Mobile Applications for Stroke: A Survey and a Speech Classiﬁcation

Approach

Ariella Richardson

, Shani Ben Ari

, Maayan Sinai

, Aviya Atsmon

, Ehud S. Conley

Yohai Gat

and Gil Segev

Lev Academic Center, Jerusalem, Israel

BGSegev Ltd. (segevlabs.org), Jerusalem, Israel

Keywords:

Digital Health, m-Health, Mobile, Stroke, Speech, Cardiovascular, Machine Learning, Data Mining.

Abstract:

Strokes are a cause of serious long-term disability and create an immense burden on healthcare. Among the

sea of mobile applications for health, some target stroke patients, and most require active user cooperation.

Our proposed application, collects data, without user intervention. We apply data mining methods to create

personal feedback to the patient or doctor. We provide a survey of applications for mobile or wearables,

speciﬁcally for stroke. We also survey papers that apply data mining to stroke. In addition to the survey, we

present a feasibility study on using speech for classiﬁcation of stroke patients. We created a new data set of

unstructured speech recordings, increasing applicability. We present experimental results on classiﬁcation of

stroke patients. Our study provides promising insight to detecting stroke patients using a mobile application

without requiring active user participation.

1 INTRODUCTION

Strokes produce immense health and economic bur-

dens and are a leading cause of serious long-term dis-

ability. Projections show that by 2030, there will be a

20.5% increase in prevalence from 2012 (Centers for

Disease Control and Prevention, USA, , 2009; Writ-

ing et al., 2016). Between 2012 and 2030, total direct

medical stroke-related costs are projected to triple,

from $71.6 billion to $184.1 billion, with the majority

of cost increase arising from those 65 to 79 years of

age (Ovbiagele et al., 2013).

Health related applications for mobile phones and

smart-watches are capable of improving health mon-

itoring and detection for general health and speciﬁ-

cally for stroke. The number of health related apps

available is astounding, approximating 40,000 apps in

2013 (Boulos et al., 2014) and 165,000 in 2015 (Terry,

2015). Applications that target stroke patients specif-

ically, and are aimed at managing risk factors (Seo

et al., 2015), rehabilitation (Zhang et al., 2015; Mi-

callef et al., 2016) and telemedicine (Nam et al., 2014;

Demaerschalk et al., 2012; Mitchell et al., 2011).

Other conditions are covered in several survey stud-

ies (Ozdalga et al., 2012; Boulos et al., 2014; Dobkin

and Dorsch, 2011; Patel et al., 2012; Pantelopoulos

and Bourbakis, 2010; King and Sarrafzadeh, 2017).

Everyday adoption of health apps is sometimes

compounded by factors such as confusion regarding

which app to use, slow adaptation of the traditional

healthcare community, the lack of integration with

electronic health records etc. (Crockett and Eliason,

2016; Terry, 2015). Despite these challenges, estab-

lishing patient self-monitoring with tools such as mo-

bile apps is important for improving patient health

(Dobkin and Dorsch, 2011; Writing et al., 2016). The

diversity of conditions that are covered ranges from

everyday diet apps (Recio-Rodriguez et al., 2016) to

critical oncology apps (de Bruin et al., 2015) and

touches on psychiatric symptoms (Place et al., 2017).

Among the surplus of medical apps, including

those that target stroke, many require the users to ac-

tively interact with the application in order to achieve

medical feedback, for example (Seo et al., 2015;

Zhang et al., 2015; Nam et al., 2014). In this study, we

demonstrate how an app will be able to provide mean-

ingful information, without requiring user actions for

data collection and input.

During regular phone usage, applications can col-

lect data that will be periodically analyzed and used

for monitoring patient health. Data can be collected

from a variety of sources, for example, sensors such

as the accelerometer or gyroscope, alongside key-

board usage. This data can be collected while the

Richardson, A., Ben Ari, S., Sinai, M., Atsmon, A., Conley, E., Gat, Y. and Segev, G.

Mobile Applications for Stroke: A Survey and a Speech Classiﬁcation Approach.

DOI: 10.5220/0007586901590166

In Proceedings of the 5th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2019), pages 159-166

ISBN: 978-989-758-368-1

159

device is used for regular unrelated tasks. This fact

enables us to collect data in a way that is transparent

to the user and does not impose on his daily phone us-

age. The application may run in the background and

then provide a service such as creating periodical re-

ports for self monitoring or for sharing with a health

provider. The data may also be used for applications

that detect and alert the user in case of an emergency

or deterioration (BGSEGEV, 2018).

We present a study that provides a proof of con-

cept for an application that performs stroke detec-

tion without the user having to actively provide input.

Our study, demonstrates stroke detection using data

mining performed on samples obtained with a smart-

phone/watch during regular use.

One of the main issues that affect stroke patients

is speech impairment. As speech impairments play

an important role in stroke detection and rehabili-

tation we decided to use data mining to determine

whether voice recordings can be classiﬁed as belong-

ing to stroke patients as opposed to healthy subjects.

We also study whether we can differentiate between

different types of speech impairments that may result

from having a stroke.

We study two types of speech impairments that

are caused by strokes: aphasia and dysarthria. Apha-

sia, is the loss or impairment of language (Berthier,

2005), whereas dysarthria is a speech neuro-motor

control problem (Sellars et al., 2005) (see Section 3.1

for more details). The data mining module that we

studied is proposed for use in an app for monitoring

stroke patients (patent request submitted (BGSEGEV,

2018)).

For classiﬁcation of speech impairments we use

voice recordings. Voice recordings are of interest as

they are obviously available for collection on mobile

phones. By using voice data generated throughout

normal phone usage (i.e. during phone calls) we are

able to develop an app that can monitor stroke patients

without intervention in daily activity. For example an

application could periodically record the users voice

during phone calls. To the best of our knowledge there

is no such application for stroke.

Deciding on the source and type of data to ana-

lyze is complex. Many studies used controlled exper-

iments where patients are asked to repeat the same

passage or some predetermined sentences or words

(Frid et al., 2014; Sakar et al., 2013) or only use vow-

els (Hazan et al., 2012; Sakar et al., 2013). Our study

uses a different, more general approach. We studied

voice recordings without limiting the nature of the

recording. In our recordings there is a mix of free

speech and predeﬁned speech tasks.

As part of this study, in order to study free speech,

for stroke detection we created a new data set. Our

data is obtained from a variety of online sources. The

criteria for video collection was that they are labeled

as belonging to stroke patients. We extracted the

voice signal from the videos we selected, and use it

to build our data set.

The data is recorded under a variety of conditions

and there is no known structure to what the subjects

say. Although this might sound like a strange strategy,

the motivation to using this type of data is that in a

real application subjects are expected to speak freely,

and classiﬁcation should be possible without the bias

of a predeﬁned set of words. Although our use of

free speech makes the classiﬁcation problem harder, it

has the advantage of higher applicability to real world

data collected from a phone.

We performed classiﬁcation of the speech data us-

ing several algorithms as detailed in Section 3.2. Our

successful comparative study provides promising in-

sight to detecting stroke patients from normal phone

usage. The results of this study demonstrate that we

can differentiate between stroke patients to healthy

subjects. The voice samples are of short recordings

enabling the detection of stroke in real-time. Further-

more, we successfully differentiate between the two

stroke conditions of aphasia and dysarthria. The abil-

ity to perform a ﬁner classiﬁcation could be useful in

monitoring stroke rehabilitation.

The contributions of this study are:

• We present a broad survey on stroke related apps

and on data mining for stroke.

• We build a data set of unrestricted speech of stroke

patients. This data set can be used in further stud-

ies, and also demonstrates the strengths of using

unstructured data.

• We compare several models for classiﬁcation of

speech and show that we successfully classify

stroke patients vs. healthy subjects, along with

the ability to differentiate between different stroke

speech conditions.

The structure of the paper is as follows: We ﬁrst

present a broad survey on mobile apps and data min-

ing for stroke. Next we present our study on classiﬁ-

cation of stroke speech impairments. We describe the

data set we built alongside an experimental evaluation

of classiﬁcation methods and a discussion of results.

Finally, we conclude our paper.

2 RELATED WORK SURVEY

This section presents a survey on applications and

data mining for stroke. We begin by discussing mo-

ICT4AWE 2019 - 5th International Conference on Information and Communication Technologies for Ageing Well and e-Health

160

bile applications and other technologies and proceed

to cover studies on data mining for stroke.

2.1 Applications

We take look at applications targeting stroke patients

speciﬁcally. Mobile apps can be used for monitoring

and rehabilitation of patients after stroke. (Seo et al.,

2015) tested the feasibility of a mobile app for pa-

tients who had suffered a stroke. The app was aimed

at managing risk factors for stroke such as blood pres-

sure and diabetes management. The study was aimed

at testing adherence to app usage and concluded that

more work must be done in order to encourage adher-

ence. Another application, (Zhang et al., 2015), ac-

companies the patient by encouraging exercises, fol-

lowing up on taking pills, and logging mood reports.

(Micallef et al., 2016) introduced another exercise ori-

ented application for post stroke patients. This ap-

plication is aimed to help patients remember to exer-

cise more frequently. The application was evaluated

on a smartphone, tablet and smart-watch. The au-

thors found that stroke survivors seem to prefer smart-

phones compared to other mobile devices due to their

ease of use, usability, familiarity and being easier to

handle with one arm. An interesting study, (Beeson

et al., 2013) and views the cell phone as a device for

writing on, this case study demonstrates how writing

can be used as treatment for aphasia in a post stroke

patient.

Some of the apps are aimed at bridging the dis-

tance between the patients to medical assistance (Nam

et al., 2014; Demaerschalk et al., 2012; Mitchell et al.,

2011). (Nam et al., 2014) provide a stroke screen-

ing application. The application shows a set of car-

toons representing stroke symptoms. Potential pa-

tients can follow the cartoons and try to determine

whether they may be suffering from a stroke. Sugges-

tions of nearby hospitals providing appropriate treat-

ment are provided by the app. (Mitchell et al., 2011)

bridge the gap by providing a tele-radiology system

that enables a doctor to interpret a CT scan. This en-

ables diagnosis when the hospital does not have a ex-

pert on call. (Demaerschalk et al., 2012) introduce a

similar app that provides high-quality video telecon-

ferencing. Diagnosis is the aim of the app presented

in (Shin et al., 2012). Their app uses a smartphone

to perform the pronator drift test that is used to diag-

nose stroke. Mobile phones are tied on to the patients

wrists. The app uses the accelerometer in order to

measure changes in drift and test arm weakness.

Aside from discussing mobile phone apps we

must also consider other related technologies for

stroke. We refer the interested reader to this survey:

(Nam et al., 2013). The survey covers existing tech-

nologies applied to stroke patients (some appear in

studies described above), these include:

• Remote diagnosis by doctors watching video and

audio recorded from stroke patients (Roine et al.,

2001; Demaerschalk et al., 2012).

• Teleradiology - doctors can view and interpret CT

scans from afar (Mitchell et al., 2011; Park and

Nam, 2009).

• Pre-hospital notiﬁcation arrival, enabling the hos-

pital to prepare for the patient and be ready for

treatment on arrival. Or a mobile stroke unit con-

nected to hospital to administer treatment before

arrival (Gonzalez et al., 2011; Kim et al., 2009).

• Communication between stroke team members

within the healthcare facility (Nam et al., 2007).

• A decision support system for stroke classiﬁcation

(Nam et al., 2012).

• Tele-rehabilitation for rehabilitation of stroke pa-

tients (Krpic et al., 2013).

These studies demonstrate various types of apps

that have been developed for stroke patients. They all

require patient input and active participation. This is

in contrast to our proposed approach where the mon-

itoring is performed without user intervention.

2.2 Data Mining

Data mining in healthcare is a topic of substantial in-

terest and has even been described as ”increasingly

popular, if not increasingly essential” (Koh et al.,

2011). However, despite the advances in healthcare,

data mining incorporation into everyday healthcare

practice is slow (Crockett and Eliason, 2016). Exam-

ples to domains within healthcare where Data min-

ing is used are Parkinson’s (Little et al., 2009; Tsanas

et al., 2010; Hazan et al., 2012; Sakar et al., 2013), oc-

cupational therapy (Richardson et al., 2008) and Med-

ical Imaging (Guo et al., 2016).

The previous section reviewed various applica-

tions for stroke. Many of these applications collect

and analyze data. This section will discuss various

studies that apply data mining methods to stroke data.

Some of the studies we cover acquired data using mo-

bile applications, some use data from other sources.

It is important to note, that the data mining studies we

found on stroke were very different to our study.

Brain CT scans are of major importance in stroke

detection, and as shown above, received attention

in mobile applications that enable viewing the scans

from afar (Mitchell et al., 2011; Demaerschalk et al.,

2012). Bently et al. (Bentley et al., 2014) take the

Mobile Applications for Stroke: A Survey and a Speech Classiﬁcation Approach

161

next step and explore whether machine learning can

be applied to brain scans in order to automatically

detect thrombolysis. They use SVMs to differenti-

ate between patients who developed symptomatic in-

tracranial hemorrhage to those who did not. The im-

age voxels were used as feature vectors. Results us-

ing 10-Fold cross validation are reported reaching an

AUC of up to 0.744. This study provides a ﬁrst step in

assisting doctors by automatic analysis of CT scans.

Khosla et al. (Khosla et al., 2010) use data from

the cardiovascular Health Study data set. Their study

pays special attention to the issue of missing values

in health records. They compare Cox regression to

SVM and Margin Based Sensor Regression (MCR)

for predicting stroke. 10-fold cross validation is used

on 5 random trials, average AUC is reported as high

as 0.774 for SVM, and 0.777 for MCR, while Cox

regression only reaches 0.747. A broad discussion of

variations on features selected and appropriate results

is described in the paper.

A slightly different type of study is presented by

(Mans et al., 2008). They use two sets of data be-

longing to stroke patients. The ﬁrst refers to the clini-

cal course of stroke patients during hospital stay. The

second, refers to pre-hospital behavior. Both data sets

are used to perform Process Mining in order to con-

struct process models. Interesting results included de-

tecting differences treatment strategies between hos-

pitals, and causes of delay in arriving at hospital for

treatment - critical to stroke outcome.

3 STROKE SPEECH

CLASSIFICATION STUDY

In this section we present the study we performed on

classiﬁcation of stroke speech samples. The study

shows how unstructured speech can be used for the

detection of stroke patients, and for the differentia-

tion between speech impairments that are common in

stroke. The results provide a step towards integrating

data mining into an application for stroke detection or

management in an unobtrusive manner. We describe

the data set that we built, the experimental setup used

and ﬁnally, results are presented and discussed.

3.1 Data Collection

In order to increase the generalization of our study,

we build a new data base. We use data found freely

on public sources. The variability in an uncontrolled

setting raises the applicability of our results. Our

data set was obtained from videos of stroke patients

with aphasia or dysarthria. Aphasia is the loss or im-

pairment of language caused by brain damage. It is

one of the most devastating cognitive impairments of

stroke (Berthier, 2005). Aphasia is present in 21 -

38% of acute stroke patients and is associated with

high morbidity, mortality and expenditure. Dysarthria

is a speech problem which can be caused by a num-

ber of brain disorders including conditions such as

stroke and head injury (Sellars et al., 2005). Typical

features of dysarthria include slurring of speech and

quiet voice volume. Psychological distress is often

experienced by people with dysarthria.

The videos we collected include speech from both

stroke patients and healthy subjects. We extracted

the audio from the videos, as segments of unin-

terrupted speech. Each segment is labeled stroke-

aphasia, stroke-dysarthria or healthy. Although we

have labeling of the data from the notations provided

with the videos, the patients may suffer from different

intensities of the labeled condition. The advantage of

using the data from the same videos for healthy and

stroke subjects is that they all come from the same dis-

tribution of background noise and quality. We were

careful to note the patient ID for each audio sample,

this is important in the experimental setup. The exact

ages of our subjects is unknown, but the videos show

that they range from approximately 15 to 70. Our data

set consists of 16 stroke patients, of whom 8 are la-

beled with aphasia, and 8 are labeled with dysarthria

and 12 healthy subjects. 17 subjects are female and

11 male. For every patient there are many samples,

as we cut the videos into segments of uninterrupted

speech. In total there are There are 269 segments from

healthy subjects, 1902 segments for stroke subjects,

split as 987 aphasia and 915 dysarthria. The lengths

of audio vary from several seconds to several minutes.

Some of the samples were too small to generate fea-

tures from as they contained almost silent parts of the

speech. These samples were discarded.

3.2 Experimental Setup

We ran two sets of experiments, results are described

below in Section 3.3. The ﬁrst experiment is aimed at

classifying stroke patients in the general population.

The data is split into two sets, and labeled stroke or

healthy. The second experiment uses data belonging

to stroke patients (no healthy samples) and is split into

two sets, labeled aphasia or dysarthria and differen-

tiates between the two speech impairments.

For all sets of experiments we used 4-fold cross

validation. Meaning, that we divided our data into 4

even sets, and ran 4 sets of tests. In each test we used

3 sets of the data for training and 1 set for testing. We

ICT4AWE 2019 - 5th International Conference on Information and Communication Technologies for Ageing Well and e-Health

162

did this 4 times, each time leaving out a different set

for testing. The results presented are the average of 4

runs. Although it is usually standard practice to use

10 folds, we chose 4 as this provided a good split for

the numbers of subjects we had in the various classes.

It is very important to note that when we split the data

we were careful that all the data belonging to a single

subject was kept in the same group. For example if

the data from subject A is in the testing data, then all

the data from this subject is in the test, none of it in

the training data. This is a point often unnoticed when

using cross validation, experiments are often run with

random cross validation, thus contaminating the pro-

cess. The results must reﬂect our ability to diagnose a

new patient, meaning a patient that has not been sam-

pled, and is not used for the model creation. This is

an important restriction to adhere to.

Each audio segment is used for a single vector in

the data set. We use signal processing methods in

MATLAB as suggested in (Bunkheila, 2018). We ex-

tract standard audio features from the audio signal to

create the input vector for our classiﬁcation algorithm.

Some of the features are: mean, median, standard de-

viation, skewness, kurtosis, Shannon’s entropy, spec-

tral entropy, dominant frequency (value, magnitude

and ratio), wavelet features and Mel-frequency cep-

stral coefﬁcients.

We report on the the following algorithms: (im-

plemented in Mathworks Classiﬁcation Learner App.)

Decision Trees (C4.5) (Quinlan, 1993), K- Nearest

Neighbor (KNN) (Altman, 1992), Logistic regres-

sion (Walker and Duncan, 1967), Support Vector Ma-

chines (SVM) (Platt, 1999) and AdaBoost (Freund

et al., 1999) with decision tree learners. Evalua-

tion is presented using several measures: Accuracy

is the ﬁrst measure, it simply presents the percentage

of correctly classiﬁed samples, from both classes of

the data. Although this is an informative measure, it

does not always provide a complete evaluation of the

model. Precision and Recall provide the next layer of

evaluation. Precision deﬁnes the number of correctly

classiﬁed positive samples from among the samples

classiﬁed as being positive, In terms of our work, how

many samples of patients that are classiﬁed as having

stroke, actually had a stroke. Recall is the the number

of correctly classiﬁed positive samples from among

the samples labeled positive. In our case, from among

the samples belonging to patients that had a stroke

how many did we ﬁnd. As there often exists a trade-

off between precision to Recall two other measures

are commonly used. The F1-measure (F1 Score) is

the weighted average of Precision and Recall. The last

measure we use is the Area under the curve (AUC)

that considers the trade-off between the Recall to the

False positive rate (the proportion of negative samples

(healthy) that are mistakenly considered as positive

(stroke)). Measures run between 0 to 1 (1 is better).

3.3 Experimental Results and

Discussion

The ﬁrst set of experiments uses data from stroke pa-

tients and from healthy patients and builds a classiﬁ-

cation model for detection of stroke patients. Results

are shown in Table 1. Each row describes a classi-

ﬁcation model, and results for the various measures

(described above) are displayed across columns.

As shown in Table 1 the best results were obtained

using AdaBoost. These results show clearly that

stroke patients can be classiﬁed from voice record-

ings. A closer look at the results shows variations

in classiﬁcation ability between models. For exam-

ple KNN (k=10) has a recall of 1 meaning all stroke

samples were discovered. However this comes at a

price, the precision in this case is only 0.88, mean-

ing that some healthy samples are mistakenly diag-

nosed as stroke. These differences are important

when transferring the academic study to technology,

as is done when using this technology in a proof-

of-concept (POC) (BGSEGEV, 2018). For each spe-

ciﬁc application it is important to decide whether one

prefers missing a few stroke samples over diagnosing

healthy samples as stroke, and selecting the appropri-

ate algorithm accordingly.

Our second experiment uses data only from stroke

but labels them as suffering from aphasia or dysarthria

(see Section 3.1 for an explanation) and builds a clas-

siﬁcation model for classiﬁcation of these two impair-

ments. Results are shown in Table 2. Aphasia is la-

beled as positive, dysarthria as negative. Results show

the highest accuracy for the Decision Tree. The other

algorithms have lower accuracy rates. Taking a look

at the result table provides insight as to why. For most

algorithms presented the precision and the AUC are

high but the recall is low. The decision tree does not

suffer from this problem and as can be seen precision,

recall and AUC are all high. In this example the F1

measure is perhaps the best measure to use as it cap-

tures the trade-off between recall and precision and

clearly singles out the decision tree as building the

best model for this data.

What the results mean is that when the models

detect aphasia it is actually with a high probability

aphasia (tagged as positive). However, a large num-

ber of the aphasia samples are classiﬁed as dysarthria.

This may explained by the fact that the conditions of-

ten overlap, and some practitioners consider aphasia

more severe. The milder cases of aphasia may be sim-

Mobile Applications for Stroke: A Survey and a Speech Classiﬁcation Approach

163

Table 1: Results of classiﬁcation for stroke vs. healthy.

Accuracy Precision Recall F1 AUC

Decision Tree 0.87 0.88 0.98 0.93 0.77

KNN(k=1) 0.89 0.90 0.99 0.94 0.80

KNN(k=10) 0.88 0.88 1.00 0.94 0.88

Logistic Regression 0.87 0.89 0.97 0.93 0.82

SVM 0.89 0.90 0.99 0.94 0.90

AdaBoost (tree) 0.94 0.96 0.96 0.96 0.96

Table 2: Results of classiﬁcation for aphasia vs. dysarthria.

Accuracy Precision Recall F1 AUC

Decision Tree 0.95 0.96 0.95 0.95 0.97

KNN(k=1) 0.86 0.93 0.78 0.84 0.97

KNN(k=10) 0.86 0.93 0.78 0.84 0.99

Logistic Regression 0.87 0.94 0.79 0.85 0.98

SVM 0.87 0.94 0.79 0.85 0.99

AdaBoost (tree) 0.88 0.96 0.80 0.85 0.92

ilar to dysarthria causing the model to confuse them.

These experiments demonstrate that data collected

from an unrestricted environment, where speech is

not limited to the use of vowels, or predeﬁned pas-

sages can be used for the classiﬁcation of stroke or

even speciﬁc speech impairments. This step is one

step needed towards providing a mobile application

for the detection or monitoring of stroke in an unob-

trusive manner. During usual phone use, given ad-

vance permission by the user, any speech collected on

the phone such as phone calls can be used for detec-

tion of stroke. This could for example be very useful

for high risk patients, and automatically emergency

contacts if deterioration in speech is detected. Aside

from the emergency usage a mobile application could

be used to monitor progression during rehabilitation.

These results motivate future expansion of the data set

to other impairments, other diseases etc.

4 CONCLUSION

We presented a study on the use of mobile applica-

tions for stroke. We performed a broad survey on

studies related to the use of mobile applications for

stroke. We also surveyed work on the use of data min-

ing for stroke.

Aside from the survey study, we present an exper-

imental study on the use of speech samples for the

classiﬁcation of stroke. This study provides a step

towards building a mobile application for stroke de-

tection and monitoring such as those developed by

(BGSEGEV, 2018). Our study provides a proof of

concept for an app that uses non intrusive data collec-

tion for medical detection (in this case stroke). As part

of the experimental study we built a database of voice

samples. The data base is special as it uses free speech

collected in unrestricted environments and provides a

good source for studying real data.

Our experiments were run in two settings. One

used to classify stroke patients from a group of

healthy subjects. And the other to differentiate be-

tween two types of stroke speech impairments aphasia

and dysarthria. Both experiments show high success

rates and indicate that using free speech samples in

unrestricted environments for detection is feasible.

Future work will look at expanding the data base

and enriching it with other types of speech impair-

ments, and forwarding the classiﬁcation study on the

new data. We will also study other types of sensor

data such as typing and walking. As we expand our

work we will also look at other cardiovascular dis-

eases such as Peripheral vascular diseases.

ACKNOWLEDGEMENTS

First and foremost we thank BGSegev Ltd.

(SegevLabs.org) for leading the technological

aspects of this study. Thanks to Shani Hilkiya,

Odaya Krombein, Sara Tayer, Chen Danino, Yehudit

Matzliach and Ruchama Amsalem, students from the

Lev Academic Center, for their assistance. We also

thank Gabriele Bunkheila, Ronen Cohen and Roy

Fahn from Systematics MathWorks for their support

and guidance in adapting tools for this study Website:

http://www.sunrise-setting.co.uk

ICT4AWE 2019 - 5th International Conference on Information and Communication Technologies for Ageing Well and e-Health

164

REFERENCES

Altman, N. S. (1992). An introduction to kernel and nearest-

neighbor nonparametric regression. The American

Statistician, 46(3):175–185.

Beeson, P. M., Higginson, K., and Rising, K. (2013).

Writing treatment for aphasia: A texting approach.

Journal of Speech, Language, and Hearing Research,

56(3):945–955.

Bentley, P., Ganesalingam, J., Jones, A. L. C., Mahady, K.,

Epton, S., Rinne, P., Sharma, P., Halse, O., Mehta, A.,

and Rueckert, D. (2014). Prediction of stroke throm-

bolysis outcome using ct brain machine learning. Neu-

roImage: Clinical, 4:635–640.

Berthier, M. L. (2005). Poststroke aphasia. Drugs & aging,

22(2):163–182.

BGSEGEV (2018). BGSegev STROKETOR - stroke detec-

tor. https://stroketor.segevlabs.org/.

Boulos, M. N. K., Brewer, A. C., Karimkhani, C., Buller,

D. B., and Dellavalle, R. P. (2014). Mobile medical

and health apps: state of the art, concerns, regula-

tory control and certiﬁcation. Online journal of public

health informatics, 5(3):229.

Bunkheila, G. (retrieved Jan 2018). Signal processing

and machine learning techniques for sensor data an-

alytics. https://www.mathworks.com/videos/signal-

processing-and-machine-learning-techniques-for-

sensor-data-analytics-107549.html.

Centers for Disease Control and Prevention, USA, (2009).

Prevalence and most common causes of disability

among adults–united states, 2005. MMWR: Morbidity

and Mortality weekly report, 58(16):421–426.

Crockett, D. and Eliason, B. (2016). What is data mining in

healthcare. Insights: Health Catalyst.

de Bruin, J. S., Schuh, C., Seeling, W., Luger, E., Gall, M.,

utterer, E., Kornek, G., Ludvik, B., Hoppichler, F.,

and Schindler, K. (2015). Assessing the feasibility

of a mobile health-supported clinical decision support

system for nutritional triage in oncology outpatients

using arden syntax. Artiﬁcial intelligence in medicine.

Demaerschalk, B. M., Vegunta, S., Vargas, B. B., Wu, Q.,

Channer, D. D., and Hentz, J. G. (2012). Reliability

of real-time video smartphone for assessing national

institutes of health stroke scale scores in acute stroke

patients. Stroke, 43(12):3271–3277.

Dobkin, B. H. and Dorsch, A. (2011). The promise of

mhealth: daily activity monitoring and outcome as-

sessments by wearable sensors. Neurorehabilitation

and neural repair, 25(9):788–798.

Freund, Y., Schapire, R., and Abe, N. (1999). A short in-

troduction to boosting. Journal-Japanese Society For

Artiﬁcial Intelligence, 14(771-780):1612.

Frid, A., Safra, E. J., Hazan, H., Lokey, L. L., Hilu, D.,

Manevitz, L., Ramig, L. O., and Sapir, S. (2014).

Computational diagnosis of parkinson’s disease di-

rectly from natural speech using machine learning

techniques. In Software Science, Technology and En-

gineering (SWSTE), 2014 IEEE International Confer-

ence on, pages 50–53. IEEE.

Gonzalez, M. A., Hanna, N., Rodrigo, M. E., Satler, L. F.,

and Waksman, R. (2011). Reliability of prehospital

real-time cellular video phone in assessing the simpli-

ﬁed national institutes of health stroke scale in patients

with acute stroke. Stroke, 42(6):1522–1527.

Guo, X., Yu, Q., Li, R., Alm, C. O., Calvelli, C., Shi, P., and

Haake, A. (2016). Intelligent medical image grouping

through interactive learning. International Journal of

Data Science and Analytics, 2(3-4):95–105.

Hazan, H., Hilu, D., Manevitz, L., Ramig, L. O., and Sapir,

S. (2012). Early diagnosis of parkinson’s disease via

machine learning on speech data. In Electrical &

Electronics Engineers in Israel (IEEEI), 2012 IEEE

27th Convention of, pages 1–4. IEEE.

Khosla, A., Cao, Y., Lin, C. C.-Y., Chiu, H.-K., Hu, J.,

and Lee, H. (2010). An integrated machine learn-

ing approach to stroke prediction. In Proceedings of

the 16th ACM SIGKDD international conference on

Knowledge discovery and data mining, pages 183–

192. ACM.

Kim, D.-K., Yoo, S. K., Park, I.-C., Choa, M., Bae,

K. Y., Kim, Y.-D., and Heo, J.-H. (2009). A mobile

telemedicine system for remote consultation in cases

of acute stroke. Journal of telemedicine and telecare,

15(2):102–107.

King, C. E. and Sarrafzadeh, M. (2017). A survey of

smartwatches in remote health monitoring. Journal

of Healthcare Informatics Research, pages 1–24.

Koh, H. C., Tan, G., et al. (2011). Data mining applica-

tions in healthcare. Journal of healthcare information

management, 19(2):65.

Krpic, A., Savanovic, A., and Cikajlo, I. (2013). Telere-

habilitation: remote multimedia-supported assistance

and mobile monitoring of balance training outcomes

can facilitate the clinical staff’s effort. International

Journal of Rehabilitation Research, 36(2):162–171.

Little, M. A., McSharry, P. E., Hunter, E. J., Spielman, J.,

Ramig, L. O., et al. (2009). Suitability of dysphonia

measurements for telemonitoring of parkinson’s dis-

ease. IEEE transactions on biomedical engineering,

56(4):1015–1022.

Mans, R., Schonenberg, H., Leonardi, G., Panzarasa, S.,

Cavallini, A., Quaglini, S., and van der AALST, W.

(2008). Process mining techniques: an application to

stroke care. Studies in health technology and infor-

matics, 136:573.

Micallef, N., Baillie, L., and Uzor, S. (2016). Time to ex-

ercise!: an aide-memoire stroke app for post-stroke

arm rehabilitation. In Proceedings of the 18th Inter-

national Conference on Human-Computer Interaction

with Mobile Devices and Services, pages 112–123.

ACM.

Mitchell, J. R., Sharma, P., Modi, J., Simpson, M., Thomas,

M., Hill, M. D., and Goyal, M. (2011). A smartphone

client-server teleradiology system for primary diagno-

sis of acute stroke. Journal of medical Internet re-

search, 13(2).

Nam, H., Cha, M.-J., Kim, Y., Kim, E., Park, E., Lee,

H., Nam, C., and Heo, J. (2012). Use of a hand-

held, computerized device as a decision support tool

Mobile Applications for Stroke: A Survey and a Speech Classiﬁcation Approach

165

for stroke classiﬁcation. European journal of neurol-

ogy, 19(3):426–430.

Nam, H. S., Han, S. W., Ahn, S. H., Lee, J. Y., Choi, H.-Y.,

Park, I. C., and Heo, J. H. (2007). Improved time in-

tervals by implementation of computerized physician

order entry-based stroke team approach. Cerebrovas-

cular Diseases, 23(4):289–293.

Nam, H. S., Heo, J., Kim, J., Kim, Y. D., Song, T. J., Park,

E., and Heo, J. H. (2014). Development of smartphone

application that aids stroke screening and identifying

nearby acute stroke care hospitals. Yonsei medical

journal, 55(1):25–29.

Nam, H. S., Park, E., and Heo, J. H. (2013). Facilitating

stroke management using modern information tech-

nology. Journal of stroke, 15(3):135.

Ovbiagele, B., Goldstein, L. B., Higashida, R. T., Howard,

V. J., Johnston, S. C., Khavjou, O. A., Lackland, D. T.,

Lichtman, J. H., Mohl, S., Sacco, R. L., et al. (2013).

Forecasting the future of stroke in the united states.

Stroke, 44(8):2361–2375.

Ozdalga, E., Ozdalga, A., and Ahuja, N. (2012). The smart-

phone in medicine: a review of current and potential

use among physicians and students. Journal of medi-

cal Internet research, 14(5).

Pantelopoulos, A. and Bourbakis, N. G. (2010). A survey on

wearable sensor-based systems for health monitoring

and prognosis. IEEE Transactions on Systems, Man,

and Cybernetics, Part C (Applications and Reviews),

40(1):1–12.

Park, E. and Nam, H. S. (2009). A service-oriented medical

framework for fast and adaptive information delivery

in mobile environment. IEEE transactions on infor-

mation technology in biomedicine, 13(6):1049–1056.

Patel, S., Park, H., Bonato, P., Chan, L., and Rodgers, M.

(2012). A review of wearable sensors and systems

with application in rehabilitation. Journal of neuro-

engineering and rehabilitation, 9(1):21.

Place, S., Blanch-Hartigan, D., Rubin, C., Gorrostieta, C.,

Mead, C., Kane, J., Marx, B. P., Feast, J., Deckers-

bach, T., et al. (2017). Behavioral indicators on a

mobile sensing platform predict clinically validated

psychiatric symptoms of mood and anxiety disorders.

Journal of medical Internet research, 19(3).

Platt, J. (1999). Using sparseness and analytic QP to speed

training of support vector machines. In Advances in

Neural Information Processing Systems, pages 557–

563.

Quinlan, J. R. (1993). C4.5: Programs for Machine Learn-

ing. Morgan Kaufmann Publishers Inc., San Fran-

cisco, CA, USA.

Recio-Rodriguez, J. I., Agudo-Conde, C., Martin-

Cantera, C., Gonz

alez-Viejo, M. N., Fernandez-

Alonso, M. D. C., Arietaleanizbeaskoa, M. S.,

Schmolling-Guinovart, Y., Maderuelo-Fernandez, J.-

A., Rodriguez-Sanchez, E., Gomez-Marcos, M. A.,

et al. (2016). Short-term effectiveness of a mobile

phone app for increasing physical activity and adher-

ence to the mediterranean diet in primary care: A ran-

domized controlled trial (evident ii study). Journal of

medical Internet research, 18(12).

Richardson, A., Kraus, S., Weiss, P. L., and Rosenblum,

S. (2008). COACH - cumulative online algorithm for

classiﬁcation of handwriting deﬁciencies. In IAAI’08

Proceedings of the 20th national conference on In-

novative applications of artiﬁcial intelligence, pages

1725–1730.

Roine, R., Ohinmaa, A., and Hailey, D. (2001). Assess-

ing telemedicine: a systematic review of the literature.

Canadian Medical Association Journal, 165(6):765–

771.

Sakar, B. E., Isenkul, M. E., Sakar, C. O., Sertbas, A., Gur-

gen, F., Delil, S., Apaydin, H., and Kursun, O. (2013).

Collection and analysis of a parkinson speech dataset

with multiple types of sound recordings. IEEE Jour-

nal of Biomedical and Health Informatics, 17(4):828–

834.

Sellars, C., Hughes, T., and Langhorne, P. (2005). Speech

and language therapy for dysarthria due to non-

progressive brain damage. Cochrane Database Syst

Rev, 3.

Seo, W.-K., Kang, J., Jeon, M., Lee, K., Lee, S., Kim, J. H.,

Oh, K., and Koh, S.-B. (2015). Feasibility of using

a mobile application for the monitoring and manage-

ment of stroke-associated risk factors. Journal of Clin-

ical Neurology, 11(2):142–148.

Shin, S., Park, E., Lee, D. H., Lee, K.-J., Heo, J. H., and

Nam, H. S. (2012). An objective pronator drift test

application (ipronator) using handheld device. PLoS

One, 7(7):e41544.

Terry, K. (2015). Number of health apps soars but use does

not always follow. Medscape Medical News.

Tsanas, A., Little, M. A., McSharry, P. E., and Ramig, L. O.

(2010). Accurate telemonitoring of parkinson’s dis-

ease progression by noninvasive speech tests. IEEE

transactions on Biomedical Engineering, 57(4):884–

893.

Walker, S. H. and Duncan, D. B. (1967). Estimation of the

probability of an event as a function of several inde-

pendent variables. Biometrika, 54(1-2):167–179.

Writing, G. M., Mozaffarian, D., Benjamin, E., Go, A., Ar-

nett, D., Blaha, M., Cushman, M., Das, S., de Fer-

ranti, S., Despr

es, J., et al. (2016). Heart disease and

stroke statistics-2016 update: A report from the amer-

ican heart association. Circulation, 133(4):e38.

Zhang, M. W., Yeo, L. L., and Ho, R. C. (2015). Harness-

ing smartphone technologies for stroke care, rehabili-

tation and beyond. BMJ innovations, 1(4):145–150.

ICT4AWE 2019 - 5th International Conference on Information and Communication Technologies for Ageing Well and e-Health

166