet of healthcare professionals at the point of evaluating the suitability of incorporating assistive technology into the daily life of PwD (e.g., in a memory clinic setting). During a patient assessment, information of the identied relevant features of a patient could be entered into the predictive model and the healthcare professional informed immediately as to the likelihood of success if technology were to be introduced. Such a predictive model can therefore maximize the opportunity of using technology to support independent living. Also, if the model can successfully recognize PwD unsuitability to such technology, the cost can be minimized and potential negative impact on the quality of life of the PwD avoided. A number of attempts to address the notion of prediction have been reported in the literature, for example, the Technology Acceptance Model (TAM) [6], [7] or the psychosocial impact of

2168-2194 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

376

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 1, JANUARY 2014

assistive devices scale (PIADS) [8]. The TAM is based on the theory of reasoned action and states that the behavior intention is inuenced by perceived usefulness and perceived ease of use. This has been shown to have a direct effect on actual behaviour, for example, in [9]. A common approach when considering likely features to adoption is to separate the features into external environmental features, such as social structures, the regulatory environment, and the infrastructure, in addition to internal personal features such as utility perception, expectations, and self-esteem [8]. The PIADS is an extension to TAM that is focused on personal features and acknowledges the existence of external features, such as social networks and a larger society, which may also have an impact on usage and on the self-image. The PIADS is targeted specically toward assistive technology. The assessment of embarrassment and other negative connotations has been included in the PIADS in order to assess the psychosocial impact of assistive technology on three levels: competence, adaptability, and self-esteem. This 26-item scale, however, requires that a person is able to reect and provide feedback on their perceptions; for PwD such a reection may be difcult. Both the TAM and PIADS models have been criticized in the literature due to their questionable heuristic value and the lack of explanatory and predictive power [10]. Research has been focused on the issues related to general technology adoption and the relevant factors. Generic factors for the use and abandonment of assistive devices for elderly people were researched in [11] with a focus on the elderly population with functional disability and impairment. Also, focusing on people with disabilities, four factors were found to be signicantly related to abandonment of assistive technologies in [12], namely 1) lack of consideration of user opinion; 2) easy device procurement; 3) poor device performance; and 4) changes in user needs or priorities; this prompted the suggestion that better service design can enhance consumer satisfaction with assistive technology and reduce device abandonment. Factors that inuence the response to healthcare robots were studied in [13] in order to facilitate robot assistance of older people in their own homes for improved health and safety, from the perspective of individual factors as well as robot factors. However, in the literature, little attention has been given to investigating technology adoption for the fast growing population group of PwD and their carers. In addition, studies have not made use of the identied factors for the prediction of long-term success of adoption. Also, the evaluation of the effectiveness of telehealth over a 12 month period reported in [14] concluded with the ndings that based on the perhaps less than positive ndings from the study, future work would consider the predictors for participants of early removal of the technology-based solutions. It is, therefore, the aim of the current study to investigate the development of a more transparent approach to technology adoption based on the readily available information that can be gleaned easily during consultation with the PwD and their carer. The organisation of this paper is as follows. Section II provides a synopsis of experimental setup and data collection. Section III describes the learning framework, including the process of feature selection, discretisation, and key aspects in building a predictive model, including classication algorithms, feature

set selection, and handling class imbalance. This is followed by the model evaluation in terms of prediction performance, prediction robustness, class bias, and model usability in Section IV. Conclusions and future work are provided in Section V. II. DATA COLLECTION The MPVS system is comprised of three components. First, there is a mobile phone-based component, which has been modied to support easy interaction for PwD and is used to deliver personalized video messages as reminding prompts. Upon receiving reminders, users are required to press a large button on the device, which acknowledges receipt of the reminder and causes playback of a prerecorded video. The second element of the system provides caregivers with a touch screen and associated software application for recording and scheduling video reminders. The third component of the system is a backend server, which manages the storage, communication, and the transmission of data between the caregiver application and the mobile phone-based application. The MPVS system has been evaluated and improved iteratively. Throughout this iterative process a range of evaluations has been undertaken with a cohort of 40 PwD and their respective caregivers. The technological platform has been updated in accordance with feedback from these evaluations [2][5]. The development of the MPVS system and its pretrial evaluation was reported in [2] on the design and usability of the system along with its technical stability based on nine recruited participants. The rst trial of the updated system was conducted with nine recruited patients with memory impairment along with their carers, using an ABA experimental design, to validate the ability of the system to satisfy user needs as well as to attempt to investigate the potential effect of the technology on the care burden [3]. A longitudinal trial was subsequently carried out over a technology usage period of 5 weeks on patients with AD to assess the possibility of fully incorporating the technology into everyday life [4]. A validation study on the possibility of assessing compliance of performance of a task following reminders using low cost, off the shelf sensors was conducted in a lab-based environment with ten healthy young participants [5]. In an attempt to identify the relevant features pertaining to adoption or abandonment of the system, interviews were performed with the various members of the research and development team, which included: biomedical engineers, computer scientists, research nurses, and geriatric consultants. Together, they identied a range of features potentially relevant to the adoption of the MPVS system, such as appropriate infrastructure, the cognitive and physical ability to manage the system, previous experience with technology, and the perceived utility of the assistive technology. In addition, the role of the caregiver was identied as being important, with features such as caregiver burden, encouragement for the PwD, perception of utility, and their technology experience being identied. Following these interviews and discussions, an initial inuence diagram was created, indicating which features may have an impact on other features with respect to the PwD and also to the family caregiver in relation to adoption of the technology. A subsequently rened inuence diagram (see

ZHANG et al.: PREDICTIVE MODEL FOR ASSISTIVE TECHNOLOGY ADOPTION FOR PEOPLE WITH DEMENTIA

377

Fig. 1.

Inuence diagram of features impacting on technology adoption.

Fig. 1) was established following the presentation and discussion of the initial diagram in a workshop with formal caregivers of PwD (n = 8). The inuence diagram reveals the relationships between the features themselves and more importantly, the relationship with the home-based assistive technology adoption. The diagram contains independent features (within the rectangles with thin lines) and summary features (within the rectangles with thick lines). Though summary features are inuenced by the independent features, they may also be a feature in their own right. For example, in the diagram, age and previous profession are likely to have an impact on technology experience; however, technology experience is a feature in its own right. Initially, only independent features were considered such as age or gender. The impact of these features on the others was discussed in terms of whether they were likely to have a direct impact on technology adoption. Most of the features could be summarized as having an impact on, for example, the ability to perform a task, or the attitude toward technology. Although carer burden was a standalone feature, it turned out that many other features would have an impact on carer burden. Above a certain threshold, these features then implicitly gain an impact on technology adoption. On

the other hand, encouragement had not initially been identied as an important feature but after realizing how many features inuence how a carer would encourage or discourage the use of technology, it became obvious that encouragement itself was an important feature. Consequently, the data from the aforementioned evaluations, from the 40 participant pairs were collated from databases and participant visit logs. The features under consideration are listed in Table I. The level of adoption was described by a 4-item Likert scale recorded by the research nurse, based on previously recorded notes, which indicated whether the dyad had dropped out, was noncompliant or compliant or even eager to keep the technology. Later this scale was reduced to two classes of nonadopter and adopter, where the nonadopter class includes users in the categories of both dropped out and noncompliant and the adopter class contains users in the categories of compliant and eager to keep the technology. III. LEARNING FRAMEWORK In order to develop the most suitable model for the prediction of assistive technology adoption, a range of popular data mining

378

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 1, JANUARY 2014

TABLE I FEATURES COLLECTED FROM PATIENTS THROUGH INTERVIEWS AND QUESTIONNAIRES

algorithms were assessed for their suitability to the prediction task. Also, due to an imbalance of class sizes in the collected data, we investigated the benet of using a resampling technique for building the prediction models. Each individual data mining algorithm was then applied to data with different feature sets, on both scenarios with and without data resampling to handle the data imbalance. A. Feature Selection First, we highlight the characteristics that could have an inuence on technology adoption. Feature selection was performed to select relevant features for the output class, Adoption. In the rst instance, a pair-wise signicance test was performed on an individual feature against the output class to select the directly relevant features. The Chi-square or Fishers exact test [15], in the case of small numbers, was used for categorical features and the MannWhitney test was used for continuous data-based features. All tests were performed using IBM SPSS Statistics 20.0.0. A conventional p-value of 0.05 was used for the signicance threshold. Test results revealed four signicantly relevant features for the Adoption output class, namely MMSE (p = 0.021), Age (p = 0.005), Living_Arrangement (Chi-square = 7.435, p = 0.006), and Broadband (Chi-square = 7.556, p = 0.006). The remaining features were not found to be signicantly related to Adoption on the basis of a pair-wise test. However, we further used principal component analysis as a multivariate approach to assess the possible predictors of the Adoption output class. The rst principle component included three additional features: carer_involvement, gender, and mobile_reception as well as the four aforementioned. The three additional features were subsequently included in the reduced feature set in our analysis to allow for the possibility that they

might be indirectly related to adoption. This feature set, represented by V1 was, therefore, taken into the investigation for the development of adoption predictive model where V1 = {Gender, Living_Arrangement, MMSE, Broadband, Age, MobileRec, and Carer}. Next, we investigate the possibility of reduction of this feature set V1 for the possible benet of both cost related to feature collection and the computational complexity of classication models in relation to less features. In statistics, stepwise regression was widely used for feature reduction [16]. Stepwise regression is a greedy algorithm, where the algorithm selects the best feature to add, or the worst to delete, during each iteration, depending on whether it is a forward-step or backward-step. In our work, the seven relevant features in V1 in relation to the output class were included in a logistic regression model with conditional backward elimination, where removal is based on the probability of the likelihood-ratio statistic based on parameter estimates. Three of the seven features were derived from the backward stepwise logistic regression model, represented by V2 where: V2 = {Gender, Living_ Arrangement and MMSE}. Both the feature sets V1 and V2 were taken into account during the investigation into the development of a suitable adoption predictive model. Following the process of feature selection, the continuous features of Age and MMSE were discretized. The discretization of the MMSE score follows the guidelines from the Alzheimer Society [17] which stated that: MMSE scores of 27 or above (out of 30) are considered normal, mild Alzheimers disease equates to an MMSE score of 2126, moderate Alzheimers disease equates to an MMSE score between 1020, and severe

ZHANG et al.: PREDICTIVE MODEL FOR ASSISTIVE TECHNOLOGY ADOPTION FOR PEOPLE WITH DEMENTIA

379

TABLE II DETAILS OF CLASSIFICATION ALGORITHMS AND PARAMETER SETTINGS

Alzheimers disease equates to an MMSE score of less than 10. Given the inclusion criterion of recruitment of PwD with an MMSE score over 18 in the study conducted, there was only a small number of patients in the category of Moderate and none within the Severe category. The distribution of Age generally follows a normal distribution with a mean value of 72.5 years. Based on the age histogram and to avoid possible overtting based on a small sample size, the discretization was based on a division using the two values 65 and 75 years. B. Prediction Models Learning A range of classication models, learned for technology adoption prediction, was employed in the investigation for both their performance and suitability. Three key aspects were considered in building a predictive model, as follows. 1) Classication algorithms: a range of popular data mining algorithms for the classication task were selected for our investigation [18], [19], namely neural network (NN), C4.5 decision tree (DT), support vector machine (SVM), Na ve Bayes (NB), adaptive boosting (AB), CART, and k -Nearest-Neighbour (k NN). The details of the selected classication algorithms are shown in Table II. 2) Feature set selection: after the feature selection process, described in the previous section, data with a full relevant feature set V1 and a reduced feature set V2 were taken into consideration when building the predictive model. The corresponding datasets are denoted by D1 and D2 , respectively. Models were developed on the data for each of the two feature sets separately. 3) Handling imbalanced classes: the size distribution of the two classes of technology adopters and nonadopters, from collected data, is 70% and 30%. The imbalance between such class data can have an impact on some classication algorithms, typically with bias toward the majority class prediction. Data resampling techniques can be employed

to handle this imbalance. In particular, a dataset is resampled by applying the synthetic minority oversampling technique (SMOTE). Given these three key aspects, a range of predictive models was learned during the investigation into the most suitable model for the prediction of MPVS adoption by PwD. The listed classication algorithms were applied to the data using seven features, set V1 , and three features, set V2 , respectively, with and without using the SMOTE to deal with the issue of imbalance between adopters and nonadopters. Each derived model is denoted by a combination of each of the three aspects relevant to building the model. For example, a model derived using k NN classication on data with feature set V2 , and SMOTE for data resampling, is denoted as k NN_3_S; similarly, a model built without data resampling is denoted as k NN_3. IV. MODEL EVALUATION Model performance was evaluated in terms of class prediction, model robustness, prediction bias among classes, and ease of use. Model prediction and robustness performance reveal how reliably the model can recognize an adopter or a nonadopter. Bias measures whether the model has similar prediction performance for the two classes of adopter and nonadopter. Due to a data imbalance between the two classes, there is a tendency that a model will be biased toward the majority class prediction, therefore causing a high prediction error for the other class. If the model fails to reject a potential nonadopter that is in the minority data class, a nonadoption solution is inappropriately prescribed. Consequently, not only will there be nancial implications, the failure to interact with the device can affect the mood of the PwD and subsequently have a negative impact on the quality of life of both the PwD and their caregiver. The reason to evaluate whether the model is easy to understand or not is that the end users of the prediction tool will typically be healthcare professionals. The outputs of the tool need to be rationalized, clinically understood,

380

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 1, JANUARY 2014

Fig. 2. Prediction accuracies of models learned from a range of prediction algorithms on data with different feature sets.

Fig. 3. Prediction accuracies of models learned from a range of prediction algorithms on data with different feature sets and data resampling SMOTE.

and easily used by such healthcare professionals, especially if they do not have a technical or computational background. The models were evaluated on the overall prediction accuracy, the standard deviation of the prediction accuracy and the difference between the two types of errors (false positive and false negative classications). The evaluation of model performance within the current study was, therefore, performed using a four-fold cross validation. Both the training and test data should be representative samples of the underlying problem. To meet this criterion, a stratication process was applied to ensure that each class was properly represented in both the training and test datasets for each fold. To assist in obtaining reliable results, the four-fold cross-validation process was repeated ten times with different stratied random sampling to produce reliable results. Finally, the model performance was averaged over the ten fourfold cross-validation repetitions as was the standard deviation. A. Model Prediction Performance In this rst scenario, models were derived on the original collected data without handling the data imbalance. The prediction performances were compared (see Fig. 2) between models derived using the same data mining algorithm, on data with the two different features sets. To handle the data imbalance, the SMOTE was applied initially only on the training dataset. Given the data distribution (70%, 30%) over the adopter class and the nonadopter class, in the SMOTE process, the percentage of data to create for the minority nonadopter class was set to 100%. The models were still evaluated on the test data that was not preprocessed, thus retaining the original distribution of the two classes. Similarly to the methodology discussed in the previous Section, models were derived using the selected classication algorithms over data with two sets of feature sets. Prediction performances were then compared, as shown in Fig. 3. From the classication results shown in Figs. 2 and 3, the top performances stemmed from models derived using the four classication algorithms: DT, NN, SVM, and k NN, across all four scenarios in the setting comprising two feature sets and whether or not to employ data resampling to deal with the data imbalance issue. In terms of the performance of the models with

Fig. 4. Comparison of accuracies by models learned from the three feature set, with and without data resampling SMOTE, for a range of algorithms.

respect to the number of features, common outcomes from the comparisons are shown in Figs. 2 and 3. Given the same classication algorithm and data sampling strategy, the model trained on dataset D2 with three features was almost always better than the model learned from dataset D1 .with seven features. This nding applies to both the use of original data for training, as illustrated in Fig. 2, and when employing resampling SMOTE for data imbalance, as illustrated in Fig. 3. This result is encouraging, as with a smaller number of features in the predictive model, not only will it be cheaper to collect three features than seven features, but it also reduces the models computational complexity in learning and more importantly in making a prediction. Next, we analyzed the effect of oversampling minority class data to handle data imbalance, in terms of the prediction performances of the models. Fig. 4 presents an pair-wise comparison between the two models learned on dataset D2 with three features using the same data mining algorithm, with one of the models employing SMOTE. As shown in Fig. 4, in terms of the prediction accuracy, the models trained using SMOTE on data with an adjusted probability distribution over the two classes generally performed worse than the models trained using the data with the original distribution. This outcome applies to the

ZHANG et al.: PREDICTIVE MODEL FOR ASSISTIVE TECHNOLOGY ADOPTION FOR PEOPLE WITH DEMENTIA

381

TABLE III PREDICTION ERROR DIFFERENCE COMPARISON WITH AND WITHOUT SMOTE FOR MODELS TRAINED ON DATA WITH 7 AND 3 FEATURES, RESPECTIVELY

TABLE IV MODELS PREDICTION PERFORMANCE RANK TABLE FOR THE BEST MODELS OF EACH OF THE CLASSIFICATION ALGORITHMS

model prediction performance comparison derived from dataset D1 . The outcome is expected, as the models were tested on data that followed the original distribution. The benet of using data resampling mostly lies in the reduction of misclassication of data in the minority class and the difference between the prediction error rates of false positive and false negative. We will further investigate the performance of the models on other evaluation criteria in the following Section. B. Model Prediction Bias and Usability In this Section, we evaluate the model prediction bias on the two imbalanced classes and analyze the usability of the models, in terms of their ease of use for nontechnical healthcare professional end users. Model prediction bias toward the majority class can be an important issue, especially in a situation where the cost associated with a prediction error on the minority class is high and greatly exceeds the cost of a misclassication error for the majority class case. In our project, if the model fails to reject a potential nonadopter, which is the minority data class, and a technical solution is inappropriately prescribed, not only will there be nancial implications, the mood of the PwD can also be affected by the inability to properly handle the device. This may subsequently have a negative impact on the quality of life for them and their caregivers. On the other hand, in the situation of failing to recognize a potential adopter, a signicant opportunity is missed, in terms of the potential impact that using an assistive technology may have in terms of allowing the PwD to remain in their home for longer. The cost estimation for misclassication goes beyond the focus of the current work. Nevertheless, the difference between the two types of prediction errors is an indication of the prediction bias and the rst step toward an investigation into minimizing the total cost of misclassication within future work. Table III provides a comparison of average prediction errors obtained between models trained on data with and without SMOTE, for both dataset D1 and D2 with seven and three features, respectively. As shown in Table III, in both scenarios, the prediction bias toward the majority class has been reduced using the data resampling approach on the training data. The targeted end users of this tool are healthcare professionals or decision makers, who are often nontechnical. Therefore, criteria relevant to the ease of use, understanding, and interpretation are also highly required features for such healthcare-based

applications. DTs are especially useful in the healthcare-based applications, given that the decision making process is transparent and can be presented visually as trees. Models derived using DT-based algorithms, i.e., C4.5 DT and CART, are easy to use and interpret. Such features are appreciated by healthcare professionals, thereby permitting the outputs from the tool to be rationalized and clinically understood, especially if they do not have a technical or computational background. Similarly, predictions by the k NN-based models are based on one or more nearest neighbours of the new case, which indicates the similarity between the new case and its neighbours. It is, therefore, easier to interpret and understand the prediction results as the decision is made from known experience on the acceptance of such assistive technologies from people with similar features to the new case. On the other hand, the understanding of the decision making process of more complicated models, such as NNs and SVMs, can be a challenge for nontechnical professionals. C. Comparison of Models We chose the best prediction model for each classication algorithm from the models trained on data with different feature sets and whether or not to handle data imbalance. Each models average prediction performances from the repeated cross-validation evaluations are listed in Table IV in rank order, along with other evaluation criteria, the standard deviations, and the bias, in terms of the differences between the two types of error and model usability. Overall, the model k NN_7_S trained using the k NN classication algorithm on dataset D1 with seven features and SMOTE applied, has the best prediction performance, is robust, has a reasonable bias against the two classication errors and can be easily understood by the healthcare professionals. At the same time, the DT model DT_3, though not particularly good in prediction accuracy, is robust in its prediction performance with little bias observed. Furthermore, the model can be easily demonstrated visually, making the decision making process more transparent. Thus, it is easier for healthcare professionals to understand and to use it both appropriately and condently. As an example, we present the predictive model (see Fig. 5) derived using the C4.5 entropy-based decision tree algorithm to illustrate the screening process to identify the adopters or

382

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 1, JANUARY 2014

Fig. 5. DT model DT_3 for technology adoption, where the dark grey nodes marked NA represents non-adopter class and the light gray nodes marked A represents Adopter class.

nonadopters of assistive technology by PwD. As shown in Fig. 5, the DT model provides an intuitive and easily understood tool to facilitate easy decision making, by following the path from the root node, given the conjunction of feature values to reach to a class node. For example, if PwD are not living alone and have mild Alzheimers disease, then female patients are predicted as a technology adopter and male patients, on the other hands, are predicted as nonadopters. Compared with the collected raw data, all female patients in this category were technology adopters after nishing the trial. For the male patients in this category, four out of ve dropped out of the trials or rejected the technology at the end of the trial. Nevertheless, in some circumstances, the predictive model cannot provide strong evidence for decision making. For example, for the PwD who is not living alone although (s)he has moderate memory impairment, the prediction of whether the patient will adopt the assistive technology is unclear with no strong evidence of the person being either an adopter or a nonadopter. In the collected data, there are two instances in this category with one adopter and one nonadopter. There is not sufcient information to produce a decision in such circumstances so additional features may need to be investigated to support a more condent decision and/or additional data may be required. V. CONCLUSION AND FUTURE WORK The acceptance of assistive technologies is crucial for healthcare professionals in the provision of such technologies to PwD. In this paper, we characterized PwD features that are relevant to assistive technology adoption. Based on these features, an optimal predictive model was developed through the investigation of a range of classication algorithms, different feature sets, and data resampling to handle class imbalance. The models were evaluated using the multiple criteria of model predictive performance, prediction robustness, bias toward two types of errors, and usability by healthcare professionals. Overall, the model trained using the k NN classication algorithm on data collected from seven features best addressed the four criteria for model evaluation. This predictive model can maximize the opportunity

of using assistive technology in order to allow people to stay in their home for longer, thus minimizing the risk of negative impacts on mood and the quality of life for PwD, and nancial implications for inappropriate deployment to nonsuitable technology adopters. A limitation to our work is the amount of data available. It was both expensive and time-consuming to collect such data from the PwD using the technology. Questionnaires about the PwD and their user experience were particularly time consuming to administer; caregivers could face additional work checking if the PwD was handling the device well; trials required weeks to complete in order to allow the users to become familiar with the device before deciding whether to adopt it or not. Consequently such trials may be intimidating for the PwD. Nevertheless, a collaborative project is currently underway, which will allow our current approach to be extended to a larger sample size. This collaboration is based around the Cache County Study on Memory in Aging, a large database containing genetic and environmental factors associated with risk for Alzheimers disease and other forms of dementia. Another interesting future direction is to embed the cost of the two types of error into the classication model in order to minimize the total cost of misclassication. REFERENCES
[1] R. Orpwood, C. Gibbs, T. Adlam, R. Faulkner, and D. Meegahawatte, The design of smart homes for people with dementiaUser interface aspects, Univ. Acc. Inform. Soc., vol. 4, no. 2, pp. 156164, Dec. 2005. [2] M. P. Donnelly, C. D. Nugent, S. Mason, S. I. McClean, B. W. Scotney, A. P. Passmore, and D. Craig, A mobile multimedia technology to aid those with Alzheimers disease, IEEE Multimedia, vol. 17, no. 2, pp. 42 51, Apr. 2010. [3] S. A. ONeill, S. C. Mason, G. Parente, M. P. Donnelly, C. D. Nugent, S. I. McClean, B. W. Scotney, and D. Craig, Video reminders as cognitive prosthetics for people with dementia, Aging Int., vol. 36, no. 2, pp. 267 282, Jun. 2011. [4] C. D. Nugent, S. A. ONeill, M. P. Donnelly, G. Parente, M. P. Beattie, S. I. McClean, B. W. Scotney, S. Mason, and D. Craig, Evaluation of video reminding technology for persons with dementia, in Proc. Int. Conf. Select. Topics, Montreal, Canada, 2011, pp. 153160. [5] S. A. ONeill, G. Parente, M. P. Donnelly, C. D. Nugent, M. P. Beattie, S. I. McClean, B. W. Scotney, S. Mason, and D. Craig, Assessing task compliance following mobile phone-based video reminders, in Proc. IEEE Eng. Med. Biol. Society, Boston, MA, USA, 2011, pp. 52955298. [6] F. D. Davis, A technology acceptance model for empirically testing new end user information systems: Theory and results, Ph.D. dissertation, Sloan School of Management, MIT, Cambridge, MA, USA, 1986. [7] F. D. Davis, Perceived usefulness, perceived ease of use, and user acceptance of information technology, MIS Quart., vol. 13, no. 3, pp. 319340, Sep. 1989. [8] H. Day and J. Jutai, Measuring the psychosocial impact of assistive devices: The PIADS, Can. J. Rehabilitat., vol. 9, no. 2, pp. 159168, 1996. [9] D. C. Yen, C. Wu, F. Cheng, and Y. Huang, Determinants of users intention to adopt wireless technology: An empirical study by integrating TTF with TAM, Comput. Human Behavior, vol. 26, no. 5, pp. 906915, Sep. 2010. [10] M. Chuttur, Overview of the technology acceptance model: Origins, developments, and future directions, Sprouts: Working Papers Inf. Syst., vol. 9, Article 37, 2009. [11] L. N. Gitlin, Why older people accept or reject assistive technology, Generations: J. Amer. Soc. Aging, vol. 19, no. 1, pp. 4146, Mar. 1995. [12] B. Phillips and H. Zhao, Predictors of assistive technology abandonment, Assist Technol., vol. 5, no. 1, pp. 3645, 1993. [13] E. Broadbent, R. Q. Stafford, and B. A. MacDonald, Acceptance of healthcare robots for the older population: review and future directions, I. J. Social Robotics, vol. 1, no. 4, pp. 319330, Nov. 2009. [14] M. Cartwright, S. P. Hirani, L. Rixon, M. Beynon, H. Doll, P. Bower et al., Effect of telehealth on quality of life and psychological outcomes

ZHANG et al.: PREDICTIVE MODEL FOR ASSISTIVE TECHNOLOGY ADOPTION FOR PEOPLE WITH DEMENTIA

383

[15] [16] [17]

[18]

[19] [20] [21]

over 12 months (Whole Systems Demonstrator telehealth questionnaire study): nested study of patient reported outcomes in a pragmatic, cluster randomised controlled trial, BMJ, vol. 346, no. f653, Feb. 2013. P. W. Zehna, Probability Distributions & Statistics. Boston, USA: Allyn & Bacon, 1970. H. Z. An, D. Huang, W. Q. Yao, and C. H. Zhang. (2008). Stepwise searching for feature variables in high-dimensional linear regression. [Online]. Available: http://stats.lse.ac.uk/q.yao/qyao.links/paper/ahyz08.pdf Alzheimers Society. (2012). Factsheet 436: The mini mental state examination (MMSE). Alzheimers Society. [Online]. Available: http:// www.alzheimers.org.uk/site/scripts/documents_info.php?documentID = 121 X. D. Wu, V. Kumar, J. R. Quinlan, J. Ghosh, Q. Yang, H. Motoda, G. J. McLachlan, A. Ng, B. Liu, P. S. Yu, Z. H. Zhou, M. Steinbach, D. J. Hand, and D. Steinberg, Top 10 algorithms in data mining, Knowl. Inf. Syst, vol. 14, no. 1, pp. 137, Jan. 2008. B. Yegnanarayana, Articial Neural Networks. New Delhi, India: PHI Learning Pvt. Ltd., 2006, pp. 88141. J. R. Quinlan, C4.5: Programs for Machine Learning. San Mateo, CA, USA: Morgan Kaufmann, 1993, pp. 1755. J. C. Platt, Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods, in Advances in Large Margin Classiers, A. J. Smola, P. Bartlett, B. Scholkopf, and D. Schuurmans, Eds. MA, USA: MIT Press, 2000, pp. 6174.

Mark P. Donnelly received the Bachelors degree and Ph.D. degree in computer science from the University of Ulster, Ulster, U.K., in 2004 and 2008, respectively. He is currently a Lecturer with the School of Computing and Mathematics and a member of the Smart Environments Research Group, the University of Ulster. He has more than 80 publications spanning a range of applied research topics in biomedical engineering, medical informatics and the integration of assistive technologies within smart environments. His research interests include the use of mobile technology to support remote monitoring of user behaviours and activities. Dr. Donnelly he has been serving on the board of directors for European Society for Engineering and Medicine since 2011.

Shuai Zhang received the B.Sc.degree in computer science from Heilongjiang University, Harbin, China, the M.Phil. degree in visual arts data mining from the University of Bradford, Bradford, U.K., and the Ph.D. degree in intelligent data analysis from the University of Ulster, Ulster, U.K. She is currently a Lecturer in the School of Computing and Mathematics at the University of Ulster. Her research interests include intelligent data analysis on semantically heterogeneous aggregate data in a distributed environment, learning inhabitants behavioural patterns from unreliable low-level sensor data in smart environment to support assisted living, modelling for connected health applications, and health technology assessment. Sally I. McClean (M98) received the M.A. degree in mathematics from Oxford University, Oxford, U.K., the M.Sc. degree in mathematical statistics and operational research from Cardiff University, Wales, U.K., and the Ph.D. degree in mathematics (stochastic modelling) from the University of Ulster, Ulster, U.k. She is a Professor of Mathematics and Leader of the Information and Communications Engineering Research Group at the University of Ulster, Coleraine, U.K. She has more than 250 publications. Her research interests include mathematical modelling, applied probability, multivariate statistical analysis, and applications of mathematical and statistical methods to computer science, particularly database, telecommunications, and sensor technology. Dr. McClean is a Fellow of the Royal Statistical Society, Fellow of the Operational Research Society, Associate Fellow of the Institute of Mathematics and its Applications. In addition, she is the Past President and Current Treasurer of the Irish Statistical Association. Chris D. Nugent (S96A99M03) was born in 1973. He received the B.Eng. degree in electronic systems and the D.Phil. degree in biomedical engineering from the University of Ulster, Jordanstown, U.K., in 1995 and 1998, respectively. In 1998, he took the post of Research Fellow at the University of Ulster and is currently the Professor of Biomedical Engineering. His research interests include Internet-based healthcare models and the design and evaluation of smart environments for ambient-assisted living applications.

Leo Galway received the B.Sc. in computer science from Queens University, Belfast, U.K., and the M.Sc degree in computing and intelligent systems and the Ph.D.degree in computing and information engineering both from the University of Ulster, Ulster, U.K. He is currently a Lecturer in the School of Computing and Mathematics at the University of Ulster. His current research activity focuses on the application of articial intelligence techniques within pervasive computing environments and digital entertainment. His research interests include computational intelligence, evolutionary computing, and mobile software development. He is currently a member of the Articial Intelligence and Applications Research Group at the University of Ulster.

Bryan W. Scotney received the B.Sc. degree in mathematics from the University of Durham, Durham, U.K., and the Ph.D. degree in mathematics from the University of Reading, Reading, U.K. He is currently a Professor of Informatics and Director of the Computer Science Research Institute at the University of Ulster, Coleraine, U.K. He has more than 200 publications. His research interests include mathematical computation, especially in digital image processing and computer vision, pattern recognition and classication, statistical databases, reasoning under uncertainty, and applications to healthcare informatics, ofcial statistics, biomedical and vision sciences, and telecommunications network management. Dr. Scotney is currently the President of the Irish Pattern Recognition and Classication Society, and Member of Council of the International Federation of Classication Societies.

Ian Cleland received the B.Sc. degree in biomedical engineering and the Ph.D. degree in computer science both from the University of Ulster, Ulster, U.K., in 2010 and 2012, respectively. He was subsequently awarded a studentship from the UKs ESRC to support the Ph.D. degree. He is currently a Research Associate within the Smart Environments Research Group, University of Ulster, where he is involved in the EPSRC funded Multidisciplinary Assessment of Technology Centre for Healthcare project. His research interests include the application of wireless sensor networks, wearable technologies, and pervasive computing within healthcare, with a particular focus on activity recognition.