You are on page 1of 11

Focus Theme – Original Articles 197

Learning to Detect Cognitive


Impairment through Digital Games

Pervasive Health
and Machine Learning Techniques*
A Preliminary Study
Sonia Valladares-Rodriguez1; Roberto Pérez-Rodriguez1; J. Manuel Fernandez-Iglesias1;
Luis E. Anido-Rifón1; David Facal2; Carlos Rivas-Costa1
1Department of Telematic Engineering, University of Vigo, Vigo, Pontevedra, Spain;
2
Department of Developmental Psychology, Facultade de Psicoloxia, Universidade de Santiago de Compostela,
Santiago de Compostela, Spain

Keywords Results: This first pilot study offered initial 1. Introduction


Alzheimer’s disease, early detection, machine evidence about psychometric validity, and
learning, data analytics, serious games, more specifically about construct, criterion Cognitive functions make it possible to

Downloaded by: Trinity College Dublin. Copyrighted material.


Classification and Regression Trees, logistic and external validity. After an analysis using navigate the world around us, actively con-
regression, random forest, Support Vector machine learning techniques, findings show trolling the reception, selection, processing,
Machine a promising 100% rate of success in classifi- storage, elaboration and retrieval of in-
cation abilities using a subset of three games formation. Presently, their evaluation is
Summary in the battery. Thus, results are encouraging supported by validated tests or neuropsy-
Objective: Alzheimer’s disease (AD) is one of as all healthy subjects were correctly dis- chological batteries, which will be referred
the most prevalent diseases among the adult criminated from those already suffering AD as “classical tests” [1] in the rest of this
population. The early detection of Mild Cog- or MCI. paper. The usual format for this type of
nitive Impairment (MCI), which may trigger Conclusions: The solid potential of digital tests consists of a collection of tasks or ex-
AD, is essential to slow down the cognitive serious games and machine learning for the ercises to be solved using pen and paper. At
decline process. early detection of dementia processes is the end of the test, a measurable outcome
Methods: This paper presents a suit of seri- demonstrated. Such a promising perform- or score on the subject’s cognitive status
ous games that aims at detecting AD and ance encourages further research to event- can be computed from the subject’s re-
MCI overcoming the limitations of traditional ually introduce this technique for the clinical sponses.
tests, as they are time-consuming, affected diagnosis of cognitive impairment. Classical tests are time consuming as
by confounding factors that distort the result they require a personalized and compre-
and usually administered when symptoms hensive attention from the professionals
are evident and it is too late for preventive responsible for administering such tests.
measures. The battery, named Panoramix, as- Besides, some subjects perceive classical
sesses the main early cognitive markers (i.e., tests as intrusive tools, and more specifi-
memory, executive functions, attention and cally as an exam of their intelligence, which
gnosias). Regarding its validation, it has been may influence their motivation, alertness
tested with a cohort study of 16 seniors, in- or effort, and thus influence the outcome.
cluding AD, MCI and healthy individuals. In addition, classical tests are dependent on
confounder variables [2] that may alter the
final result. Age, educational level [3] or
Correspondence to: Methods Inf Med 2018; 57: 197–207
Sonia Valladares-Rodriguez https://doi.org/10.3414/ME17-02-0011
geo-cultural aspects may have relevant im-
University of Vigo received: July 28, 2017 pact on the outcomes. Finally, they are ad-
Department of Telematic Engineering accepted: July 6, 2018 ministered when the symptoms of cogni-
Campus Lagoas-Marcosende Vigo, Galicia Vigo Funding tive decline are evident for the person or
Vigo Pontevedra 36310 The present work was partially funded by the govern-
Spain ment of Galicia (Xunta de Galicia, Spain), which cover-
their relatives [4] and too late for efficient
E-mail: soniavr@det.uvigo.es ed the travel expenses to participants‘ homes during rehabilitation or palliative measures.
the pilot study (grant # 2016/236), and also by ‘Rede Alternatives to classical tests are there-
Galega de Investigación en Demencias’ (IN607C- fore very much in demand. The introduc-
* Supplementary material published on our 2017/02) funded by Axencia Galega de Innovación
website https://doi.org/10.3414/ME17-02-0011 GAIN – Xunta de Galicia.
tion of information and communication
technologies (ICT) for neuropsychological

© Georg Thieme Verlag KG 2018 Methods Inf Med 4/2018


198 S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML

evaluation brings some relevant benefits comprises 16 different scenes and relevant coding, storing and syncing different spa-
[5], such as a better standardization of ad- events happening at those places. Subjects tial representations.
ministration; increased accuracy of timing have to carry out cognitive tasks such as Finally, some contributions supporting
presentation and response latencies; an ea- free recall, visuospatial recall, delayed free the evaluation of prospective memory are
Pervasive Health

sier administration and collection of data; recall and recognition. Virtual Reality Shopping Task [23], Virtual
the possibility of capturing and monitoring With respect to the assessment of atten- Week [24] and Banville et al. [25]. The first
dynamic indicators (e.g., cognitive markers tion, the videogame AULA [14, 15, 16], re- example reproduces a virtual reality en-
such as reaction time latency[6] or other produces a virtual environment depicting a vironment in which participants have to
not cognitive indicators such as gait pat- classroom, which is targeted to the diag- perform an ongoing task (i.e., purchasing a
terns[7]); and better randomization of the nosis of the Attention Deficit Hyperactivity list of items) and three event-based pros-
presentation of stimuli through repeated Disorder (ADHD). The cognitive tasks in- pective memory tasks. The second one is a
administrations. volved are omission and commission er- computerized board game that embeds
A promising approach is the use of seri- rors, reaction time, motor activity, differ- event-based and time-based prospective
ous games. They are “games that do not entiation between No-X and X tasks, com- memory tasks. And finally the last
have enjoyment, entertainment or fun as parison of sensorial modalities, and quality example, involves the assessment of pros-
their primary purpose” [8]. They are being of attention focus. pective memory by requesting the player to
introduced in many application areas, such We can also find some videogames tar- remember to perform some actions such as
as education, rehabilitation, or military geting the assessment of working memory. feeding a fish and switching off a fan at
training among many others. Serious The Virtual Radial Arm Maze (VRAM) points of time in the future.

Downloaded by: Trinity College Dublin. Copyrighted material.


games include a playful component as a [17], reproduces a virtual room that con- The examples above illustrate how seri-
motivation element, and have a non rec- tains a maze with six arms, three of which ous games, and specifically those based in
reational aim or motivation such as learn- contain a treasure at their ends. The goal of virtual reality environments, are being in-
ing, evaluation or training. Moreover, for the game is to find such treasures. As an- troduced in the field of cognitive evalu-
older adults playing serious games, differ- other example, Shapebuilder [18] displays a ation. These solutions are presented as
ent game aspects must be taken into ac- 4×4 grid and four sets of basic colored better alternatives to classical tests, as they
count, including positive effect, fun, and shapes at the sides of the grid. The goal is are more ecological because they typically
challenge [9]. Thus, the serious game con- to replay the series of stimuli presented reproduce real-life situations, and they are
cept encompasses approaches such as (i.e., ordering and spatial position of a set not perceived as an intrusive tool. How-
simulations, educational games, or even of colored shapes). ever, more research is still needed to build
books. The next paragraphs provide addi- There are some proposals to evaluate reliable and valid serious game-based in-
tional insight into this approach to cogni- executive functions. The Scavenger Hunt struments ready to be used in clinical neur-
tive evaluation. [19] displays a set of randomly placed dots opsychological assessment [10]. The main
Several contributions exploring the in- that the player connects by clicking them challenge to be addressed is the lack of an
troduction of serious games to assess dif- in a specific sequence. Data captured in- adequate psychometric validation that
ferent cognitive areas can be found in the clude recall time, search time, motor time would confirm their potential as a serious
literature [10]. For example, regarding the and number of steps. Finally, the Virtual alternative to classical neurophsychological
evaluation of episodic memory, VR Driv- Action Planning Supermarket (VAP-S) tests, following the Standards for Edu-
ing [11] emulates a route through a city or [20] reproduces a virtual environment dis- cational and Psychological Testing [26].
town that partcipants are requested to playing a supermarket. Data captured in For example, the works referenced lack a
complete. They also have to memorize de- this case include the total distance tra- comprehensive validation of all properties
tails of the scenes they encounter while versed, total time, number of items pur- related to their psychometric validity. Fur-
driving the route. Another example is the chased, number of correct actions, number thermore, another relevant limitation
VR-based home [12], which displays a vir- of incorrect actions, number of pauses, found is the limited introduction of state-
tual apartment. Participants are distributed combined duration of pauses and time to of-the-art machine learning techniques for
into two groups, active and passive. Partici- pay. the prediction of cognitive states, as they
pants in the passive group just watch a pre- Attending the visuospatial ability’s are based on mainstream statistical tech-
viously recorded video of the apartment, evaluation, a relevant contribution is the niques only. The Discussion section below
while active participants are requested to Geoshooter application [21], where partici- offers a comparative study of the work in-
actively explore each room and carefully pants are requested to plan routes in four troduced in this paper and previous con-
look at the objects in order to try to recall different conditions in a virtual environ- tributions in the field.
them. Finally, as a last example, Jebara et al. ment (i.e., real, hybrid, virtual, and vir- This paper presents the design and pre-
[13] propose a virtual reality environment tual+condition). Another example is liminary validation of Panoramix, a suite of
that displays a town with buildings, people, NeuroVirtual 3D [22], which uses virtual serious games to detect MCI and AD in
cars, street objects such as lampposts, and reality for assessing the abilities in en- senior adults. Unlike previous approaches
several intersections. The environment [10], Panoramix has been designed accord-

Methods Inf Med 4/2018 © Georg Thieme Verlag KG 2018


S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML 199

ing to the standard recommendations [26] 2. Methods 2.1.1 Test Subjects


to guarantee the required psychometric
properties that will pave the way for a fu- 2.1 Design and Development of A first pilot study was carried out with 16
ture clinical validation solution to comple- Serious Games subjects (12 women and 4 men) from Gali-

Pervasive Health
ment or even replace classical tests. In The design process was centred on the end cia, Spain, in May and June, 2016. The co-
order to classify adults in two groups (i.e., user. For this, two focus group sessions [31] hort study (cf. ▶ Table 1) was composed by
healthy or MCI/AD), we introduced a were conducted with senior adults to (1) 8 individuals with no cognitive impair-
novel application of machine learning [10] gather information on the specific needs ment as a control group (average age
in this field. This process is discussed in that should cover such games. Further- 68.3±8.88 years); (2) 5 AD patients (aver-
Sect. Methods where the design and devel- more, the research team had a multidisci- age age 75.8±5.36 years, and (3), 3 MCI
opment methodology is introduced from plinary composition by including experts patients (average age 75±6.08 years). All
both a technological and psychometric per- in the field of cognitive neuropsychology participants signed informed consent be-
spective. and computer technologies. For each seri- fore participation, in accordance with the
Panoramix is composed of six serious ous game, design was carried out according Declaration of Helsinki. The pilot study
games addressing the following relevant to (1) the selection of cognitive area to was approved by the Autonomic Research
areas to assess the cognitive status [27]: epi- be assessed; (2) identification of the most Ethics Committee of Galicia (Spain).
sodic memory, attention, semantic mem- appropriate design approach, and finally Participants provided data through (1)
ory, working memory, procedural memory (3) identification of variables to be col- an initial interview where several classical
and gnosias. This suite aims to reproduce lected throughout game playing. Finally, tests were administered to gather golden

Downloaded by: Trinity College Dublin. Copyrighted material.


the abilities of classic neuropsychological from a technological perspective, we chose standard data and to make sure that they
batteries to assess cognition abilities in sen- Unity, which has a broad background in were correctly classified according the
ior adults and to detect potential cognitive the development of 3D applications to be standard clinical practice; (2) a perception
impairment. Cognition abilities addressed run in multiple systems and devices. More- survey on the use of technology in general
include early markers of cognitive alter- over, games were designed to run on main- and serious games in particular; (3) data
ations [28, 29], particularly episodic mem- stream computers and touch devices sup- collected from Panoramix games and (4) a
ory and attention, driving to MCI and po- porting accessible control interfaces such as final questionnaire on their experiences
tentially AD. Due to the central role played joysticks, gaming mice or gestural remotes during the pilot study. The pilot was car-
by episodic memory evaluation, a previous when required. The Panoramix suite offers ried out at participants’ homes in two
work [30] addressed the design procedure muti-language support and all stimuli are sessions. With this, a relaxing, non-intru-
and a preliminary validation of that spe- also presented in textual and audio formats sive and close to real-life environment was
cific area. The research discussed in this to enhance accessibility. In other words, guaranteed. In other words, cognitive
paper complements and completes a first these serious games provide additional ver- evaluation was performed in the most eco-
psychometric validation of Panoramix satility, interactivity, and personalization logical environment possible.
aimed to overcome the limitations detected capabilities when compared with pen-and- The size of the sample for this prelimary
in related work. The Panoramix tool is in- paper tests, as they are software systems de- study was estimated according to main-
troduced in Sect. Results. signed under such premises. stream statistical methods to calculate
Our final aim is to provide an answer to
the following research question: Can seri-
ous games and machine learning algo-
rithms be used to detect MCI or AD? As a Table 1 Test subjects’ characteristics (general and cognitive).
first step towards the validation of this hy- Parameters AD (N=5) MCI (N=3) HC (N=8)
pothesis we carried out a first experiment
Age (M±SD) 75.8±5.36 75±6.08 68.3±8.88
with 16 older adults, including MCI and
AD patients, and healthy adults. The out- Women/Men (n) 4/1 2/1 6/2
comes of this pilot study, focusing in con- aSchooling years (M±SD) 7.6±3.13 10±8.54 12.12±5.89
struct, criterion and external validation (cf. bFreq. Games (M±SD) 1.4±0.54 2.7±0.58 1.37±0.74
Sect. Results), are presented in Sect. Dis- cFreq. ICT (M±SD) 1±0.00 2.7±1.53 3.3±1.58
cussion. Finally, conclusions are presented
dMMSE (M±SD) 21±4.95 25.33±0.58 28.37±1.60
about the potential of using digital serious
games and classification algorithms for the dMAT (M±SD) - 44±1.41 46±3.40
early detection of dementia processes. aSchoolingyears: adapted from ISCED, UNESCO [32]
b”Frequency of use of video games” and
c”Frequency of use of technologies”: 5-point Likert scale [1(never) to 5 (very frequently)]
dMMSE score
dMAT score. It was not administered to AD group to avoid frustration.

© Georg Thieme Verlag KG 2018 Methods Inf Med 4/2018


200 S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML
Pervasive Health

Downloaded by: Trinity College Dublin. Copyrighted material.


Figure 1 Methodology: cross-sectional and leave-one-out validation strategy.

sample sizes in small populations, and also, 2.1.2.1 Construct Validity 2.1.2.2 Criterion Validity
according to existing recommendations Construct validity is the degree to which a Criterion validity reflects the extent to
[33; 34] for the estimation of Spanish adult test measures what it claims to be measur- which a measure is related to an outcome, a
population. According to existing data ing. There are several ways to assess con- classification into health, MCI or AD sub-
about the Spanish adult population and the struct validity, including the correlation of jects in our case. It is typically measured
prevalence of dementia in Spain, the a measure of the construct and designated through concurrent validity (i.e., with re-
sample selected —n=16— is a represen- variables; factor analysis; among others. In spect to diagnostic categories) and predic-
tative sample of the Spanish senior popu- our study, correlations were computed be- tive validity (i.e., predicting performance or
lation with a 95% confidence interval and tween the classical tests administered – behaviour). In this case, data captured
an error of 7.5%. Mini Mental State Examination (MMSE) from games only was processed using ma-
[37]; Memory Alteration Test (MAT) [38]; chine learning techniques to classify par-
Spanish version of California Verbal Learn- ticipants in healthy individuals or people
2.1.2 Psychometric Validation
ing Test (CVLT) [39] and Informant Ques- with cognitive impairments (i.e., MCI or
As the main goal of Panoramix is to emu- tionnaire on Cognitive Decline in the AD in our case). In other words, we aim to
late a classical neuropsychological battery, Elderly (IQCODE) [40] –, the correspond- compute a cognitive assessment or cogni-
the same validation process of such bat- ing games, and static general variables – tive status of senior adults, only from their
teries shall be followed to guarantee psy- gender, age, schooling years, socialize (i.e., interaction with the games in the Panora-
chometric validity. As a reference for this level of interaction with other people), fre- mix suit.
procedure, we relied on the Standards for quency of use of video games, frequency of
Educational and Psychological Testing use of technologies, etc.. Note that classical 2.1.2.3 External Validity
[26]. Besides, from a methodological per- tests provided our golden standard truth Finally, external validity indicates whether
spective a cross-sectional or prevalence data. Despite their limitations, these tools the test and its results can be generalized to
study [35] was performed, combined with a have clinical validity and are widely ac- different subjects, people, places, experi-
one-left-out validation strategy [36] (cf. cepted in medical practice. Thus, they con- menters, etc. For this, we designed both the
▶ Figure 1). In our case, data were ob- stitute our ground truth in order to develop digital games and the pilot experiment ac-
tained to assess construct, criterion and ex- and to validate an improved cognitive digi- cording to these parameters: (1) we used
ternal validity. tal tool based on serious games. sufficiently heterogeneous sample, even

Methods Inf Med 4/2018 © Georg Thieme Verlag KG 2018


S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML 201

though it was small, to guarantee popu-


lation validity; (2) the experimental en-
vironment was as natural as possible (i.e.,
to reinforce the environmental validity) by

Pervasive Health
taking place at participants’ homes; and
finally (3) we conducted the experiment at
different times for each user in order to
avoid temporary/daily aspects that may in-
terfere with the administration of digital
games. Finally, an adaptation of the Game
Experience Questionnaire (GEQ) [9, 41],
was completed by those participating in the
pilot study, to capture data about this issue.
Particularly, a questionnaire – based on a
5-point Likert scale – about perception of
Panoramix (cf. ▶ Online Appendix A), fa-
cilitated the assessment of players’ feelings
(i.e., perception extracted from Q4, Q5 and
Q6) and thoughts (i.e., perception extract-

Downloaded by: Trinity College Dublin. Copyrighted material.


ed from Q3 and Q7), that is, some game
experience factors.
Figure 2 Screenshot of pilot experiment playing Attentix game.
2.2 Classification Algorithms
From an analytics perspective, we relied on cy of use of video games, frequency of use 3. Results
Logistic Regression (LR), as it is a tradi- of technologies, etc.) and all data collected
tional statistical classification method ex- from interacting with the serious games. In 3.1 The Panoramix Suit
tensively used in medical classification addition, we applied the classification met- 3.1.1 Episodix Serious Game
problems. Additionally, we applied other rics below [43, 44], to offer an accuracy Episodic memory is a cognitive system that
machine learning classifiers such as Sup- value for each algorithm: allows human beings to recall past experi-
port Vector Machines (SVM) and Classifi- • True Positive (TP): rate of people with ences [47]. The evaluation of episodic
cation and Regression Trees (CART) [42], cognitive problems correctly identified memory is paramount to MCI or AD de-
which were recently introduced to support as people with MCI or AD. tection, as it is the most reliable indicator
medical prediction and classification tasks • True Negative (TN): rate of healthy of the likelihood of conversion of MCI into
[43, 44]. people correctly identified as healthy. AD [29, 48, 49, 50]. Episodix (cf. ▶ Figure
For each algorithm, we used a training/ • False Positive (FP): rate of healthy 2) is based on the gamification of the Cali-
testing ratio based on the one-left-out people incorrectly identified as people fornia Verbal Test (CLVT) [39], which is
strategy [36] – each algorithm is trained with MCI or AD. widely used to assess this type of memory.
with the whole sample except one, and • False Negative (FN): rate of people with In order to obtain a more ecological tool,
then the same algorithm is used to predict cognitive problems incorrectly ident- we introduced some changes concerning
the classification of the one left out; repeat- ified as healthy. CVLT’s administration procedure. Firstly,
ing this process as many times as there are • F-measure (β): β-weighted harmonic instead of reproducing an imaginary shop-
samples. We could apply the algorithm to average of precision and recall. Its opti- ping list verbally, a virtual walk is carried
all variables or just to a particular subset of mal value is β = 1 and the worst value is out where the objects’ list to learn is nat-
the most informative of them. To do it, we β = 0. Parameter β determines the urally present on the street. Secondly, we
used internally the Random Forest algo- weight of precision in the combined redefined the categories and words, so that
rithm [45, 46] to automatically select the precision-recall score. In our study, re- they are easier to recognize visually, but al-
most informative variables. sults were computed for β = 1 and addi- ways maintaining the same frequency of
On the other side, the classification cri- tionally for β = 1.5. In the later case, use as in the original test. For example, cat-
terion was modelled as a dichotomous more weight is given to the β par- egories of vehicles and stores are included,
variable: healthy controls or subjects with ameter to reduce the influence of false instead of herbs and tools. At the end, the
cognitive problems – MCI and AD pa- negatives. Note that this is a common variables captured from Episodix (cf.
tients. Moreover, neuropsychological pre- criterion in medical studies. ▶ Online Appendix A) are similar to the
dictors included demographic variables F-measure (β) = (1 + β2) × precision ones in the classical test (e.g., yes/no recog-
(i.e., age, gender, schooling years, frequen- × recall / (β2 × precision) + recall nition, free recall, short/long delayed recall,

© Georg Thieme Verlag KG 2018 Methods Inf Med 4/2018


202 S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML

recency, primacy, semantic clustering, re- input” [52]. One of the classical methods to consists of 4 trials lasting 15 seconds where
sponse to inhibitions, etc.), and also, hits, evaluate this cognitive area is the Corsi’s different aspects of the rotor are modified,
repetitions, guesses, omissions, errors, and Cubes test [53]. This test involves placing including the rotational speed and the
total time spent in each phase (e.g., nine cubes stuck randomly on a board. The radius of the different elements composing
Pervasive Health

RCL_CP2: long-term recall with 2nd sem- examiner plays a sequence of cubes in a the rotor (cf. ▶ Online Appendix A).
antic clues; RCL_CP: long-term recall with specific order and the subject under study
clues; RL_CP: long-term recall; R1_B1: im- must reproduce that sequence (i.e., two se-
3.1.6 Gnosix Serious Game
mediately recall, etc.). quences of 3, 4, 5 and 6 span). In the case of
assessing working memory, the input se- Gnosia [55] is defined as the human capac-
quence must be repeated in reverse order. ity to develop, interpret and assign mean-
3.1.2 Attentix Serious Game
If we replied in the same order, attention ing to the information captured by the
This game assesses the attention capabil- and not working memory would be the senses. There are different categories of
ities of players, as the way they select in- cognitive area involved. In other words, gnosia, depending on the sense through
formation for its conscious processing and this test – and therefore the designed seri- which information is captured: visual,
action, maintaining the state of alertness ous game – supports the evaluation of auditory, tactile, olfactory, gustatory, and
required for attentive processing at the working memory or attention span, ac- body scheme.
same time. We used the popular Simon cording to the order in which the subject In particular, the Gnosix game focuses
Says game as an inspiration. Like in the must reproduce the sequence. Workix on visual gnosias, that is, the ability to vis-
original game, Attentix reproduces a color gamifies this test to be run from a digital ually recognize different elements to assign

Downloaded by: Trinity College Dublin. Copyrighted material.


sequence that the player must repeat. The environment. Variables considered to com- a meaning to them. To do this, four activ-
length of the sequence is incremented by pute a score include the total length of the ities are proposed (i.e., recognition of
one after a correct answer, for a maximum game, average response time, execution shapes, colors, expressions or feelings, or a
length of 10 colors. There is also a select- mode, sequence reproduced indicating the random combination of them). The vari-
able difficulty for the color palette (4 or 5 number of successful cubes (cf. ▶ Online ables captured to compute a score include
different colors). Appendix A). the number of correct sequences, number
Finally, the variables captured to com- of sequences failed, task type (i.e., shapes,
pute a score value are average response colors, feelings or random), average re-
3.1.5 Procedurix Serious Game
time, total time of play, hits, errors, and sponse time, total time spent, response
level of difficulty of the sequence to be rep- The procedural system is instrumental in time and number of attempts (cf. ▶ Online
licated (cf. ▶ Online Appendix A). Note performing activities of daily living, that is, Appendix A).
that it will not be mandatory to perform all it learns and retrieves all schemas or rou-
phases to avoid frustration in some players. tines of thought and action [54], without 3.2 Preliminary Validation
which people would not be able to develop Outcomes
a normal life. Like in the case of Semantix
3.1.3 Semantix Serious Game 3.2.1 Construct Validity
and Workix, design was based on the ga-
Semantix is based on the gamification of mification of a classical test by embedding The correlation study performed between
the Pyramids and Palm-trees test [51]. It it in a computerized environment. In par- variables collected from serious games and
consists of 52 sets of chips with three ticular, the Pursuit Rotor Task test [32] was classical tests according to each cognitive
images: the stimulus given, the target used as a foundation. In this test, the sub- group shows that:
stimulus, and the distraction stimulus. The ject has to track a rotating circle as accu- • Episodix: best correlations are found
player must try to correctly match the rately as possible, using the mouse or other for: MMSE (e.g., time duration for both
stimulus given to one of the other two op- integrated control element in the digital HC – time duration RCL_CP2 (r=0.78,
tions. The assessment of the state of sem- version. Thus, this game measures pro- ρ=0.01) – and MCI – time duration
antic memory is obtained from the analysis cedural memory (cf. ▶ Figure 2) along with RCL_CP (r=0.64, ρ=0.25); guesses –
of the total playing time, number of hits, motor coordination and executive func- guesses RL_CP (r=0.78, ρ=0.02) – and
number of errors, number of omissions, tions. omissions – omissions RL_CP (r=−0.78,
and the total number of right answered For this, the variables of interest that ρ=0.02) – for AD; and finally socialize –
chips (cf. ▶ Online Appendix A). eventually allowed us to obtain a cognitive socialize (r=0.91, ρ=0.03) for MCI – and
assessment of procedural memory include age – (r=0.86, ρ=0.01) for AD – as
total playing time, deviation from the el- examples of static variables); MAT (e.g.,
3.1.4 Workix Serious Game
ement to be tracked (i.e., percentage of time duration for HC – time_du-
Working memory, also known as short- deviation during the first trial), initial re- ration_RCL_CP2 (r=0.78, ρ=0.01) –
term memory, “registers and retains in- sponse time, and total time on track (i.e., and MCI; guesses and schooling years
coming information in a highly accessible total time over the ball during the second for MCI; etc.); IQCODE (e.g., time du-
form for a short period of time after the trial; idem for others trials). Each session ration for HC – time duration RL_CP

Methods Inf Med 4/2018 © Georg Thieme Verlag KG 2018


S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML 203

(r=0.62, p=0.46) – and socialize – frequency of use of games (r=−1, ρ=0), 3.2.3 External Validity
(r=0.77, ρ=0.51) – for AD); and finally frequency of use of ICT (r=−1, ρ=0),
CVLT (e.g., time duration for HC – and subjects’ age (r=−1, ρ=0), for HC The core module of the GEQ questionnaire
time_duration_R1_B1 (r=0.73, ρ=0.03) subjects and MMSE and IQCODE. [9, 41]was utilized to collect actual feelings

Pervasive Health
and time_duration_RCL_CP2 (r=0.96, Idem for MAT but direct correlation. and thoughts during game playing. In
ρ=0)). addition, we applied the Cronbach’s alpha
• Attentix: the higher correlations are ob-
3.2.2 Criterion Validity
algorithm [56] to measure the internal con-
served between the following game vari- sistency of results from this questionnaire
ables and MMSE, MAT and CVLT tests Several trials were run using all game vari- (cf. ▶ Table 3 and ▶ Online Appendix B,
results: average time – (r=−0.76, ρ= ables, or just a subset of them including Table 7 as supplementary information).
0.01) – and failures – (r=−0.69, ρ=0.05). those deemed as most representative (e.g., The main results obtained are:
Idem for IQCODE, but with a positive 100%, the best 50% or the best percentage • About their feelings – α=.85±1.79 for all
correlation. for each method). As a preliminary step to sample and α=.89±2.41 for participants
• Semantix: in this case, the best corre- apply the algorithm, we constructed the with cognitive problems –, a 93.4% per-
lation is observed for MCI patients: dataset and preprocessed it by interpolat- ceived the games as entertaining; a
average time (r=−1, ρ=0), time duration ing median values for the missing features, 93.4% as not boring; and a 93.4% as not
(r=−1, ρ=0), socialize (r=−1, ρ=0) and as well as normalizing them as required by being irritating or unpleasant.
age (r=−1, ρ=0) for both MMSE and SVM. • Regarding their thoughts – α=.61±1.8 in
MAT. A similar correlation is main- Due to the limitations caused by fast general and α=.40±1.49 for subjects

Downloaded by: Trinity College Dublin. Copyrighted material.


tained with CVLT (r=−0.98, ρ=0). Be- input stimuli, mainly in the case of AD with MCI and AD –, a 53.4% consider-
sides, we encountered proper positive subjects, Semantix, Procedurix and Episo- ed they played properly; and a 35.7%
correlation with number of points dix were the only games played by all 16 considered they had to work hard to
(r=−0.73, ρ=0), percentage of points users in the sample. Thus, the criterion/ play games.
over the total (r=−0.73, ρ=0), and so- prediction study focused on the analysis of
cialize (r=0.77, ρ=0) for AD people and these three games. The main results ob- Finally, we collected subject’s perception
IQCODE. tained are summarized in ▶ Table 2. before and after interacting with the games
• Workix: best correlations are obtained Considering the most sensitive par- developed. Firstly, the motivation to use vi-
for MMSE and MAT. Specifically, the ameter (i.e., false negatives) and after ap- deogames increases from 6.3% to 60%, and
time duration (r=0.60, ρ=0.001) for HC. plying all algorithms to the games indepen- secondly, the percentage of participants as-
By contrast, the correlation with IQ- dently and using all the variables, the most signing the highest score, to consider vi-
CODE test is poor. On the other hand, relevant result is a FN rate of 6.25% for deogames as a proper tool for cognitive
we found a good correlation with Semantix using LR and CART; 0% for evaluation, increases from 12.5% to 40%.
gender (r=−0.91, ρ=0) and schooling Procedurix using SVM; and 9.09% for Epi-
years (r=0.88, ρ=0). sodix using SVM. Repeating the same
• Procedurix: best correlations are found analysis but using only 50% of the most in- 4. Discussion and
for the time that the cursor stays on the formative variables, the false negative rate Conclusions
ball for the second – it2_TOB (r=0.80, is reduced to 6.25% for Semantix with LR
ρ=0.001) for MCI subjects and MMSE and SVM; 0% for Procedurix with CART This work discusses the design and evalu-
and it2_TOB (r=0.89, ρ=0.001) for AD and SVM; and finally, 4.55% for Episodix ation of the Panoramix serious game bat-
patients and IQCODE; and the devi- using SVM. In addition, using the optimal tery, which is aimed to assess the main
ation with respect to the center of the percentage of most informative variables, cognitive areas considered the most rel-
ball, for the first and second iterations similar results are obtained (i.e., 6.25% – evant diagnostic indicators of cognitive im-
(e.g., it1_deviation (r=−0.82, ρ=0.001) Semantix using the best 32% of data; 0% – pairments – episodic memory, attention,
for HC people and MAT). In the case of Procedurix using the best 42%; and 9.09% semantic memory, working memory, pro-
static variables, socialize (r=0.83, – Episodix using only the best 5.5.%). cedural memory and gnosias.
ρ=0.001) presents best correlation for Finally, the algorithms were applied Firstly, the design approach followed –
the HC group and MMSE. jointly to Procedurix, Semantix and Episo- gamification of existing classical tests or
• Gnosix: variables having a larger posi- dix, using all the variables captured from the emulation in the digital realm of tasks
tive correlation in this game are the the interaction with the three games – a clearly linked to the cognitive area to be
average general time (r=−0.90, ρ=0.003) data set of 89 instances/triplets. In this case, studied –, guarantees that the games evalu-
for MAT and MCI people; the reaction prediction is 100% correct for all algo- ate exactly the areas for which they were
time in the second task (r=0.60, ρ=0.03) rithms: LR, CART and SVM. In other designed. Furthermore, the focus group
for both MMSE and MAT, and inverse words, there were no false negatives. sessions led us to gather information on the
correlation for IQCODE. With regard specific needs by seniors that should be
to static data, the best results are with covered by the Panoramix suite. Thus, de-

© Georg Thieme Verlag KG 2018 Methods Inf Med 4/2018


204 S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML

Table 2 Classification of individuals using Panoramix.


Games Variables Algorithm TP+TNa FNb FPc F-measured F-measured
Accuracy (β=1) (β=1.5)
Semantix 100% LR 81.25% 6.25% 12.5% 0.86 0.87
Pervasive Health

Data set 100% CART 75% 6.25% 18.75% 0.82 0.85


(16 triples)
100% SVM 75% 12.5% 12.5% 0.80 0.80
Best 50% LR 75% 12.5% 12.5% 0.80 0.80
Best 50% CART 81,25% 6,25% 12,5% 0.86 0.87
Best 50% SVM 87.5% 6.25% 6.25% 0.90 0.90
Best % (32%) LR 81,25% 6,25% 12,5% 0.86 0.87
Best % (32%) CART 81,25% 6,25% 12,5% 0.86 0.87
Best % (32%) SVM 87.5% 6.25% 6.25% 0.90 0.90
Procedurix 100% LR 73.33% 11.67% 15% 0.73 0.74
Data set 100% CART 95% 1.67% 3.33% 0.95 0.96
(60 triples)
100% SVM 93.33% 0% 6.67% 0.94 0.96

Downloaded by: Trinity College Dublin. Copyrighted material.


Best 50% LR 90% 3.33% 6.67% 0.90 0.91
Best 50% CART 95% 0% 5% 0.95 0.97
Best 50% SVM 93.33% 0% 6.67% 0.94 0.96
Best % (42%) LR 80% 10% 10% 0.79 0.79
Best % (42%) CART 96.67% 0% 3.33% 0.97 0.98
Best % (42%) SVM 93.33% 0% 6.67% 0.94 0.96
Episodix 100% LR 50% 27.27% 22.73% 0.56 0.55
Data set 100% CART 50% 22.7% 22.3% 0.62 0.62
(22 triples)
100% SVM 81.82% 9.09% 9.09% 0.85 0.85
Best 50% LR 50% 27.27% 22.73% 0.56 0.55
Best 50% CART 50% 22.7% 22.3% 0.62 0.62
Best 50% SVM 90.91% 4.55% 4.55% 0.92 0.92
Best % (5.5%) LR 86.36% 9.09% 4.55% 0.88 0.87
Best % (5.5%) CART 77.27% 9.09% 13.64% 0.81 0.83
Best % (5.5%) SVM 77.27% 9.09% 13.64% 0.81 0.83
E+S +Pe 100%, Best 50% LR 100% 0% 0% 1.00 1.00
Data set & Best %
(89 triples) CART 100% 0% 0% 1.00 1.00
SVM 100% 0% 0% 1.00 1.00
aTP (True Positive) and TN (True negative) = rate of num. correct classifications or Accuracy
bFN (False Negative rate)
cFP (False Positive rate)
dF-measure (β) = is the weighted harmonic mean of precision and recall – its optimal value at 1 and its worst value at 0
eEpisodix + Semantix + Procedurix

Note: different subsets of most informative data are obtained using internally Random Forest (RF)

sign was carried out to adequately cover external validity of the battery. With re- CVLT – mainly with Episodix, but partially
the cognitive spectrum needed to detect spect to the first one, that is, the ability to with other games, such as Attentix or Pro-
MCI or AD. measure cognitive areas that Panoramix is cedurix. This is due to the existing inter-
With respect to the validation of this re- intended to measure, all games had statisti- relations among cognitive areas, which the
search, the pilot study gave a preliminary cally significant correlations with classical games would also make apparent. In gen-
assessment of the construct, criterion and tests (e.g., MMSE, MAT, IQCODE and eral terms, all games showed good corre-

Methods Inf Med 4/2018 © Georg Thieme Verlag KG 2018


S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML 205

lations – correlation factor r close to unity Table 3 Cronbach’s alpha scores for GEQ [9, 41].
and significance ρ<0.05. Especially relevant Parametersc Game experience factora α±SD α±SD α±SD
is the correlation obtained for MCI patients (all) (HC) (MCI/AD)b
using the Semantix game. In sum, the cor- Feelings Positive affect, Negative 0.85±1.75 0.38±0.71 0.89±2.41

Pervasive Health
relation analysis shows that the serious affect & Annoyance
games implemented cover the cognitive
Thoughts Competence & Challenge 0.61±1.8 * 0.4±1.49
areas for which they were designed in the
same way as their corresponding classical aImmersion and flow answers were discard, due to we have detected difficulties to understand by
tests do (i.e., construct validity). senior participants
bSubjects with cognitive problems were grouped, due sample size
Regarding the ability of classifying par-
ticipants according their cognitive status as ▶
c Please refer to Online Appendix B for further details about the perception questionnaire and
game experience factors assessment (e.g., feelings and thoughts)
healthy controls or people with cognitive
impairment (i.e., criterion validity), we
conducted a separate analysis on the pre- the classification accuracy of the Pano- as no incidences were detected related to
diction capabilities of Episodix, Semantix ramix suite, in line with previous studies the characteristics of the sample (i.e., genre,
and Procedurix, as all participants played [43, 44]. At the same time, most inform- educational level, training on ICT, etc.), or
these three games. As a first result, classifi- ative variables present a high correlation, environmental issues (i.e., natural location
cation offered by each game using all vari- which led us to confirm both construct and at participants’ homes), or behavioral ones
ables is correct in more than 75% of cases. criterion validity. For instance, time du- (i.e., not detecting any intrusive feeling in

Downloaded by: Trinity College Dublin. Copyrighted material.


Particularly relevant is the case of Pro- ration RCL_CP2 (5th most informative their daily activities).
cedurix using SVM, which predicts cor- variable) and time duration RCL_CP4 (8th On the other hand, results obtained
rectly 93.33% of subjects’ cognitive status, most informative variable) and socialize about the Panoramix battery, and taking as
resulting excellent F-values above 0.94, (1st); average time (1st) and time duration a reference the previous contributions dis-
both for equal precision and recall weights, (2nd) for Semantix; it2_TOB (2nd) and so- cussed in Sect. 1, contribute to the advance
and and for recall weight > precission cialize (1st) for Procedurix. the state of art in this field. Note that Pano-
weight. Concerning the external validity of ramix supports the concurrent evaluation
In general, prediction of each game im- Panoramix, the results obtained through os several cognitive areas, while previous
proves when we perform the analysis with the assessment of the experience of the par- contributions are focused on some specific
a set of most informative variables. Using ticipants were very positive, as a very high area. Furthermore, the main limitation of
the best 50%, classification results improve, percentage perceived the games as enter- previous studies was the lack of compre-
in particular for the three games using taining, not boring, and not irritating (i.e., hensive psychometric validation. In our
CART. In addition, results already provide 93.4% agree or strongly agree). Besides, case, evidence of the spychometric validity
promising prediction/classification levels about GEQ’s internal consistency, the best of Panoramix is provided from several per-
using just the best specific subset of vari- findings were related to feelings, both in spectives such as construct validity (i.e., ad-
ables for each method (i.e., Semantix = 32%, general and in subjects with cognitive dressed in 57.14% of the works analysed
Procedurix = 42% and Epiosdix = 5.5%), in problems (α = 0.89). By contrast, thoughts, except [19], [21], [24] and [25]); criterion
all cases but specially for CART and SVM. mainly in the HC group, show a low inter- validity (i.e., addressed only in [18], [21],
To sum up, individually, the best classifi- nal consistency, but we consider that they [24] and [25]); and finally, external validity
cation is performed by Procedurix, fol- are influenced by some difficulties detected (addressed only in the 14.29% of studies
lowed by Semantix and finally Episodix. in understanding and expressing opinions [20, 57]).
Finally, using all the variables captured about competence in video games; and With respect to the introduction of ma-
from the three games simultaneously, a also, an excessive negative perception chine learning techniques to detect cogni-
total correct classification is achieved. about their technological abilities. Even so, tive impairment as an innovative method
Those results are supported by the general results show a proper internal con- to study criterion validity, Panoramix pro-
F-measure values obtained, particularly sistency and GEQ led us to obtain adequate posess a novel approach beyond main-
when limiting the influence of false information about external validity of stream statistical techniques based on clas-
negatives (β=1.5). Panoramix. Other very relevant inform- sifier algorthims such as Support Vector
We can conclude that a traditional ation was the subject’s perception about Machines (SVM) and Classification and
analysis method such as LR offers a good technology. For that, we asked them about Regression Trees (CART). Besides, prelimi-
prediction, but it can be improved by ap- motivation to use games and also to use nary content, face and ecological validation
plying emerging machine learning meth- them as a tool to perform cognitive assess- was addressed in a previous work [30].
ods such as CART and specially SVM, ment. Both opinions dramatically improve Thus, we can conclude that Panoramix
which offers some better results. Moreover, after playing the games. Finally, there is has been subject to a comprehensive psy-
the different analysis approaches carried promising evidence about this tool being chometric validation process. Besides the
out show consistent results, which stresses applicable to the general senior population, validation aspects addressed, this research

© Georg Thieme Verlag KG 2018 Methods Inf Med 4/2018


206 S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML

extends our previous work [30] with addi- were played by all users in the sample. No therapy’s impact on trust in veterans with PTSD: a
pilot study. Bull Menninger Clin 2014; 78(4):
tional parameters such as the Pearson cor- matter the introduced machine learning 335–348.
relation for all games, the introduction of algorithms (i.e., LR, CART and SVM) pro- 3. Cordell CB, Borson S, Boustani M, Chodosh J,
additional quality metrics (e.g., accuracy; vided promising results, ,this future re- Reuben D, Verghese J, Thies W, Fried LB, Medi-
Pervasive Health

weighted harmonic mean of precision and search may also be extended to additional care Detection of Cognitive Impairment Work-
group. Alzheimer’s Association recommendations
recall, for both β=1 and β=1.5; and the machine learning models. A larger data set for operationalizing the detection of cognitive im-
complete confusion coefficient matrix), consequence of a larger population sample pairment during the Medicare Annual Wellness
and an assessment of the gaming experi- may facilitate the investigation of addi- Visit in a primary care setting. Alzheimer’s De-
ence using the Cronbach’s alpha scores for tional machine learning approaches that ment 2013; 9(2): 141–150.
4. Holtzman DM, Morris JC, Goate AM. Alzheimer’s
GEQ. better adapt to the new data model. disease: the challenge of the second century. Sci
Some limitations were detected during Finally, the first pilot study provided Transl Med 2011; 3(77): 77sr1.
the pilot study. Firstly, the sample size is evidence of the construct and external psy- 5. Parsons TD. Virtual Teacher and Classroom for
evidently small, no matter its design has chometric validity of the Panoramix bat- Assessment of Neurodevelopmental Disorders. In:
Brooks A, Brahnam S, Jain L, editors. Technologies
been rigorous from a methodological (i.e., tery, and also encouraging evidence about of Inclusive Well-Being. Berlin, Heidelberg:
cross-sectional study), morfological (i.e., the criterion validity of a combination of Springer; 2014. p. 121–137.
sample’s heterogeneity), and ethical/legal Semantix, Procedurix and Episodix, as data 6. Luciana M. Practitioner review: computerized as-
sessment of neuropsychological function in
perspectives (i.e., it was approved by the obtained from playing with these three
children: clinical and research applications of the
Galician institutional ethics committee and games correctly discriminated all healthy Cambridge Neuropsychological Testing Auto-
was conducted in accordance with the subjects from those already suffering AD

Downloaded by: Trinity College Dublin. Copyrighted material.


mated Battery (CANTAB). J Child Psychol Psy-
provisions of the Declaration of Helsinki, or MCI in the context of this first pilot chiatry 2003; 44(5): 649–663.
7. Hsu Y-L, Chung P-C, Wang W-H, Pai M-C, Wang
as revised in Seoul 2008). However, the study. Thus, no matter the limited sample, C-Y, Lin C-W, Wu H-L, Wang J-S. Gait and bal-
sample size is representative from a re- the potential of serious games and machine ance analysis for patients with Alzheimer’s disease
search perspective [34], taking into con- learning algorithms to support the early using an inertial-sensor-based wearable instru-
sideration the Spanish adult population. detection of dementia processes was con- ment. IEEE J Biomed Health Inform 2014; 18(6):
1822–1830.
Particularly, the selected sample represents firmed. Furthermore, discrimination was 8. Michael DR, Chen SL. Serious games: Games that
the target Spanish population within a performed in a nonintrusive, non-frustrat- educate, train, and inform. Muska & Lipman/Pre-
aceptable error margin for reserarch works ing way, which is especially relevant when mier-Trade; 2005.
(i.e., 0.05 < ε=0.075 < 0.10) [33, 34] and a targeting elder adults that may suffer cog- 9. Nap HH, Diaz-Orueta U, González MF, Lozar-
Manfreda K, Facal D, Dolničar V, Oyarzun D,
confidence interval of 95%, that can be nitive impairments. Ranga M-M, de Schutter B. Older people’s percep-
deemed as acceptable for a preliminary tions and experiences of a digital learning game.
study. Gerontechnology 2014; 13(3): 322–331.
Acknowledgment 10. Valladares-Rodriguez S, Pérez-Rodriguez R,
On the other hand, regarding usability,
Anido-Rifón L, Fernández-Iglesias M. Trends on
and no matter participants’ perceptions are The authors would like to thank the valu- the application of serious games to neuropsycho-
positive in general, in most cases they able comments and suggestions provided logical evaluation: A scoping review. J Biomed In-
would prefer tactile control devices, instead by experts during design phase of digital form 2016; 64: 296–319.
11. Plancher GG, Barra J, Orriols E, Piolino P. The in-
of keyboard and mouse. The velocity of games: Dr. Arturo X. Pereiro Rozas and
fluence of action on episodic memory: a virtual
stimulus presentation should also be re- Dr. Cristina Lojo Seoane from the Depart- reality study. Q J Exp Psychol (Hove) 2013; 66(5):
vised, mainly in the case of Attentix and ment of Developmental Psychology at the 895–909.
Workix, to facilitate their use by AD sub- University of Santiago de Compostela, and 12. Sauzéon H, Arvind Pala P, Larrue F, Wallet G,
Déjos M, Zheng X, Guitton P, N’Kaoua B. The use
jects. On the other hand, some data incon- M.D. M. Jose Moreno Carretero from the of Virtual Reality for episodic memory assessment:
sistencies were found in the answers for Neurology service at the Alvaro Cunqueiro Effects of active navigation. Exp Psychol 2012;
immersion and flow in one participant. University Hospital (Vigo, Galicia, Spain). 59(2): 99–108.
Eventually, both questions were discarded We also appreciate the support of the As- 13. Jebara N, Orriols E, Zaoui M, Berthoz A, Piolino P.
Effects of enactment in episodic memory: a pilot
due to the difficulties detected to under- sociation of Relatives of Alzheimer’s Pa- virtual reality study with young and elderly adults.
stand them by senior participants. Finally, tients and other Dementias of Galicia Front Aging Neurosci 2014; 6: 338.
additional research is needed to evaluate (AFAGA), which facilitated the selection of 14. Díaz-Orueta U, Garcia-López C, Crespo-Eguílaz
the complete psychometric properties of participants among its members. N, Sánchez-Carpintero R, Climent G, Narbona J,
Diaz-Orueta U, Garcia-Lopez C, Crespo-Eguilaz
this novel system of neuropsychological N, Sanchez-Carpintero R, Climent G, Narbona J.
evaluation, mainly in order to provide AULA virtual reality test as an attention measure:
clinical-grade evidence by means of an References Convergent validity with Conners Continuous
statistically significant population sample. Performance Test. Child Neuropsychol 2014;
1. Selnes OA. A Compendium of Neuropsychological 20(3): 328–342.
More specificall, the analysis of criterion Tests: Administration, Norms, and Commentary. 15. Iriarte Y, Diaz-Orueta U, Cueto E, Irazustabarrena
validity shall be extended to the complete Oxford: Oxford University Press; 1991. P, Banterla F, Climent G. AULA – Advanced vir-
Panoramix battery, as only three games 2. Williams W, Graham DP, McCurry K, Sanders A, tual reality tool for the assessment of attention:
Eiseman J, Chiu PH, King-Casas B. Group psycho-
(i.e., Episodix, Semantix and Procedurix)

Methods Inf Med 4/2018 © Georg Thieme Verlag KG 2018


S. Valladares-Rodriguez et al.: Detection of Cognitive Impairment Using Games and ML 207

Normative study in Spain. J Atten Disord 2016; plaints in primary care centres. Dement Geriatr 44. Lehmann C, Koenig T, Jelic V, Prichep L, John RE,
20(6): 542–568. Cogn Disord 2012; 33(4): 226–232. Wahlund L-O, Dodge Y, Dierks T. Application and
16. Shriki L, Weizer M, Pollak Y, Weiss PL, Rizzo AA, 30. Valladares-Rodriguez S, Perez-Rodriguez R, Facal comparison of classification algorithms for recog-
Gross-Tsur V. The utility of a continuous perform- D, Fernandez-Iglesias MJ, Anido-Rifon L, Mour- nition of Alzheimer’s disease in electrical brain ac-
ance test embedded in virtual reality in measuring iño-Garcia M. Design process and preliminary tivity (EEG). J Neurosci Methods 2007; 161(2):

Pervasive Health
the effectiveness of MPH treatment in boys with psychometric study of a video game to detect cog- 342–350.
ADHD. Harefuah 2010; 149(1): 18–23, 63. nitive impairment in senior adults. PeerJ 2017; 5: 45. Archer KJ, Kimes R V. Empirical characterization
17. Lee J-Y, Kho S, Yoo H Bin, Park S, Choi J-S, Kwon e3508. of random forest variable importance measures.
JS, Cha KR, Jung H-Y. Spatial memory impair- 31. Liamputtong P. Focus group methodology: Prin- Comput Stat Data Anal 2008; 52(4): 2249–2260.
ments in amnestic mild cognitive impairment in a ciple and practice. Sage Publications; 2011. 46. Pal M. Random forest classifier for remote sensing
virtual radial arm maze. Neuropsychiatr Dis Treat 32. UNESCO Institute for Statistics [Internet]. Inter- classification. Int J Remote Sens 2005; 26(1):
2014; 10: 653–660. national Standard Classification of Education: 217–222.
18. Atkins SM, Sprenger AM, Colflesh GJ, Briner TL, ISCED 2011. UIS, Montreal, Quebec; 2012. Avail- 47. Tulving E. Episodic memory: from mind to brain.
Buchanan JB, Chavis SE, Chen SY, Iannuzzi GL, able from: http://uis.unesco.org/sites/default/files/ Annu Rev Psychol 2002; 53(1): 1–25.
Kashtelyan V, Dowling E, Harbison JI, Bolger DJ, documents/international-standard-classification- 48. Albert MS, DeKosky ST, Dickson D, Dubois B,
Bunting MF, Dougherty MR. Measuring working of-education-isced-2011-en.pdf. Feldman HH, Fox NC, Gamst A, Holtzman DM,
memory is all fun and games: A four-dimensional 33. Hajian-Tilaki K. Sample size estimation in diag- Jagust WJ, Petersen RC, Snyder PJ, Carrillo MC,
spatial game predicts cognitive task performance. nostic test studies of biomedical informatics. J Thies B, Phelps CH. The diagnosis of mild cogni-
Exp Psychol 2014; 61(6): 417–438. Biomed Inform 2014; 48: 193–204. tive impairment due to Alzheimer’s disease: rec-
19. Hagler S, Jimison HB, Pavel M. Assessing execu- 34. Suresh KP, Chandrashekara S. Sample size esti- ommendations from the National Institute on
tive function using a computer game: computa- mation and power analysis for clinical research Aging-Alzheimer’s Association workgroups on
tional modeling of cognitive processes. IEEE J studies. J Hum Reprod Sci 2012; 5(1): 7. diagnostic guidelines for Alzheimer’s disease. Alz-
Biomed Health Inform 2014; 18(4): 1442–1452. 35. Barnett K, Mercer SW, Norbury M, Watt G, Wyke heimers Dement 2011; 7(3): 270–279.

Downloaded by: Trinity College Dublin. Copyrighted material.


20. Werner P, Rabinowitz S, Klinger E, Korczyn AD, S, Guthrie B. Epidemiology of multimorbidity and 49. Serino S, Pedroli E, Cipresso P, Pallavicini F, Alba-
Josman N. Use of the Virtual Action Planning implications for health care, research, and medical ni G, Mauro A, Riva G. The role of virtual reality
Supermarket for the Diagnosis of Mild Cognitive education: a cross-sectional study. Lancet 2012; in neuropsychology: The virtual multiple errands
Impairment. Dement Geriatr Cogn Disord 2009; 380(9836): 37–43. test for the assessment of executive functions in
27: 301–309. 36. Soyka F, Giordano PR, Barnett-Cowan M, Bulthoff Parkinson’s disease. Intell Syst Ref Libr 2014; 68:
21. van der Ham IJM, Faber AME, Venselaar M, van HH. Modeling direction discrimination thresholds 257–274.
Kreveld MJ, Löffler M. Ecological validity of vir- for yaw rotations around an earth-vertical axis for 50. Campos-Magdaleno M, Díaz-Bóveda R, Juncos-
tual environments to assess human navigation arbitrary motion profiles. Exp Brain Res 2012; Rabadán O, Facal D, Pereiro AX. Learning and ser-
ability. Front Psychol 2015; 6: 637. 220(1): 89–99. ial effects on verbal memory in mild cognitive im-
22. Serino S, Morganti F, Di Stefano F, Riva G. Detect- 37. Cockrell JR, Folstein MF. Mini-Mental State pairment. Appl Neuropsychol Adult 2015; 23(4):
ing early egocentric and allocentric impairments Examination (MMSE). Psychopharmacol Bull 1–14.
deficits in Alzheimer’s disease: An experimental 1987; 24(4): 689–692. 51. Howard D, Patterson KE. The Pyramids and Palm
study with virtual reality. Front Aging Neurosci 38. Rami L, Bosch B, Sanchez-Valle R, Molinuevo JL. Trees Test: A test of semantic access from words
2015; 7: 88. The memory alteration test (M@ T) discriminates and pictures. Thames Valley Test Company; 1992.
23. Canty AL, Fleming J, Patterson F, Green HJ, Man between subjective memory complaints, mild cog- 52. Schacter DL, Tulving E. Memory systems 1994.
D, Shum DHK. Evaluation of a virtual reality pros- nitive impairment and Alzheimer’s disease. Arch ilustrada. Mit Press; 1994.
pective memory task for use with individuals with Gerontol Geriatr 2010; 50(2): 171–174. 53. Kaplan E, Fein D, Morris R, Delis D. WAIS-R NI
severe traumatic brain injury. Neuropsychol Reha- 39. Delis DC, Kramer JH, Kaplan E, Thompkins BAO. Manual. San Antonio: Psychological Corporation;
bil 2014; 24(2): 238–265. CVLT: California Verbal Learning Test-Adult Ver- 1991.
24. Henry JD, Terrett G, Altgassen M, Raponi- sion: Manual.San Antonio, Texas: Psychological 54. Schank RC, Abelson RP. Scripts, plans, goals, and
Saunders S, Ballhausen N, Schnitzspahn KM, Ren- Corporation; 1987. understanding: An inquiry into human knowledge
dell PG. A Virtual Week study of prospective 40. Jorm AF. The Informant Questionnaire on cogni- structures. Psychology Press; 2013.
memory function in autism spectrum disorders. J tive decline in the elderly (IQCODE): a review. Int 55. Lezak MD. Neuropsychological assessment. Ox-
Exp Child Psychol 2014; 127: 110–125. Psychogeriatrics 2004; 16(3): 275–293. ford: Oxford University Press; 2004.
25. Banville F, Nolin P, Lalonde S, Henry M, Dery 41. IJsselsteijn WA, De Kort YAW, Poels K. The Game 56. Gliem JA, Gliem RR. Calculating, interpreting,
M-P, Villemure R. Multitasking and prospective Experience Questionnaire: Development of a self- and reporting Cronbach’s alpha reliability coeffi-
memory: Can virtual reality be useful for diag- report measure to assess the psychological impact cient for Likert-type scales. In: Midwest Research-
nosis? Behav Neurol 2010; 23(4): 209–211. of digital games. Forthcoming. to-Practice Conference in Adult, Continuing, and
26. AERA, APA, NCME. The Standards for Edu- 42. De’ath G, Fabricius KE. Classification and regres- Community Education; The Ohio State University,
cational and Psychological Testing. AERA; 2014. sion trees: a powerful yet simple technique for eco- Columbus: 2003.
27. Petersen RC. Mild cognitive impairment as a diag- logical data analysis. Ecology 2000; 81(11): 57. Plancher G, Tirard A, Gyselinck V, Nicolas S,
nostic entity. J Intern Med 2004; 256(3): 183–194. 3178–3192. Piolino P. Using virtual reality to characterize epi-
28. Facal D, Guàrdia-Olmos J, Juncos-Rabadán O. 43. Maroco J, Silva D, Rodrigues A, Guerreiro M, sodic memory profiles in amnestic mild cognitive
Diagnostic transitions in mild cognitive impair- Santana I, de Mendonça A. Data mining methods impairment and Alzheimer’s disease: influence of
ment by use of simple Markov models. Int J Ge- in the prediction of Dementia: A real-data com- active and passive encoding. Neuropsychologia
riatr Psychiatry 2015; 30(7): 669–676. parison of the accuracy, sensitivity and specificity 2012; 50(5): 592–602.
29. Juncos-Rabadán O, Pereiro AX, Facal D, Rodri- of linear discriminant analysis, logistic regression,
guez N, Lojo C, Caamaño JA, Sueiro J, Boveda J, neural networks, support vector machines, classifi-
Eiroa P. Prevalence and correlates of cognitive im- cation trees and random forests. BMC Res Notes
pairment in adults with subjective memory com- 2011; 4(1):299.

© Georg Thieme Verlag KG 2018 Methods Inf Med 4/2018

You might also like