AI-driven Approach for Automatic Synthetic Patient Status Corpus Generation

Full text

AI-driven Approach for Automatic Synthetic Patient Status
Corpus Generation
Boris Velichkov
Faculty of Mathematics and
Informatics, Soa University
"St. Kliment Ohridski", Soa, Bulgaria
bobby.velichkov@gmail.com
Kristina Ivanova
Faculty of Mathematics and
Informatics, Soa University
"St. Kliment Ohridski", Soa, Bulgaria
kiivanova735@gmail.com
Valeri Hristov
Faculty of Mathematics and
Informatics, Soa University
"St. Kliment Ohridski", Soa, Bulgaria
valerihristov96@gmail.com
Ivan Borisov
Faculty of Mathematics and
Informatics, Soa University
"St. Kliment Ohridski", Soa, Bulgaria
ivbborisov@gmail.com
Alexander Peychev
Faculty of Mathematics and
Informatics, Soa University
"St. Kliment Ohridski", Soa, Bulgaria
a.peychev@yahoo.com
Ivan Koychev
Faculty of Mathematics and
Informatics, Soa University
"St. Kliment Ohridski", Soa, Bulgaria
koychev@fmi.uni-soa.bg
Svetla Boytcheva
Institute of Information and
Communication Technologies,
Bulgarian Academy of Sciences, Soa,
Bulgaria
svetla.boytcheva@gmail.com
ABSTRACT
Medical data for patients is sensitive personal information and
therefore to be used in the original form is unacceptable. On the
other hand, in order to be able to do various studies and analysis,
we need such data. In many cases, such data even anonymized,
by removing the personal identiers, which are not suitable to be
shared. Therefore we decided to create a corpus of synthetic statuses
of patients that GPs place when performing a general examination.
Each status consists of several sentences, each sentence describing
the condition of an organ, system or part of the patient’s body.
We divided the status into its constituent sentences and then each
sentence was classied based on the organ it refers to. We build a
gold standard of manually classied sentences into list of human
body organs and systems. Then we use it to train a neural network
classier of sentences that reaches almost 99% accuracy. Finally,
from the all classied sentences we generate synthetic statuses,
composed according to statistics in the available real statuses and
medical domain constrains. The proposed approach can be easily
adapted to other languages.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
AIVR2020, October 23–25, 2020, Kumamoto, Japan
©2020 Association for Computing Machinery.
ACM ISBN 978-1-4503-8799-6/20/10. . . $15.00
https://doi.org/10.1145/3439133.3439141
CCS CONCEPTS
•Computing methodologies
;
•Natural language processing
;
•Neural networks
;
•Machine learning
;
•Applied computing
;
•Health informatics;
KEYWORDS
Health Informatics, AI technologies, Big Data, Machine Learning,
Text Generation, Synthetic Data, Corpus creation
ACM Reference Format:
Boris Velichkov, Kristina Ivanova, Valeri Hristov, Ivan Borisov, Alexander
Peychev, Ivan Koychev, and Svetla Boytcheva. 2020. AI-driven Approach
for Automatic Synthetic Patient Status Corpus Generation. In 2020 4th Inter-
national Conference on Articial Intelligence and Virtual Reality (AIVR2020),
October 23–25, 2020, Kumamoto, Japan. ACM, New York, NY, USA, 7 pages.
https://doi.org/10.1145/3439133.3439141
1 MOTIVATION
In the age of digitalization and e-society, sharing data that can
contribute to improving our lives is becoming increasingly valuable.
Unfortunately, there is a lot of sensitive data or personal data that
is not possible to be shared. The main obstacle in sharing such
resources relates to ethical and legal restrictions such as GDPR
legislation and privacy concerns of the users. A typical example
of such data is healthcare data. In the current COVID-19 outbreak
there was a huge demand on machine learning and text analytics
tools that can be used to process clinical narratives, more over for
non-English language. The availability of big data is crucial for
training machine learning methods in healthcare.
We propose an innovative method for automatic generation of
synthetic clinical notes, based on a combination of rule-based, statis-
tical and machine learning methods. This allows generation of big
open data repositories of clinical narratives, which is breakthrough
29

AIVR2020, October 23–25, 2020, Kumamoto, Japan Boris Velichkov et al.
in corpora creation in the healthcare domain. The generated data
are valid and more over the synthetic corpora can contain huge va-
riety of patient status data, which can be quite useful for modelling
dierent phenotype and to be used in the research.
The current paper is organized as follows: Section 2 overviews
approaches for synthetic data generation in Healthcare; Section 3
describes the real data and their preprocessing, on the top of which
are generated the synthetic data; Section 4 presents in detail the
proposed AI-based approach for automatic generation of synthetic
patient status data; and Section 5 concludes and sketches some
directions for further work.
2 RELATED WORK
Various methods for synthetic data generation have been developed.
The most important and desired feature of such repositories is the
validity of data and minimization of the information loss.
The problem with the availability and access of the research data
in the healthcare sector is addressed by Shamsuddin, Rittika, et al
[
10
]. They also oer a Virtual Patient Model (VPM), a process that
combines the use of an optimization algorithm, statistical analysis,
and machine learning techniques to generate synthetic time series
data and report their eectiveness in predictive models. The pro-
posed model is validated by implementing a genetic algorithm that
captures important characteristics of real-time patient data (origi-
nal). As a result of applying constraints in the generative process,
the algorithm outputs a best t synthetic candidate solution of the
original time series data. Experimental results using statistical tests
to verify both the synthetic and original datasets show that the
synthetic dataset retains features from the original dataset. The au-
thors of the paper claim that by using machine learning prediction
models in combination with synthetic data as a training dataset,
they manage to achieve promising results in the class recognition
compared to the training with the original data. The rst simulation
framework which allows the generation of realistic 3D synthetic
cardiac US and MR (both cine and tagging) image sequences from
the same virtual patient is proposed by Zhou, Yitian, et al [
14
]. The
approach for VPM repositories generation is explored in [
6
] as well.
There are several attempts for generation of realistic synthetic
Electronic Health Records (EHRs) [
7
], [
5
], [
3
]. Walonoski, Jason,
et al [
12
] developed Synthea, an open-source software package
that simulates the lifespans of synthetic patients, modeling the
10 most frequent reasons for primary care encounters and the 10
chronic conditions with the highest morbidity in the United States.
It is expected that by engaging a growing community of users, the
synthetic data generated will become increasingly comprehensive,
detailed, and realistic over time. Also the authors mention that syn-
thetic patients can be simulated with models of disease progression
and corresponding standards of care to produce risk-free realistic
synthetic health care records at scale. Zand, Ramin, et al [
13
] de-
scribe the importance of building synthetic patient populations and
customized, predictive models using heterogeneous datasets, such
as electronic health records (EHRs). This is a pioneering work on
the use of in silico clinical trials to accelerate the development of
new drugs.
Another important aspect that was addressed by Chen et al [
4
]
is the validity of synthetic clinical data and formal methods for
assessment.
Baytas, Inci M., et al [
1
] propose a novel LSTM unit called Time-
Aware LSTM (T-LSTM) to handle irregular time intervals in longitu-
dinal patient records. In their paper they claim that experiments on
synthetic and real world datasets show that the proposed T-LSTM
architecture captures the underlying structures in the sequences
with time irregularities. Begoli et al [
2
] propose a method for syn-
thetic generation of mental health characteristics reports, based
on deep learning generative methods applied over literature and
public statistical models. Several applications have already been
developed using synthetic data corpora, for example development
of family history annotation guidelines presented by Rama et al
[9].
We focus on only a small part of the outpatient record generation,
namely the patient status. The proposed approach is based on a
natural language generation method using statistical and rule-based
templates.
3 DATA
Clinical patient data cannot be used for research, due to ethical
and legal restrictions, because they contain sensitive personal in-
formation. Even anonymizing patient data does not guarantee their
security. There are cases in which anonymized data is de-identied.
The data preparation task consists of three subtasks (see Figure
1): (1) repository creation from real data; (2) patient status template
modelling and (3) preparation of small training corpus for classi-
cation of sentences according to anatomical organs and systems
they describe.
3.1 Repository Creation
For this research we used a repository of real anonymized outpatient
records (ORs) described by General Practitioners (GPs), from the
Bulgarian National Diabetes Register [
11
]. In order to prevent any
data breach the ORs were splitted on sentences and was generated
repository of individual sentences used in real ORs. In this study
we focused on sentences that are used to describe patient status
only.
Each patient status description consists of several sentences,
each sentence describing the condition of an anatomical organ,
system or part of the patient’s body. A patient status description
can range from 4-5 to 12-13 sentences, in which some anatomical
organs almost always present in ORs, whereas other organs are
less frequently described.
The rst subtask is the repository generation from the real
anonymized ORs. One of the most complex and challenging tasks
is to process the real ORs - namely the sentence splitting of the
patient status data. The diculty in this case comes from the fact
that in the sentences themselves there are used many abbreviations.
The patient status presentation is more in telegraphic style, rather
than using complete sentences, which implies a lack of punctuation.
This makes almost impossible to determine where the description
of the condition of one anatomical organ ends and where the next
one begins.
30

AI-driven Approach for Automatic Synthetic Patient Status Corpus Generation AIVR2020, October 23–25, 2020, Kumamoto, Japan
Figure 1: Data preparation: (1) repository creation from real data; (2) patient status template modelling and (3) preparation of
small training corpus for classication of sentences according to anatomical organs and systems they describe.
Figure 2: Synthetic patient status data template.
Some additional specic sentence splitting rules were
introduced. For instance, in the sentence "Корем-б.о.
Черендробислезка-б.о. Succ. ren: –, Периференилимфнивъзли-
несепалпиратувеличени" (“Abdomen-w.c. Liver and spleen-w.c.
succ.ren:—, Peripheral lymph nodes-Not palpable".), the phrase "б.о." (
“without complications".) is a signaling for the end of an anatomical
organ description and can be used for sentence splitting rule.
After splitting the patient status by sentences from 10,000 pa-
tients ORs we generated a repository of about 1 million individual
sentences. From the generated repository were removed duplicated
sentences, by preserving information about the frequency. The
result repository contain 100,000 unique sentences.
3.2 Synthetic Patient Status Data Templates
Modelling
The next subtask is templates modelling for the synthetic patient
status data. The real anonymized ORs were used for desk research
only in modelling the templates. Based on the frequency analysis
were modelled dierent patient status description patterns (see
Figure 2).
3.3 Training Corpus with Categorized Patient
Status Descriptions
One of the main subtasks in synthetic data generation is to train text-
based classication method for individual sentences classication
to categories representing dierent anatomical organ, systems and
parts of human body.
Since there is lack of language resources for medical domain in
Bulgarian language, the initial step was to create training corpus.
Some representative subsets of dierent sentences were selected
from the generated repository. The data cleaning was applied and
then the result set of clear sentences was labeled manually. Web
based data curation tool was developed to assist this activity.
For this initial subtask were used the word frequency in the
repository. The most commonly described anatomical organs are
such as heart, lungs, liver, limbs, skin, eyes, nose, throat, nervous
system, etc. Another important feature is the use of many Latin
terms in the Bulgarian clinical narratives. The Bulgarian medical
terminology is a mixture between medical terms in Bulgarian, medi-
cal terms in Latin and Latin medical terms transliteration in Cyrillic.
Adding to this the abbreviations for all three cases, categorizing an
anatomical organ becomes a challenging task.
The real data contain a lot of noise - typos, concatenated words,
homoglyphs, etc. Some preliminary cleaning of the repository was
performed. A balanced training set for all anatomical organs and
systems was selected, where only sentences representing infor-
mation for single anatomical organ/system were selected in the
training corpus in order to reduce the ambiguity.
The created training corpus contains 17,000 manually classied
sentences to 25 dierent organs and systems. Some examples of
sentences that refer to only one organ can be seen in Table 1. In
this case the examples are for heart and lung.
31

AIVR2020, October 23–25, 2020, Kumamoto, Japan Boris Velichkov et al.
Table 1: An example of sentences that refer to only one organ (in this case the organs are heart and lung)
Anatomical Organ/
System
Sentence
Lung БЯЛДРОБ–чистовезикуларнодишане,безхрипованаходка(LUNGS - pure vesicular respiration, no wheezing)
Lung Пулмо:отслабеновезикуларнодишане(Pulmo: weakened vesicular respiration)
Lung Pulmo–вез.дишане,судълженоиздишване,ед.сухисв.хрипове(Pulmo - weighted breathing, prolonged
exhalation, single dry wheezing)
Heart СЪРЦЕ–ритмичнасърдечнадейност,глухитонове(HEART - rhythmic heart activity, deaf tones)
Heart СЪРЦЕ–Б.О(HEART– without complaints)
Heart КОР–РСД(COR– rhythmic heart activity)
Figure 3: Synthetic patient status data generation.
4 METHODS
The synthetic corpus generation is based on statistical models, that
use the templates created over real patient status data, and text-
based classication methods applied on the preprocessed repository
of sentences.
4.1 Sentence Preprocessing
The division of the statuses into separate sentences appears to be
a quite complex task. The diculty in this case comes from the
fact that in the sentences themselves there are many abbreviations
of words, dierent abbreviations of organs and systems are used,
sometimes no punctuation marks and it becomes almost impossible
to determine where the description of the condition of an organ ends
and where the next one begins. This subtask contains the classical
preprocessing pipeline steps like tokenization, text normalization,
stop words removal, stemming, etc.
Tokenization:
As a token we consider words only. Thus all nu-
merical values are skipped, as well as white spaces, and punctuation.
The omitted data do not bring any additional value to the task of
anatomic organ/system classication.
Text normalization:
Since the outpatient record is created dur-
ing the patient’s examination, general practitioners usually use
plenty of abbreviations and acronyms due to the time shortage.
We have compiled a list of abbreviation and their corresponding
extended meaning. This is an essential step of the sentence catego-
rization because very often the abbreviation of a system of organs
consists of the organ to which we want to categorize the sentence.
For example: "ССС" (cardio vascular system)
•"Сърдечносъдовасистема" (cardio vascular system)
•
“Сърдечно” (cardiac) has the same root as the word (heart)
“сърце” which can help in the comparison step.
Stop words removal:
This step brought some benets on the
performance. As a resource was used available list of stop word for
Bulgarian language provided by Bultreebank
1
. As it left fewer but
only meaningful tokens, it decreased the size of the dataset which
resulted in speeding up the classication process and increasing
the accuracy.
Examples:
•
(abdomen without features) Коремб.о.
→
Корембо
→
Корем
•
Sentences describing patient’s general and mental condition:
These sentences are ignored for categorization, later on, on
synthetic status generation they are added with the respec-
tive probability of encounters.
Stemmig:
Each word of the sentence is stemmed by inectional
stemmer for Bulgarian BulStem [
8
]. All the available stemming
rules are imported but due to the specicity and terminology of
1http://bultreebank.org/bg/resources/
32

AI-driven Approach for Automatic Synthetic Patient Status Corpus Generation AIVR2020, October 23–25, 2020, Kumamoto, Japan
Table 2: An example of sentences from some of the clusters
Cluster Sentences
General Status Общосъстояние-Доброобщосъстояние,Безпромянавобщотосъстояние(General condition - Good general
condition, No change in general condition)
Body Regions Регионинатялото-ГЛАВАиШИЯ-Безособености,КОРЕМ-Мек,палпаторнонеболезнен,Крайници-
безотоци(Body Regions - HEAD and NECK- without complications, Abdomen - Soft, palpably painless, Limbs - no
swelling)
Respiratory
system
Дихателнасистема-БЯЛДРОБ-чистовезикуларнодишане,Пулмо-везикуларнодишанебезхрипове,Гърло:
Нормално,сливици-безаномалии(Respiratory system - LUNGS-pure vesicular respiration, Pulmonary-vesicular
respiration without wheezing, Throat: Normal, tonsils-without abnormalities)
Cardio Vascular
system
Сърдечносъдовасистема-СЪРЦЕ-ритмичнасърдечнадейност,Ритмичнасърдечнадейност-яснитонове,
безприбавенишумове, Cor - р.с.д., fr 72 умин. , RR 130 80 (Cardiovascular system - HEART-rhythmic heartbeat,
Rhythmic heartbeat - clear tones, without added noise, Cor - rhythmic heartbeat, fr 72 HR , RR 130 80)
Digestive
system
Храносмилателнасистема-ЧЕР.ДРОБиСЛЕЗКА-Неувеличени,ЕЗИК-Влаженнеобложен,
Черендробислезканесепалпират(Digestive system - LIVER AND SPLEEN-Not enlarged, TONGUE-Moist uncoated,
Liver and spleen are not palpable)
Haemic and
Immune
Systems
ХемичнаиИмуннасистема-ПЕРИФЕРНИЛИМФНИВЪЗЛИ-Неувеличени,Далак:Несепалпира,Слезка-
неувеличена(Haemic and Immune System - PERIPHERAL LYMPH -Not enlarged, Spleen: Not palpable, Spleen –
not enlarged)
the medical terminology and the prevalence of Latin terms, the
stemmer’s rules are not enough, and a small amount of the words
is actually stemmed. Moreover, in some cases the word’s length
goes nearly two or three characters which results in one more step
in the preprocessing, because words of this type are not eligible for
comparison.
Examples: (afebrile) афебрилен→афебрилен
4.2 Classication of Sentences into
Human-body Organs and Systems
4.2.1 Lexicon Based Classification. The available data about organs
and systems is held in two main resources:
•GPs lexicon
with an anatomical organs/system, which are
most common in general practitioner’s examinations in our
resource bank.
•Human body structure dictionary
with most of the or-
gans in the human body, according to the medical terminol-
ogy standard classications.
This allows searching also for categories descriptions that are
out of the vocabulary of our ORs repository.
Searching in GPs lexicon:
The searching is happening using
TF, (term frequency) as this data is manually extracted, the results
that TF is giving are sucient. Very often the general practitioners
are describing organs which are semantically connected as liver and
spleen. Because of that we have a list with most of the described
pairs of anatomical organs in the ORs. After choosing the most
appropriate category candidate then in a list of anatomical organ
pairs is searched. If a result is found, then the sentence is annotated
with the pair match else the initial organ becomes the sentence’s
category.
4.2.2 Searching in The Human Body Structure Dictionary. If the
searching in GPs lexicon does not return a result, then in the list of
all organs is searched.
Figure 4: LSTM training over 17 thousand sentences divided
into 25 classes(organs), in 5 epochs. The achieved accuracy
is about 98.9%.
In order to nd an organ there a slight customization of co-
sine metric is done - if a potential category is an n-gram and it’s
contained in the sentence then a maximal score is given, else the
standard implementation is applied. The category with the highest
score is checked in the pairs list and then used for nal category.
The ignored sentences in the preprocessing step about the mental
and general state of the patient are added with the same frequency
of encounter and annotated with default category "General Status",
so later on they can be added in the status generation.
4.2.3 LSTM Multi-Class Classification. From the gold standard of
manually classied sentences, we decided to train a neural network
to be used in the further classication of sentences and to compare
the results obtained. For this purpose, we used the LSTM model of
a recurrent neural network for multiple classication. For the real-
ization of the recurrent neural network we used Python and Keras,
using the following parameters - activation function
=
"softmax" and
optimizer
=
"Adam". Because it is a multi-class classication task,
33

AIVR2020, October 23–25, 2020, Kumamoto, Japan Boris Velichkov et al.
Table 3: Table with the specic number of sentences, unsuit-
able for classication by organ category and description of
the reason. Below they are taken out of the gold standard
and the categorization algorithm is rated on this set as well
Attribute Count Consideration
blood pressure 1722 Organ description/status
weight 163 Organ description/status
height 104 Organ description/status
temperature 172 Organ description/status
lymph nodes 193 Organ pair
pulse 355 Organ description/status
head 271 Organ pair
Table 4: Table with classication results of the manually an-
notated list of sentences to only comparable ones obtained
by taking in account the described above considerations
# All sentences Comparable
Sentences
Number of sentences
12404 9015
Correct classied
(count)
6045 6045
Incorrect classied
(count)
6350 2970
Correct classied (%)
48.73% 67.05%
Incorrect classied
(%)
51.26% 32.95%
categorical crossentropy is used as the loss function. By submitting
the gold standard of about 17 thousand sentences divided into 25
classes (organs), in 5 epochs of training we achieved about 98.9%
accuracy (see Figure 4).
4.2.4 Algorithm Evaluation. The categorization algorithm is evalu-
ated against 12 404 manually annotated sentences.
4.2.5 Considerations. In the synthetic status generation and data
preparation steps, some categories are divided into smaller pieces
and follow the exemplary hierarchy: Cardiovascular system
→
heart →blood pressure, pulse
As pulse and blood pressure are not organs but a description of
such, this leads to false negative results in the comparison. It’s the
same for sentences which have the description of two organs at
the same time like lymph nodes and thyroid gland. Also, the organ
dataset is referring to head and neck as one organ, but the manual
annotation is referring to them as two - separate ones.
4.2.6 Evaluation. Dierences between the manually annotated and
the algorithm result are present in a total of 6359 sentences, which
are exactly 51.27%.
If we have the considerations in account then the incorrect clas-
sied sentences are 2970 in total, for more details see Table 3 and
Table 4. As the analysis of the false negative categorization is done
manually it is possible for some cases to be missing.
4.3 Synthetic Data Generation
After successful classication of sentences to anatomical organs,
pair of organs and systems, comes the generation of the synthetic
patient status corpora (see Figure 3). Since sentences are categorized
to organs, pair of organs or systems we iterate and split them in
clusters of sentences each cluster representing higher level human
body systems which are the constituents of an ambulatory status.
As mentioned in the data processing section we identied a not so
large number of organs and systems that present in the description
of the statuses from the whole corpus, created a list of them and
identied the human body systems that each organ belongs to. The
reorganization of sentences into dierent clusters was realized us-
ing already categorized sentences to organs and the map of organ
to system. Even though we have relatively big list of human body
systems, only 11 of them will take part in the templates for genera-
tion as the other have zero sentences categorized to them. On the
one hand the reorganization of sentences into clusters will enable
us to execute frequency analysis of the sentences in each cluster, on
the other we follow Stratied Random Sampling approach which
requires to group sentences and skip overlapping. Some examples
of sentences from some of the clusters are shown in Table 2.
Since there are guidelines that general practitioners follow, but
they are not strict giving some freedom to the doctor, we decided to
build a standard for our synthetic statuses. A status will consist of
no more than 11 sentences, each referring to unique human body
system, ordered in a predened way generally followed by doc-
tors. The ordered categories are: “General status”, “Body regions”,
“Respiratory System”, “Cardiovascular System”, “Digestive System”,
“Urogenital System”, “Endocrine System”, “Nervous System”, “Sense
Organs”, “Musculoskeletal System”, “Haemic and Immune Systems”.
Some statuses can consist of the maximum number of sentences
which in our case is 11, but most of the statuses will consist of 7,
8 sentences or even less, for this purpose systems are split into
obligatory and optional sections and each optional section has as-
signed probability, based on how often it is included by the doctors
in the original statuses. For example, the rst section of a status is
“General status” which is obligatory and always present, as opposed
to other optional categories like “Sense Organs” which can be found
only in 10percent of the generated synthetic statuses. The number
of sentences, respectively the number of human body systems in
a status is dynamic, it always includes the obligatory systems :
“General status”, “Respiratory System”, “Cardiovascular System”,
“Digestive System”, the rest of them are optional sections and the
number of optional sections included in a status is generated at
random. In Table 5 are shown some examples of synthetic statuses,
one that includes all sections and one that includes only a few.
5 CONCLUSION AND FURTHER WORK
In this paper we proposed a method for generating a corpus of syn-
thetic statuses of patients that GPs place when performing a general
examination. The approach uses a corpus of sentences taken form
real statuses, where each sentence describing the condition of an
organ, system or part of the patient’s body. We divided the status
into its constituent sentences and then each sentence was classied
based on the organ it refers to. We develop an approach for division
of the statuses into separate sentences, which appears to be to be a
34

AI-driven Approach for Automatic Synthetic Patient Status Corpus Generation AIVR2020, October 23–25, 2020, Kumamoto, Japan
Table 5: Examples of synthetically generated patient statuses, one that includes all sections and some that includes optional
sections
# Status with obligatory with 2 optional sections and with all optional sections
1Вдоброобщосъстояние.ГЪРЛО-Безособености.СЪРЦЕ-ритмичнасърд.дейност.ЕЗИК-Влаженнеобложен. (In good overall
condition. THROAT-No features. HEART-rhythmic heart activity. LANGUAGE-Wet uncoated.)
2Доброобщосъстояние.БЯЛДРОБ-Лекоотслабеновезикуларнодишане. Cor - р.с.д,.
Черендробислезканесепалпиратувеличени.ОДА-б.о.ПЕРИФЕРНИЛИМФНИВЪЗЛИ-Неувеличени.
(In good overall condition. LUNGS-Slightly weakened vesicular respiration. Cor - rhythmic heartbeat, Liver and spleen are not
palpated enlarged. Musculoskeletal System - without complaints PERIPHERAL LYMPH -Not enlarged. )
3Тегло72кг.КОРЕМ-Мекпалпаторнонеболезнен.БЯЛДРОБ-чистовезикуларнодишане.ПУЛС-68. Черендробислезка-
неувеличени.Сук.реналисотр.ЩитовиднажлезанеувеличенаНервнасистема:Безневролог.Зрение-запазено.
ОПОРНО-ДВИГ.АПАРАТ-болкавдветеколеннистави.. ЛигавицибледорозовиПЛВнесепалпиратувеличени.
(Weight 72kg. ABDOMEN - Soft palpably painless. LUNGS - pure vesicular respiration. Heart rate - 68. Liver and spleen - not
enlarged. Succusio renalis negative. Thyroid gland not enlarged Nervous system: Without a neurologist. Vision-preserved.
MUSCULOSKELETAL SYSTEM-pain in both knee joints. Mucous membranes pale pink PERIPHERAL LYMPH are not palpated
enlarged.)
quite challenging task. The approach proposes techniques to deal
with: abbreviations of words; abbreviations of organs and systems;
missing punctuation marks, which makes almost impossible to de-
termine where the description of the condition of an organ ends and
where the next one begins. We build a gold standard of manually
classied sentences into list of human body organs and systems.
Then we use it to train a neural network classier of sentences that
reaches almost 99% accuracy. Finally, from all classied sentences
we generate synthetic statuses, composed according to statistics in
the available real statuses taking in to account medical domain con-
strains. The proposed approach is as possible language independent
by design and can be easily adapted to other languages. As further
works we are planning to extend our corpus and to explore the
applicability of recent language model such us Bert for generation
of synthetic patient statuses.
ACKNOWLEDGMENTS
This research is partially funded by the Bulgarian National Sci-
ence Fund, grant DN-02/4-2016 ’Specialized Data Mining Methods
Based on Semantic Attributes’ (IZIDA). We also acknowledge the
provided access to thee-infrastructure of the Centre for Advanced
Computing and Data Processing, with the nancial support by the
Grant No BG05M2OP001-1.001-0003, nanced by the Science and
Education for Smart Growth Operational Program (2014–2020) and
co-nanced by the European Union through the European struc-
tural and Investment funds.
REFERENCES
[1]
Inci M Baytas, Cao Xiao, Xi Zhang, Fei Wang, Anil K Jain, and Jiayu Zhou. 2017.
Patient subtyping via time-aware LSTM networks. In Proceedings of the 23rd
ACM SIGKDD international conference on knowledge discovery and data mining.
65–74.
[2]
Edmon Begoli, Kris Brown, Sudarshan Srinivas, and Suzanne Tamang. 2018.
SynthNotes: A Generator Framework for High-volume, High-delity Synthetic
Mental Health Notes. In 2018 IEEE International Conference on Big Data (Big
Data). IEEE, 951–958.
[3]
Anna L Buczak, Steven Babin, and Linda Moniz. 2010. Data-driven approach
for creating synthetic electronic medical records. BMC medical informatics and
decision making 10, 1 (2010), 59.
[4]
Junqiao Chen, David Chun, Milesh Patel, Epson Chiang, and Jesse James. 2019.
The validity of synthetic clinical data: a validation study of a leading synthetic data
generator (Synthea) using clinical quality measures. BMC medical informaticsand
decision making 19, 1 (2019), 44.
[5]
Kudakwashe Dube and Thomas Gallagher. 2013. Approach and method for gen-
erating realistic synthetic electronic healthcare records for secondary use. In
International Symposium on Foundations of Health Informatics Engineering and
Systems. Springer, 69–86.
[6]
Uri Kartoun. 2016. A methodology to generate virtual patient repositories. arXiv
preprint arXiv:1608.00570 (2016).
[7]
Scott H Lee. 2018. Natural language generation for electronic health records. NPJ
digital medicine 1, 1 (2018), 1–7.
[8]
Preslav Nakov. 2003. BulStem: Design and evaluation of inectional stemmer
for Bulgarian. In Workshop on Balkan Language Resources and Tools (Balkan
Conference in Informatics).
[9]
Taraka Rama, Pål Brekke, Øystein Nytrø, and Lilja Øvrelid. 2018. Iterative de-
velopment of family history annotation guidelines using a synthetic corpus of
clinical text. In Proceedings of the Ninth International Workshop on Health Text
Mining and Information Analysis. 111–121.
[10]
Rittika Shamsuddin, Barbara M Maweu, Ming Li, and Balakrishnan Prabhakaran.
2018. Virtual patient model: an approach for generating synthetic healthcare time
series data. In 2018 IEEE International Conference on Healthcare Informatics
(ICHI). IEEE, 208–218.
[11]
Dimitar Tcharaktchiev, Zhivko Angelov, Svetla Boytcheva, and Galia Angelova.
2018. Automatic generation of a national diabetes register from outpatient records.
Mathematical Modeling 2, 4 (2018), 163–166.
[12]
Jason Walonoski, Mark Kramer, Joseph Nichols, Andre Quina, Chris Moesel,
Dylan Hall, Carlton Duett, Kudakwashe Dube, Thomas Gallagher, and Scott
McLachlan. 2018. Synthea: An approach, method, and software mechanism for
generating synthetic patients and the synthetic electronic health care record.
Journal of the American Medical Informatics Association 25, 3 (2018), 230–238.
[13]
Ramin Zand, Vida Abedi, Raquel Hontecillas, Pinyi Lu, Nariman Noorbakhsh-
Sabet, Meghna Verma, Andrew Leber, Nuria Tubau-Juni, and Josep Bassaganya-
Riera. 2018. Development of synthetic patient populations and in silico clinical
trials. In Accelerated Path to Cures. Springer, 57–77.
[14]
Yitian Zhou, Sophie Giard-Roisin, Mathieu De Craene, Sorina Camarasu-Pop,
Jan D’Hooge, Martino Alessandrini, Denis Friboulet, Maxime Sermesant, and
Olivier Bernard. 2017. A framework for the generation of realistic synthetic
cardiac ultrasound and magnetic resonance imaging sequences from the same
virtual patients. IEEE transactions on medical imaging 37, 3 (2017), 741–754.
35