Expertise versus Data: Comparison of Expertise Based Process Models

to Process-Mining Models of Surgery Under General Anesthesia

Hugo Boisaubert

1 a

, Chlo

e Grivaud

, Antoine Bouchet

, Corinne Lejus-Bourdeau

2,3

and

Christine Sinoquet

Nantes Universit

e, CNRS, LS2N UMR 6004, 2 Chemin de la Houssini

ere, 44300 Nantes, France

Nantes Universit

e, LESiMU, 9 rue Bias, 44000 Nantes, France

Department of Anaesthesia and Intensive Care, CHU Nantes, 1 place Alexis Ricordeau, 44000 Nantes, France

Keywords:

Process-Mining, Anesthesia, Healthcare, Model Comparison, Machine Learning, Expertise, Personal Data.

Abstract:

Digital tools accessible for healthcare are often based on models representing a medical process and learned

from medical data. Unfortunately, those data are protected by privacy regulation and therefore are quite rare.

This rarity leads to process models mainly based on the expertise of caregivers. Those expertise-based model

and data-based models are rarely compared to show their common characteristics and differences. When both

model can be produced for the same situation multiple questions arise. Should the expertise-based model

be invalidated if it is not in full conformity to the data-based model ? Are those models’ characteristics the

same? In this article, we present a comparison of expertise-based models and data-based models produced for

a surgery under general anesthesia with 204 real cases. We conducted a process mining algorithm performance

comparison on our speciﬁc real data to identify the most promising learning method. Then we compared the

produced data-based models to the expertise-based models with some metrics. The comparison results show

strong differences between the two types of models, the expertise-based model is very much smaller than

the data-based model, but we have noticed that the expert-based model is included in a data-based model.

Therefore, the main difference between the two models appears to be on a level of abstraction.

1 INTRODUCTION

Anesthesia is a common practice in occidental

medicine to facilitate speciﬁc care or medical ges-

tures. More than 300 million anesthesias are induced

every year in the world (Gao et al., 2022).

The development in digital sciences plays a role in

this improvement with new tools for caregivers, ﬁrst

with decision aid software (O’Connor et al., 1999)

for anesthesia consultation, with the involvement of

digital simulation of patients in training of caregivers

(Kononowicz et al., 2019), and most recently with

digital twins (Katsoulakis et al., 2024).

Many of those tools are based on process models

that represent the medical situation where the tool is

intended to be used, from organization process mod-

els for a full hospital to medical treatment process

models for the care of a patient (Kaymak et al., 2012).

However, before giving access to caregivers of this

kind of tool, those models must be accurately pro-

https://orcid.org/0000-0002-1855-0727

duced. This can be easily achieved by process min-

ing and machine learning techniques based on real

medical data. Unfortunately, those kinds of data are

quite rare, as they are highly sensitive patient data

and therefore protected by privacy regulations like the

GPDR

. Consequently, models are usually built with

the help of caregivers and experts based on their ex-

pertise.

This work is born inside a collaboration between

the Laboratory of Digital Sciences of Nantes Uni-

versity and the Nantes University Hospital. During

this collaboration, aiming to develop new methods for

caregivers in general anesthesia training with the help

of digital tools, medical data should have been used to

learn process models by machine-learning techniques

and process mining approaches.

However, because of the private and sensitive na-

ture of those data, access was not easily granted.

Thus, a modeling effort, involving anesthesia experts,

trainers, and caregivers, has been carried out, re-

The General Data Protection Regulation (GPDR), is

the European Union regulation on the usage of data.

742

Boisaubert, H., Grivaud, C., Bouchet, A., Lejus-Bourdeau, C. and Sinoquet, C.

Expertise versus Data: Comparison of Expertise Based Process Models to Process-Mining Models of Surgery Under General Anesthesia.

DOI: 10.5220/0013262300003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 742-749

ISBN: 978-989-758-731-3; ISSN: 2184-4305

sulting in expertise-based process models of general

anesthesia.

When access to the precious health data was,

at least, granted, the question of the conformity of

expertise-based process models to real data produced

by the real medical procedure has arisen.

The aim of this work is to compare different

expertise-based process models to data-based process

models in healthcare situations, more precisely in a

general anesthesia situation. To our knowledge, this

work of comparison has not yet been done.

This article will ﬁrst present our medical situation,

then our application context and approach, and ﬁnally

our results and an analysis where we will try to deter-

mine how both types of models compare to, hopefully,

lead us to a proposition of usage.

2 A SPECIFIC MEDICAL

CONTEXT

Application of process mining approaches for health

is now a full research domain (known as PM4H) with

distinct characteristics and challenges (Munoz-Gama

et al., 2022). Experimental results and speciﬁc ap-

proaches to process mining for health have been pub-

lished in the last decade (Rojas et al., 2016)(Guzzo

et al., 2022), including multiple results on anesthesia

(Kaymak et al., 2012).

In the case of surgery performed under general

anesthesia, the process is directly dependent on the

speciﬁcities of the patient. Although there is a com-

mon framework for all interventions under general

anesthesia, there is no reference process model to

serve as the ground truth, just recommendations for

good practices from health authorities or national

anesthesia society.

2.1 General Anesthesia

General anesthesia consists of the use of powerful

analgesics, such as drugs from the opioid family, like

morphine, on an unconscious patient. It is recom-

mended when the patient cannot consciously endure

a care. The loss of consciousness in a patient in-

volves the use of speciﬁc medications, such as hyp-

notics - substances capable of inducing and/or main-

taining sleep - and specialized medical procedures to

initiate and maintain it.

Depending on the depth of sedation - the loss of

vigilance and consciousness through the use of med-

ication - the patient may lose their respiratory reﬂex.

It is therefore necessary to intubate them and connect

them to a mechanical respirator to provide respira-

tory assistance. Throughout the operation, anesthe-

sia is maintained by continuous or regular injection

of drugs.

A general anesthesia involves many risks. The use

of curare is the cause of the majority of allergic risks

for that kind of care, but the use of certain hypnotics

can lead, in very rare cases, to serious and generally

unpredictable allergic reactions.

The high variability of these procedures leads to

high variability in the data they produce and there-

fore hinder the ability of process mining algorithms

to generate models. Handling this variability is one of

the challenges for PM4H (Munoz-Gama et al., 2022).

2.1.1 A Semi-Rigid Structure

General anesthesia is administered in four main steps:

• Patient Entry: the patient arrives in the operating

room and is prepared for the procedure.

• Anesthesia Induction: analgesia is performed

and loss of consciousness is induced. The patient

is placed on respiratory assistance.

• Surgical or Medical Procedure: the actions

planned during the procedure are performed. Loss

of consciousness and analgesia are maintained.

• Patient Exit: maintenance of anesthesia is

stopped and the patient is prepared for awakening.

The four major steps of an anesthesia procedure

correspond to different moments of activity of the in-

tervention and are always present. For each of these

major steps, speciﬁc steps are carried out by care-

givers in a determined and ﬁxed order. This is the

rigid component of the structure of anesthesia. Steps

for the ﬁrst main step are presented in Table 1.

Table 1: First Main Step, Steps and Sub-steps of a general

anesthesia.

Main step Step Sub-step

Patient Enter in room

set up Set up on table

Heart rate

Patient entry Monitoring Blood pressure

set up O2 saturation

Curare level

Bair Hugger

Preparation Venous route

Prophylaxis

Each of these actions, relating to care, is com-

posed of sub-steps that correspond to different activ-

Expertise versus Data: Comparison of Expertise Based Process Models to Process-Mining Models of Surgery Under General Anesthesia

743

ities dedicated to medical care. They are carried out

by the medical team and vary according to the inter-

vention and the patient’s proﬁle. This is the variable

component of the structure of anesthesia.

2.2 Anesthesia Event Logs

The various actions carried out by the medical team

during the main steps, steps and sub-steps of an anes-

thesia procedure are recorded as event logs during the

monitoring of the intervention.

Since 1994 (Journal Ofﬁciel, 1994), in France

all anesthesia procedures must be monitored, and all

events and physiological time series of the patients

must be archived for forensic reasons.

Since 2000, forensic archiving has been increas-

ingly digitalized in more and more hospitals. Those

archived data form an important database of real-

world anesthetic data. For this work, only the event

log of the surgery has been used as raw data. Those

event logs are composed of the timestamp of the

event, the event name, and sometimes a comment

added by the medical team. Some examples of event

log tuples are shown in Table 2.

Table 2: Example of an anesthetic event log.

Timestamp Event name

21/02/2022 09:30:27 Patient set up

21/02/2022 09:30:57 Heart rate monitoring

21/02/2022 09:31:27 Blood pressure monitoring

3 COMPARED MODELS

In this work, we seek to represent general anesthe-

sia situations during surgery in the form of a process

model. Whether these process models come from

machine learning techniques or from the expertise of

caregivers, they aim to represent the same situations.

3.1 Process Models

A process model is a representation of a process. This

kind of modeling is used for multiple tasks, such

as discussing processes, document workﬂow, work-

ﬂow veriﬁcation, performance analysis, etc. (Van

Der Aalst, 2016).

3.1.1 Concepts

Each process model is composed of a set of activities

A. With a process model, we aim to represent which

activities will be executed and the sequential order of

execution during the process.

Therefore, a process model is also composed of a

set of transitions T that deﬁne the sequential order of

execution. Process models are represented as an ori-

ented graph where activities and transitions are nodes

linked by arcs. Some activities or chains of activi-

ties may be optional in the processes, or concurrent;

we can therefore identify speciﬁc transitions into two

categories: splits and gateway. Split transitions deﬁne

multiple possible split-choices after an activity:

• AND-split imposes multiple next activities;

• XOR-split imposes to select one of the multiple

next activities (the choice is made after the evalu-

ation of speciﬁcally deﬁned conditions);

• OR-split OR-split permits to select one or more

in the multiple next activities (here too, the choice

is made after the evaluation of speciﬁcally deﬁned

conditions).

The join transitions (i.e., gateways) work in a sym-

metric way to deﬁne an activity joining multiple pos-

sible preceding activities:

• AND-join requires multiple preceding activities

to be completed before the next activity can be

executed;

• XOR-join XOR-join requires one of the previ-

ous activities to be completed before executing the

next activity;

• OR-join requires one or more of the previous ac-

tivities to be executed before the next activity can

take place.

3.1.2 Notation

Process model notation is usually a graphical repre-

sentation of the model; those notations are plethorous,

and it is relatively easy to translate a process model

in a speciﬁc notation to another one (van Der Aalst

et al., 2003) with tools dedicated to this task in ref-

erence process mining toolkit ProM (Verbeek et al.,

2010).

The most used notations are Petri Net and BPMN

(Business Process Model and Notation), as process

mining has a good proximity with business modeling.

3.2 Expertise-Based Process Models

The models obtained from experts were constructed

based on various discussion sessions with experts

combined with reference sources, such as biblio-

graphic sources, recommendations from anesthesia

societies, or internal good practice references from

hospitals.

HEALTHINF 2025 - 18th International Conference on Health Informatics

744

The models were iteratively improved from ses-

sion to session using a top-down approach. A ﬁrst

model corresponding to a generic anesthesia situation

was formulated, and this ﬁrst model was then spec-

iﬁed by successively adding cases speciﬁc to certain

patients, certain medical procedures, or medications.

3.3 Data-Based Process Models

Data-based processes are process models constructed

using machine learning techniques from real data,

speciﬁcally the event log of an anesthetic procedure.

Those techniques are part of the process mining ﬁeld.

3.3.1 Datasets

All the process models have been learned on anes-

thetic data from the Nantes University Hospital.

The used dataset is composed of the event log ex-

tracted from a forensic archive. The cohort of pa-

tients used to produce this dataset is composed of 204

male patients, aged between 30 and 50 years, with no

speciﬁc medical history or comorbidities, receiving a

surgery under general anesthesia for the curation of

an inguinal hernia under laparoscopy.

A dataset consisting of 204 cases may seem too

small for any machine learning approach. How-

ever, this is a constraint when using real medical

datasets. Synthetic data could be used, but that would

have undermined the value of our comparative ap-

proach. Handling ”small” datasets, in regard to ha-

bitual datasets in machine learning, is one of the

challenges of PM4H algorithms (Munoz-Gama et al.,

2022). Different characteristics of the dataset are

shown in the Table 3

Table 3: Multiple characteristics of the used dataset.

Characteristic Values

Number of Cases 204

Number of Activities 144

Mean Activities per event log 34.8

Min/Max Activities per event log 27/45

Average Trace Time (seconds) 8166

Min/Max Trace Time (seconds) 2048/18654

3.3.2 Selection of Some Process Mining Methods

There are many algorithmic approaches to learn pro-

cess models by process mining techniques. Algo-

rithms dedicated to process discovery are numer-

ous. In order to determine which approaches of pro-

cess discovery would be most suitable for anesthetic

event logs, a small comparative study was carried

out. We selected 6 different process discovery al-

gorithms, based on their usage in literature or per-

formance in healthcare process mining (Guzzo et al.,

2022)(Munoz-Gama et al., 2022). We will introduce

them in more detail in the next subsections.

All those algorithms are implemented in Java,

mostly inside the reference tool ProM (Verbeek et al.,

2010), except for the Split Miner algorithm. For some

of them, the PM4PY(Berti et al., 2023) library has

been used as a Python interface.

The Alpha+ algorithm (De Medeiros et al., 2004)

is an upgrade of the original alpha algorithm, with

the ability to mine short loops (De Medeiros et al.,

2004) . The Alpha+ algorithm is a formal approach

to process mining. This approach presupposes a per-

fect event log (Weijters et al., 2006) where the log

is complete, which means that if an activity can fol-

low another activity directly, the log should contain

an example of this behavior. Besides, the log contains

no noise, meaning everything is correctly recorded.

Real-life event logs are very rarely noise-free and/or

complete. The use of this algorithm with our data, if

computation is possible, will probably show its lack

of performance with real data. However, as it is a ref-

erence algorithm, we have selected it for this study.

The Heuristics Miner algorithm (Weijters et al.,

2006) is an upgrade of the alpha+ algorithm, using the

frequency of a behavior in an event log to be able to

deal with noise and low-frequent behavior. The main

advantage of the heuristics miner is its ability to deal

with noise. However, it is only able to express the

main behavior registered in an event log, not all de-

tails and exceptions.

The Inductive Miner (Leemans et al., 2014)uses

a divide-and-conquer approach to discover a process

model from an event log. The log is split into sublogs

in order to discover an operator between them (Lee-

mans et al., 2014). The process model discovered is

a process tree, easily convertible into a Petri-Net or a

BPMN notation. For this algorithm, a noise threshold

must be selected; the values chosen for this parameter

are discussed in the result section.

The ILP Miner (Integer Linear Programming

Miner) (Van der Werf et al., 2008) uses a region-based

approach. The main goal of this algorithm is to search

for as many places as possible, such that the resulting

Petri net is consistent with the log (i.e able to replay

the log).

The ETM Miner (Evolutionary Tree Miner)

(Buijs, 2014) uses an evolutionary approach to create,

evaluate, and change candidate solutions. The evalua-

tion is always done using the four quality dimensions

used in process discovery (see 3.3.3). This approach

is quite ﬂexible because it allows for the selection of

evaluation criteria that we want to maximize.

The Split Miner (Augusto et al., 2017) algorithm

is designed to generate understandable and accurate

Expertise versus Data: Comparison of Expertise Based Process Models to Process-Mining Models of Surgery Under General Anesthesia

745

Table 4: Comparison of the selected process discovery algorithms.

Algorithms Noise Duplicate tasks Hidden tasks No-free choice Loops

sensitivity detection detection construct detection

Alpha+ Algorithm Sensitive No No No Small loops

Heuristic Miner Less sensitive No No Partially Yes

Inductive Miner Robust Yes No Yes Yes

ILP Miner Very sensitive No Partially Yes Yes

ETM Miner Robust Yes Yes Yes Yes

Split Miner Robust Yes No Yes Yes

process models. It stands out for its ability to han-

dle complex transitions and concurrent behaviors by

splitting the event log into partitions based on depen-

dency relations, then creating local models for each

partition, and ﬁnally integrating the local models into

a global process model using logical operators.

A comparison of the process discovery algorithms

selected for this study is presented in Table 4.

3.3.3 Process Mining Model Evaluation Criteria

The models generated using the different algorithms

presented earlier were evaluated using four different

metrics : ﬁtness, precision, generalization, and sim-

plicity.

The ﬁtness metric calculates how many behaviors

from the event log are covered by the model. The

token-based replay ﬁtness method returns the percent-

age of traces that are complete in the model (Berti and

van der Aalst, 2019).

The precision metric involves replaying different

parts of the event log on the model to check if the

model is underﬁtting the log. To compute precision,

we used the heuristic-based (Munoz-Gama and Car-

mona, 2010) method, faster but not exact.

The generalization metric measures whether the

elements of a model are sufﬁciently visited when the

event log is replayed on the model. The generalization

measure is the one given in (Buijs et al., 2014).

The simplicity metric is calculated using an in-

verse arc degree method described by (Blum, 2015).

3.4 Comparison of Process Models

In this work, we are taking a similar approach to the

search for conformity. However, it is not a question

of comparing a learned model with a reference model

assumed to be true (ground truth) but of comparing

two models in order to highlight their differences and

deduce the impact on the possible use of this model.

There are several approaches to checking the con-

formance of a process model: alignments, comparing

footprints (Van der Werf et al., 2008). However, those

approaches are mainly used to ﬁnd commonalities and

discrepancies between a process model and an event

log to then improve the process model.

Other approaches are more user-friendly to com-

pare two different process models. In this work we

will use two metrics of distinguishability : Support

and Conﬁdence (Kuo and Chen, 2012). Those metrics

are complementary and aim to quantify the similarity

of the compared models.

The Support metric aims to quantify the resem-

blance, from a relational point of view, of the activi-

ties of both process models; For two process models

a and b, this metric is computed as:

Support(a, b) =

a∩b

a∪b

a∩b

a∪b

(1)

where :

• A

a∩b

is the number of the common activities in

process models a and b ;

• A

a∪b

is the number of activities in process models

a and b ;

• T

a∩b

is the number of the common transitions in

process models a and b ;

• T

a∪b

is the number of transitions in process mod-

els a and b.

The Conﬁdence metric computes a ratio of se-

quential activities (i.e a transition) common to both

models with the number of transitions in a speciﬁc

model. For two process models a and b, this metric is

computed as:

Con f idence(b|a) =

a∩b

(2)

where :

• T

a∩b

is the number of the common transitions in

process models a and b ;

• T

is the number of transitions in process models a.

4 RESULTS

This section presents the results of our comparative

process mining study. We then introduce the process

models produced, and we present the results of their

quality evaluation. Lastly, the expert-based process

HEALTHINF 2025 - 18th International Conference on Health Informatics

746

model and the data-based process model are com-

pared.

4.1 Comparative Study of Algorithms

All the selected algorithms have been applied to the

inguinal hernia dataset presented in 3.3.1.

4.1.1 Algorithm Parameters

Most of the process discovery algorithms we selected

do not require speciﬁc parameter settings. However,

the Inductive Miner and the ETM Miner do require

speciﬁc parameters.

For the Inductive Miner algorithm, a noise thresh-

old must be selected. By default, it should be set to

0.2 (Leemans et al., 2014). However, depending on

the quality of the data, it is noted by the authors that

it might be useful to increase this threshold. This is

why we have decided to test three noise threshold val-

ues when testing on real data: 0.2, 0.25, and 0.3. Our

goal with these settings is to determine whether these

variations can improve the metrics’ values and lead to

the best possible model, as medical data are known to

be noisy (Munoz-Gama et al., 2022).

The ETM miner algorithm requires many param-

eters. Since this is a model that takes a long time to

run, compared to the other approaches, we have cho-

sen to keep the default parameters to not increase the

computing time (Buijs, 2014).

4.1.2 Computing and Result Transformation

As the dataset used is composed of personal medical

data, the computing of all selected process discovery

algorithms has been done inside the digital infrastruc-

ture of the University Hospital.

A computing pipeline has been created with

Python to select data from the dataset, then compute

the different process models in one job with the help

of PM4PY Python library to interface the ProM tool.

However, this was not possible for the ETM algorithm

and Split Miner; those computations have been made

manually. Computing times for all the algorithms are

presented in Table 5.

Some of the selected process mining algorithms

produce a process model using the Petri Net notation;

others use the BPMN notation. All the produced pro-

cess models have been converted to BPMN using the

conversion tools of ProM(Verbeek et al., 2010) to en-

sure uniform notation.

4.1.3 Evaluation of the Obtained Process Models

Two of the selected process discovery algorithms, the

Alpha+ and ILP miner, did not produce results due to

Table 5: Computing times (in seconds) on the dataset for

the selected process mining algorithms.

Algorithm Computing Time

Inductive Miner (T h = 0.20) 2.723

Inductive Miner (T h = 0.25) 1.582

Inductive Miner (T h = 0.30) 1.419

Heuristic Miner 2.237

Split Miner 2.178

ETM Miner 6060

too long computation times. 6 models were produced,

with 3 models for Inductive Miner, one for each noise

threshold. These models have been evaluated using

the different metrics described in 3.3.3, and the results

are shown in Table 6.

Table 6: Evaluation result for the models obtained with

Inductive Miner, Heuristics Miner, Split Miner and ETM

Miner.

Metric Inductive Miner Heur. ETM Split

0.2 0.25 0.3

Fitness 0.983 0.883 0.841 0.785 0.353 0.764

Precision 0.387 0.816 0.867 0.797 0.995 0.997

Generalization 0.798 0.759 0.745 0.344 0.400 0.549

Simplicity 0.562 0.595 0.617 0.407 0.530 0.58

As we can see, of all produced process models, ﬁt-

ness and precision values are lower than what may be

expected. The performance of the ETM Miner is quite

disappointing given the computing time involved. If

the Heuristics Miner and Split Miner produce good

models from the point of view of ﬁtness and preci-

sion, those models lack simplicity and generalization.

It may be the result of noise in the dataset.

As predicted, the multiples value of the noise

threshold for the inductive miner shows an associa-

tion with the relaxation of this value.

Based on these results, we have selected the pro-

cess model produced by the Inductive Miner algo-

rithm (th = 0.3), to be used as our data-based process

model as it ﬁts with the best equilibrium for the dif-

ferent evaluated metrics.

4.2 Comparison of Models

The expert-based process model obtained after mul-

tiple iterations with the expert and caregivers for the

cure of an inguinal hernia under laparoscopy has been

validated by a group of external experts. This process

model has been then compared with the selected data-

based process model.

The two process models are too big to be inte-

grated in this article. The main characteristics of those

models are presented in Table 7.

The two metrics of distinguishability (Kuo and

Expertise versus Data: Comparison of Expertise Based Process Models to Process-Mining Models of Surgery Under General Anesthesia

747

Table 7: Characteristics of the two compared process mod-

els: data-based process model (dd) and expertise-drive pro-

cess model (ed).

Expertise-based Data-based

process model process model

Activities 35 95

Transitions 45 138

Arcs 74 210

Chen, 2012) presented in 3.4 have been computed

between the data-based process model (db) and

expertise-based process model (eb) and are presented

in table 8.

Table 8: Metrics of distinguishability for the two compared

models : data-based process model (db) and expertise-

based process model (eb).

Metric Value (%)

Support(db, eb) 8.56 %

Con f idence(db|eb) 10.24 %

Con f idence(eb|db) 83.58 %

A comparison of the characteristics of the two

models involved shows their differences in terms of

the number of activities, transitions, and arcs, which

inevitably leads to a rather weak support metric.

If it is consequently expected that the conﬁdence

between the data-based model and the expert-based

model is low, it is interesting to note that the op-

posite, the conﬁdence between the expert-based and

data-based models, is very high.

We can easily deduce that the expert-based model

is included in the data-based model, but that they still

differ in some parts. It may be due to the impact of

noise in the data (whose impact we see in the eval-

uation metrics) and by the intrinsic character of the

expertise-based model in terms of level of abstraction.

5 DISCUSSION

These elements lead us to several analyses and cri-

tiques concerning the results obtained, the biases they

may highlight as well as the generalizable and repeat-

able nature of this work.

5.1 Some Criticisms of the Results

As often described in the literature, medical data are

often very noisy (Munoz-Gama et al., 2022), and this

has an impact on the values of the different evalua-

tion metrics, where one could expect higher ﬁtness

and precision values. Therefore, the process mod-

els obtained with state-of-the-art algorithms are sub-

performing with our dataset.

In addition, the technical constraints related to car-

rying out the experiments in the digital environment

of the hospital added a constraint in the implemen-

tation of the experiment and led us to use a Python

library as a roundabout way to carry out the study.

This constrained experiment did not allow us to use

the different settings to their full potential. Work on

reﬁning the parameters should be considered.

5.2 Experts’ Reﬂexive Analysis

The expert-based and data-based models are conven-

tionally opposed; on the one hand, one comes from a

reﬂexive analysis process carried out by experts dur-

ing their gain in experience; on the other hand, the

data-based models are based exclusively on data, and

the current approaches to process discovery do not al-

low this reﬂexive analysis.

Process mining approaches using expert knowl-

edge for the learning of model may be a solution to

explore.

5.3 Repeatability and Generalization

If this work focuses on a speciﬁc intervention, the

treatment of an inguinal hernia, in a speciﬁc con-

text, general anesthesia, it can be extended beyond

this medical context. This comparative approach of

expert-based and data-based models can be applied to

more situations. If an expert is capable of creating a

model, and if a process mining algorithm can learn it

from traces, this approach can be employed.

Unfortunately, our speciﬁc result cannot be repro-

duced easily as the real data are unavailable to the

public for privacy reasons. However, this approach

applied to situations where it can be assumed that

traces and the experts’ point of view differ in their

levels of abstraction, should give similar results.

6 CONCLUSIONS

To conclude, we can consider different elements. The

experimental results that we have produced indicate

that an expert-based model can be included in a data-

based model. The differentiation between the two

models appears to be on a level of abstraction.

Thus, the use of these two types of models can be

interesting for health in that they provide two differ-

ent points of view on the same situation. The expert-

based model brings a high-level abstraction point of

view and therefore a rationalization of the process,

whereas the data-based model brings a point of view

HEALTHINF 2025 - 18th International Conference on Health Informatics

748

oriented by the data and therefore by the execution

of the process, which inherently implies all the varia-

tions and anomalies that can occur in the real world.

The models have different characteristics that op-

pose realism and rationalization. A relevant use of

these two types of models could be to evolve the ex-

pert point of view towards more realism and to evolve

the learning models with inclusions of expertise.

6.1 Future Work and Improvement

Those results are quite preliminary and need more

work to consolidate and reﬁne our conclusion.

The critiques of our result must be addressed in

future work to reduce noise in the event log and im-

prove the quality of the models produced.

As stated by (Munoz-Gama et al., 2022) pro-

cess mining for healthcare is confronted with mul-

tiple challenges, and using classical approaches of

process discovery on this kind of data leads to still-

interpretable but sub-performing results. Therefore,

a future intended work is to propose new approaches

and algorithms for process discovery in medical data

and expertise-based process model comparison with

data-based process models and inclusion of expertise

to enhance the process discovery.

Process models learned from traces of differ-

ent periods of times can show changes in healing

practices and therefore contribute to evidence-based

medicine. In the same way, process models combined

with patient clustering can offer a precious decision

aid tool for personalized medicine to choose the best

process for a patient.

Better process model learning algorithm for health

data is the common goal of solving those challenges.

REFERENCES

Augusto, A. et al. (2017). Split miner: Discovering accurate

and simple business process models from event logs.

In 2017 IEEE international conference on data mining

(ICDM), pages 1–10. IEEE.

Berti, A. and van der Aalst, W. M. (2019). Reviving token-

based replay: Increasing speed while improving diag-

nostics. ATAED@ Petri Nets/ACSD, 2371:87–103.

Berti, A., van Zelst, S., and Schuster, D. (2023). Pm4py: A

process mining library for python. Software Impacts,

17:100556.

Blum, F. R. (2015). Metrics in process discovery. Tech. Rep.

Technical Report TR/DCC-2015–6, Computer Science

Dept., University of Chile.

Buijs, J. C. (2014). Flexible evolutionary algorithms for

mining structured process models. PhD thesis, Tech-

nische Universiteit Eindhoven. Research TU/e / Grad-

uation TU/e), Mathematics and Computer Science.

Buijs, J. C., van Dongen, B. F., and van der Aalst, W. M.

(2014). Quality dimensions in process discovery. In-

ternational Journal of Cooperative Information Sys-

tems, 23(01):1440001.

De Medeiros, A. A. et al. (2004). Process mining: extend-

ing the alpha-algorithm to mine short loops. Working

papers, 113.

Gao, M., Liu, W., et al. (2022). Global trends in anesthetic

research over the past decade: A bibliometric analysis.

Annals of Translational Medicine, 10(10).

Guzzo, A., Rullo, A., and Vocaturo, E. (2022). Process

mining applications in the healthcare domain: A com-

prehensive review. Wiley Interdisciplinary Reviews:

Data Mining and Knowledge Discovery, 12(2):e1442.

Journal Ofﬁciel, R. F. (1994). D

ecret n°94-1050 du 5

ecembre 1994.

Katsoulakis, E., Wang, Q., et al. (2024). Digital twins

for health: a scoping review. NPJ Digital Medicine,

7(1):77.

Kaymak, U., Mans, R., Van de Steeg, T., and Dierks, M.

(2012). On process mining in health care. In 2012

IEEE international conference on Systems, Man, and

Cybernetics (SMC), pages 1859–1864. IEEE.

Kononowicz, A. A., Woodham, L. A., et al. (2019). Virtual

patient simulations in health professions education.

Journal of medical Internet research, 21(7):e14676.

Kuo, M.-H. and Chen, Y.-S. (2012). A method to identify

the difference between two process models. J. Com-

put., 7(4):998–1005.

Leemans, S. J. et al. (2014). Discovering block-structured

process models from event logs containing infrequent

behaviour. In Business Process Management Work-

shops: BPM 2013, pages 66–78. Springer.

Munoz-Gama, J. and Carmona, J. (2010). A fresh look at

precision in process conformance. In International

Conference on Business Process Management, pages

211–226. Springer.

Munoz-Gama, J., Martin, N., Fernandez-Llatas, C., et al.

(2022). Process mining for healthcare: Characteristics

and challenges. Journal of Biomedical Informatics,

127:103994.

O’Connor, A. M. et al. (1999). Decision aids for patients

facing health treatment or screening decisions: sys-

tematic review. Bmj, 319(7212):731–734.

Rojas, E., Munoz-Gama, J., et al. (2016). Process mining in

healthcare: A literature review. Journal of biomedical

informatics, 61:224–236.

Van Der Aalst, W. (2016). Data science in action. Springer.

van Der Aalst, W. M., Ter Hofstede, A. H., Kiepuszewski,

B., and Barros, A. P. (2003). Workﬂow patterns. Dis-

tributed and parallel databases, 14:5–51.

Van der Werf, J. M. E. et al. (2008). Process discovery using

integer linear programming. In International confer-

ence on applications and theory of petri nets, pages

368–387. Springer.

Verbeek, H., Buijs, J., Van Dongen, B., and van der Aalst,

W. M. (2010). Prom 6: The process mining toolkit.

Proc. of BPM Demonstration Track, 615:34–39.

Weijters, A. J., van Der Aalst, W. M., and De Medeiros,

A. A. (2006). Process mining with the heuristicsminer

algorithm. Working papers, 166.

Expertise versus Data: Comparison of Expertise Based Process Models to Process-Mining Models of Surgery Under General Anesthesia

749