Decision Making Support in the Scheduling of Chemotherapy Coping

with Quality of Care, Resources and Ethical Constraints

Christophe Ponsard

, Renaud De Landtsheer

, Yoann Guyot

Franc¸ois Roucoux

and Bernard Lambeau

CETIC Research Centre, Charleroi, Belgium

Grand H

opital de Charleroi, Charleroi, Belgium

Enspirit, Sombreffe, Belgium

Keywords:

Clinical Pathway, On-line Scheduling, Oncology, Workﬂow Enactment, Constraint-Based Local Search,

Agent System.

Abstract:

The scheduling of clinical pathways such as oncological treatments involves a tricky decision process because

the therapeutic regimens require to respect strict timing constraints with possibly limited resources such as

beds and caregivers availability with an increasing number of patients. Such constraints must be met simul-

taneously for every patient treated at the same time, by making the best use of limited hospital resources.

The scheduling must also be robust in case of adverse events such as unexpected delays or partial treatment

deliveries due to their toxicity. In this paper, we show how such a decision process can be driven by care

quality indicators to ensure all the dimensions. We demonstrate how constraint-based local search techniques

can cope with real-world size chemotherapy pathways and efﬁciently adapt to changes. We also share some

ethical concerns about the way the objective function is expressed and more generally about how the tool

integrates in the medical decision process.

1 INTRODUCTION

In Western countries, due to progress in medical care

and ageing of the population, hospitals have to man-

age increasingly complex and multidisciplinary med-

ical procedures over a growing pool of patients. In

the worst case, this results in a decrease in the quality

of care received by patients, which does not always

match the recommended care process yet prescribed.

A survey of 30 pathologies ranging from osteoarthri-

tis to breast cancer, observed that, on average, half of

the patients received the recommended medical care

(McGlynn et al., 2003).

In order to reduce the variability in clinical pro-

cesses and improve the care quality, a level of stan-

dardisation was proposed through clinical pathways

(or care pathways). A clinical pathway is deﬁned

as a multi-disciplinary speciﬁcation of the treatment

process required by a group of patients presenting

the same medical condition with a predictable clini-

cal course (Campbell et al., 1998). It describes con-

crete treatment activities for patients having identical

diagnoses or receiving the same therapy.

This standardization results usually in less delays,

higher quality assurance and in reduced costs. As they

are strongly oriented on the process description, clin-

ical pathways also maintain a global view on the pa-

tients overall journey, instead of individual doctors

having a view exclusively limited to their medical

speciality (van Dam et al., 2013).

Figure 1: A typical chemotherapy workﬂow.

The use of clinical pathways have been reported

as successful in many therapies, such as arthroplasty

(Walter et al., 2007) and breast cancer (van Dam et al.,

2013). Clinical pathways in oncology involve a pre-

cise description of the therapeutic workﬂow and all

its ancillary activities. Such a partial workﬂow about

460

Ponsard, C., Landtsheer, R., Guyot, Y., Roucoux, F. and Lambeau, B.

Decision Making Support in the Scheduling of Chemotherapy Coping with Quality of Care, Resources and Ethical Constraints.

DOI: 10.5220/0006360504600470

In Proceedings of the 19th International Conference on Enterprise Information Systems (ICEIS 2017) - Volume 1, pages 460-470

ISBN: 978-989-758-247-9

the chemotherapeutic aspect of a treatment is illus-

trated in Figure 1. Of course, implementing clinical

pathways requires involvement. A number of success

factors have been reported, like continued clinician

acceptance, top management support and a dedicated

team of case managers, nurses and paramedical pro-

fessionals (Choo, 2001).

On the IT side, the computerization of workﬂows,

guidelines, and care pathways is also reported as a

key step for process-oriented health information sys-

tems (Gooch and Roudsari, 2011). This allows these

processes to be managed by hospital information sys-

tems or in dedicated workﬂow management systems.

Such autonomous workﬂow management systems can

indeed use clinical workﬂows as a process model de-

scription (Mauro et al., 2010). A key component in

this evolution is to provide efﬁcient tools to support

the scheduling of these workﬂows. While scheduling

the pathway of a single patient or scheduling the activ-

ities of a speciﬁc medical department are not that dif-

ﬁcult, scheduling a large pool of patients in an hospi-

tal with limited resources raises a lot of trade-off con-

cerns (Marynissen and Demeulemeester, 2016). Ide-

ally, such concerns should not impact the quality of

care of individual patient. More, basic ethical princi-

ples state that every patient deserves optimal care re-

gardless of his medical condition or prognosis. Given

that the patients ﬂow is continuous and that a number

of unforeseen events require postponing or adaptation

of treatment sessions, schedules need to be adjusted

on the go. These adjustments should of course com-

ply with already conﬁrmed appointments. The ob-

served practice is that treatment scheduling still usu-

ally rely on human-operated manual tools such as

spreadsheets or scheduling templates (Ahmed et al.,

2011).

In this paper, we address the problem of schedul-

ing treatment appointments in clinical pathways. In

order to ensure care quality over a large pool of pa-

tients within available resources, we put care quality

indicators at the heart of the scheduling algorithms.

The latter are implemented using Constraint-Based

Local Search (CBLS), a technique known for its abil-

ity to scale on large scheduling problems (Van Hen-

tenryck and Michel, 2009). Our algorithms can

also efﬁciently provide new schedules in reaction to

changes in some patient constraints (on-line prob-

lems). In order to provide a concrete validation,

we have focused our efforts on the scheduling of

the chemotherapy part of oncological pathways and

more precisely on cancer chemotherapy (breast can-

cers, brain cancers, lymphoma ...) for which a key

treatment quality indicator called RDI (Relative Dose

Intensity) has been deﬁned (Lyman, 2009). This case

study is relevant because proper enforcement of time

constraint is critical to maximize the chances of sur-

vival of the patients. The availability of the RDI for

these treatments, and its proven correlation with sur-

vival rates, enables us to quantify the enforcement of

time constraints, so that corrective actions might be

considered in case of a deviation.

This paper is organised as follows. In section 2,

we present domain background about chemotherapy

pathways and technical background on constraint-

based local search. Section 3 gives a clear statement

of the problem tackled in the paper. Section 4 dis-

cusses key design aspects of the solution while Sec-

tion 5 details its implementation. The validation is

carried out in Section 6 and relies on an environment

simulator. Related work is discussed in Section 7. Fi-

nally Section 8 concludes and highlights future work.

2 BACKGROUND

This section ﬁrst presents the problem domain of

chemotherapy pathways before introducing local

search frameworks, and then focusing on the neces-

sary vocabulary of CBLS framework used in the re-

maining of the paper, based on the OscaR.cbls engine.

2.1 Chemotherapy Pathways

The typical workﬂow for a chemotherapy is a se-

quence of drugs deliveries or cures, typically adminis-

tered in day hospital. Each cure is followed by a rest-

ing period at home that lasts for a few days to a few

weeks. A minimal interval between cures is required

because chemotherapy drugs are toxic agents and the

body needs some time to recover between two drugs

deliveries. When following the ideal treatment proto-

col, the number of cancerous cells are progressively

reduced, hopefully to reach a full healing or cancer

remission as shown in Figure 2.

Figure 2: Optimal chemotherapy cycles.

Decision Making Support in the Scheduling of Chemotherapy Coping with Quality of Care, Resources and Ethical Constraints

461

If for some reason, chemotherapy cures do not

closely follow the intended periodicity or if doses are

signiﬁcantly reduced, the treatment efﬁciency may be

suboptimal. In such conditions, cancerous cells may

multiply again, which can result in a cancer relapse as

shown in Figure 3

Figure 3: Delayed chemotherapy cycles.

In order to measure the quality of chemotherapeu-

tic cares, a quantiﬁable indicator called the “Relative

Dose Intensity” (RDI) (Lyman, 2009) was deﬁned.

It captures both the the fact that the required dose is

administered and the timing of the delivery, on a scale

from 0% (no treatment) to 100 % (total conformance).

RDI =

planned dose

delivered dose

real duration

planned duration

Figure 4: Probability of relapse-free survival vs. RDI.

Medical literature has shown, for a number of can-

cers, that the relapse-free survival is strongly corre-

lated with the RDI. For instance, for breast cancer, a

key threshold value is 85 % as illustrated in Figure 4

(Piccart et al., 2000). Hence this indicator can be seen

has a gauge that should be carefully managed across

the whole clinical pathway.

2.2 Local Search Frameworks

Local search frameworks aim at making the devel-

opment of algorithmic local search solutions much

simpler than traditional coding. To this end, they

provide different degrees of implementation support,

from problem modelling to the elaboration of a search

procedure. We use such a framework to develop a

schedule optimizer for clinical pathways.

Among the general local search frameworks, Ea-

syLocal++ is a well known and representative one

that requires a dedicated model to be developed

from scratch using ad-hoc algorithms. It mainly

provides support for declaring the search procedure

(Di Gaspero and Schaerf, 2003). It does not provide

as much assistance in the development of a model as

a CBLS framework would. Notably it does not allow

the developer to package efﬁcient global constraints

that can be instantiated on demand.

Besides OscaR, the framework we have used and

described in the next section, there are a few tools

supporting constraint-based local search speciﬁcally:

• Comet is the seminal system for constraint-based

local search (Van Hentenryck and Michel, 2009).

It features a differentiation facility that is not im-

plemented in OscaR.cbls. OscaR.cbls relies rather

on partial propagation to provide a comparable ef-

ﬁciency. Besides, differentiation as provided by

Comet cannot handle intricate models where con-

straints are posted on variables controlled by in-

variants. Comet is a proprietary system available

under a commercial licence.

• LocalSolver is a commercial solver implement-

ing CBLS. It supports boolean and ﬂoating point

variables. It does not require the user to specify

neighbourhoods or meta-heuristics (Benoist et al.,

2011).

• Kangaroo provides a partial propagation feature

that is more selective than OscaR.cbls (Newton

et al., 2011).

2.3 CBLS, the OscaR Way

Among the different solvers, OscaR.cbls was se-

lected. Since this contribution has been done in the

context of the OscaR.cbls tool, we further introduce

the basic concepts of CBLS using the vocabulary of

OscaR.cbls.

As usual in local search, solving a problem in-

volves specifying a model and a search procedure.

The model is composed of incremental variables

(integers and sets of integers at this point), and in-

variants which are incremental directed constraints

maintaining one or more output variables according

to the atomic expressions they implement (e.g. Sum:

the sum of inputs). Constraints are special invari-

ants that maintain their violation as an output variable.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

462

They are Lagrangian relaxations of their speciﬁcation.

Besides, they also maintain some information about

which variable causes the violation.

The search procedure is expressed using neigh-

bourhoods, which can be queried for a move, given

the current state of the model, an acceptance crite-

rion, and an objective function. Combinators are a set

of operators on neighbourhoods that combine them

and incorporate several metaheuristics, so that a com-

plex search strategy can be represented by a compos-

ite neighbourhood totally expressed in a declarative

way (De Landtsheer et al., 2015). A library of combi-

nators is available for specifying standard metaheuris-

tics (e.g. simulated annealing, restart, hill climbing),

for managing solutions (e.g. when to save the current

state, or restore a saved state), and for expressing stop

criteria.

In order to set up the ﬂoor for designing a schedul-

ing solution, we give details on how the model is rep-

resented and updated during the search.

The data structure behind a model is a graph,

called the propagation graph, which we can approxi-

mate to a directed acyclic graph, where the nodes are

variables and invariants. Variables have an associ-

ated type and implement speciﬁc algorithms related

to their type. Invariants have speciﬁc deﬁnitions, and

implement this deﬁnition mostly through incremental

algorithms. Edges in the graph represent data ﬂows

from variables to listening invariants and from in-

variants to controlled variables. The directed acyclic

graph starts with input (a.k.a. decision) variables, and

typically ends at a variable whose value is maintained

to be the one of the objective function. Figure 5 il-

lustrates a propagation graph for a simple warehouse

location problem.

Figure 5: Propagation graph on a warehouse location prob-

lem.

In such engine, propagation is about propagating

updates along the propagation graph in such a way

that a node is reached at most once by the update

wave, and only if one of its inputs has changed and

if needed by the model update. OscaR.cbls manages

this wave by sorting the nodes based on the distance

from the decision variables. The propagation is co-

ordinated through a dedicated heap that aggregates

nodes at the same distance in a list. This offers a

slightly better time complexity than the classical ap-

proach based on topological sort initially presented in

(Van Hentenryck and Michel, 2009).

The search starts from an initial solution and ex-

plores the speciﬁed neighbourhood. Each neigh-

bour solution is examined by modifying the input

variables, and querying the objective function of the

model which is updated through propagation.

During propagation, variables notify each invari-

ant listening to them about their value change. For

integer variables, a notiﬁcation carries a reference to

the variable, and the old and new value of the variable.

For set variables, it carries a reference to the variable,

the old value of the variable, the new value of the vari-

able, and both the set of values that have been added

and removed from the variable. All values transmitted

by variables, through notiﬁcation or through queries

to the variables are immutable, to make the imple-

mentation of algorithms in invariants easier.

3 PROBLEM STATEMENT

The problem considered here is to continuously opti-

mize the scheduling on an evolving set of patients fol-

lowing a speciﬁc chemotherapy process as described

in Section 2. The goals of the scheduling optimization

are the following:

• maintain the best quality of care (i.e. achieve the

best RDI) by avoiding delay for all the patients in

the pool

• meet the resources constraints: available treat-

ment rooms and nurses.

• respect service opening days (weekends, holi-

days) and hours.

• take into account strong unavailabilities of pa-

tients, when known.

• when possible, distribute the workload evenly

over time to avoid work peaks.

When entering his chemotherapy pathway, a pa-

tient can be given an indicative optimal schedule

based on what is known at that time. However the

global scheduling can be impacted by many events

such as:

• the cancellation of treatment delivery, because of

patient no shows or medical no-go (e.g. too low

white blood cells detected in last blood test),

• the delivery of a partial dose, due to degraded con-

dition induced by chemotherapy toxicity,

• other patients entering and leaving the pathway.

Decision Making Support in the Scheduling of Chemotherapy Coping with Quality of Care, Resources and Ethical Constraints

463

These events are communicated by different ac-

tors to the system (e.g. nurses monitoring the

drugs delivery, doctors checking the patient condi-

tion, administrative staff registering the arrival or non-

attendance of a patient).

To maintain optimality, the occurrence of such

events will trigger a re-scheduling. Consequently, the

considered scheduling is an on-line problem which

should meet the following additional constraints:

• the recorded past is of course irreversible: this

makes any deviation to the ideal care delivery

schedule hard to reverse.

• conﬁrmed appointments for other patients should

preferably not be changed because it requires ad-

ministrative work and can induce a cascading ef-

fect.

A key actor in charge of activity re-planning is

the administrative nurse. She is frequently in con-

tact with the patient and acts as a relay between the

patient and the system. She is also in charge of ne-

gotiating and conﬁrming the future treatment delivery

dates between the patient and the system:

• in the ideal case, all dates initially computed are

respected.

• in case of problems in the patient’s pathway, the

patient schedule is adapted. This can impact the

schedules of other patients. This is not important

if the schedule shift is minor and concerns dates

that are not yet conﬁrmed.

• if the patient wants to delay a treatment delivery,

the system shall estimate the impact of such delay

in terms of degradation of the chances of healing.

This degradation shall be reported to the patient,

e.g. by strongly insisting on the importance to

comply with the proposed date if a good RDI is

compromised.

Figure 6 gives a complete contextual view of the

information ﬂow between the patient and the schedul-

ing system and the information the system need an

access to.

Figure 6: Problem context diagram.

4 SOLUTION DESIGN

In this section, we ﬁrst propose a global architecture

for the scheduling solution and then consider the more

tricky problem of specifying a fair objective function.

4.1 Global Scheduling vs. First Come

First Serve

Our approach is about scheduling the care of all pa-

tients together in such a way that some global time

constraints are enforced. The actual situation in most

day hospitals is that patients are scheduled on a ﬁrst-

come ﬁrst-serve basis. With such a policy, in case of

resource shortage (beds, nurses), the treatment of a

patient might be postponed by some days. For some

patients, such a delay can result in great harm in terms

of chance of healing.

In contrast, our solution avoids resource shortage

by smartly spreading over time the start date of the

chemotherapy pathways. However, if resources were

still limited, the system will smartly select patients

to postpone by limiting the impact on their time con-

straints and thus their RDI.

4.2 Proposed Architecture

An agile prototype-based approach was applied to de-

sign our scheduling solution. The architectural design

of our solution quickly evolved towards the agent-

based architecture depicted in Figure 7 and composed

of the following agents:

• the Orchestrator is the central agent. It ensures

that the system behaviour is consistent with the

input received and that the information generated

by the system is dispatched to the end user.

• the User interface captures relevant patient infor-

mation and gives comprehensive views over the

pathways at different levels of detail.

• the Simulator is used for validation purpose (see

Section 5).

• the Persister is is charge of managing the state in-

formation about the patients involved in the clin-

ical pathways. It provides a domain representa-

tion to the orchestrator and relies on a relational

database for persistent storage.

• the Oncoplanner is responsible for proposing so-

lutions matching the domain constraints sent by

the Orchestrator when a change requires to com-

pute an updated solution. It relies on the Os-

caR.cbls framework.

This architecture has the following beneﬁts:

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

464

Figure 7: Global architecture.

• it ensures a clear separation between the real

world (user side) and the computer world.

• it allows one to plug different kinds of user inter-

faces easily: ﬁrst, a basic command line user in-

terface, and later a web-based one using the same

communication protocol.

• it also enables to easily integrate a full environ-

ment simulator which can play complex scenarios

that are able to test the system under high stress.

• it provides good integration capabilities with hos-

pital systems, e.g. to retrieve information from

available databases through speciﬁc agents.

In this architecture, the scheduler can also work as

a background service constantly trying to improve the

solution in the open future (i.e. beyond all conﬁrmed

appointments), while the orchestrator can take care of

reporting when some change occurs in the real world.

In case a change makes the current computation irrel-

evant, the orchestrator can ask the scheduler to stop

his work and launch a new computation based on the

updated constraints.

We started with a simpliﬁed model combining

the chemotherapy workﬂows models resulting from

a rigorous analysis process (Damas et al., 2014),

resource constraints and possible interfering events.

An appointment scheduler was developed along with

key companion tools such as a scenario repository,

a graphical interface to manage appointments and a

simulator of patient-related events. This greatly made

easier the validation described in Section 6. The fol-

lowing features were progressively addressed to reach

a model that is now realistic enough to consider a val-

idation at day hospital :

• simple resource model, expressed in bed/nurse

hours evolving to a ﬁner grained model where

each nurse/bed is explicitly allocated.

• service opening days and hours.

• treatment plans, modelled as sequences of steps

(day of cure, resting periods) parameterized with

doses, durations and involved resources.

• constraints on treatment plan instances: earli-

est/latest start date, patient unavailability, set ap-

pointment (past or conﬁrmed).

4.3 Modelling the Objective Function

The objective function to maximize is the global RDI

over the pool of patients. We have developed two

global criteria:

A ﬁrst criterion was to maximize the minimal RDI

among all patients. It is implemented by minimiz-

ing the schedules makespan among all patients us-

ing iFlatRelax (Michel and Hentenryck, 2004). The

schedule of a patient is an interleaving of appoint-

ments and resting periods, followed by a “stub” activ-

ity at the end. This stub is needed because all patients

do not start their treatment at the same time. That stub

activity enables us to consider their treatment duration

instead of reasoning on their ending date. This crite-

ria may look fair but patients with the highest “healing

chances at start” (e.g. with no dose reduction) could

be considered as “neglected”.

A second criterion was to maximize the summed

RDI. This can be modelled as a tardiness problem,

i.e. overshot of a given point in time (patient depen-

dent) multiplied by a constant. This problem is widely

studied and was solved using a task swapping neigh-

bourhood starting from a solution provided by iFla-

tRelax because it was tightly packed and computed

very quickly.

5 IMPLEMENTATION OF

AGENTS

The implementation of the architecture detailed in the

previous section relies on web services technologies:

our agents communicate through a RESTful protocol

relying on the JSON format for exchanging the re-

quired information. This section highlight key imple-

mentation issues of the agents, except for the simula-

tor which is detailed in Section 6.

5.1 Scheduler Agent

A strong requirement was to cope with large patient

sets, typically involving hundreds of patients simulta-

neously at various stages of their own clinical path-

way. In order to scale to such size, we used lo-

Decision Making Support in the Scheduling of Chemotherapy Coping with Quality of Care, Resources and Ethical Constraints

465

cal search-based approaches, mainly iFlatRelax for

scheduling, and in a later phase, BinPacking for day-

level reasoning. Both algorithms were implemented

using the CBLS engine of OscaR (De Landtsheer and

Ponsard, 2013)(OscaR Team, 2012). They are further

described in the next subsections.

Based on this techniques, our prototype is able

to schedule chemotherapy appointments over roughly

ﬁve hundred of patients in a few seconds and supports

interactive adjustments.

5.1.1 Iterative Flattening-Relaxation Search

The algorithm implemented in our prototype is the

one already presented in our previous report, namely:

the iFlatRelax. This algorithm has been improved

with the following new features:

• The possibility to deﬁne non-moveable tasks, that

is: tasks that already have a given date. We need

to represent these, because some appointments of

treatments might already be ﬁxed, and we do not

want them to be moved, since they were commu-

nicated to the patient.

• The possibility to deﬁne forbidden zones for some

activities, that is: a set of points in time where the

activity cannot take place. The activity is hence-

forth moved forward in time until a proper posi-

tion is found. We will use this feature to represent

patient unavailability.

• A more ﬂexible model of resources, that will

enable us to represent bin-packing resources, as

mentioned in the next section.

5.1.2 Bin Packing

Beds and nurses need to be modelled as they are in

the real world: a patient needs to occupy a bed for a

certain amount of time, and more beds allocated for

a patient will not reduce his time spent in a speciﬁc

bed. The consequence is that we cannot represent the

bed resources as a single integer value in terms of bed

hours available in a certain day. Instead, we need to

model beds as done in a bin packing problem: each

bed is a bin whose size is the duration of the day, and

each patient of the day must be put in a bed selected

among the available ones. This allocation shall com-

ply to the fact that the sum of duration of each patient

occupying the same bed is lower or equal to the dura-

tion of the day. Patients will occupy the same bed in

sequence, of course.

We therefore needed to represent a so-called “bin

packing resources” exhibiting this behaviour. A bin

packing resource in a scheduling problem is a re-

source divisible into bins of given size. Each activ-

ity using this resource declares a certain amount of

usage. The activities using the same resource at the

same time must be scheduled in proper bins, so that

the sum of each activity using the same bin do not

exceed the capacity of that bin.

So far, we have developed a model of the bin-

packing problem, and a solver for bin-packing prob-

lems. We still need to integrate this solver into our

scheduling engine.

5.2 Persister Agent

The persister agent provides a service-based imple-

mentation of the normalised conceptual model de-

picted in Figure 8.

Figure 8: Data model for a chemotherapy pathway.

It can also be considered as a domain speciﬁc lan-

guage for clinical pathways and is composed of the

following concepts:

• Patient: models the information about patients

treated and monitored by the tool.

• TreatmentPlan: captures a reusable treatment

template which is composed of one or more

SchemaSteps, either RestStep or DeliveryStep.

• Treatment: captures a processing instance, for

a given patient. It involves a number of moni-

tored events such as Prescription, Delivery and

Appointment. These are linked to drug injection

steps through DeliveryStep.

• PatientUnavailability, NurseAvailability, and Be-

dAvailability respectively capture the availability

of beds, nurses and patients for the scheduling of

appointments.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

466

5.3 Agenda-based User Interface Agent

The user interface is a browser-based HTML/-

JavaScript application interacting with the orchestra-

tor. It is implemented using AngularJS. The interface

is depicted in Figure 9. It provides:

• the visualisation of the clinical pathways of the

whole set of patients, allowing to spot the past

chemotherapy deliveries, the future deliveries

with a distinction between the conﬁrmed ones (the

scheduler will not alter them but the nurse could

move them provided the patient is in the loop)

and unconﬁrmed ones (these can be moved by the

scheduler until they become conﬁrmed). Service

and patient unavailability are also displayed.

• control over the pool of patient, e.g. adding a new

patient and encoding speciﬁc characteristics of his

workﬂow instance (earliest start date, regimen pe-

riodicity and duration, target dose, etc.).

• encoding of delivery related information, e.g. par-

tial dose delivery, cancellation, no-show, . . .

Figure 9: Agenda user interface.

6 VALIDATION

6.1 Validation Approach

Validation of care pathways in a real day hospital en-

vironment will be considered in advanced research

phases. Until now we used a simulation-based ap-

proach because it helps to understand the system be-

haviour over long periods and under stressed condi-

tions that are difﬁcult to experience in the real world.

It also has the ability to step inside processes and pro-

vide a good understanding of problematic scenarios.

In order to enable an earlier validation, we de-

signed the simulator to:

• Provide a reactive and complete user interface,

clearly illustrating the key characteristics of the

algorithms: quality of the solutions, speed of re-

calculation, taking into account complex events

(no-show, partial dose, report ...)

• Integrate an environment simulator able to gener-

ate ﬂows of planned and unforeseen events that

are experienced by the targeted hospital services

and that can be used to consolidate the required

indicators proving the value of our tool, especially

in relation to the quality of care in terms of com-

pliance with RDI and load management.

6.2 Simulation Framework

As shown by Figure 7, the simulator agent is directly

connected to the orchestrator through the same proto-

col as the user interface. The simulator has two main

components:

• a “driver” responsible for simulating the interac-

tions a user would normally have with the orches-

trator using the UI. It relies on exactly the same

communication protocol (JSON via HTTP) as the

UI.

• a “control panel” to have the control over the

running simulation, depicted in Figure 10. Two

modes are available: (1) step-by-step mode. This

mode enables to have a deep understanding of a

speciﬁc run. It can also be used to introduce spe-

ciﬁc events manually using the usual user inter-

face. (2) fully automated mode, running at ma-

chine speed. This mode relies on the generation

of events based on probability distributions. It can

be used to spot speciﬁc problems and also to as-

sess the performance over a large number of runs

using a Monte-Carlo process.

• a component collecting and displaying the evolu-

tion of relevant indicators, such as the RDIs and

the service load.

Figure 10: Simulator control panel.

Decision Making Support in the Scheduling of Chemotherapy Coping with Quality of Care, Resources and Ethical Constraints

467

The state machine for the patient is illustrated in

Figure 11: between treatment deliveries, it should be

in an “appointment ﬁxed” state and during delivery, it

will be in “Delivery” state (if correctly reported). The

“waiting state” should only be transitory: either just

after a delivery or in case of no show. Those states

should be tracked for minimal duration. The simu-

lator generated events covering those transitions with

given probabilities, e.g. for no show events.

Figure 11: Possible patient states and transitions.

6.3 Results Validation

Several simulations sessions were organised both

with the technical team and in sessions with oncology

practitioners involving three hospitals (UCL/Cancer

Institute, Grand Hospital of Charleroi and UZ Leu-

ven). The feedback from doctors was quite positive

about our contribution to ensure both quality of care

and the smarter use of resources. Figure 12 shows

that the RDI is kept above the 0.85 and above 0.90

in many cases. It is decreasing over time as the re-

sult of minor delays or partial doses delivered due to

chemotherapy toxicity. The day hospital load is also

exhibiting a smooth curve meaning that appointments

can be evenly dispatched over time.

Figure 12: RDI reached in a typical oncology unit.

The prototype also raised ethical concerns, such

as the capacity of the tool to choose to favor some

patients rather than others in case of resource short-

age. Our conclusion is that the system should report

such situations ahead of time to allow the day hospi-

tal team to take corrective measures, like a transient

increase of stafﬁng. In order to keep the medical team

in control, the developed graphical display was also

a huge practical improvement. Some interesting fea-

ture such as the visualization of allocation windows

ensuring a good RDI level deﬁnitely helped oncolo-

gists and nurses in charge of appointments updates.

7 RELATED WORK

A complete literature review on integrated hospi-

tal scheduling problems, including pathways, was

recently published by (Marynissen and Demeule-

meester, 2016). Although concepts such as clini-

cal pathways or diagnosis related groups have been

around for more than 20 years, the study reveals that

most of the relevant work is quite recent. Besides

progress in methods and tooling, the main trigger fac-

tor is that hospitals are facing the necessity to break

barriers across services for dealing with the perfor-

mance and capacity challenges they have to face. Cur-

rently, off-line scheduling approaches are more often

reported than on-line methods because hospitals want

to provide the best possible solutions which are re-

lated to the largest possible exploration of the state

space. This rules outs methods which are mostly

based on (meta-) heuristics. However very good (and

in some cases optimal) results of using local search

for outpatient scheduling has been reported in the

literature (Kaandorp and Koole, 2007) and are con-

ﬁrmed by the quality of our simulation results. A

key point is that our CBLS engine ensures fast ex-

ploration of the search space, resulting in a good cov-

erage. Moreover our design also allows the system

to continue optimizing the time horizon for appoint-

ments not yet conﬁrmed to patients and thus without

strong time constraints.

A methodology to design appointment systems for

outpatient clinics and diagnostic facilities that offer

both walk-in and scheduled service is presented in

(Kortbeek et al., 2011). The proposed schedule has

two levels: a global level stating the number of ap-

pointments per day and a day level detailing when

each appointment should be scheduled on a given day.

Each level is managed by a speciﬁc model and the

two models are connected by an algorithm. Our ap-

proach also proposes two levels with a coarser gran-

ularity at the global level (e.g. we consider global

bed-hours availabilities) while at the day level a ﬁne

grained model is used (i.e. we assign patient to avail-

able beds). Our current validation is however cur-

rently limited to the global level.

Scheduling has also been successfully applied in

other hospital areas. In radiotherapy, several ap-

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

468

proaches of operation research ranging from strategic

capacity management to operational scheduling lev-

els are reported in (Vieira et al., 2016). It highlights

that many improvements regarding the waiting times

and resource utilization can be achieved. A substan-

tive attention is also devoted to the scheduling of the

operation room because it accounts for up 40% of re-

source costs in a hospital. However it differs from on-

cology pathways by a far more important level of un-

predictability. Reported results shows that about 30%

more patients can be scheduled than in actual practice

and the operating room utilization rate is increased by

20% (Barbagallo et al., 2015).

Multiple algorithms and software tools to generate

qualitative surgery schedules on the tactical and op-

erational level are also reported in (Demeulemeester

et al., 2007). This work actually also shows a wider

impact on the whole hospital, since these operation

rooms are interrelated to many other departments or

organizational problems like nurse scheduling or bed

levelling. It also points out the necessity of a good vi-

sualization capabilities because they help health man-

agers to have good insights and they also guide them

in testing different scenarios.

Regarding the computer tools used, the observed

practice is that tool support is often still relying on

manual or basic tools such as spreadsheets or schedul-

ing templates (Ahmed et al., 2011).

8 CONCLUSION

This paper presented an approach for assisting the

scheduling of drug deliveries for treatments where

such deliveries must comply to time constraints un-

der relatively limited resources, namely chemother-

apy regimens. A chemotherapy is delivered in several

deliveries or cures; some timing constraints must be

enforced between these cures. If the cures are spaced

too much, the efﬁciency of the treatment gradually de-

creases. If the cures are too close in time, there is a

high risk of side effect. These side effects should be

avoided as much as possible.

When it comes to mass medicine, we must con-

sider not only individual patients, but pools of pa-

tients. Every patient that is treated by chemotherapy

depends on the resource of the hospital for the deliv-

ery of his cures. These resources consist mostly in

bed time, nurse time, and doctor time. Since they are

shared resource, there exists a form of competition -

in the mathematical sense - to access these resources.

In practice, such competition is usually solved in

a ﬁrst come ﬁrst serve way: every patient sets its ap-

pointments for deliveries, according to its constraints,

and to the available resources. The level of critical-

ity of time enforcement is patient-dependent as it de-

pends on their past history, the status of their disease,

their actual chemotherapy, etc. An indicator has been

proposed, called the RDI to measure the timeliness of

a chemotherapy regimen, and it has been shown to be

correlated with the survival rate of patients for some

cancer types.

Our approach suggests setting the dates of treat-

ment deliveries based on this RDI indicator, by max-

imizing a global measure of the RDI among the con-

sidered pool of patients.

By combining the use of a scalable open source

CBLS scheduling technology with visualization and

simulation components, we were able to show the

feasibility of quality indicator-driven scheduling of a

large pool of patients.

At some point, the following provocative question

for operations research practitioners was raised: is it

really a good or a fair idea to install an optimizing

engine in such a critical setting? As usual the answer

is not in the technology but in the way it is used and

controlled. For instance, the global formula of RDI

is very critical in the sense that it might set global

policies for deciding some life-or-death trade-offs in

the pool of patients. If a patient has a very poor RDI

for its past treatment deliveries, should he get a very

high priority for his coming deliveries or should the

hospital resources be allocated to patients that have

higher probabilities of survival, thus abandoning the

ones that have lower chances?

Clinical pathways involve an intricate decision

making process and our experience shows that the

scheduler can deﬁnitely support the medical actors in

their work.

Our next step is to conduct on-site validation

based on an extended prototype. A major request is to

achieve ﬁner tasks management, i.e. within each day.

This requires to rework our algorithms to integrate the

bin packing solver into the scheduling engine. On site

validation should be carefully designed to ensure that

at no time the global process will depend solely on

non-validated tooling. As such, a ﬁrst step is typically

to run such a tool in parallel with the existing process

and compare their behaviour and outcomes. The next

step is to transmit scheduling suggestions from our

tool to the existing process to check with practition-

ers that the suggestions proposed are applicable and

contribute to improve the quality of care.

Decision Making Support in the Scheduling of Chemotherapy Coping with Quality of Care, Resources and Ethical Constraints

469

ACKNOWLEDGEMENTS

This research was partly funded by the Walloon Re-

gion by the PIPAS (nr. 1017087) and PRIMa-Q (nr.

1610088) projects. We thanks UCL/Cancer Institute,

Grand Hospital of Charleroi and UZ Leuven for their

valuable feedback in key elicitation, design and vali-

dation phases of our work.

REFERENCES

Ahmed, Z., T., E., and S., B. (2011). Developing an efﬁcient

scheduling template of a chemotherapy treatment unit:

A case study. Australas Med J., 4(10):575–88.

Barbagallo, S. et al. (2015). Optimization and planning of

operating theatre activities: an original deﬁnition of

pathways and process modeling. BMC Medical Infor-

matics and Decision Making, 15(1):38.

Benoist, T., Estellon, B., Gardi, F., Megel, R., and Nouioua,

K. (2011). Localsolver 1.x: a black-box local-search

solver for 0-1 programming. 4OR, 9(3):299 – 316.

Campbell, H., Hotchkiss, R., Bradshaw, N., and Porteous,

M. (1998). Integrated care pathways. British Medical

Journal, pages 133–137.

Choo, J. (2001). Critical success factors in implementing

clinical pathways/case management. Ann Acad Med

Singapore 30 (4 Suppl), 7:17–21.

Damas, C., Lambeau, B., and van Lamsweerde, A. (2014).

Analyzing critical decision-based processes. IEEE

Trans. Software Eng., 40(4):338–365.

De Landtsheer, R., Guyot, Y., Ospina, G., and Ponsard, C.

(2015). Combining neighborhoods into local search

strategies. In Proceedings of MIC’2015.

De Landtsheer, R. and Ponsard, C. (2013). Oscar.cbls :

an open source framework for constraint-based local

search. In Proceedings of ORBEL’27.

Demeulemeester, E., Sermeus, W., Belin, J., and Cardoen,

B. (2007). Clinical pathways and operations manage-

ment. it takes two to tango. Review of Business and

Economic Literature, LII(3):451–470.

Di Gaspero, L. and Schaerf, A. (2003). EASYLOCAL++:

an object-oriented framework for the ﬂexible design

of local-search algorithms. Software: Practice and

Experience, 33(8):733–765.

Gooch, P. and Roudsari, A. (2011). Computerization of

workﬂows, guidelines, and care pathways: a review of

implementation challenges for process-oriented health

information systems. Journal of the American Medi-

cal Informatics Association, 18(6):738–748.

Kaandorp, G. C. and Koole, G. (2007). Optimal outpatient

appointment scheduling. Health Care Management

Science, 10(3):217–229.

Kortbeek, N., Zonderland, M. E., Boucherie, R. J., Litvak,

N., and Hans, E. W. (2011). Designing cyclic appoint-

ment schedules for outpatient clinics with scheduled

and unscheduled patient arrivals.

Lyman, G. (2009). Impact of chemotherapy dose intensity

on cancer patient outcomes. J Natl Compr Canc Netw,

pages 99–108.

Marynissen, J. and Demeulemeester, E. (2016). Literature

review on integrated hospital scheduling problems.

Working Papers Department of Decision Sciences and

Information Management 555258, KU Leuven, Fac-

ulty of Economics and Business.

Mauro, C., Happle, T., Sunyaev, A., M.Leimeister, J., and

Krcmar, H. (2010). From medical processes to work-

ﬂows: modeling of clinical pathways with the uniﬁed

modeling language. In Int. Conf. on Health Informat-

ics (HealthInf), Valencia, Spain. 161 (9-10).

McGlynn, E. A. et al. (2003). The quality of health care

delivered to adults in the United States. New England

Journal of Medicine, 348(26):26352645.

Michel, L. and Hentenryck, P. V. (2004). Iterative relax-

ations for iterative ﬂattening in cumulative schedul-

ing. In 14th International Conference on Automated

Planning & Scheduling (ICAPS).

Newton, M. A. H., Pham, D. N., Sattar, A., and Maher, M.

(2011). Kangaroo: an efﬁcient constraint-based local

search system using lazy propagation. In Proceedings

of CP’11, pages 645–659.

OscaR Team (2012). OscaR: Operational research in

Scala. Available under the LGPL licence from

https://bitbucket.org/oscarlib/oscar.

Piccart, M., Biganzoli, L., and Di Leo, A. (2000). The im-

pact of chemotherapy dose density and dose intensity

on breast cancer outcome: what have we learned? Eur

J Cancer., 36(Suppl 1).

van Dam, P. A. et al. (2013). A dynamic clinical pathway

for the treatment of patients with early breast cancer

is a tool for better cancer care: implementation and

prospective analysis between 2002–2010. World Jour-

nal of Surgical Oncology, 11(1):70.

Van Hentenryck, P. and Michel, L. (2009). Constraint-

based Local Search. MIT Press.

Vieira, B. et al. (2016). Operations research for resource

planning and -use in radiotherapy: a literature re-

view. BMC Medical Informatics and Decision Mak-

ing, 16(1):149.

Walter, F., Bass, N., Bock, G., and Markel, D. (2007). Suc-

cess of clinical pathways for total joint arthroplasty

in a community hospital. Clin Orthop Relat Res,

7:457:133.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

470