With a Little Help from My Conversational Agent:

Towards a Voice Assistant for Improved Patient Compliance and

Medication Therapy Safety

Jan Schulte to Brinke

, Christian Fitte

, Eduard Anton

, Pascal Meier

and Frank Teuteberg

Smart Enterprise Engineering, German Research Center for Artificial Intelligence,

Parkstraße 40, 49080 Osnabrueck, Germany

Accounting and Information Systems, Osnabrueck University, 49069 Osnabrueck, Germany

Keywords: Conversational Agent, Medical Assistant, Patient Compliance, Medication Therapy Safety.

Abstract: The chronically ill and the elderly often need to take several drugs, which increases the complexity of

medication management. This frequently results in a decrease in patient compliance and raises the risks of

their drug therapy. To support patients in medication management, we developed a multimodal assistant that

includes a conversational agent supplied with data from a database managed by healthcare professionals via

a web service. The developed artifact analyzes medication plans, identifies adverse drug reactions and side

effects, and reminds patients to take their medication correctly and on time. Applying the design science

research paradigm, we systematically identified 16 issues, derived eight meta-requirements, and elaborated

three design principles. Based on this, the artifact was implemented and evaluated by three experienced

pharmacists, who highlighted the usefulness of the solution and provided feedback for further improvements.

Finally, we present an evaluation concept for potential users and discuss the implications of the medication

assistant. Overall, the medical assistant comprises valuable functionalities to support patients, and it increases

medication therapy safety and patients’ compliance.

1 INTRODUCTION

Given global demographic change, the share of the

elderly is increasing worldwide. Senior citizens are

much more likely to suffer from chronic or multiple

diseases and rely on several medications

simultaneously (Peters et al. 2010). With high

quantities of drugs, the complexity of medication

management rises, thereby increasing the risk of

medication errors and patient non-compliance

(Schäfer 2011; Vrijens et al. 2008). Taking the wrong

medication, an incorrect dosage, or a combination of

several drugs can lead to dangerous adverse drug

reactions (ADRs) (Reimers and Klein 2015). An

ADR is a “response to a medicinal product which is

noxious and unintended” (EMA 2017, p. 8), causing

severe side effects and even leading to 25,000 annual

deaths in Germany (Dormann et al. 2017). In addition

https://orcid.org/0000-0002-2637-1198

https://orcid.org/0000-0002-5676-710X

https://orcid.org/0000-0003-3870-6642

to ADR-induced medical implications, economic

consequences are mirrored in the costs of ADR-

related hospital admissions, which are estimated for

only Germany at around 434 million euros per year

(Rottenkolber et al. 2011), excluding the associated

indirect costs (Stark, John, and Leidl 2011).

Yet, most ADRs are considered avoidable

(Schurig et al. 2018), particularly with patients’

adhering to their prescribed medication plans

(Tafreshi et al. 1999). In many cases, patients do not

take their drugs because they forget to take them (Kim

et al. 2018), they fear side or adverse effects (Flávio

Ferreira et al. 2014; W. T. Hsieh et al. 2018; Teixeira

et al. 2017), or they are ignorant to the benefits of the

medication (Flávio Ferreira et al. 2014; Sebillo et al.

2017; Teixeira et al. 2017). Therefore, to achieve

effective patient compliance, the provision of

information on drugs and their medication is of

central importance (Grube, Dehling, and Sunyaev

Schulte to Brinke, J., Fitte, C., Anton, E., Meier, P. and Teuteberg, F.

With a Little Help from My Conversational Agent: Towards a Voice Assistant for Improved Patient Compliance and Medication Therapy Safety.

DOI: 10.5220/0010411707890800

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 789-800

ISBN: 978-989-758-490-9

789

2017). In this context, eHealth technologies can

support patients to comply with their therapy

(Mertens et al. 2015; Sedlmayr 2018). In particular,

conversational agents (CAs) offer promising

potential for improving medication self-management

and reducing intake errors (EmmaHome 2020; Flavio

Ferreira et al. 2013; Jesús-Azabal et al. 2020; Teixeira

et al. 2017). A CA in the healthcare domain is an

“artificial intelligence program that can conduct an

intelligent conversation via auditory or textual

methods regarding healthcare issues” (Wang and Siau

2018, p. 1). The use of CAs for medication

management can allow for avoiding medication

errors by providing the necessary information in a

more natural human-computer interaction, thus

improving the usability and accessibility, especially

for the elderly who have less technological know-how

(Teixeira et al. 2017). However, available solutions

often offer only reminder functions without providing

comprehensive information about the medication

(EmmaHome 2020; Jesús-Azabal et al. 2020).

Moreover, the functionalities often primarily utilize

information that is based on data managed by the

patient (MyTherapy 2020; Sebillo et al. 2017; Silva

et al. 2013). Thus, most CAs reflect the error-prone

input of the patient, without any monitoring by

healthcare professionals who can assure data quality

and the safety of drug therapy. Drawing on these

shortcomings in assuring the safety of drug therapy

for patients, we pose the following research question:

RQ: How Can CAs Be Designed and Implemented to

Increase Medication Therapy Safety for Patients?

To answer this research question, we design and

implement a multimodal assistant comprising a CA to

increase medication therapy safety by allowing a

patient to use voice commands to inquire about the

side effects or reactions of their medication. The

system is complemented by a web portal for

healthcare professionals which receives input from a

central pharmacy data service and allows for the

management of patients’ medication plans. Driven by

the design science paradigm (DSR) (Gregor and

Hevner 2013), we structure the remainder of this

paper as follows: We introduce our DSR multistep

research approach and the applied methods for the

development and evaluation of the application in

Section 2. Then, outlining the theoretical background,

we summarize the related work and derive issues for

medication assistants in Section 3. Based on the

identified issues, we derive meta-requirements and

elaborate design principles in Section 4. The

subsequent Sections 5 is devoted to describing the

designed and developed artifact and the evaluation

concept. Finally, the paper concludes by discussing

the implications for research and practice, presenting

the limitations, and providing an outlook for future

research.

2 RESEARCH APPROACH

The design, development, and evaluation of the

artifact in response to our research question follows

the design science research (DSR) methodology by

Peffers et al. (2007). DSR aims to address important

real-world problems that remain unsolved or require

further investigation via a technological artifact

(Hevner et al. 2004). In this paper, we address the

issues of the current available solutions for

medication management to increase patient

compliance and medication therapy safety by

developing a multimodal assistant comprising a CA

and a web service that handles patient queries by

accessing a managed medical database. The process,

in pursuit of realizing the solution, involves the

identification of an observed problem, followed by

the design, implementation, and evaluation of the

artifact (cf. Figure 1) (Peffers et al. 2007).

Figure 1: DSR Approach of this Study (Peffers et al. 2007).

As a starting point for problem identification, we

conducted a systematic literature analysis (vom

Brocke et al. 2009). We queried titles, abstracts, and

keywords in the interdisciplinary databases

SpringerLink, ScienceDirect, AISeL, IEEE Xplore,

Emerald insight, JSTOR, and EBSCOhost using the

search string ("requirements" OR "design principles"

OR "Anforderungen") AND ("medication assistant"

OR "health assistant" OR "medical agent" OR

"Gesundheitsassistent" OR "Medikationsassistent").

Ever since Apple introduced the voice assistant Siri

in 2011, the research and general interest in CAs have

increased significantly (Luger and Sellen 2016);

hence, we included papers from 2011 onward and

filtered for the languages English and German. The

resulting 621 publications were analyzed based on the

following inclusion and exclusion criteria: We

considered any theoretical and practical work

involving medication assistants or their development

and excluded studies that have a broader research

focus in the eHealth field without focusing on drug

Scale-IT-up 2021 - Workshop on Scaling-Up Healthcare with Conversational Agents

790

management. After applying the inclusion and

exclusion criteria, the remaining eight studies served

for a forward and backward search that yielded

another nine studies, resulting in a literature corpus of

17 studies. An analysis of the literature yielded the

identification of 16 issues (Is), from which we derived

nine meta-requirements (MRs) for the solution, which

we finally combined into three design principles

(DPs). Based on the systematically derived DPs, we

developed a multimodal system consisting of three

sub-systems, as presented in Section 5. Finally, we

evaluated the artifact by interviewing three

experienced pharmacists regarding their assessment

of its usefulness and their evaluation of areas for

improvement of the artifact (Myers and Newman

2007). We analyzed the interviews according to

Mayring (2010), and incorporated the experts’

suggestions for improvements in order to prepare the

CA for the next evaluation cycle with patients, which

will be presented as an experimental design.

3 THEORETICAL

BACKGROUND

3.1 Complexity of Patient Compliance

and Medication Therapy

Management

In Germany alone, more than 100,000 approved

medications exist, which may contain a variety of

ingredients (ABDA 2019). Nearly half of these

medications are available only by prescription (Rx).

In contrast, over-the-counter medications (OTC) are

available to every customer without restriction.

Approximately 23% of German citizens take more

than three medications simultaneously (ABDA

2019). This can lead to a high risk of ADRs,

especially if patients are treated by several healthcare

professionals without a central medication plan

(Reimers and Klein 2015), thus necessitating that the

patients notify the treating physician of their

medications themselves. Patients might forget to

mention medications, assume they are not relevant, or

deliberately withhold information out of shame.

In addition to the high risks of polymedication,

many patients must change their way of life due to

their drugs. The challenge of integrating their

medication intake into a daily routine poses a barrier

for many people causing them to forget or even refuse

the medication. This results in a reduction in patient

compliance, which “[...] describes the extent to which

a person's behavior (in terms of taking medications,

following diets or executing lifestyle changes)

coincides with medical or health advice” (Haynes

1979). The concept of compliance has evolved from

a directive model (obedience to therapy) to a passive

model (loyalty to therapy) and from an active model

(cooperation in therapy) to an interactive model, in

which the patient participates in the treatment and

cooperates with the doctor as a means of

empowerment (Schäfer 2011). Patients can be

divided into three categories when assessing

compliance: If patients follow at least 80% of their

therapy guidelines, they are described as “compliant”

(Schäfer 2011). A compliance level between 20% and

80% categorizes a patient as “partially compliant.” It

can be assumed that this patient group can best be

persuaded to adhere to therapy measures. Compliance

levels below the threshold of 20% are called “non-

compliant.” Non-compliance may be intentional

(e.g., due to religious reasons, exaggerated fears,

general distrust of medicine, reservations about side

effects, or convenience), or it can occur

unintentionally by accidentally choosing the wrong

dosage or changing the duration or frequency

(Petermann and Mühlig 1998).

3.2 Issues for Medication Assistant

Application

Based on the literature review described in Section 2,

we derived issues of and requirements for medication

assistant applications. Of the 17 identified studies, we

found eight papers that focus specifically on

applications developed for seniors, while nine papers

present solutions engineered for patients in general.

The user interfaces of the identified applications

include either CAs, web interfaces, specific hardware

components, or a mixture of multiple components.

We identified 16 issues to be addressed using our

medical assistant. First, patients are often confronted

with an information overload (I1) (e.g., in form of

endless information from leaflets), which renders it

difficult to find relevant information regarding the

proper medication intake (Dehling and Sunyaev

2013; Tiwari et al. 2011). On the other hand,

medication errors can occur due to missing

information (I2) (e.g., concerning the medication

storage) (Flávio Ferreira et al. 2014; Mira et al. 2014;

Sebillo et al. 2017; Teixeira et al. 2017). Regardless

of the information quantity, a language that is too

complex or contains technical terminology (I3) can

cause misinterpretations (Chang et al. 2019; Dehling

and Sunyaev 2013; Farhadyar and Safdari 2018;

Flavio Ferreira et al. 2013; Flávio Ferreira et al. 2014;

Jesús-Azabal et al. 2020; Kim et al. 2018; Teixeira et

With a Little Help from My Conversational Agent: Towards a Voice Assistant for Improved Patient Compliance and Medication Therapy

Safety

791

al. 2017). This can then lead to situations in which the

identification of medications is based on appearance

or storage rather than the technical product name. A

too-small illustration of information (I4) causes

comparable problems (Chang et al. 2019; Dehling

and Sunyaev 2013; Farhadyar and Safdari 2018;

Flavio Ferreira et al. 2013; Flávio Ferreira et al. 2014;

Silva et al. 2013; Teixeira et al. 2017; Tiwari et al.

2011). Numerous studies indicate that medical

assistants are generally too complex to use (I5)

(Flávio Ferreira et al. 2014; P. J. Hsieh 2016; Tiwari

et al. 2011). In particular, a complex initial set-up or

a required adjustment of the medication plan can

reduce acceptance among users (Sebillo et al. 2017).

Additional issues arise in the registration of

medications (I6) (Dayer et al. 2013; Flavio Ferreira et

al. 2013; Flávio Ferreira et al. 2014; W. T. Hsieh et

al. 2018; Sebillo et al. 2017; Silva et al. 2013;

Teixeira et al. 2017). The manual integration of

information can produce mistakes and might be

perceived by the patient as requiring too much effort

(Dayer et al. 2013; Silva et al. 2013). A prerequisite

for acceptance and thus an efficient application is for

patients to trust the CA without worrying about

misinformation (I7) (Flavio Ferreira et al. 2013;

Flávio Ferreira et al. 2014; Sneha and Varshney 2012;

Teixeira et al. 2017; Tiwari et al. 2011). In this

context, data security concerns (I8) may hinder such

acceptance (Dehling and Sunyaev 2013; Kim et al.

2018; Santo et al. 2016; Sneha and Varshney 2012).

In addition, many patients suffer from unintended

drug interactions due to polymedications (I9), which

can lead to serious health problems and avoidable

hospital admissions (Dayer et al. 2013; Dehling and

Sunyaev 2013; W. T. Hsieh et al. 2018; Kim et al.

2018; Mira et al. 2014; Silva et al. 2013). Further

health problems can occur due to side effects (I10)

(Dehling and Sunyaev 2013; Farhadyar and Safdari

2018; Flávio Ferreira et al. 2014; W. T. Hsieh et al.

2018; Teixeira et al. 2017; Tiwari et al. 2011).

Moreover, uncertainty of patients often results in a

false dosage of medications (I11) (Chang et al. 2019;

Dayer et al. 2013; Farhadyar and Safdari 2018; Mira

et al. 2014; Sebillo et al. 2017; Silva et al. 2013;

Sneha and Varshney 2012; Tang et al. 2011; Teixeira

et al. 2017). This could be addressed by increasing the

transparency and traceability of doctors’ visits (I12)

(Chang et al. 2019; Dayer et al. 2013). Fourteen

papers mention problems related to polymedications

(I13), and 16 of the 17 relevant papers identified that

patients forget or deny medication intake (I14).

Chronic patients often require long-term medication

(I15) (Chang et al. 2019; W. T. Hsieh et al. 2018;

Jesús-Azabal et al. 2020; Mira et al. 2014; Santo et al.

2016; Sneha and Varshney 2012; Teixeira et al. 2017;

Tiwari et al. 2011) which requires strong discipline

and well-organized medication management (I16)

(Chang et al. 2019; Dayer et al. 2013; Farhadyar and

Safdari 2018; Kim et al. 2018; Santo et al. 2016;

Sebillo et al. 2017; Tiwari et al. 2011).

Existing applications focusing on the safety of

drug therapy include hanahealth, Mytherapy, Emma

Home, and mediteo (EmmaHome 2020; hana health

2019; Mediteo 2020; MyTherapy 2020). Applications

such as SapoMed and Sedato, for the elderly and rural

areas, have been highlighted in the scientific literature

(Flavio Ferreira et al. 2013; Jesús-Azabal et al. 2020;

Sebillo et al. 2017; Silva et al. 2013). However, to the

best of our knowledge, the existing solutions focus on

specific target groups or offer rather limited

functionalities for patient safety and compliance, such

as medication reminders. This study aims to create an

age-independent assistant for the management of

multiple medications, which considers side effects,

ADRs, and medication reminders and enables a

combination of medication data with continously

measured vital signs. Our artifact represents an

enhancement of the FeelFit platform (Meier et al.

2019). FeelFit aggregates measured vital parameters

of several devices (e.g., weight, pulse, and blood

pressure) and visualizes the medical record for the

patient, authorized physicians, and pharmacists.

4 DESIGN PRINCIPLES

We categorized the identified issues into groups,

namely usability, information processing, and

medical issues to derive MRs and consolidate DPs

according to Gregor et al. (2020). By following these

guidelines, we ensure to consider aim, context, and

mechanism within the design of our artifact (Gregor,

Kruse, and Seidel 2020). The linked Is, MRs and DPs

are depicted in Figure 2.

To counteract the issues of information overload

(I1) and missing information (I2), the medical

assistant should present the information in a context-

sensitive manner (MR1) by emphasizing only

relevant information (e.g., appearance of the pill and

intake information). If necessary, it should provide

the user with additional information upon request.

Furthermore, terminology that is too specialized or

technical (I3) and a representation that is too small

(I4) should be avoided. Therefore, we derive MR2,

for which the information should be presented in an

understandable and clear way. To enable the intuitive

operation of the application (MR3), complex

interfaces that are difficult to use (I5) should be

Scale-IT-up 2021 - Workshop on Scaling-Up Healthcare with Conversational Agents

792

Figure 2: Issues, Meta Requirements, and Design Principles.

avoided alongside I3 and I4. As a result, the

medication assistant should be easy to use by users

who have little technical experience and during

exceptional situations. MRs 1-3 were assigned to the

category usability and can be consolidated into the

first design principle:

DP1: To provide users with comprehensive

information, provide an application with relevant

information in a context-sensitive and

understandable way, because this sparse and

manageable display of the information protects

the user from being overwhelmed and enables an

easy and efficient application of the relevant

information

Problems with the registration of medication (I6)

and false information provided by the medication

assistant (I7) should be avoided. For this purpose, the

input of incorrect data should be prevented through

the design of the application (MR4). Data entry

should be designed to be as simple as possible and

managed by medical professionals. Furthermore, to

avoid incorrect entries (I7), it must be ensured that the

basic information on the drugs is correct. Therefore,

a trustworthy drug database must be used (MR5).

Furthermore, the data security concerns of users must

be considered (I8), which requires responsible

handling of private health data (MR6). This implies

parsimonious data storage so that only the necessary

user data is saved in the databases. Furthermore, this

data should be managed securely. The guideline for

information processing can be summarized as

follows:

DP2: To enable trustworthy information

processing, provide an application that ensures

integrity, secure and reliable data management,

because health data are considered to be

particularly "sensitive" and are subject to special

protection and, in addition, false or incorrect data

can lead to health consequences.

The application is intended to prevent problems

pertaining to interactions (I9), side effects (I10) and

dosage (I11). To achieve this, the medication

assistant should warn patients about possible ADRs

(MR7). On one hand, new drugs must be checked for

interactions with already-recorded medications. In

this way, possible ADRs can be identified during the

prescription process. On the other hand, ADRs should

be identified as a possible cause of the existing

symptoms by analyzing recorded drugs based on their

side effects. In addition, issues of tracking visits to the

doctor (I12) and difficulties in taking several

medications (I13) should be avoided by helping

patients to take their medication correctly (MR8) by

raising the transparency of information. Moreover,

the application should make it easier to distinguish

drugs, and it should provide information regarding

the correct type and quantity of medication. In

addition, forgetting or refusing to take the medication

(I14), non-compliance with long-term medication

(I16), and problems in medication management (I17)

should be avoided. Therefore, users should be

With a Little Help from My Conversational Agent: Towards a Voice Assistant for Improved Patient Compliance and Medication Therapy

Safety

793

supported in the long term through taking all

medications at the right time (MR9). In this way, the

user should receive important information at regular

intervals and throughout the entire medication period.

Furthermore, the application should remind the user

to take the medication on time and obtain follow-up

prescriptions. From these MRs follows the DP3,

which focuses on the medical issues:

DP3: To protect users from medical problems,

provide an application that supports chronic

disease management by identifying ADRs and

reminding patients to take the right medications

at the right time, because these features help to

improve medication management and patient

compliance.

5 MEDICAL ASSISTANT

5.1 Software Architecture

An overview of the software architecture of the

multimodal assistant is provided in Figure 3. The

arrows depicted in the overview indicate the direction

in which data is transferred. The three main modules

of the application are the web application (Med-

Portal), the RESTful web service (Med-REST), and

the CA (Med-CA).

Med-REST: This subsystem serves as the interface

to the Med-Portal and Med-CA and functions as a

data communication link between the components.

An essential component of the web service is the

medication data provided by ABDATA Pharma-

Daten-Service, a division of Avoxa - Mediengruppe

Deutscher Apotheker GmbH. The data comprises all

available Rx and OTC medications in Germany with

various drug information including the ingredients,

storage and application methods, side effects and

possible interactions with other drugs. In addition, the

database includes economic and legal information

entailing the medications’ price, distribution

channels, or information on dispensing conditions

(Pharma-Daten-Service 2019). As we emphasize

drug safety and patient compliance in this study, we

neglect the economic and legal drug information. To

enable an enhanced query of the data, we imported

the relevant data into a relational database (ABDA-

DB) using the open-source H2 Database Engine

(Database 2019). Since we are determined to prevent

changes in the ABDA database to avoid

misinformation, we implement only HTTP-GET

methods for the individual resources of the web

service. In addition, we created an authentication

method for security purposes that acts as the interface

for using the web service. With each call, the system

first checks whether a valid token of the Med-Portal

or the Med-CA has been transferred, so that only

registered users can access the web service. If

authentication is successful, the request is forwarded

to the corresponding class in the application. All data

is transmitted encrypted via the Hypertext Transfer

Protocol Secure (HTTPS).

Med-Portal: To allow medical professionals to use

the Med-Rest interface and manage the medications

of their patients, we created an HTML-page. First, we

added a search field that allows for searching the web

service (Med-REST) for medication information. To

do so, a GET-function of the service is called by the

client. We implemented the possibility of using

wildcards (“%”) to enhance the search usability. The

method returns the trade name, potency, and

manufacturer of the queried drugs, displayed in a

table.

If a new drug must be added for a particular

patient, another GET-function is applied to check

whether possible interferences between the new drug

and the patient's existing medication plan can be

identified. If possible interactions occurred, a

Figure 3: Software Architecture of the Medication Assistant.

Scale-IT-up 2021 - Workshop on Scaling-Up Healthcare with Conversational Agents

794

Figure 4: User Interface: Med-Portal (left) and Med-CA (right).

warning message displays the trade names of the

interacting drugs, any recommended actions, and a

description of the effect (cf. Figure 4, left image).

Based on this, the medical professional can decide

whether to cancel the process or enter the drug despite

the possible interactions. To add the new drug, the

ingredients, trade name, dosage and shape of the drug

are retrieved from the web service via a dedicated

function. Furthermore, the physician can manually

enter the intake time (morning, noon, evening, night),

unit quantity, notes, reason for ingestion, reminder

time (morning, noon, evening, night), and remaining

stock for the patient.

Med-CA: This subsystem represents the module used

by the patient (cf. Figure 4, right image). The Med-

CA utilizes Google’s Dialogflow platform for speech

processing (Google 2020). The voice commands

issued by the patient are performed by the

corresponding functions on the server side on

Google’s App Engine platform. These functions are

fed by general medication information via the Med-

REST service and can be supplemented by patient-

specific data from a medical database (Portal-DB)

connected to the open API. As mentioned, we used

the FeelFit database (Meier et al. 2019) as the Portal-

DB for testing purposes, which stores various patient-

specific data such as blood pressure or patient

activity. In addition, a software development kit

(SDK) is used to access another database from the

App Engine (CA-DB). This database stores data that

is relevant only for the functions of the Med-CA and

serves as the foundation for the ReminderPushJob

application. The ReminderPushJob checks whether

there are reminders defined in the database for a given

time, which triggers the sending of a push notification

containing the medication information to the user.

The information comprises the medicine (product

name), doses, color of the drug, method, and intake

advice derived from the ABDA database. The

remaining stock stored in the Portal-DB is

automatically adjusted based on the applied

medication. If the remaining stock drops below a

threshold of 10 units, a warning message is stored in

the CA-DB and is issued to the patient.

5.2 Evaluation

The evaluation is aimed to validate whether the

system meets the proposed design and to analyze the

systems performance (Kuechler and Petter 2012). In

the first evaluation cycle, we interviewed three

experienced pharmacists. The second evaluation

cycle is supposed to generate insights on user’s

acceptance. To do so, we present an experimental

design for a future long-term study.

First Iteration: Expert Interviews.

To evaluate the medication assistant regarding its

usefulness, the quality of the information, and the

potential for improvement, we interviewed three

qualified, self-employed pharmacists based in

Germany (E1: 25, E2: 20 and E3: 29 years of

experience). After presenting the medical assistant

and the functionalities of the system to the

pharmacists, we conducted a semi-structured

interview (Myers and Newman 2007).

According to the experts, the target group of the

Med-CA are patients with chronic diseases suffering

from asthma, diabetes, or cardiovascular diseases

(E1, E2, E3). The CA could support patients in

organizing their medication management,

particularly by checking the remaining stocks and

With a Little Help from My Conversational Agent: Towards a Voice Assistant for Improved Patient Compliance and Medication Therapy

Safety

795

reminding them to reorder medications (E1). In

addition, reminders for medication intake and

respective advice (intake: oral, injection, liquid, and

so on) could increase patients’ compliance.

According to E2, an improvement in functionality

would be the confirmation of the patient’s drug intake

following a reminder. Furthermore, E2 remarks that

if a patient is not at home when receiving a

notification, the patient should be reminded on a loop

until the final intake is confirmed. The pharmacist E2

stressed that notifications of all types should avoid

complex and professional terminology, which our

artifact based on I3 considered. Therefore,

information on the medication intake, storage, and

side effects from the ABDA-DB should be translated

into generally understandable language. The process

of identifying certain drugs could be supported by

providing pictures of the respective pill on the

patient’s smartphone (E2). In addition to notifications

concerning medication intake, patients could also be

reminded to measure their vital signs, such as blood

pressure (E2). These measurements should be stored

directly in the patient’s profile to enable the

pharmacist or other healthcare professionals to check

whether the patient’s medication is correctly adjusted.

All pharmacists agreed that they could integrate

the Med-Portal into their daily work routine.

However, they indicated that additional effort should

be minimized (e.g., by interfaces to already-used

software) (E2, E3). In particular, compatibility with

nation-wide standardized medication plans should be

ensured (E3). Furthermore, pharmacists require

additional individual advice options on the dosage or

storage of medicines (e.g., in the form of notes to a

particular entry) (E1, E3). In case of questions,

patients should also have the ability to contact their

pharmacist directly via chat (E3). In addition to

automatic ADR analyses, pharmacists should be able

to perform regular manual assessments of the

medication plan to identify potential redundant

prescriptions or obsolete drugs (E3).

Regarding Med-REST, the experts mentioned

that the web service should feed on data that goes

beyond the ABDA-database, as not all potential

interactions might be listed, and patients could take

additional substances that are relevant to their

medication plan but are not included in the

pharmaceutical database (e.g., nutritional

supplements) (E1). Complementary health-related

data such as vital signs, blood-, liver- and kidney-

values, weight, allergies or food intolerances should

be integrated into the patient’s profile as they might

be relevant for the effective use of certain

medications (E1).

Overall, the basic functions of the assistant were

assessed by the experts as beneficial in terms of

supporting users in drug management and improving

patient compliance. The features for further

improvement can be categorized in data and

functional improvements and are summarized in

Table 1.

Table 1: Feedback of Interviewed Pharmacists.

No. Data Improvements Fre

uenc

1 Easy Language 1

Detailed advice to medication

intake

3 Information about stora

e 1

4 Pictures of drugs 1

Additional data: vital signs,

nutrition supplements, allergies,

intolerances, wei

No. Functional Improvements Fre

uenc

6 Interfaces to existing software 1

7 Notifications for vital signs

measurement

8 Inte

ration of vital si

ns 3

9 Integration of laboratory values 2

10 Consideration of chronic diseases 1

11 Confirmation of intake 1

12 Notification-loo

13 Chat 1

Second Iteration: Experimental Design.

After evaluating the CA with experienced

pharmacists, the system needs to be tested and

evaluated by potential users. To prepare the second

evaluation cycle, we conceptualize a suitable

experimental design which will be operationalized in

a future study. The study design is supposed to be an

experiment. It will be held in a smart home

showroom, which creates a comfortable atmosphere

in which participants feel at ease. Thereby we can

avoid biases due to a laboratory setting.

First, we will survey participants in regards to

their demographics, pre-experience and general

attitude towards technology. Second, they will be

introduced to a scenario involving a specific person,

who suffers from multiple diseases, takes several

medications, lives alone and is not able to visit the

pharmacy on his or her own. Nevertheless, the person

obtains his or her medication supply through home

deliveries. In addition to the delivery service, the

pharmacy that supplies the person offers a CA for the

medication management. After being introduced to

the scenario, the study participants will be asked to

conduct the following three tasks:

Scale-IT-up 2021 - Workshop on Scaling-Up Healthcare with Conversational Agents

796

I. Ask the CA, which medications must be taken

today and ask a follow-up question about the

appearance of the pill.

II. Ask the CA, when you need to reorder

medication according to your therapy plan. If

necessary, order new medication from your

local pharmacy.

III. Report side effects to the CA, resulting from

taking medication, e.g. headache, and find out

causes for the symptoms.

After completing the tasks, participants have to

answer a user experience questionnaire. We will

apply the user experience questionnaire based on

Schrepp et al. (2014). Besides initial questions

towards the acceptance and the trust towards the

technology, participants have to share their

impression of the system by rating 26 items on a scale

of two contradictory features, e.g., good/bad,

slow/fast or enjoyable/annoying. The user experience

questionnaire provides insights in terms of novelty of

the artifact, stimulation, dependability, efficiency,

perspicuity and attractiveness. In addition, it allows to

establish a benchmark with comparable artifacts.

6 DISCUSSION

The management of the risks associated with

polymedication is of great medical and economic

importance (Taylor et al. 2013). Therefore, patients in

Germany, who take more than three different

prescribed medications have been eligible for a

standardized medication plan since 2016. In addition,

some pharmacies manage customer profiles to track

customers’ individual medication records (Reimers

and Klein 2015). However, most patients do not

request such a medication plan, as they are unaware

of their right to a managed plan and the possible

ADRs of their medication. Furthermore, medication

plans often do not consider OTC drugs, and ADRs

might just as well occur with fewer than three

prescribed drugs. To support medication

management, digital assistants such as CAs can be

found in many popular app stores and are designed to

support patients in their medication therapy

(EmmaHome 2020; MyTherapy 2020). However,

these applications often support only a reminder

function for medication intake (EmmaHome 2020;

Jesús-Azabal et al. 2020) and rely on the error-prone

input of the patient to create a medication plan

(MyTherapy 2020; Sebillo et al. 2017; Silva et al.

2013). As a result, the safety of medication therapy is

not emphasized in these particular solution

approaches. The CA presented in this study enters

into dialog with patients using simple language to

provide them with important information concerning

the medication to be ingested, and it considers side

effects and ADRs with other drugs. The system

enables patients to self-manage their medications and

provide access to selected healthcare professionals,

who can access their medication plan via a web

service. Through the integration of healthcare

professionals such as pharmacists or doctors, it is

possible to strengthen the intersectoral cooperation, to

send secured notifications about medication intake to

the patient and to reorder medications directly,

thereby simplifying the integration of therapy plans

into everyday life and increasing patient compliance

on several levels. Moreover, the ADR check function

and the associated notification of interferences and

information concerning the composition of drugs,

side effects, and symptoms increases the health

competence and health awareness of patients.

All interviewed pharmacists emphasized the

relevance of vital parameters for medication therapy

safety. We connected the FeelFit database to our

system to integrate vital signs from various devices

(Meier et al. 2019). On one hand, this enables

pharmacists to check whether the medication doses

are correctly adjusted. On the other hand, a

continuous visualization of vital signs can

demonstrate the consequences of non-compliance to

the patient (Meier et al. 2019). This is especially

relevant for non-compliant patients who do not take

their chronic diseases seriously, intentionally refuse

adherence to their therapy plan, or secretly deny

medications because of inconvenience (Petermann

and Mühlig 1998).

Implications for Research and Practice: All the

pharmacists involved in the evaluation agreed that our

multimodal prototype has a high value for medication

management safety and patient compliance. From the

perspective of a practitioner who works in the

healthcare sector, the application could provide more

transparency into the ongoing therapy of patients and

thereby intervene in the treatment more quickly and

proactively, thereby increasing the overall quality of

healthcare supply. This is particularly important,

since ADRs could often be avoided by increasing the

transparency for healthcare professionals and thus

reducing healthcare expenses caused by avoidable

hospital admissions due to medication errors (Salvi et

al. 2012; Taylor et al. 2013). Furthermore, the active

management of patients’ data (e.g., by pharmacists)

can increase the importance of local healthcare

experts and enhance the relationship with patients

(Mossialos et al. 2015). Given the challenges of

With a Little Help from My Conversational Agent: Towards a Voice Assistant for Improved Patient Compliance and Medication Therapy

Safety

797

increasing e-commerce and the loss of rural

infrastructure due to urbanization and the shortage of

physicians, the number of German pharmacies has

decreased over the past 10 years (ABDA 2019). Our

artifact could provide a new interface in the patient-

pharmacist relationship (Volland 2015). In this

context, the combination with blistering business

models might be a valuable future service for local

pharmacies. From the patient’s perspective, the

medication assistant can not only support the safety

of drug therapy but also increase the integration into

everyday life and thus promote patient compliance.

As a result, this facilitates healing and supports health

prevention. However, to establish integrity to our

system, it requires a trustworthy and reliable data

management and compliance standards for all

involved users.

Our findings contribute to theory, as we shed light

on guidelines to design and develop multimodal

medical assistants for medication management. With

the derived DPs in this study, we provide context-

oriented guidelines for medical assistants that

complement more generic DPs of CAs. With our

modular approach, we enable the connection to other

applications, such as FeelFit and thus the

enhancement of functionality in the realm of personal

medical health assistants. On this basis, further

research can be initiated, and additional

functionalities and improvements can be developed

based on our evaluation.

Limitations: As with any study, our research is

subject to limitations. First, we conducted only one

iteration of the DSR cycle. The feedback of the expert

interviews must be implemented into the prototype

during the next iteration cycle. Especially the

readability of a standardized medication plan by

scanning the QR code, should be added in the next

development cycle. Second, we have not yet

evaluated the CA form the perspective of potential

patients. Since evaluations with patients require a

comprehensive study design, we presented an

evaluation concept which will be operationalized in a

future study. Third, we have not yet analyzed the

integration of the system into existing healthcare

initiatives or patient portals. Future studies with

patients should investigate the connection to current

systems like advanced patient portals or virtual health

coaches. Finally, our CA is partially focused on the

German healthcare system. The prototype and its

evaluation might be biased by legal, structural, and

cultural influences. Nevertheless, the demand for

medical assistants rises internationally, and the results

of our study can be transferred to any other healthcare

system.

7 CONCLUSION

Our study aimed to develop a multimodal assistant

that supports and secures patients in their medication

management, thereby increasing safety and

facilitating adherence to therapy plans. Within a DSR

project, we identified 16 Is to describe problem areas

that must be considered when developing an

application to enhance medication therapy safety, and

we derived nine MRs and consolidated them into the

three DPs focusing on usability, information

processing and medical issues. Those were built in a

prototype consisting of three components.

First is the Med-Portal component, which can be

accessed by medical professionals to manage

patients’ medication plans and provide individual

advice. The second is the Med-REST web service,

which is the interface of the German ABDA database

that contains all available Rx and OTC medications.

Last is the Med-CA component, which enables

patients to follow their medication plan and obtain

additional information regarding the correct intake,

storage, and side effects. Notifications about

medication intake and reordering can support patients

in integrating therapy management into their

everyday lives. An evaluation with experienced

pharmacists has demonstrated the high relevance and

usefulness of our developed medication assistant.

Finally, an evaluation with patients is conceptualized

for the application within a future study. So far, the

designed artifact offers the potential to increase

patients’ compliance and medication therapy safety

and to improve the relationship between patients and

pharmacists.

REFERENCES

ABDA. 2019. ‘DIE APOTHEKE - Zahlen - Daten - Fakten

2019’. : 84.

vom Brocke, Jan et al. 2009. ‘Reconstructing the Giant: On

the Importance of Rigour in Documenting the

Literature Search Process’. 17th European Conference

on Information Systems 9: 2206–2217.

Chang, Danni, Zhenyu Gu, Fan Li, and Rong Jiang. 2019.

‘A User-Centric Smart Product-Service System

Development Approach: A Case Study on Medication

Management for the Elderly’. Advanced Engineering

Informatics 42: 100979.

Database, Engine H2. 2019. H2 Documentation.

Dayer, Lindsey et al. 2013. ‘Smartphone Medication

Adherence Apps: Potential Benefits to Patients and

Providers’. Journal of the American Pharmacists

Association 53(2): 172–81.

Scale-IT-up 2021 - Workshop on Scaling-Up Healthcare with Conversational Agents

798

Dehling, Tobias, and Ali Sunyaev. 2013. ‘Improved

Medication Compliance Through Health IT: Design

and Mixed Methods Evaluation of the Application

EPill’. ICIS 2013 Proceedings.

Dormann, Harald, Michael Bangemann, Hans-Ulrich

Prokosch, and Jürgen Zerth. 2017. ‘Digitalisierte

Arzneimittelversorgung Am Beispiel Des

Bundeseinheitlichen Patientenbezogenen

Medikationsplans – Eine Frage Der

Stakeholderakzeptanz’. In Digitale Transformation von

Dienstleistungen Im Gesundheitswesen I, Springer

Fachmedien Wiesbaden, 149–64.

EMA. 2017. ‘Guideline on Good Pharmacovigilance

Practices (GVP) Annex I - Definitions (Rev 4)’. Heads

of Medicine Agencies (October): 1–33.

EmmaHome. 2020. ‘Emma - EmmaHome - Emma’.

https://www.emma-hilft.com/emma-home/.

Farhadyar, Kiana, and Reza Safdari. 2018. ‘Requirements

of MHealth-Based Medication Management Systems’.

International Journal of Innovative Research in

Computer Science & Technology 6(2): 12–17.

Ferreira, Flavio et al. 2013. ‘Multimodal and Adaptable

Medication Assistant for the Elderly: A Prototype for

Interaction and Usability in Smartphones’. In Iberian

Conference on Information Systems and Technologies,

CISTI,.

Ferreira, Flávio et al. 2014. ‘Elderly Centered Design for

Interaction - The Case of the S4S Medication

Assistant’. In Procedia Computer Science, Elsevier

B.V., 398–408.

Google. 2020. ‘Dokumentation Zu Dialogflow | Google

Cloud’. https://cloud.google.com/dialogflow/docs/.

Gregor, Shirley, and Alan R Hevner. 2013. ‘Positioning and

Presenting Design Science Research for Maximum

Impact’. MIS quarterly: 337–55.

Gregor, Shirley, L Chandra Kruse, and Stefan Seidel. 2020.

‘The Anatomy of a Design Principle’. Journal of the

Association for Information Systems.

Grube, Anton, Tobias Dehling, and Ali Sunyaev. 2017.

‘How Do Patients Expect Apps to Provide Drug

Information?’ Hawaii International Conference on

System Sciences 2017 (HICSS-50).

hana health. 2019. ‘Medikation Neu Gedacht’.

https://www.hana-health.de/index.html (August 11,

2020).

Haynes, R Brian. 1979. ‘Determinant of Compliance: The

Disease and the Mechanics of Treatment’. Compliance

in health care.

Hevner, Alan, S.T. March, J. Park, and S. Ram. 2004.

‘Design Science in Information Systems Research’.

MIS Q 28: 75–105.

Hsieh, Pi Jung. 2016. ‘An Empirical Investigation of

Patients’ Acceptance and Resistance toward the Health

Cloud: The Dual Factor Perspective’. Computers in

Human Behavior 63: 959–69.

Hsieh, Wen Ting, Yung Cheng Su, Hsin Lien Han, and

Ming Yuan Huang. 2018. ‘A Novel MHealth Approach

for a Patient-Centered Medication and Health

Management System in Taiwan: Pilot Study’. JMIR

mHealth and uHealth 6(7): e154–e154.

Jesús-Azabal, Manuel et al. 2020. ‘Voice Assistant to

Remind Pharmacologic Treatment in Elders’. In

Communications in Computer and Information

Science, Springer, 123–33.

Kim, Ben Y B et al. 2018. ‘Consumer Mobile Apps for

Potential Drug-Drug Interaction Check: Systematic

Review and Content Analysis Using the Mobile App

Rating Scale (MARS)’. JMIR mHealth and uHealth

6(3): e74–e74.

Kuechler, Bill, and Stacie Petter. 2012. ‘Design Science

Research in Information Systems’. (1): 1–66.

Luger, Ewa, and Abigail Sellen. 2016. ‘Like Having a

Really Bad PA: The Gulf between User Expectation

and Experience of Conversational Agents’. In

Proceedings of the 2016 CHI Conference on Human

Factors in Computing Systems, ACM, 5286–97.

Mayring, Philipp. 2010. ‘Qualitative Inhaltsanalyse’.

Handbuch qualitative Forschung in der Psychologie:

601–13.

Mediteo. 2020. ‘Mediteo App |

Medikamentenerinnerungen Und Pillenwecker’.

https://www.mediteo.com/de/.

Meier, P. et al. 2019. ‘FeelFit – Design and Evaluation of a

Conversational Agent to Enhance Health Awareness’.

In Proceedings International Conference on

Information Systems (ICIS 2019), Munich.

Mertens, Alexander et al. 2015. ‘Influence of Mobile ICT

on the Adherence of Elderly People with Chronic

Diseases’. Lecture Notes in Computer Science

(including subseries Lecture Notes in Artificial

Intelligence and Lecture Notes in Bioinformatics) 9194:

123–33.

Mira, José Joaquín et al. 2014. ‘A Spanish Pillbox App for

Elderly Patients Taking Multiple Medications:

Randomized Controlled Trial’. Journal of Medical

Internet Research 16(4): e99–e99.

Mossialos, Elias et al. 2015. ‘From “Retailers” to Health

Care Providers: Transforming the Role of Community

Pharmacists in Chronic Disease Management’. Health

Policy 119(5): 628–39.

Myers, Michael D., and Michael Newman. 2007. ‘The

Qualitative Interview in IS Research: Examining the

Craft’. Information and Organization 17(1): 2–26.

MyTherapy. 2020. ‘Tabletten Erinnerung Leichtgemacht –

MyTherapy App’. https://www.mytherapyapp.com/de.

Peffers, Ken, Tuure Tuunanen, Marcus A Rothenberger,

and Samir Chatterjee. 2007. ‘A Design Science

Research Methodology for Information Systems

Research’. Journal of management information systems

24(3): 45–77.

Petermann, Franz, and Stephan Mühlig. 1998. ‘Grundlagen

Und Möglichkeiten Der Compliance-Verbesserung’.

Petermann, F.(Hg.): Compliance und

Selbstmanagement. Göttingen: Hogrefe: 73–102.

Peters, E, R Pritzkuleit, F Beske, and A Katalinic. 2010.

‘Demografischer Wandel Und

Krankheitshäufigkeiten’. Bundesgesundheitsblatt -

Gesundheitsforschung - Gesundheitsschutz 53(5): 417–

26.

With a Little Help from My Conversational Agent: Towards a Voice Assistant for Improved Patient Compliance and Medication Therapy

Safety

799

Pharma-Daten-Service, ABDATA. 2019. ABDA-

Datenbank.

Reimers, Kai, and Stefan Klein. 2015.

Arzneimitteltherapiesicherheit Im Spannungsfeld von

Vollständiger Medikationsübersicht, Mündigem

Patienten Und Individualisierter Medikation. 3rd ed.

Cuvillier Verlag Göttingen.

Rottenkolber, Dominik et al. 2011. ‘Adverse Drug

Reactions in Germany: Direct Costs of Internal

Medicine Hospitalizations’. Pharmacoepidemiology

and Drug Safety 20(6): 626–34.

Salvi, Fabio et al. 2012. ‘Adverse Drug Events as a Cause

of Hospitalization in Older Adults’. Drug Safety 35:

29–45.

Santo, Karla et al. 2016. ‘Mobile Phone Apps to Improve

Medication Adherence: A Systematic Stepwise Process

to Identify High-Quality Apps’. JMIR mHealth and

uHealth 4(4): e132–e132.

Schäfer, Christian. 2011. Patientencompliance : Durch

verbesserte Therapietreue Effizienzreserven

ausschöpfen TT - Patient compliance: Using efficacy

reserves by improvement of treatment compliance.

Schäfer, Christian: Universität Mainz; Gutenberg

School of Management and Economics (Germany):

Gabler.

Schrepp, Martin, Andreas Hinderks, and Jörg

Thomaschewski. 2014. ‘Applying the User Experience

Questionnaire (UEQ) in Different Evaluation

Scenarios’. In International Conference of Design,

User Experience, and Usability, Springer, 383–92.

Schurig, A. Marlen et al. 2018. ‘Adverse Drug Reactions

(ADR) and Emergencies-the Prevalence of Suspected

ADR in Four Emergency Departments in Germany’.

Deutsches Arzteblatt International 115(15): 251–58.

Sebillo, Monica, Giuliana Vitiello, Danilo Cuciniello, and

Serena Carrabs. 2017. ‘Human-Centered Design of a

Personal Medication Assistant - Putting Polypharmacy

Management into Patient’s Hand!’ In Lecture Notes in

Computer Science, Springer Verlag, 685–99.

Sedlmayr, Martin. 2018. ‘EHealth Als Schlüssel Für

Bessere Patientencompliance-Technische

Möglichkeiten Und Medizinische Herausforderungen’.

GesundheitsRecht 17(1): 17.

Silva, Bruno M et al. 2013. ‘A Mobile Health Application

for Outpatients Medication Management’. In IEEE

International Conference on Communications, Institute

of Electrical and Electronics Engineers Inc., 4389–93.

Sneha, Sweta, and Upkar Varshney. 2012. ‘Strategies

Towards Chronic Disease Management via Medication

Compliance’. AMCIS 2012 Proceedings.

Stark, Renee G, Jürgen John, and Reiner Leidl. 2011.

‘Health Care Use and Costs of Adverse Drug Events

Emerging from Outpatient Treatment in Germany: A

Modelling Approach’. BMC Health Services Research

11.

Tafreshi, Mohammad J, Michael J Melby, Keith R Kaback,

and Teresa C Nord. 1999. ‘Medication-Related Visits

to the Emergency Department: A Prospective Study’.

Annals of Pharmacotherapy 33(12): 1252–57.

Tang, Lei et al. 2011. ‘MHS: A Multimedia System for

Improving Medication Adherence in Elderly Care’.

IEEE Systems Journal 5(4): 506–17.

Taylor, Robert, Joseph V Pergolizzi, R Amy Puenpatom,

and Kent H Summers. 2013. ‘Economic Implications of

Potential Drug-Drug Interactions in Chronic Pain

Patients’. Expert Review of Pharmacoeconomics and

Outcomes Research 13(6): 725–34.

Teixeira, António et al. 2017. ‘Design and Development of

Medication Assistant: Older Adults Centred Design to

Go beyond Simple Medication Reminders’. Universal

Access in the Information Society 16(3): 545–60.

Tiwari, Priyesh et al. 2011. ‘Feasibility Study of a Robotic

Medication Assistant for the Elderly’. In Conferences

in Research and Practice in Information Technology

Series, , 57–66.

Volland, Dirk. 2015. Spring ‘Extending Pharmacist-Patient

Communication with ICT’. St. Gallen.

Vrijens, Bernard et al. 2008. ‘Adherence to Prescribed

Antihypertensive Drug Treatments: Longitudinal Study

of Electronically Compiled Dosing Histories’. Bmj

336(7653): 1114–17.

Wang, Weiyu, and Keng Siau. 2018. ‘Trust in Health

Chatbots’.

Scale-IT-up 2021 - Workshop on Scaling-Up Healthcare with Conversational Agents

800