lookfordiagnosis(1).pdf

(245 KB) Pobierz
Multilingual Assistant for Medical Diagnosing and Drug Prescription
Based on Category Ranking
Fernando Ruiz-Rico
University of Alicante
frr@alu.ua.es
Jose-Luis Vicedo
University of Alicante
vicedo@dlsi.ua.es
Mar´a-Consuelo Rubio-S´ nchez
ı
a
University of Alicante
mcrs7@alu.ua.es
Abstract
This paper presents a real-world applica-
tion for assisting medical diagnosis and
drug prescription, which relies on the
exclusive use of machine learning tech-
niques. We have automatically processed
an extensive biomedical literature to train
a categorization algorithm in order to pro-
vide it with the capability of matching
symptoms to MeSH descriptors. To in-
teract with the classifier, we have devel-
oped a multilingual web interface so that
professionals in medicine can easily get
some help in their decisions about di-
agnoses (lookfordiagnosis.com) and pre-
scriptions (lookfortherapy.com). We also
demonstrate the effectiveness of this ap-
proach with a test set containing several
hundreds of real clinical histories.
1 Introduction
Text categorization consists of automatically as-
signing documents to pre-defined classes. It has
been extensively applied to many fields and in par-
ticular, some efforts have been focused on MED-
LINE abstracts classification (Ibushi and Tsujii,
1999). However, as far as we are concerned, it
has never been used to assist multilingual medical
diagnosing and drug prescription by using the tex-
tual information provided by biomedical literature
together with patient histories.
Every year, thousands of documents are added
to the
National Library of Medicine
and the
Na-
c 2008.
Licensed under the
Creative Commons
Attribution-Noncommercial-Share Alike 3.0 Unported
li-
cense (http://creativecommons.org/licenses/by-nc-sa/3.0/).
Some rights reserved.
tional Institutes of Health
databases
1
. Most of
them have been manually indexed by assigning
each document to one or several entries in a con-
trolled vocabulary called MeSH
2
(Medical Subject
Headings). The MeSH tree is a hierarchical struc-
ture of medical terms which are used to define the
main subjects that a medical article or report is
about. Due to the wide use of this terminology, we
can find translations into several languages such as
Portuguese and Spanish (i.e. DeCS
3
- Health Sci-
ence Descriptors). This paper focuses on both the
diseases sub-tree (from C01 to C23) and drugs sub-
tree (from D01 to D20). The first one defines on its
own more than 4,000 pathological states, and also
offers the chance to search for documented case re-
ports related to each of them. The drugs sub-tree
provides the capability of arranging around 8,000
active principles, which can be directly matched to
commercial drugs.
Our proposal tries to estimate a ranked list of di-
agnoses and possible prescriptions from a patient
history. To tackle this problem, we have selected
an existing categorization algorithm, and we have
trained it using the textual information provided
by lots of previously reported cases and labora-
tory findings. This way, a detailed symptomatic
description is sufficient to obtain a list of possible
diseases and prescriptions, along with an estima-
tion of probabilities and bibliography.
We have not used binary decisions from binary
categorization methods, since they might leave
some interesting MeSH entries out, which should
probably be taken into consideration. Instead, we
have chosen a category ranking algorithm to obtain
an ordered list of all possible diagnoses and pre-
1
http://www.pubmed.gov
2
http://www.nlm.nih.gov/mesh
3
http://decs.bvs.br/I/homepagei.htm
169
Coling 2008: Companion volume – Posters and Demonstrations,
pages 169–172
Manchester, August 2008
scriptions so that the user can finally decide which
of them better suits the clinical history.
In this paper, first of all, we will explain the way
we have developed our experiments, including a
full description of the sources and methods used to
get both training and test data. Secondly, we will
provide an example of a patient history and both
the expected and provided diagnoses. We will also
show the suggested drugs recommended by the al-
gorithm for a common disease. And we will finish
by showing and commenting several evaluation re-
sults on.
that lets us perform updates easily with no sub-
stantial performance degradation after increment-
ing the number of categories or training samples.
The restrictive complexity of other classifiers such
as SVM could derivate to an intractable problem,
as stated by (Ruch, 2005).
To evaluate how worth our suggestion is, we
have measured accuracy through three common
ranking performance measures (Ruiz-Rico et al.,
2006): Precision at recall = 0 (P
r=0
), mean aver-
age precision (AvgP) and Precision/Recall break
even point (BEP). Sometimes, only one diagno-
sis is valid for a particular patient. In these cases,
P
r=0
let us quantify the mistaken answers, since it
indicates the proportion of correct topics given at
the top ranked position. To know about the qual-
ity of the full ranking list, we use the AvgP, since
it goes down the arranged list averaging precision
until all possible answers are covered. BEP is the
value where precision equals recall, that is, when
we consider the maximum number of relevant top-
ics as a threshold. To follow the same procedure as
(Joachims, 1998), the performance evaluation has
been computed over the top diseases level.
2.2 Drug Prescription
Multilingual drug prescription can be achieved
through the international active principles, which
are the constituents of drugs on which the charac-
teristic therapeutic action of the substance largely
depends. The appropriate nomenclature for the ac-
tive principles can be found translated to several
languages from MeSH, and can lead to the final
commercial medicaments in most of the countries
around the world.
To train the algorithm for this new purpose, we
have launched the following query to the
PubMed
database:
(“Plant Families and Groups”[majr] OR “Inorganic
Chemicals”[majr] OR “Organic Chemicals”[majr] OR
“Heterocyclic Compounds”[majr] OR “Polycyclic Com-
pounds”[majr] OR “Macromolecular Substances”[majr]
OR “Hormones, Hormone Substitutes, and Hormone An-
tagonists”[majr] OR “Enzymes and Coenzymes”[majr] OR
“Carbohydrates” OR “Lipids”[majr] OR “Amino Acids, Pep-
tides, and Proteins”[majr] OR “Nucleic Acids, Nucleotides,
and Nucleosides”[majr] OR “Complex Mixtures”[majr])
AND “therapeutic use”[sh] NOT (“adverse effects”[sh] OR
“contraindications”[sh] OR “poisoning”[sh] OR “radiation
effects”[sh] OR “toxicity”[sh])
2 Procedures
2.1
Medical Diagnosis
We have extracted the training data from the
PubMed
database
1
by selecting every case re-
ports on diseases written in English including ab-
stract and related to human beings. These docu-
ments were extracted by using the “diseases cat-
egory[MAJR]” query, where [MAJR] stands for
“MeSH Major Topic”, asking the system for re-
trieving only documents whose subject is mainly a
disease. The query provided us with 483,726 doc-
uments
4
leading us to 4,024 classes with at least
one training sample each.
With respect to the test set, we have used 400
medical histories from the School of Medicine
of the University of Pittsburgh (Department of
Pathology
5
). Although, so far the web page con-
tains more than 500 histories
4
, not all of them are
suitable for our purposes. There are some which
do not provide a concrete diagnosis but only a dis-
cussion about the case, and some others do not
have a direct matching to the MeSH tree. We
have used from each document both the title and
all the clinical history, including radiological find-
ings, gross and microscopic descriptions, etc. To
get the expected output, we extracted the top level
MeSH diseases categories corresponding to the di-
agnoses given on the titles of the “final diagnosis”
files (dx.html).
As the ranking algorithm, we have chosen the
Sum of Weights (SOW) approach (Ruiz-Rico et
al., 2006), that is more suitable than the rest for its
efficiency, accuracy and incremental training ca-
pacity. Since medical databases are frequently up-
dated and they also grow continuously, we have
preferred using a fast and unattended approach
4
Data
obtained on February 14th 2007
5
http://path.upmc.edu/cases
After filtering only articles written in English
which have abstract, a total amount of 540,235
4
training documents are left.
170
Figure 1: Example of the first level of a hierarchical diagnosis
2.3
Multilingual Environment
Since all training data is written in English, ev-
ery symptom provided to the algorithm must also
be written in English. For this purpose, an au-
tomatic translation tool is used for input data in
other languages than English. We also promote the
translation by using the MeSH vocabulary, which
has been delivered by human experts, and pro-
vides a reliable correspondence of thousands of
non phrases in many language pairs. Although
the automatic translation method is not accurate
enough for natural speaking, it may be capable
of giving quite good results for independent noun
phrases (Ruiz-Rico et al., 2006), which are the
pieces of information the ranking algorithm uses.
2.4
Availability and Requirements
Figure 2: Output example after manual expansion of high
ranked topics (up) and by selecting the flat diagnosis mode
(down)
rheumatoid arthritis
No special hardware nor software is neces-
sary to interact with the assistant.
Just an
Internet connection and a standard browser
are enough to access on-line through the fol-
lowing sites:
www.lookfordiagnosis.com
and
www.lookfortherapy.com.
By using a web interface and by presenting re-
sults in text format, we allow users to access from
many types of portable devices (laptops, PDA’s,
etc.). Moreover, they will always have available
the latest version, with no need of installing spe-
cific applications nor software updates.
3 A Couple of Examples
3.1
Medical Diagnosis
One of the 400 histories included in the test set
looks as follows:
Case 177 – Headaches, Lethargy and a Sel-
lar/Suprasellar Mass
A 16 year old female presented with two months
of progressively worsening headaches, lethargy
and visual disturbances. Her past medical his-
tory included developmental delay, shunted hydro-
cephalus, and tethered cord release ...
The final diagnosis expected for this clinical his-
tory is: “Rathke’s Cleft Cyst”, which is a syn-
Figure 3: Example of the drug prescription suggestions for
rheumatoid arthritis
(up) and the final medicament (down)
found through the
drugs
link provided by the assistant.
171
onym of the preferred term “Central Nervous Sys-
tem Cysts”. Translating this into one or several
of the 23 top MeSH diseases categories we are
led to the following entries: “Neoplasms”, “Ner-
vous System Diseases” and “Congenital, Heredi-
tary, and Neonatal Diseases and Abnormalities”.
In hierarchical mode, our approach provides au-
tomatically a first categorization level with ex-
panding possibilities as shown in figure 1. We pro-
vide navigation capabilities to allow the user to go
down the tree by selecting different branches, de-
pending on the given probabilities and his/her own
criteria. Moreover, a flat diagnosis mode can be
activated to directly obtain a ranked list of all dis-
eases, as shown on the lower part of figure 2.
After an individual evaluation of this case, we
have obtained the following values:
P
r=0
=
1,
AvgP
=
0.92, and
BEP
=
0.67, since the right top-
ics in figure 1 are given at positions 1, 2 and 4.
3.2
Drug Prescription
Table 1: Averaged performance for both text categorization
and diagnosis
Corpus
OHSUMED
Case reports and
patient histories
Algor.
SVM
SOW
SOW
P
r=0
-
-
0.69
AvgP
-
-
0.73
BEP
0.66
0.71
0.62
is in charge of many sites containing drug com-
pendiums (vademecum.es, vidal.fr, cddata.co.uk,
etc.). We have already performed preliminary tests
by using the symptoms and diseases in the MeSH
tree as the input data, and an arranged list of active
principles as the output data. We have reached an
AvgP
around 0.9.
5 Conclusions and Further Work
We believe that category ranking algorithms may
help in multilingual medical diagnosing and drug
prescription from clinical histories. Although the
output of the categorization process should not be
directly taken as a medical advice, the accuracy
achieved could be good enough to assist human ex-
perts. However, due to the large amount of new ar-
ticles continuously added to biomedical literature,
it becomes quite difficult for a practitioner to keep
up to date. Further works are focused on providing
bibliographic references for each suggestion of the
classifier. We pretend to select from the PubMed
database those entries most related to the patholog-
ical states entered by the user.
As an example for drug prescription, figure 3
shows the suggestions that the ranking algorithm
provides for
rheumatoid arthritis,
where the user
obtains a ranked list of active principles. Fi-
nally, we reach the name of one of the possible
medicaments containing the selected active princi-
ple, along with particular recommendations from
pharmacists (secondary effects, etc).
4 Results
Last row in table 1 shows the performance mea-
sures calculated for each medical history and its
diagnosis, averaged afterwards across all the 400
decisions. P
r=0
indicates that we get 69% of the
histories correctly diagnosed with the top ranked
MeSH entry. AvgP value means that the rest of the
list also contains quite valid topics, since it reaches
a value of 73%.
First row in table 1 provides a comparison be-
tween SVM (Joachims, 1998) and sum of weights
(Ruiz-Rico et al., 2006) algorithms using the well
known OHSUMED evaluation benchmark. Even
using a training and test set containing different
document types, BEP indicates that the perfor-
mance is not far away from that achieved in text
classification tasks, meaning that category ranking
can also be effectively applied to our scenario.
Regarding drug prescription tests, we are still
working under the evaluation process, colaborat-
ing with companies such as
CMPMedica,
which
References
Ibushi, Katsutoshi, Collier-Nigel and Jun’ichi Tsujii.
1999. Classification of medline abstracts.
Genome
Informatics, volume 10,
pages 290–291.
Joachims, Thorsten. 1998. Text categorization with
support vector machines: learning with many rel-
evant features. In
Proceedings of ECML-98, 10th
European Conference on Machine Learning,
pages
137–142.
Ruch, Patrick. 2005. Automatic assignment of
biomedical categories: toward a generic approach.
Bioinformatics, volume 22 no. 6 2006,
pages 658–
664.
Ruiz-Rico, Fernando, Jose Luis Vicedo, and Mar´a-
ı
Consuelo Rubio-S´ nchez. 2006. Newpar: an au-
a
tomatic feature selection and weighting schema for
category ranking. In
Proceedings of DocEng-06, 6th
ACM symposium on Document engineering,
pages
128–137.
172

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika:

Zgłoś jeśli naruszono regulamin