Data-driven Web-based Intelligent Decision Support System for Infection

Management at Point-Of-Care: Case-Based Reasoning Beneﬁts and

Limitations

Bernard Hernandez

, Pau Herrero

, Timothy M. Rawson

, Luke S. P. Moore

, Esmita Charani

Alison H. Holmes

and Pantelis Georgiou

Centre for Bio-Inspired Technology, Imperial College London, U.K.

Health Protection Unit in Healthcare Associated Infections and AMR, Imperial College London, U.K.

Keywords:

Antimicrobial Resistance, Infection Diseases, Antibiotics, Decision Support System, Case-Based Reasoning,

Machine Learning, User Interface, Point of Care.

Abstract:

Antimicrobial Resistance (AMR) is a major patient safety issue. Attempts have been made to palliate its

growth. Misuse of antibiotics to treat human infections is a main concern and therefore prescription behaviour

needs to be studied and modiﬁed appropriately. A common approach relies on designing software tools to

improve data visualization, promote knowledge transfer and provide decision-making support. This paper

explains the design of a Decision Support System (DSS) for clinical environments to provide personalized,

accurate and effective diagnostics at point-of-care (POC), improving continuity, interpersonal communication,

education and knowledge transfer. Demographics, biochemical and susceptibility laboratory tests and individ-

ualized diagnostic/therapeutic advice are presented to clinicians in a handheld device. Case-Based Reasoning

(CBR) is used as main reasoning engine to decision support for infection management at POC. A web-based

CBR-inspired interface design focused on usability principles has also been developed. The proposed DSS is

perceived as useful for patient monitoring and outcome review at POC by expert clinicians. The DSS was rated

with a System Usability Scale (SUS) score of 68.5 which indicates good usability. Furthermore, three areas of

improvement were identiﬁed from the feedback provided by clinicians: thorough guidance requirements for

junior clinicians, reduction in time consumption and integration with prescription workﬂow.

1 INTRODUCTION

Antimicrobials are drugs that kill or stop the growth

of microbes (e.g. bacteria or viruses), thereby are

commonly used to treat infections. Recently, An-

timicrobial Resistance (AMR) has been reported to

be a leading public health and safety problem (Wise

et al., 1998; ONeill, 2014) with the inappropriate

use of antibiotics in humans identiﬁed as a leading

driver (Holmes et al., 2016). Microbes are contin-

uously evolving and unnecessary antibiotic prescrip-

tion, particularly within infection diseases, are a com-

mon concern in critical care and infection manage-

ment, which are observing and suffering the conse-

quences of an increased rate of AMR. In addition,

failure to recognize and respond to the early stage in-

fections is considered a major cause of avoidable mor-

tality. Thus, it is needed to develop guidelines and

software tools that facilitate healthcare professionals

to treat their patients at the patient bedside by collect-

ing and visualizing laboratory test results while pro-

viding a support system to assist in decision-making.

Antibiotic resistance is most likely to develop in

areas with a considerable concentration of sick pa-

tients and high risk of infection where antimicrobials

are used extensively. Henceforth, the Intensive Care

Unit (ICU), where proportion of inappropriate antibi-

otic prescription ranges from 41% to 66%, is targeted

in our preliminary studies. Handheld Decision Sup-

port Systems (DSSs) including local antibiotic guide-

lines have proved to reduce antibiotic prescribing in

the ICU (Sintchenko et al., 2005). Despite their ben-

eﬁts, factors as hardware availability or interface de-

sign (Tsopra et al., 2014) obstruct the acceptance of

DSSs in clinical environments. To promote their use,

it is necessary to determine the best way to present

the information (Moxey et al., 2010). The Health-

care Information and Management Systems Society

(HIMMS) stressed the beneﬁts of designs based on

usability principles (Belden et al., 2009). In addi-

Hernandez B., Herrero P., Rawson T., Moore L., Charani E., Holmes A. and Georgiou P.

Data-driven Web-based Intelligent Decision Support System for Infection Management at Point-Of-Care: Case-Based Reasoning Beneﬁts and Limitations.

DOI: 10.5220/0006148401190127

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 119-127

ISBN: 978-989-758-213-4

119

tion, a well designed DSS sustains the advanced al-

gorithms implemented in its core with a fully com-

prehensible representation to support conﬁdence in

doctors. Since medical knowledge is voluminous, it

has to be focused on data and decision making while

providing access to electronic health records (HER)

and personal health information (PHI). Many stud-

ies show that young clinicians engage better with the

use of mobile applications, displaying great potential

to improve learning and augmenting traditional train-

ing (Boulos et al., 2014).

In this paper we postulated that an appropri-

ately designed clinical information technology sys-

tem could improve reliability and consistency of col-

lecting vital signs; their visualization, interpretation

and analysis; and the delivery of a more sophisti-

cated DSS. Therefore, health care professionals and

biomedical engineers from Imperial College of Lon-

don have designed a prototype system accessible at

the point-of-care (POC) with the speciﬁc objectives

of improving three main areas: personalization and

therefore outcomes of infection management; conti-

nuity through POC support for interpersonal commu-

nication; and education during interactions between

clinicians and infection specialists.

2 BACKGROUND

Critical care, infection management and antimicrobial

stewardship is predominantly multidisciplinary with

involvement of infection specialists being crucial. In

practice, antimicrobial prescribing frequently occurs

out of hours, and when advice is dispensed by in-

fection specialists, uptake can be variable. Current

patient management systems rarely integrate DSSs to

assist with this, or if they do, this is very basic. There-

fore, there is an evident need for an intelligent clinical

DSS.

2.1 Decision Support Systems

A clinical DSS can be deﬁned as a computer program

that is designed to analyse data to help health care

professionals make clinical decisions. They are meant

to increase quality of care, enhance health outcomes

and reduce human errors while improving efﬁciency,

cost-beneﬁt ratio and patient satisfaction. Most ba-

sic systems include assessment, monitoring and in-

formative tools in the form of computerized alerts,

reminders and electronic clinical guidelines. For ex-

ample, therapy and vital signs monitoring (McGregor

et al., 2006) or susceptibility test results visualiza-

tion (Flottorp et al., 2002). More advanced diagno-

sis and advisory tools usually rely in statistics, ma-

chine learning and artiﬁcial intelligent techniques to

provide a higher level of data extraction. For exam-

ple, diagnose and therapy advisers (Paul et al., 2006)

or infection risk assessment (Mullett et al., 2004).

Different approaches have been used to design

intelligent DSS, each one with their own beneﬁts

and drawbacks. Decision trees are popular for their

simplicity to understand and construct from logi-

cal rules and have been applied in dengue fever di-

agnosis (Tanner et al., 2008) and antibiotic selec-

tion (William Miller, 2013). The amount of comput-

ing time required for large datasets is still reasonable.

However, they do not tend to work well if decision

boundaries are smooth (Quinlan, 1986) and are not

optimal for uncorrelated variables. As a result of the

greedy strategy applied, they also present high vari-

ance and are often unstable, tending to over-ﬁt.

Probability-based approaches are emerging due

to its capacity to represent and handle uncertain-

ties (Pearl, 1988). Bayesian Networks (BN) are prob-

abilistic networks that represent a set of variables

(nodes) and their dependencies (arcs) using a graph.

Such causal dependencies, inﬂuences or correlations

are deﬁned based on the experience of clinicians.

Hence, it can be associated with a rule-based sys-

tem, which uses data to reﬁne the previously deﬁned

relationships. They have been widely exploited in

health-care (Lucas et al., 2004). Particularly, Causal

Probabilistic Networks (CPN) have been used to de-

velop DSS in diagnosis of cancer (Kahn et al., 1997),

ventilator-associated pneumonia (Lucas et al., 2003)

and sepsis (Tsoukalas et al., 2015). Bayesian Net-

works offer a natural way of representing uncertain-

ties, however an insufﬁcient understanding of their

formal meaning may give rise to modelling ﬂaws.

In particular, Causal Probabilistic Networks are best

suited to tackle very speciﬁc situations as bloodstream

infection (Paul et al., 2006). Unfortunately, treatment

recommendation is poor since they usually prescribe

broad-spectrum antibiotics (Kofoed et al., 2009). Fur-

thermore, there is a lack of guidance to report and in-

terpret their results by non experts.

The Case-Based Reasoning (CBR) methodol-

ogy (Aamodt and Plaza, 1994) has been used to tackle

problems in antibiotic therapy (Heindl et al., 1997)

and molecular biology (Jurisica and Glasgow, 2004).

The aim is to use previous experience in form of cases

to understand and solve new problems.

2.2 Case-Based Reasoning

Case-based reasoning is a widely used approach to

solve new problems based on previous experience in

HEALTHINF 2017 - 10th International Conference on Health Informatics

120

form of cases. It is considered as a methodology to

follow rather than an algorithm in itself as shown in

Figure 1. The CBR cycle is divided in four different

phases. The ﬁrst phase retrieves from the database

those cases that are relevant based on a predeﬁned

similarity measure (e.g. euclidean distance). In the

second phase, advice is commonly given by adapting

or combining the solutions from the retrieved cases

(i.e. antibiotic therapies). The proposed solution is

incorporated to the case and saved in the database.

The third phase monitors the treatment evolution to

assess its outcome (e.g. success or failure). Finally, a

decision to whether retain or not the case based on its

reusability is made.

Solved

Case

Retrieved

Similar

Cases

Current

Case

Learned

Case

Tested/

Repaired

Case

Case i

Case Base

General Knowledge

Problem

Figure 1: Diagram showing the different phases for a cycle

within the Case-Based Reasoning methodology as outlined

in (Aamodt and Plaza, 1994).

This methodology is very generic and can be par-

ticularized to tackle many different problems. Nev-

ertheless, the most important property that makes

CBR appropriate to be used in clinical environments

is the straightforward relation that can be found be-

tween cases in the CBR methodology and cases as

interpreted by clinical staff. Due to this nexus be-

tween the clinical and the scientiﬁc environments,

CBR methodology has been selected to be incorpo-

rated in the decision support system and strongly in-

ﬂuenced the design of the user interface.

3 METHODOLOGY

3.1 EPIC IMPOC

Enhanced, Personalized and Integrated Care for In-

fection Management at Point Of Care (EPIC IMPOC)

is a decision support system designed to record a com-

plete set of vital signs at the patients bedside on hand-

held computing devices while providing instant bed-

side decision-making assistance to clinical staff. It

also pulls in data from the hospital patient admin-

istration system, laboratory results and other clini-

cal information stored electronically. It can be used

anywhere in the hospital by staff with appropriate ac-

cess rights, using mobile devices or desktop comput-

ers linked to the hospital intranet. EPIC IMPOC has

been preliminarily trialled at critical care antimicro-

bial prescribing, a known reservoir for antimicrobial

resistance, and it is being extended to secondary care.

The system architecture is shown in Figure 2 were

two parts are clearly differentiated: server and client

sides. The server side processes queries, interacts

with the permanent storage and serves web pages

to the client side. The latter displays information

to users. The modules constituting the server side

are: a) CBR for history review and case compari-

son. b) Probabilistic Inference (PI) aims to provide

step-wise guidance ﬁtting the decision pathway fol-

lowed by clinicians for infection management. c) Pa-

tient engagement module. d) Personalized antibiotic

dosing. e) Visualization of Antimicrobial Resistance

related information. This paper focuses exclusively

on the CBR module.

3.2 Server Side

The server side has been implemented in Java and

uses an object-relational mapping java library (Hi-

bernate ORM) to map an object-oriented domain

model to a traditional relational database (SQL). The

Lightweight Directory Access Protocol (LDAP) ac-

complishes the authorization and authentication of

users and it is provided in all hospitals at Imperial

College Healthcare National Health Service Trust.

The server side follows the REST (Representational

State Transfer) architectural design, which suggests a

group of guidelines to create scalable, more perfor-

mant and maintainable web services.

The core CBR module implementation is based on

the JColibri framework (D

ıaz-Agudo et al., 2007) in-

cluding some improvements to achieve better gener-

alization and performance. It is used to retrieve cases

from the database based on a similarity measure. A

case is deﬁned by a compendium of parameters that

Data-driven Web-based Intelligent Decision Support System for Infection Management at Point-Of-Care: Case-Based Reasoning Beneﬁts

and Limitations

121

EPIC IMPOC

Dosing

CBR

AMR

Trends

Antibiogram

Patient

Module

API

ADM IN

P AT H

M ICRO

LAB.RESU LT S

HIST ORY

ST AT S

Server

Client

Figure 2: High-level diagram describing the main components of the DSS. The external databases that are currently being ac-

cessed are patient administration system (ADMIN), pathology laboratory tests (PATHO) and microbiology results (MICRO).

The server side has the following independent modules: Case-Based Reasoning (CBR), Probabilistic Inference (PI), Patient

module, Dosing module, antibiogram and AMR trends. All the information is accessed through an API and presented on a

handheld device to clinicians.

can be grouped in ﬁve different sets: metadata, de-

scription, solution, justiﬁcation and result. However,

only those in the description container are used to

compute the similarity scores. Some examples of

attributes used to deﬁne the case are: demograph-

ics (age, gender or weight), existing diseases (aller-

gies, HIV or diabetes), respiratory system (ventilation

support or oxygen requirements), abdomen (abdomi-

nal examination, renal support or catheter), biochemi-

cal markers (creatinine or bilirubin) and microbiology

(infectious organisms).

3.3 Client Side

The client side is a web-based application imple-

mented using HTML, CSS and Javascript which is

accessible through the browser. It follows a respon-

sive design approach to render a consistent interface

across different devices, from desktop computers to

mobile phones and tablets with different screen sizes.

An efﬁcient DSS user interface should present all

the information relevant to clinicians neatly, combin-

ing different resources of patient-related information

(i.e. demographic, pathology and microbiology data).

Since some data might be missing or not available, it

is also desired to enable clinicians to manually input

data or comments for further consideration. In addi-

tion, infections evolve with time and so do treatments.

Thus, inspection of previous symptoms, treatments

and outcomes is desired. Since there is an straightfor-

ward relation between cases as interpreted in clinical

environments and cases in the CBR methodology, a

case is considered as main unit of patient related in-

formation to be presented in the interface. A single

case is formed by several components, mostly regard-

ing the type and source of the data, and has been di-

vided in different sections (tabs in Figure 3) for visu-

alization purposes. The Resume section is read-only

and displays the most relevant information (e.g. in-

fectious micro-organisms or organs infected) while

Description shows additional information and allows

the insertion/modiﬁcation of parameters within the

case. Solution contains the antibiotic therapy pre-

scribed (including frequency, via and doses) and a

section to collect feedback from users.

Six routinely requested biochemical markers were

selected as main indicators of infection and patient

status after reviewing the scientiﬁc literature and dis-

cussion with clinicians and infectious disease experts.

The temporal evolution of such biochemical markers

is shown in Pathology (see Figure 3) as time-series

where coloured background indicates the normal ref-

erence range. Additionally, it is possible to hide/show

time-series to improve visualization.

Susceptibility testing is used to determine which

antimicrobials will inhibit the growth of micro-

organisms causing a infection and is generally per-

formed in vitro. The Microbiology section displays

the result of all the susceptibility tests requested for

the patient. The outcomes of the tests are provided

for individual pairs (infectious organism and antibi-

otic) and are categorised as resistant, intermediate and

sensitive. They are also presented during antibiotic

therapy selection for further guidance.

HEALTHINF 2017 - 10th International Conference on Health Informatics

122

Figure 3: EPIC IMPOC web-based decision support system overview. The main unit of information is the case and its

content is displayed among ﬁve different tabs (Resume, Description, Pathology, Sensitivity and Solution). The user interface

is divided in three main areas: patient selection, dashboard with current patient (top) and retrieved cases (bottom), history

review and a side bar to add/remove cases to/from such dashboard.

4 RESULTS

A working prototype of EPIC IMPOC incorporat-

ing Case-Based Reasoning methodology as decision-

support engine was preliminarily trialled in the In-

tensive Care Unit at Hammersmith Hospital in Lon-

don for a month. The predeﬁned case base contained

80 cases and information retrieval was performed

through handheld computer devices (i.e. ipads) at the

patient bed side by clinicians under the supervision of

infection specialists. From such study the following

conclusions were extracted:

• The system has potential to promote and facil-

itate communication between nurses, clinicians

and infection specialists as shown by the interac-

tion among them during the trial.

• The system improves homogeneous collection of

vital signs. Such improvement comes from the

introduction of a form in the description of the

case to input missing symptoms easily. The form

is ﬁlled automatically for those symptoms avail-

able in external databases (e.g. electronic health

records).

• The system facilitates data visualization at POC

and simpliﬁes comparison with previous similar

cases and outcomes. In addition, biochemical

markers evolution, susceptibility tests and history

review for the hospitalized patient are easily ac-

cessible and found to be very helpful at point of

care.

• The system is capable of mimicking clinicians

prescription practices in the intensive care unit. In

such trial clinicians were under the supervision of

infection specialists. As a result, therapies pre-

scribed by clinicians and therapies retrieved by

the CBR algorithm matched approximately 90%

of the times. It is especially visible in ICU where

wide-spectrum antibiotics are commonly used.

• It increases and facilitates the interaction between

clinicians and patients. Therefore it helps engag-

ing with patients and opens the possibility to ed-

ucate population on antibiotic misuse and its con-

sequences (Rawson et al., 2016b).

4.1 System Usability Scale Survey

A survey to evaluate the usability of the decision sup-

port system interface was performed. The System Us-

ability Scale (SUS) (Brooke et al., 1996) is composed

Data-driven Web-based Intelligent Decision Support System for Infection Management at Point-Of-Care: Case-Based Reasoning Beneﬁts

and Limitations

123

Table 1: The original SUS statements (Brooke et al., 1996), average agreement and SUS contribution.

SUS statements Avg. rating SUS contribution

I think that I would like to use this system frequently. 2.8 1.8

I found the system unnecessarily complex. 1.4 3.6

I thought the system was easy to use. 2.0 1.0

I think that I would need the support of a technical person to be able to use this system. 1.6 3.4

I found that the various functions in this system were well integrated. 3.0 2.0

I thought that there was too much inconsistency in this system. 0.2 4.8

I would imagine that most people would learn to use this system very quickly. 2.8 1.8

I found the system very cumbersome to use. 2.2 2.8

I felt very conﬁdent using the system. 3.0 2.0

I needed to learn a lot of things before I could get going with this system. 0.8 4.2

of 10 statements to which participants indicate their

agreement from 1 to 5, where 5 indicates strongly

agree. Predeﬁned rules for positive and negative state-

ments are used to obtain the SUS contributions. The

SUS contribution for each statement ranges from 1

to 4 where higher scores indicate better usability (see

Table 1) and their sum is multiplied by 2.5 to calcu-

late the ﬁnal SUS score. It ranges from 0 to 100 where

poor and great product usability are indicated for SUS

scores under 50 and over 80 respectively. This survey

is technology agnostic, quick, provides a single score

and is non proprietary. A free-text box was added for

additional comments and suggestions.

The SUS survey was completed by 10 different

participants (83% males) from 27 to 51 years old

where technical training in the use of the system was

not provided. The proﬁle of those participants was in-

fection specialist (two), clinician (three), nurse (four)

and other staff (one). The SUS contribution for each

statement is presented in the right column in Table 1.

The ﬁnal SUS score obtained is 68.5 which indicates

good product usability with margin to improve. Ad-

ditionally, a variety of comments were provided by

participants and have been synthesized in the follow-

ing bullet points:

• There is a common concern among experienced

clinicians and infection specialists in the use ju-

nior doctors would do of such large amount of

data displayed in the interface. The decision sup-

port system has potential to help training junior

doctors and improve their prescription practices,

but it needs to narrow the presented information

providing speciﬁc guidance.

• Clinicians consider the user interface intuitive and

helpful for patient long-term monitoring and man-

agement, however it might sometimes be time

consuming. Additionally, it does not entirely ﬁt

with the work-ﬂow followed to prescribe antibi-

otic therapies.

• They suggested the possibility of recording fur-

ther parameters, not necessarily directly related

with infections.

From the preliminary trial performed by infection

specialists and the feedback obtained from the sur-

veys, it is possible to conclude that the CBR algorithm

is able to mimic the prescription practices of users.

However, that is not enough to promote change in an-

tibiotic prescription practices. Initially, as a quick so-

lution infection specialists were keen in creating an

“ideal” case base; that is, a set of cases with opti-

mal antibiotic therapies according to infection guide-

lines and expert prescriptions. Such optimal therapies

would then be suggested by the decision support sys-

tem to further users. Hence, the knowledge would be

transferred from infection specialists to other clinical

staff (e.g. nurses and clinicians).

Unfortunately, this approach presents several

drawbacks. Creating a complete case base that covers

the whole spectrum of possibilities is nearly impossi-

ble and time consuming. In addition, infections are

often acquired in hospitals by contagious as a con-

sequence of treating more severe diseases which di-

minish the immune system (e.g. surgeries and can-

cer). Therefore, future therapies prescribed by clin-

icians and recorded in the system might not agree

with the infection guidelines reshaping the case base

and therefore altering CBR recommendations. A case

base with strictly guideline oriented therapies on the

long-term is unrealistic and limits the scope and us-

ability of the system.

After discussion with a multi-professional team

including physicians, nurses, pharmacists and non-

medical researchers, an study to map the pathway

followed by clinicians to prescribe antibiotics thera-

pies was performed (Rawson et al., 2016a). The re-

ported infection management pathway was deﬁned as

a stepwise Bayesian model of estimating probabili-

ties in which each step adds systematically informa-

tion to allow optimisation of decisions on diagnosis

and management of infection. Initially, clinicians es-

timate the risk of infection and attempt to localize its

HEALTHINF 2017 - 10th International Conference on Health Informatics

124

source by looking at patient’s physiological parame-

ters. Once clinicians construct a picture of the severity

of the infection, whether or not to initiate antimicro-

bial therapy is decided. In this step, local microbiol-

ogy guidance provided within hospitals was the most

commonly cited factor. Finally, they review and reﬁne

the treatment accordingly.

This new approach enables to produce very spe-

ciﬁc step-by-step probability-like decision support.

This would improve the guidance provided to junior

clinicians and facilitates the validation of decision-

making for each individual step. Additionally, since

reviewing of previous cases is not necessary it would

greatly reduce usability time. Since the methodology

integrates with the infection management pathway, it

will likely inﬂuence prescription practices to a higher

degree than CBR.

5 DISCUSSION

DSSs are being exploited in several areas such as

business or economics but their acceptance by clin-

ical staff is obstructing its use in hospitals and other

clinical environments (Kawamoto et al., 2005). De-

signing a DSS based on usability principles and sim-

ply providing the clinical information does not guar-

antee acceptance (Moxey et al., 2010). Other factors

as accessibility, availability, easy of use, time require-

ments and integration into the clinical workﬂow are

important and need to be considered (Tsopra et al.,

2014). Taking previous knowledge in consideration,

a DSS to support prescription of antibiotic therapies

and patient monitoring at POC exploiting the CBR

methodology was implemented.

In many circumstances, as complicated cases,

providers prefer to consult their colleges or more

specialised clinicians as infection specialists. This

consultation among different members of the clini-

cal staff was facilitated by the DSS. In addition, it

was believed to enhance decision making and homo-

geneous collection of vital signs among clinicians re-

sulting in better prescribing practices. Furthermore,

re-entering patient data to generate advices is a deter-

rent to use (Moxey et al., 2010) and integration into

existing programs (e.g. electronic medical records)

was a clear facilitator.

The usability measured through the SUS survey

was 68.5 which is about average and shows poten-

tial margin for improvement. A similar strength of

agreement was shown by participants for the third and

eight statements which had an average rating of 2.0

and 2.2 respectively. The SUS contribution for each

statement was 1.0 and 2.8 respectively, indicating that

the system is usable but not necessarily easy. Note

that some wording used by the original SUS was sug-

gested to be poorly understood by participants affect-

ing the results (Bangor et al., 2008). As an exam-

ple, the sixth statement which contributed the most

contains the wording “too much” which might be un-

clear. Additionally, users may not have a good sense

of the required complexity of such systems since there

are no commonly known competing solutions and

the wording “unnecessarily” in the second statement

might have led to a higher contribution of 3.6. There-

fore, the ﬁnal SUS score is possibly slightly less that

the one presented.

There is still much to be done to make these sys-

tem work in routine clinical practice. Measuring ab-

solute usability using a single metric is very challeng-

ing since many external factors inﬂuence the results

(e.g. technical training or accessible technology). The

feedback provided greatly helped to identify areas of

improvement and the results of the survey are a use-

ful source for future comparison to asses the beneﬁts

of including new components to tackle the identiﬁed

weaknesses.

6 CONCLUSIONS

CBR methodology was incorporated in a DSS as

decision-making engine providing similar antibiotic

therapies to those prescribed by expert clinicians in

the majority of cases in our preliminary trials. In

addition, it was widely accepted for information re-

trieval and long-term patient monitoring and manage-

ment. Three main areas of improvement were identi-

ﬁed from the feedback provided by expert clinicians:

speciﬁc guidance requirements for junior clinicians, a

need to reduce time consumption using the DSS and

a better integration into clinical workﬂow.

The reported infection management pathway was

deﬁned as a multi-step Bayesian-like approach (Raw-

son et al., 2016a) which inherently tackles most of the

weaknesses identiﬁed: speciﬁc guidance, time con-

straints and integration in the workﬂow. Therefore,

it is more suitable to support and modify prescription

practices.

The combination of both approaches into one sin-

gle decision support system is a very elegant solution

that could lead to an increase in acceptability among

clinicians. To validate its feasibility, infection risk and

source of infection inference from biochemical mark-

ers are considered as primary steps. The ﬁnal integra-

tion of such inference in the user interface has poten-

tial to reduce misuse of antibiotics and its evaluation

forms the basis for future work.

Data-driven Web-based Intelligent Decision Support System for Infection Management at Point-Of-Care: Case-Based Reasoning Beneﬁts

and Limitations

125

ACKNOWLEDGEMENT

This report is independent research funded by the

National Institute for Health Research Invention for

Innovation (i4i) programme, Enhanced, Personalized

and Integrated Care for Infection Management at

Point of Care (EPIC IMPOC), II-LA-0214-20008.

The authors would like to thank members of Impe-

rial College NHS Healthcare Trust who participated

in the study. The views expressed in this publication

are those of the authors and not necessarily those of

the NHS, the National Institute for Health Research

or the UK Department of Health.

REFERENCES

Aamodt, A. and Plaza, E. (1994). Case-based reasoning:

Foundational issues, methodological variations, and

system approaches. AI communications, 7(1):39–59.

Bangor, A., Kortum, P. T., and Miller, J. T. (2008). An

empirical evaluation of the system usability scale. Intl.

Journal of Human–Computer Interaction, 24(6):574–

594.

Belden, J. L., Grayson, R., and Barnes, J. (2009). Deﬁn-

ing and testing emr usability: Principles and proposed

methods of emr usability evaluation and rating. Tech-

nical report, Healthcare Information and Management

Systems Society (HIMSS).

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

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

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

tory control and certiﬁcation. Online journal of public

health informatics, 5(3).

Brooke, J. et al. (1996). Sus-a quick and dirty usability

scale. Usability evaluation in industry, 189(194):4–7.

ıaz-Agudo, B., Gonz

alez-Calero, P. A., Recio-Garc

ıa,

J. A., and S

anchez-Ruiz-Granados, A. A. (2007).

Building cbr systems with jcolibri. Science of Com-

puter Programming, 69(1):68–75.

Flottorp, S., Oxman, A. D., H

avelsrud, K., Treweek, S., and

Herrin, J. (2002). Cluster randomised controlled trial

of tailored interventions to improve the management

of urinary tract infections in women and sore throat.

Bmj, 325(7360):367.

Heindl, B., Schmidt, R., Schmid, G., Haller, M., Pfaller,

P., Gierl, L., and Pollwein, B. (1997). A case-based

consiliarius for therapy recommendation (icons):

computer-based advice for calculated antibiotic ther-

apy in intensive care medicine. Computer Methods

and Programs in Biomedicine, 52(2):117–127.

Holmes, A. H., Moore, L. S., Sundsfjord, A., Steinbakk,

M., Regmi, S., Karkey, A., Guerin, P. J., and Pid-

dock, L. J. (2016). Understanding the mechanisms

and drivers of antimicrobial resistance. The Lancet,

387(10014):176–187.

Jurisica, I. and Glasgow, J. (2004). Applications of case-

based reasoning in molecular biology. Ai Magazine,

25(1):85.

Kahn, C. E., Roberts, L. M., Shaffer, K. A., and Haddawy, P.

(1997). Construction of a bayesian network for mam-

mographic diagnosis of breast cancer. Computers in

biology and medicine, 27(1):19–29.

Kawamoto, K., Houlihan, C. A., Balas, E. A., and Lobach,

D. F. (2005). Improving clinical practice using clini-

cal decision support systems: a systematic review of

trials to identify features critical to success. BMJ,

330(7494):765.

Kofoed, K., Zalounina, A., Andersen, O., Lisby, G., Paul,

M., Leibovici, L., and Andreassen, S. (2009). Perfor-

mance of the treat decision support system in an envi-

ronment with a low prevalence of resistant pathogens.

Journal of antimicrobial chemotherapy, 63(2):400–

404.

Lucas, P., Schurink, C., Hoepelman, I., and Bonten, M.

(2003). A model-based approach to improved pre-

scription of antibiotics.

Lucas, P. J., van der Gaag, L. C., and Abu-Hanna, A. (2004).

Bayesian networks in biomedicine and health-care.

Artiﬁcial intelligence in medicine, 30(3):201–214.

McGregor, J. C., Weekes, E., Forrest, G. N., Standiford,

H. C., Perencevich, E. N., Furuno, J. P., and Harris,

A. D. (2006). Impact of a computerized clinical de-

cision support system on reducing inappropriate an-

timicrobial use: a randomized controlled trial. Jour-

nal of the American Medical Informatics Association,

13(4):378–384.

Moxey, A., Robertson, J., Newby, D., Hains, I., Williamson,

M., and Pearson, S.-A. (2010). Computerized clinical

decision support for prescribing: provision does not

guarantee uptake. Journal of the American Medical

Informatics Association, 17(1):25–33.

Mullett, C. J., Thomas, J. G., Smith, C. L., Sarwari, A. R.,

and Khakoo, R. A. (2004). Computerized antimi-

crobial decision support: an ofﬂine evaluation of a

database-driven empiric antimicrobial guidance pro-

gram in hospitalized patients with a bloodstream in-

fection. International journal of medical informatics,

73(5):455–460.

ONeill, J. (2014). Antimicrobial resistance: Tackling a cri-

sis for the health and wealth of nations.

Paul, M., Andreassen, S., Nielsen, A. D., Tacconelli, E.,

Almanasreh, N., Fraser, A., Yahav, D., Ram, R.,

Leibovici, L., Group, T. S., et al. (2006). Pre-

diction of bacteremia using treat, a computerized

decision-support system. Clinical Infectious Diseases,

42(9):1274–1282.

Pearl, J. (1988). Probabilistic inference in intelligent sys-

tems.

Quinlan, J. R. (1986). Induction of decision trees. Machine

Learning, 1(1):81–106.

Rawson, T., Charani, E., Moore, L. S., Hernandez, B.,

Castro-Sanchez, E., Herrero, P., Georgiou, P., and

Holmes, A. H. (2016a). Mapping the decision path-

ways of acute infection management in secondary care

among uk medical physicians: A qualitative study.

BMC medicine [In Press].

Rawson, T. M., Moore, L. S., Hernandez, B., Castro-

Sanchez, E., Charani, E., Georgiou, P., Ahmad, R.,

and Holmes, A. H. (2016b). Patient engagement with

HEALTHINF 2017 - 10th International Conference on Health Informatics

126

infection management in secondary care: a qualita-

tive investigation of current experiences. BMJ open,

6(10):e011040.

Sintchenko, V., Iredell, J. R., Gilbert, G. L., and Coiera, E.

(2005). Handheld computer-based decision support

reduces patient length of stay and antibiotic prescrib-

ing in critical care. Journal of the American Medical

Informatics Association, 12(4):398–402.

Tanner, L., Schreiber, M., Low, J. G., Ong, A., Tolfvenstam,

T., Lai, Y. L., Ng, L. C., Leo, Y. S., Puong, L. T.,

Vasudevan, S. G., et al. (2008). Decision tree algo-

rithms predict the diagnosis and outcome of dengue

fever in the early phase of illness. PLoS Negl Trop

Dis, 2(3):e196.

Tsopra, R., Jais, J.-P., Venot, A., and Duclos, C. (2014).

Comparison of two kinds of interface, based on guided

navigation or usability principles, for improving the

adoption of computerized decision support systems:

application to the prescription of antibiotics. Jour-

nal of the American Medical Informatics Association,

21(e1):e107–e116.

Tsoukalas, A., Albertson, T., and Tagkopoulos, I. (2015).

From data to optimal decision making: a data-driven,

probabilistic machine learning approach to decision

support for patients with sepsis. JMIR medical infor-

matics, 3(1):e11.

William Miller, O. (2013). A decision tree: Proper antibi-

otic selection and use.

Wise, R., Hart, T., Cars, O., Streulens, M., Helmuth, R.,

Huovinen, P., and Sprenger, M. (1998). Antimicrobial

resistance is a major threat to public health. British

medical journal, 317(7159):609–611.

Data-driven Web-based Intelligent Decision Support System for Infection Management at Point-Of-Care: Case-Based Reasoning Beneﬁts

and Limitations

127