Simulations of a Computational Model for a Virtual Medical Assistant

Aryana Collins Jackson

1 a

, Marl

ene Gilles

1 b

, Eimear Wall

, Elisabetta Bevacqua

1 c

Pierre De Loor

and Ronan Querrec

ENIB, Lab-STICC UMR 6285 CNRS, Brest, France

Health Service Executive, Ireland

Keywords:

Intelligent Agents, Embodied Conversational Agents, Human-computer Interaction, Situational Leadership,

Medicine, Simulation.

Abstract:

We propose a virtual medical assistant to guide both novice and expert caregivers through a procedure with-

out the direct help of medical professionals. Our medical assistant uses situational leadership to handle all

interaction with a caregiver, which works by identifying the readiness level of the caregiver in order to match

them with an appropriate style of communication. The agent system (1) obtains caregiver behavior during the

procedure, (2) calculates a readiness level of the caregiver using that behavior, and (3) generates appropriate

agent behavior to progress the procedure and maintain a positive interaction with the caregiver.

1 INTRODUCTION

Due to advancements in virtual agents and

telemedicine, caregiving in various capacities

can now be done at a distance. For example, virtual

agents can perform health assessments (Montenegro

et al., 2019) and act as liaisons with medical staff

(Bickmore et al., 2015). In these examples, the agent

interacts directly with the patient. However, there are

cases in which an agent should interact only with a

caregiver. When a medical procedure is necessary

and there are no trained medical staff nearby, the

individual present must assume a caregiving role, and

they will need guidance on preserving the patient’s

health. During these situations, a permanent virtual

medical assistant could be useful (Nakhal, 2017).

The assistant will communicate with the caregiver

and will also communicate to a team of medical ex-

perts standing by on a different site monitoring the

procedure. Because of the possible latency and dis-

ruptions in communication between the remote site

and the site where the experts are, the virtual assistant

must be capable of guiding the caregivers through the

entire procedure without relying on the help of the

medical experts.

A huge part of organizing a successful procedure

is maintaining a positive interaction with the caregiver

https://orcid.org/0000-0002-4385-5258

https://orcid.org/0000-0003-1806-1672

https://orcid.org/0000-0002-7117-3748

while maintaining the health of the patient (Hjort-

dahl et al., 2009; Henrickson et al., 2013; Yule et al.,

2008; Flin et al., 2010). Both the caregiver and the

agent must work towards a common goal: the com-

pleted procedure that preserves the health of the pa-

tient. To foster this positive working alliance be-

tween agent and caregiver, situational leadership is

employed (Hersey et al., 1988). In situational leader-

ship, a leader adapts their behavior to the follower ac-

cording to their experience level and expertise. Situa-

tional leadership involves four readiness levels which

directly correspond to four leadership styles. The

readiness levels involve varying degrees of ability,

which refers to a person’s competence when complet-

ing tasks, and willingness, which refers to their con-

ﬁdence and interest in completing those tasks. In this

scenario, the virtual agent assumes the role of leader,

and the caregiver assumes the role of follower. Situa-

tional leadership is covered in more depth in section 2.

Previously when situational leadership has been

studied, it has been implemented (Sims et al., 2009;

Figure 1: A basic representation of the agent framework.

Jackson, A., Gilles, M., Wall, E., Bevacqua, E., De Loor, P. and Querrec, R.

Simulations of a Computational Model for a Virtual Medical Assistant.

DOI: 10.5220/0010910700003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 1, pages 94-105

ISBN: 978-989-758-547-0; ISSN: 2184-433X

Hersey et al., 1988; Lourdeaux et al., 2019) or simu-

lated for intended use in human-human relationships

(Bosse et al., 2017; Bernard, 2020). Virtual leaders

have used different methods of guiding humans and of

choosing behavior that is personalized to the individ-

ual humans. However, our work is the ﬁrst to present

both a computational model of situational leadership

and an agent framework in which a virtual agent leads

humans (see Figure 1 for a basic representation of the

agent framework and Figure 6 for a detailed view).

We contribute an implementation of situational lead-

ership in which we (1) quantify the factors that deter-

mine readiness level, (2) create an algorithm for cal-

culating readiness level based on these criteria, and

(3) create a virtual agent system that uses readiness

level and human behavior to create personalized be-

havior.

While our work is conducted with the context of a

medical procedure in mind, the model itself is ﬂexible

enough to be applicable to any situation in which a

human being must be led through a series of steps by

a virtual assistant.

In this paper, we detail what situational leadership

is and how it works, we discuss the current state of

the art regarding virtual leaders, we detail our com-

putational model of situational leadership, we explain

how our model is implemented using Mascaret, and

we discuss our conclusions and future work.

2 SITUATIONAL LEADERSHIP

As discussed in the introduction, situational leader-

ship encompasses four follower readiness levels and

four leadership styles. The four readiness levels (de-

noted with an R and a number) are listed below

(Hersey et al., 1988):

1. R1: Low ability, low willingness;

2. R2: Low to some ability, high willingness;

3. R3: High ability, variable willingness;

4. R4: High ability, high willingness.

Readiness levels R1 and R2 need the leader’s help

while readiness levels R3 and R4 should be able to

complete tasks without the leader’s help.

When a leader employs situational leadership,

they perform both task and relationship behavior to

guide the follower. Task behavior corresponds to a

follower’s competence (Bedford and Gehlert, 2013).

Relationship behavior refers to how the leader fosters

and maintains a positive working alliance with the fol-

lower and corresponds to a follower’s willingness, or

conﬁdence (Bedford and Gehlert, 2013). The lead-

ership styles that correspond to these readiness levels

are (Hersey et al., 1988):

1. Directing: high task, low relationship behavior;

2. Coaching: high task, high relationship behavior;

3. Supporting: low task, high relationship behavior;

4. Delegating: low task, low relationship behavior.

Note that low relationship behavior is required for

followers with low ability and low willingness be-

cause the person may not be interested in being en-

couraged. For followers with high ability and low to

some willingness, the agent motivates the person be-

cause they have the skills necessary to complete the

procedure successfully.

Follower behavior that can indicate a particular

readiness level includes both prior experience and

knowledge and non-anatomical behavior such as the

number of questions asked, the amount of hesitation

after receiving tasks, etc. (Collins Jackson et al.,

2019; Bosse et al., 2017; Hersey et al., 1988):

Situational leadership is an effective method of

managing followers because it accounts for follow-

ers with both high and low ability and willingness.

This model has been proven effective in an educa-

tional environment in which medical instructors im-

plemented leadership styles in their classes and with

individual students (Sims et al., 2009), work environ-

ments in which managers adopted leadership styles

with their direct reports (Hersey et al., 1988), and in

a clinical medical supervision context (Bedford and

Gehlert, 2013). Situational leadership has also been

implemented in a virtual environment to help train

medical staff (Lourdeaux et al., 2019).

There is very limited work on computational mod-

els using situational leadership, and existing work in-

volves human-human interaction only (Bosse et al.,

2017; Bernard, 2020). Rarely are agents in positions

of leadership in human-agent interaction. Our work

models the work on situational leadership and also

provides a new method in which agents and humans

can interact by allowing an agent to lead a human.

Our model permits an agent to maintain a positive in-

teraction with a follower and interact with the human

in a way that is most beneﬁcial for them in addition to

maintaining the health of the patient.

In the following section, we discuss previous work

on virtual agent leaders.

3 VIRTUAL LEADERS

This work involves a virtual agent acting as a med-

ical assistant who leads a caregiver through a medi-

cal procedure. Within the medical domain, there has

Simulations of a Computational Model for a Virtual Medical Assistant

been a lot of work regarding human-human relation-

ships in the emergency room. The person leading a

medical procedure holds an important role in that they

manage both the procedure tasks, the health of the pa-

tient, and the interaction with the caregiver(s). Thus

having the trust of the caregivers and being compe-

tent at their work are two of the most important qual-

ities a leader can have (Hjortdahl et al., 2009). Addi-

tional qualities that a medical leader should embody

include communication, negotiation, autonomy, cre-

ativity, and appreciation of the caregivers (Araszewski

et al., 2014). Many of these qualities are therefore in-

cluded in medical professional behavior taxonomies

(Henrickson et al., 2013; Yule et al., 2008; Flin et al.,

2010).

In terms of virtual agents in medicine, there are a

variety of examples of virtual agents created for use

speciﬁcally in medical situations: agents which act

as liaisons between patients and physicians using pre-

scripted speech (Bickmore et al., 2015), agents which

are meant to connect emotionally with patients for

mental health beneﬁts (Yang and Fu, 2016), agents

which act as companions for the elderly (Montenegro

et al., 2019), and agents which conduct psychiatric in-

terviews (Philip et al., 2020). In each of these exam-

ples, user engagement with the agent was dependent

on whether they trusted the agent.

The concept of developing trust is inherent in all

types of human-computer interactions, regardless of

domain, as it is a prerequisite to a positive interaction

(Hoegen et al., 2019; Kulms and Kopp, 2016; Lee

et al., 2021). Additionally, trust leads to greater ef-

ﬁciency when the human is completing tasks (Kulms

and Kopp, 2016). Situational leadership was devel-

oped knowing that followers’ trust of their leader is

hugely important to successful interactions, and that

trust can be built and maintained when leaders interact

in an appropriate manner to followers (Hersey et al.,

1988). Thus employing situational leadership itself is

a method of building and maintaining trust between a

human and an agent.

Because our work involves an agent leading a hu-

man, we examined previous work in which agents

exist in pedagogical scenarios. We again found that

trust was important to student-tutor relationships and

was a product of a supporting environment (Castel-

lano et al., 2013; Saerbeck et al., 2010). These rela-

tionships are most successful when the agent adapts

to the learner by responding based on his or her previ-

ous experience and level of knowledge (Pecune et al.,

2010; Taoum et al., 2018; Cisneros et al., 2019; Quer-

rec et al., 2018).

As mentioned in the introduction, few computa-

tional models on situational leadership have been de-

veloped (Bosse et al., 2017; Bernard, 2020). While

these models have involved human-human interaction

only, they have provided a foundation for computing

readiness level and therefore determining leadership

style from various behavioral parameters. One model

is based entirely on follower behavior and allows the

leader to adapt their leadership style by monitoring

a follower’s advancement through the different readi-

ness levels (Bosse et al., 2017). This work in particu-

lar formed the backbone of our own research.

In the following section, we outline our own com-

putational model of situational leadership when a vir-

tual agent is a medical leader.

4 A COMPUTATIONAL MODEL

OF SITUATIONAL

LEADERSHIP

Readiness level is determined by examining follower

behavior. Other research has analyzed behavior such

as facial actions and head movements (Dermouche

and Pelachaud, 2019). However, we do not uti-

lize a human activity monitor in this project. While

these kinds of non-verbal behaviors are certainly use-

ful, we chose to prioritize the caregiver’s ability to

move around the environment without worrying about

whether their faces were reliably detected. Instead,

we focus on human behavior that can be monitored

and input from a keyboard.

In this section, we discuss the parameters used in

our model, we detail the follower behavior that are

used, we discuss why we chose certain parameter val-

ues, we thoroughly explain how the model works, and

we explain how our model can be used to adhere to

guidelines set by the existing work on situational lead-

ership.

4.1 Model Parameters

Our model uses several parameters and various care-

giver behaviors to determine readiness level. From

now on, we refer to individual behaviors as behavior

criteria or simply criteria.

Two of the parameters needed in this algorithm,

persistence and valence, are adapted from an exist-

ing model (Bosse et al., 2017), but the others are

our own contributions. The parameters explained be-

low are used in equations 1 and 2 (detailed in sec-

tion 4.4) to calculate a performance value which de-

scribes how high the follower’s performance is in re-

gards to each criterion. These values change depend-

ing on the procedure. Before explaining how that cal-

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

Table 1: The behavior criteria used for computing readiness level and each of their parameters. The persistence values change

depending on whether the follower has low or high ability and willingness.

Criterion Name Domain Persistence:

high

Persistence:

low

Weight Performance

threshold

1 Error: action in task ability 0.7 0.85 0.05 0

2 Error: action outside task ability 0.1 0.85 0.25 1

3 Wrong resource chosen ability 0.7 0.85 0.15 0.75

4 No resource chosen ability 0.2 0.85 0.25 1

5 Action duration too short ability 0.1 0.85 0.06 0

6 Action duration too long ability 0.1 0.85 0.07 1

7 Question for help ability 0.4 0.85 0.17 1

8 Hesitation willingness 0.9 0.9 0.5 0.5

9 Question for reassurance willingness 0.9 0.9 0.5 0

culation works, we explain what each parameter is:

Extent refers to the extent to which the follower

exhibits the criterion and is described by either 0 or 1

where 0 indicates that the follower does not embody

that criterion at all, and 1 indicates that the follower

embodies that criterion to the fullest extent possible.

Domain refers to whether the criterion is an indi-

cator of ability or willingness; criteria can either exist

in the ability domain or the willingness domain.

Persistence describes how much the behavior dur-

ing a previous task inﬂuences the current readiness

level with a ﬂoat on the interval [0, 1]. A value of 0

indicates low persistence while a value of 1 indicates

high persistence. The lower the persistence value, the

faster the performance value will rise or fall when the

extent to which the follower exhibits that criterion is

high or low respectively (Bosse et al., 2017).

Weight refers to the importance of the criterion

within the ability and willingness domains when de-

termining readiness level and is described by a ﬂoat

on the interval [0, 1].

Performance threshold refers to the lowest perfor-

mance value possible of that criterion before the fol-

lower could be considered to have low ability or low

willingness.

4.2 Follower Behavior Criteria

In previous research, the extent to which a follower

exhibits thirty-three different behaviors determines

readiness level (Bosse et al., 2017). That model is

based on an interaction between a student and a su-

pervisor, and thus many of these behaviors were either

irrelevant to a medical scenario (such as feeling over-

obligated and lacking self-esteem) or were impossible

to compute within a virtual environment (such as de-

fensive behavior and discomfort in body language).

Additionally, these thirty-three behaviors are catego-

rized by readiness level, indicating that each behavior

is indicative of only one readiness level. While our

work is based on this research, in order to adapt it to

the context of a medical procedure led by a virtual

agent, some changes had to be made.

Using that existing list of behaviors (Bosse et al.,

2017), we created our own list of behavior criteria

(see Table 1). Instead of grouping them by readiness

level, we group them by whether they are indicators of

ability or willingness. Because all four readiness lev-

els can be described as having ability and willingness

that is high, low, or somewhere in between, follow-

ers of any style can embody each of the behaviors to

varying extents. The list of criteria are explained be-

low. Note that in the procedure, there are both tasks

and actions. A task is a complete step, while an action

is a sub-step of a task. Thus each task may involve

multiple actions. A resource is a tool that can be used

to perform an action. Each action has an appropriate

duration that it is expected to take. All of this infor-

mation exists within Mascaret, our agent framework,

which is explained further in section 6.

1. Error: action in task: The follower has chosen to

do an action out of order;

2. Error: action outside task: The follower has cho-

sen to do an action that does not exist in the cur-

rent task;

3. Wrong resource chosen: The follower has taken

the wrong resource for the action;

4. No resource chosen: The follower has neglected

to take a resource when one is required or tries to

take a resource when none is required;

5. Action duration too short: The action duration is

less than 0.9 times the expected action duration

(as devised by our medical professional);

6. Action duration too long: The action duration is

more than 1.1 times the expected action duration

(as devised by our medical professional);

7. Question for help: The follower has asked a ques-

tion because they do not know how to proceed;

8. Hesitation: The follower has hesitated for more

than ﬁve seconds before beginning an action;

Simulations of a Computational Model for a Virtual Medical Assistant

9. Question for reassurance: Also referred to as a

clarifying question; the follower has asked a ques-

tion to ensure what they are doing is correct.

4.3 Expert-deﬁned Parameter Values

The values for the parameters in Table 1 were devised

by consulting with the medical doctor on our team.

The procedure that these values were created for is the

diagnosis of abdominal pain, and with these values,

we assume that the case is not urgent.

Because our context is rather speciﬁc, it was im-

portant to create our model with the guidance of our

team’s medical professional. Therefore, all of these

values are very procedure speciﬁc. The values for per-

sistence, weight, and performance threshold would

change depending on the procedure.

For a non-urgent diagnosis of abdominal pain, the

most important ability criteria are Error: action out-

side task and No resource chosen because they indi-

cate that the follower does not understand the current

task and how to reach the end goal. Each are given

the highest weight of 0.25. If the follower is consid-

ered to be high-ability, these two errors change their

performance values considerably with low persistence

values of 0.1 and 0.2 respectively.

Question for help is the next most important cri-

terion in the ability domain because it indicates that

the follower does not know how to proceed without

the agent’s help. This criterion is given a moderate

persistence of 0.4 when the follower is considered to

have high-ability because while questions for help are

indicative of more novice behavior, the follower’s his-

tory should be more important than it is for Error:

action outside task and No resource chosen.

Unlike the previous three criteria, there is room for

human error in Wrong resource chosen where a fol-

lower may accidentally reach for the wrong resource,

hence the higher persistence value of 0.7.

Error: action in task, Action duration too short,

and Action duration too long have the lowest weights.

The most important thing is that the task has been

completed, and so errors out of order do not matter

so much as long as the task is being completed. Per-

sistence for this error is set quite high. Action com-

pletion duration times are not important unless there

is a life-threatening emergency. However, the persis-

tence is set low for the action duration errors as these

can co-occur with other criteria, such as Error: ac-

tion in/outside task and Question for help and there-

fore they should be taken quite seriously. Because of

their low weight, short or long action duration times

on their own do not generally affect the overall ability

performance unless they happen repeatedly through-

out the procedure.

Finally, Hesitation and Question for reassurance

are weighted equally in the willingness domain. Each

of these are equally indicative of a follower’s willing-

ness. Additionally, the persistence for both is set quite

high at 0.9 because a follower’s willingness should be

calculated from their history rather than from a single

task.

As shown in Table 1, there are two different val-

ues for persistence, one for followers are deemed to

have low ability or willingness and one for those who

are deemed to have high ability or willingness. These

values were devised in order to account for a novice

who does everything correctly versus an expert who

makes one or two mistakes. During a medical proce-

dure, even one small mistake can lead to serious con-

sequences. When a follower displays high-ability be-

havior, the mistakes they make should have more im-

pact on the performance value. The Persistence: high

values are used when a follower begins the proce-

dure with high ability or willingness. When the abil-

ity performance drops below the ﬁrst ability threshold

of 0.8525, then the Persistence: low values are used

instead. Persistence: low values are also used for fol-

lowers who begin the procedure with low ability or

willingness.

Additionally, persistence changes based on the

difﬁculty of the current action so that a follower who

is only easily able to do easier actions and a follower

who is able to do harder actions are not considered to

be the same level. Each action in the procedure is as-

signed a value of 1 (most difﬁculty), 2 (medium difﬁ-

culty), and 3 (least difﬁculty). If the action is assigned

a 1, then the persistence value for that task decreases

by 0.05 to ensure that the persistence is never 1 or

0. If the action is assigned a 3, then the persistence

value for that task increases by 0.05. If the action is

assigned a 2, there is no change to persistence. This

ensures that more difﬁcult tasks are weighted more

when the performance value is calculated.

4.4 The Model Explained

The ﬁrst step to determining readiness level is to

calculate the performance value for each criterion.

The performance value is a ﬂoat on the interval [0, 1]

where a value of 0 indicates low ability or willingness

and a value of 1 indicates high ability or willingness.

Equation 1 is used to calculate the performance

value v at time t for criterion c, where p represents the

persistence and e represents the extent (Bosse et al.,

2017).

c,t

= (p

∗ v

c,t−1

) + ((1 − p

) ∗ (1 −e

c,t

)) (1)

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

Equation 1 requires the performance value at time

t − 1. At the very beginning of the procedure, default

performance values are used in place of v

c,t−1

. Val-

ues of 0 or 1 are established for each criterion which

correspond to the follower’s previous experience and

knowledge, called the follower proﬁle. The follower

proﬁle establishes which readiness level the individ-

ual has prior to beginning the procedure and is based

on previous procedures, external evaluations of the

follower, and self-evaluations. For a follower proﬁle

indicating low or variable ability or willingness, the

default performance value is 0, and for a follower pro-

ﬁle indicating high ability or willingness, the default

performance value is 1.

For example, if a follower proﬁle indicates that the

individual’s readiness level is style 3, then they have

high ability and variable willingness (see Figure 2).

The default performance values for all the criteria in

the ability domain are 1, and the default performance

values for the criteria in the willingness domain are 0.

Calculating the performance value v for Error: action

in task for a follower of style 3 who has made an error

at the very beginning of the procedure would look like

this: (0.1 ∗ 1)+ ((1 − 0.1) ∗ (1 − 1)) = 0.1.

The performance value v is then fed back into the

model in equation 1 the next time readiness level is

calculated.

The next step of determining readiness level is to

average all the performance values within each do-

main. Equation 2 is used where n refers to the num-

ber of criteria in the ability and willingness domains,

weight

refers to the weight that each criterion holds,

and v

c,t

refers to the performance value at time t of

each criterion as calculated by equation 1. Equation 2

results in two values representing the follower’s over-

all ability and willingness behavior O.

∑

c=1

weight

∗ v

c,t

(2)

There are two different thresholds each for deter-

mining whether a follower has low or high ability and

willingness. For the ﬁrst, the performance thresholds

shown in Table 1 are used to determine whether a fol-

lower has high or low ability and willingness. When

using the performance thresholds in equation 1 for

both e

and v

c,t−1

, and then using the resulting v

c,t

value in equation 2, we achieve the overall ability per-

formance threshold of 0.8525 and the overall willing-

ness performance threshold of 0.625. If a follower has

an ability value below 0.8525, they have low ability,

and if that value is above or equal to 0.8525, they have

high ability. If a follower has a willingness value be-

low 0.625, they have low willingness, and if that value

is at or above 0.625, they have high willingness.

The second threshold is deﬁned by ability and

willingness equal to or greater than 0.95. This value

was devised by examining example follower behavior

and determining which were able to self-lead without

the leader’s help. Followers with both an ability and

a willingness value equal to or greater than 0.95 are

considered to be in R4. See Figure 2 for a visualiza-

tion of the thresholds and readiness levels.

Figure 2: A visualization of the four readiness levels and

the performance values needed in order to move between

them.

As shown, each readiness level is color-coded

uniquely. A follower’s ability is contained on the ver-

tical axis while willingness is described on the hori-

zontal axis. Once the initial thresholds of 0.825 and

0.625 are reached for ability and willingness respec-

tively, a follower is considered to be R3. Because

R4 is designated for completely self-sufﬁcient follow-

ers, the thresholds for R4 are higher. Ensuring that

movement between the readiness levels is linear is

discussed later in this section.

For examples of how this all works, refer to Fig-

ure 3. In this example procedure, Error: action out-

side task is examined. There are ten actions with the

following action difﬁculties: 3, 3, 3, 3, 2, 2, 1, 2, 1,

and 2. A follower with an existing readiness level of

R1 and a follower with an existing readiness level of

R4 perform the exact same steps; they each try to do

an action that is outside the scope of the task for ac-

tions 5, 7, and 9. As shown, the performance value

v dips for both the R1 and R4 follower, but the er-

rors have a greater effect on the R4 follower who is

supposed to be an expert.

Note that the extent values for each criterion re-

main the same until there is another opportunity to

change that value. For example, the extent value e

of Error: action outside task only changes when the

Simulations of a Computational Model for a Virtual Medical Assistant

Figure 3: Two examples of the evolution of the performance

value v in a procedure in which an R1 and an R4 follower

perform the criterion Error: action outside task at actions

5, 7, and 9.

follower has the opportunity to make or not make an-

other error (i.e., when it is time for the caregiver to

begin a new action). The performance values v are

only calculated when e changes. The overall ability

and willingness values are updated when the individ-

ual v values change.

4.5 Adhering to Situational Leadership

Finally, there are two possible issues that should be

avoided when using this model: (1) the skipping

of readiness levels (followers that move directly be-

tween R1 and R3 as well as R2 and R4) and (2) fol-

lowers moving between levels too quickly (Hersey

et al., 1988; Bosse et al., 2017). To combat these

issues, we implement a method of artiﬁcially lower-

ing ability and lowering or raising willingness values

as needed, which is similar to dynamic range com-

pression (Kates, 2005). This allows for one domain’s

overall performance to wait while the other domain’s

performance catches up.

There are several of instances in which the abil-

ity or willingness value would need to be artiﬁcially

changed:

• An R1 follower whose ability rises faster than

willingness will need their ability value low-

ered when their willingness crosses the high-

willingness threshold to ensure they move from

R1 to R2 (see Figures 4 and 5c);

• An R3 follower whose ability drops below the

high-ability threshold but whose willingness is

still low will need their ability value lowered and

may need their willingness value raised to ensure

they move from R3 to R2 (see Figures 5j, 5k, 5m,

and 5o);

• An R2 follower whose willingness drops below

the high-willingness threshold but whose ability

is high will need their ability value lowered to en-

sure they move from R2 to R1;

• An R4 follower whose ability or willingness drops

below the R4 thresholds of 0.95 will need the

other domain’s value lowered to either 0.8525

(ability) or 0.625 (willingness) in order to ensure

that the follower moves from R4 to R3 and from

R3 to R2 smoothly (see Figure 5m.

Because our context is a medical procedure in

which a patient’s health is at risk, when in doubt, we

always assume that a follower’s ability is lower than

it might be (Hjortdahl et al., 2009). For this reason,

followers’ ability values are only artiﬁcially lowered.

As explained in section 6, the agent’s communication

content does not vary much between levels R1 and

R2 or R3 and R4, and so willingness values can be

lowered or raised when necessary.

As an example of how values might be artiﬁcially

changed, examine an R1 follower whose ability and

willingness increase at different rates. Before exam-

ining what this looks like, we can refer to Figure 2

and see that this would mean that the follower would

move directly from R1 to R3. Because a follower can

only progress from one readiness level to the next, we

artiﬁcially lower the overall ability value in order to

wait for the willingness performance to catch up. See

Figure 4 for a demonstration of how this works.

Figure 4: An example of how ability values are artiﬁcially

lowered in order to ensure the follower progresses from one

level to the next. Green represents readiness level R1 while

pink represents readiness level R2.

In Figure 4, the follower passes the threshold for

high ability at action 14. At this point, however,

the willingness value is still low, so the follower is

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

100

still considered to be R1. The procedure continues,

and at action 18, the follower’s willingness passes the

threshold for high willingness. Since their willingness

performance has increased above the threshold, the

follower can now be considered to be in R2. Readi-

ness level R2 is deﬁned by low to some ability and

high willingness, and so the follower’s ability value is

artiﬁcially decreased to 0.25, a value we chose in or-

der to allow the system to “remember” the follower’s

performance history somewhat. In Figure 4, we can

see that for actions 18-22, the follower is clearly in

R2. In this way, the follower can progress until their

ability once again passes the ability threshold. If, at

that point, willingness has remained high, then the

follower will progress to R3.

By manually lowering the ability and willingness

values as needed, the system acts similarly to dynamic

range compression in which signals are limited once

they reach a certain threshold (Kates, 2005). Unlike

compression, we lower the values in order to allow the

follower to progress themselves through to the next

readiness level. This is how we ensure that not only

do followers progress from one readiness level to the

next without skipping levels, but we also ensure that a

follower does not move between readiness levels too

quickly.

In the following section, we demonstrate how

readiness level evolves in a variety of scenarios.

5 SIMULATIONS OF FOLLOWER

PROGRESSION

In order to visualize how readiness level might

change during a medical procedure and evaluate those

changes to ensure they are indicative of the correct

readiness level in practicality, we establish four start-

ing states (R1, R2, R3, and R4 followers) and four

possible progressions: (1) ability increases or de-

creases, willingness stays the same; (2) ability stays

the same, willingness increases or decreases; (3) abil-

ity increases or decreases, willingness increases or de-

creases; and (4) ability remains the same, willingness

remains the same. This results in sixteen different sce-

narios.

As discussed in section 4.3, the criteria in each

scenario are created with help from the medical pro-

fessional on our team and are based on real-world

errors that followers in each readiness level would

make. Therefore, creating these scenarios is also a

method of testing our model and the values for the

parameters described in section 4.3. The simulations

included in this section are not exhaustive; many more

were tested in order to ensure that the values for the

parameters in our model returned accurate readiness

levels.

The scenarios use a sample of a procedure which

comprises a total of 22 actions. The ﬁrst 9 actions

have the least difﬁculty, actions 10-12 have the most

difﬁculty, actions 13-14 have the least difﬁculty, and

actions 15-22 have medium difﬁculty.

Figures 5a-5p display the 16 scenarios and are

color-coded according to Figure 2: green represents

R1, pink represents R2, blue represents R3, and yel-

low represents R4. Note that unless a follower is

making every possible error constantly throughout the

procedure, which often isn’t possible, ability and will-

ingness will increase from 0.

6 AGENT FRAMEWORK

Figure 6 describes the ﬂow of information and how

interaction is possible in our agent framework based

on Mascaret. The agent platform contains the means

to calculate readiness level and the information re-

garding the nominal procedure. The nominal pro-

cedure contains all information regarding resources,

roles, procedure steps, and also allows for changes to

be made to the procedure.

When a caregiver makes an error in Table 1, a cri-

teria event is created which describes the error. The

agent platform calculates readiness level from the hu-

man caregiver’s speech and behavior, and then readi-

ness level and criteria events stemming from care-

giver behavior are created and work together to cre-

ate communication intentions. The modality for agent

communication is both verbal and non-verbal behav-

ior, so communication intentions (verbal behavior)

are paired with appropriate non-verbal behavior in the

behavior planner (Collins Jackson et al., 2020).

The Mascaret framework permits the modeling of

semantic, structural, geometric, and topological prop-

erties of the entities in the virtual environment and

their behaviors. Mascaret also deﬁnes the notion of

a virtual agent by their behaviors, their communica-

tions, and their organisation. Essentially, it is a frame-

work in which an embodied virtual human can inter-

act with a user.

In Mascaret, human activities can be described in

the virtual environment by using predeﬁned collabo-

rative scenarios (called procedures) which represent

plans of actions for agents or instructions provided

to users for assisting them. With Mascaret, a whole

medical procedure with roles, resources, and trajecto-

ries can be formalized and virtually executed to assist

a caregiver to go through this same procedure in the

real world.

Simulations of a Computational Model for a Virtual Medical Assistant

101

(a) R1: Ability rises, will-

ingness remains low.

(b) R1: Willingness rises

above the threshold.

ness rise above the thresh-

olds.

(d) R1: Ability and willing-

ness remain low.

(e) R2: Ability rises above

the threshold, willingness

remains high.

(f) R2: Ability remains low,

willingness drops below the

threshold.

(g) R2: ability rises above

the threshold, willingness

drops below the threshold.

(h) R2: Ability remains

low, willingness remains

high.

(i) R3: Ability remains

high, willingness rises

above the threshold.

(j) R3: Ability drops be-

low the threshold, willing-

ness remains low (but is ar-

tiﬁcially raised).

(k) R3: Ability drops be-

low the threshold, willing-

ness rises above the thresh-

old.

(l) R3: Ability remains

high, willingness remains

low.

(m) R4: Ability drops be-

low the threshold, willing-

ness remains high (but is ar-

tiﬁcially lowered).

(n) R4: Ability remains

high, willingness drops be-

low the threshold.

(o) R4: Ability and willing-

ness drop below the thresh-

olds.

(p) R4: Ability and willing-

ness remain high.

Figure 5: Sixteen simulations demonstrate how readiness level evolves when ability and willingness increases, decreases, or

remains the same, four starting in each readiness level. The solid brown line indicates the overall ability performance, and the

dotted green line indicates the overall willingness performance. Green shading refers to R1, pink shading refers to R2, blue

shading refers to R3, and yellow shading refers to R4.

To allow Mascaret to interact with the user in

a natural way, we introduce an embodied conversa-

tional agent (ECA) based on the SAIBA framework

(Vilhj

almsson et al., 2007) and which has an intent

planner which generates the communicative inten-

tions of the agent, a behavior planner which trans-

lates communicative intentions into verbal and non-

verbal signals, and a behavior realizer which trans-

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

102

Figure 6: The general ﬂow of information within our agent framework. Readiness level (highlighted in yellow) informs

leadership style. Criteria event and leadership style (highlighted in green) inform the communication actions.

forms these signals in animation.

In Mascaret, we introduce the concept of follower

criteria (see Table 1). After each action is performed

by the follower, each behavior criterion is evaluated

to see if it has been triggered (e.g., if the person has

tried to do an action that is outside the task, if they

have asked a question because they are stuck, etc.).

According to the triggered criteria, the readiness level

is computed (as explained in section 4).

The readiness level is used to compute the leader-

ship style which is then stored in the intent planner

(Collins Jackson et al., 2021)). The intent planner

then generates a communicative intention each time

a new action in the procedure must be done or when

a follower criterion that needs attention has been trig-

gered (for example, the virtual agent must inform the

caregiver if they have made a mistake, must answer a

question that the caregiver has asked, etc.).

To determine the ECA’s communicative intention,

we implement a set of rules which dictates how and

what the agent will communicate according to its

leadership style to be used in conjunction with lin-

guistic rules speciﬁed in previous work (Collins Jack-

son et al., 2022):

Directing and coaching leadership (correspond-

ing to readiness levels 1 and 2):

• Agent communicates every action to do;

• Agent communicates every time there is a criteria

event.

Supporting leadership (corresponding to readi-

ness level 3):

• Agent only communicates when criteria events

are generated by criteria 1-7 and 9 have occurred

(see Table 1 for criteria numbers).

Delegating leadership (corresponding to readi-

ness level 4):

• Agent only communicates when 1, 2, 7, or 9 have

occurred.

Thus readiness level and by extension, leadership

style, criteria events, and procedure actions inform the

communication actions that are created in the intent

planner.

Note that unstructured dialogue between the care-

giver and the agent is not possible. If the caregiver

asks a question, the agent is able to respond, but the

agent’s speech is limited to only the actions and re-

sources within the procedure.

To ensure additional safety, the caregiver can de-

cline the agent’s help at any time. When the caregiver

feels competent without the agent’s assistance, they

can decline the agent’s guidance during the procedure

as a whole. In these situations, the agent will only

act as a conduit of communication from the team of

medical experts.

The agent itself is built in Unity and assumes the

appearance of a doctor. In Figure 7, the agent is

shown next to a dummy patient in augmented reality

(AR) which we use for testing. However, our frame-

work also allows for the agent to be displayed on a

computer screen or on a tablet. When testing the sys-

tem in AR, all follower behavior criteria can be mon-

itored in the virtual system. Questions can be asked

via the interface as well and are limited to requests

for help regarding actions, resources, or roles and re-

quests for clariﬁcation regarding trajectory.

Through both the virtual and augmented reality

environments, remote medical experts can visualize

the patient, the caregiver, and the procedure steps

done thus far in 3D physical space and make deci-

sions regarding procedure changes accordingly. In

augmented reality, the caregiver can interact with the

augmented reality environment by selecting which ac-

tions have been completed (as shown in Figure 7). For

Simulations of a Computational Model for a Virtual Medical Assistant

103

more information regarding the virtual environment

set-up, please refer to (Querrec et al., 2018).

Figure 7: Our agent built with Unity pictured in augmented

reality.

7 CONCLUSIONS AND FUTURE

WORK

In this paper, we have described a novel model and

the architecture of a novel system in which situational

leadership is applied in a virtual agent system. Simu-

lations of our model demonstrate how follower readi-

ness level can be calculated for use in an agent frame-

work that generates personalized agent behavior to

each unique human user. Using our model, a virtual

agent can lead a novice or expert caregiver through a

medical procedure. This interaction is possible in our

SAIBA-compliant agent framework using Mascaret,

where interaction is completed with communication

actions actuated by both caregiver behavior and infor-

mation from the formalized procedure.

Our algorithms for determining readiness level

and leadership style can be adapted to other circum-

stances outside the medical sphere as well when it

is appropriate for an agent to lead a human being

through any kind of task.

We aim to evaluate this system to determine

whether an agent employing situational leadership is

more effective (preserves the patient’s health, is more

efﬁcient, and reduces stress of the caregiver) than an

agent using the same leadership style throughout the

procedure regardless of readiness level. Measuring

trust and engagement from participants might come

from non-anatomical behavior such as speed of ac-

cepting tasks. We plan to also measure trust and en-

gagement with the agent with survey questions both

before and after the procedure.

We also plan to further explore the co-occurrence

of criteria to understand better how criteria that

happen concurrently affect overall performance and

whether these interaction effects should be taken into

account when calculating readiness level.

ACKNOWLEDGEMENTS

This work has been carried out within the French

project VR-MARS which is funded by the National

Agency for Research (ANR).

REFERENCES

Araszewski, D., Bolzan, M. B., Montezeli, J. H., and Peres,

A. M. (2014). The exercising of leadership in the

view of emergency room nurses. Cogitare Enferm,

19(1):40–8.

Bedford, C. and Gehlert, K. M. (2013). Situational super-

vision: Applying situational leadership to clinical su-

pervision. The Clinical Supervisor, 32(1):56–69.

Bernard, C. (2020). Analyse de l’activit

e de soignants

edicaux et param

edicaux sur simulateur haute-

ﬁd

elit

e lors de simulations d’urgence, en vue de

la conception d’un environnement de soins er-

gonomique. PhD thesis, University of Lorient.

Bickmore, T., Asadi, R., Ehyaei, A., Fell, H., Henault, L.,

Intille, S., Quintiliani, L., Shamekhi, A., Trinh, H.,

Waite, K., Shanahan, C., and Paasche-Orlow, M. K.

(2015). Context-Awareness in a Persistent Hospital

Companion Agent. In Proceedings of the 15th In-

ternational Conference on Intelligent Virtual Agents,

IVA, pages 332–342, Delft, The Netherlands. ACM.

Bosse, T., Duell, R., Memon, Z. A., Treur, J., and van der

Wal, N. (2017). Computational model-based design

of leadership support based on situational leadership

theory. SIMULATION: Transactions of The Society

for Modeling and Simulation International, 93(7).

Castellano, G., Paiva, A., Kappas, A., Aylett, R., Hastie,

H., Barendregt, W., Nabais, F., and Bull, S. (2013).

Towards empathic virtual and robotic tutors. In Pro-

ceedings of the 16th International Conference on Ar-

tiﬁcial Intelligence in Education, AIED, pages 733–

736, Memphis, TN, USA. Springer-Verlag.

Cisneros, F., Mar

ın, V. I., Orellana, M. L., Per

e, N., and

Zambrano, D. (2019). Design process of an intelli-

gent tutor to support researchers in training. In Pro-

ceedings of EdMedia and Innovate Learning, pages

1079–1084, Amsterdam, Netherlands. AACE.

Collins Jackson, A., Bevacqua, E., De Loor, P., and Quer-

rec, R. (2019). Modelling an embodied conversational

agent for remote and isolated caregivers on leadership

styles. pages 256–259. IVA ’19.

Collins Jackson, A., Bevacqua, E., De Loor, P., and Quer-

rec, R. (2021). A computational interaction model for

a virtual medical assistant using situational leadership.

WI-IAT, Essendon, VIC, Australia. ACM.

Collins Jackson, A., Bevacqua, E., DeLoor, P., and Quer-

rec, R. (2020). A taxonomy of behavior for a medical

coordinator by utilizing leadership styles. pages 532–

543. International Conference on Human Behaviour

and Scientiﬁc Analysis.

Collins Jackson, A., Gl

emarec, Y., Bevacqua, E., De Loor,

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

104

P., and Querrec, R. (2022). Speech perception and im-

plementation in a virtual medical assistant. ICAART.

Dermouche, S. and Pelachaud, C. (2019). Engagement

modeling in dyadic interaction. In Proceedings of the

21st International Conference on Multimodal Interac-

tion, ICMI, page 440–445, Suzhou, Jiangsu, China.

ACM.

Flin, R., Patey, R. E., Glavin, R., and Maran, N. (2010).

Anaesthetists’ non-technical skills. BJA: British Jour-

nal of Anaesthesia, 105(1):38–44.

Henrickson, P., Flin, R., McKinley, A., and Yule, S. (2013).

The surgeons’ leadership inventory (sli): a taxonomy

and rating system for surgeons’ intraoperative leader-

ship skills. BMJ Simulation and Technology Enhanced

Learning, 205(6):745–751.

Hersey, P., Blanchard, K. H., and Johnson, D. E. (1988).

Management of Organizational Behavior: Leading

Human Resources, chapter Situational Leadership,

pages 169–201. Prentice-Hall, 5 edition.

Hjortdahl, M., Ringen, A. H., Naess, A.-C., and Wisborg,

T. (2009). Leadership is the essential non-technical

skill in the trauma team - results of a qualitative study.

Scandinavian Journal of Trauma, Resuscitation and

Emergency Medicine, 17(48).

Hoegen, R., Aneja, D., McDuff, D., and Czerwinski, M.

(2019). An end-to-end conversational style matching

agent. https://arxiv.org/pdf/1904.02760.pdf.

Kates, J. M. (2005). Principles of digital dynamic-range

compression. Trends in Ampliﬁcation, 9:45 – 76.

Kulms, P. and Kopp, S. (2016). Test. In Proceedings of

the 16th International Conference, Intelligent Virtual

Agents, IVA, pages 75–84, Los Angeles, CA, USA.

ACM.

Lee, S. K., Kavya, P., and Lasser, S. C. (2021). Social

interactions and relationships with an intelligent vir-

tual agent. International Journal of Human-Computer

Studies, 150.

Lourdeaux, D., Afoutni, Z., Ferrer, M.-H., Sabouret, N.,

Demulier, V., Martin, J.-C., Bolot, L., Boccara, V., and

Lelong, R. (2019). Victeams: A virtual environment

to train medical team leaders to interact with virtual

subordinates. page 241–243. IVA ’19.

Montenegro, C., Zorrilla, A. L., Olaso, J. M., Santana, R.,

Justo, R., Lozano, J. A., and Torres, M. I. (2019). A

dialogue-act taxonomy for a virtual coach designed to

improve the life of elderly. Multimodal Technologies

and Interaction, 3(52).

Nakhal, B. (2017). Generation of communicative intentions

for virtual agents in an intelligent virtual environment

: application to virtual learning environment. PhD

thesis.

Pecune, F., Cafaro, A., Ochs, M., and Pelachaud, C. (2010).

Evaluating social attitudes of a virtual tutor. In Pro-

ceedings of the 16th International Conference on In-

telligent Virtual Agents, IVA, pages 245–255, Los An-

geles, CA, USA. ACM.

Philip, P., Dupuy, L., Auriacombe, M., Serre, F., de Sevin,

E., Sauteraud, A., and Franchi, J.-A. M. (2020).

Trust and acceptance of a virtual psychiatric interview

between embodied conversational agents and outpa-

tients. NPJ Digital Medicine, (1).

Querrec, R., Taoum, J., Nakhal, B., and Bevacqua, E.

(2018). Model for verbal interaction between an em-

bodied tutor and a learner in virtual environments.

pages 197–202.

Saerbeck, M., Schut, T., Bartneck, C., and Janse, M. D.

(2010). Expressive robots in education: varying the

degree of social supportive behavior of a robotic tutor.

In Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems, CHI, page 1613–1622,

Atlanta, GA, USA. ACM.

Sims, H. P., Faraj, S., and Yun, S. (2009). When should a

leader be directive or empowering? how to develop

your own situational theory of leadership. Business

Horizons, 52(2):149–158.

Taoum, J., Raison, A., Bevacqua, E., and Querrec, R.

(2018). An adaptive tutor to promote learners’ skills

acquisition during procedural learning.

Vilhj

almsson, H., Cantelmo, N., Cassell, J., Chafai, N.,

Kipp, M., Kopp, S., Mancini, M., Marsella, S., Mar-

shall, A., Pelachaud, C., Ruttkay, Z., Th

orisson, K.,

Welbergen, H., and Werf, R. (2007). The behav-

ior markup language: Recent developments and chal-

lenges. volume 4722, pages 99–111.

Yang, P.-J. and Fu, W.-T. (2016). Mindbot: A social-based

medical virtual assistant. In Proceedings of the 2016

International Conference on Healthcare Informatics,

ICHI, pages 319–319, Chicago, IL, USA. IEEE.

Yule, S., Flin, R., Maran, N., Rowley, D., Youngson, G., and

Paterson-Brown, S. (2008). Surgeons’ non-technical

skills in the operating room: Reliability testing of

the notss behavior rating system. World journal of

surgery, 32:548–56.

Simulations of a Computational Model for a Virtual Medical Assistant

105