Assessment of Dendritic Cell Therapy Effectiveness Based on the Feature

Extraction from Scientiﬁc Publications

Alexey Yu. Lupatov

, Alexander I. Panov

, Roman E. Suvorov

, Alexander V. Shvets

Konstantin N. Yarygin

and Galina D. Volkova

Orekhovich Institute of Biomedical Chemistry of the Russian Academy of Medical Sciences, Moscow, Russia

Institute for Systems Analysis of the Russian Academy of Sciences, Moscow, Russia

Moscow State University of Technology “Stankin”, Moscow, Russia

Keywords:

Dendritic Cells, Anticancer Vaccine, Cell Therapy, Natural Language Processing, Data Mining, Text Mining,

JSM-method, Genetic Algorithm, AQ-method.

Abstract:

Dendritic cells (DCs) vaccination is a promising way to contend cancer metastases especially in the case of

immunogenic tumors. Unfortunately, it is only rarely possible to achieve a satisfactory clinical outcome in

the majority of patients treated with a particular DC vaccine. Apparently, DC vaccination can be successful

with certain combinations of features of the tumor and patients immune system that are not yet fully revealed.

Difﬁculty in predicting the results of the therapy and high price of preparation of individual vaccines prevent

wider use of DC vaccines in medical practice. Here we propose an approach aimed to uncover correlation

between the effectiveness of speciﬁc DC vaccine types and personal characteristics of patients to increase

efﬁciency of cancer treatment and reduce prices. To accomplish this, we suggest two-step analysis of published

clinical trials results for DCs vaccines: ﬁrst, the information extraction subsystem is trained, and, second, the

extracted data is analyzed using JSM and AQ methodology.

1 INTRODUCTION

Usually cancer patients are treated by radical re-

moval of the tumor, radiotherapy or cytotoxic drugs.

However, after surgical procedure many patients have

a relapse because some cancer cells persist in the

body. Use of radio- and chemotherapy does not al-

ways give the desired result. Efﬁcacy of radiother-

apy is restricted by the type of tumor and its location.

Chemotherapy is often too toxic for the patient. Be-

sides, the resistance of cancer cells to cytotoxic drugs

can signiﬁcantly increase in the course of chemother-

apy. Consequently, despite extended treatment, re-

lapses of the disease, including distant metastases are

quite common. Thus, it is important to further de-

velop methods to prevent tumor relapses.

Immunotherapy plays increasingly important role

in cancer care. There are numerous data conﬁrming

its efﬁciency. Use of cell vaccines based on dendritic

cells (DCs) is one of the most promising methods

of anti-cancer immunotherapy (Shortman and Caux,

1997). DCs are the main type of the antigen pre-

senting cells. They incorporate, process and exhibit

the antigen on their surface membrane in the lym-

phocytes recognizable form. After that the lympho-

cytes become able to attack tumor cells. Monocytes

(Zhou and Tedder, 1996) or hematopoietic stem cells

(Welzen-Coppens et al., 2012) isolated from blood or

bone marrow of the patient are used as starting mate-

rial for DC vaccine preparation. These cells are cul-

tured with cytokines to differentiate into DCs. DCs

undergo maturation induced by loading with tumor

antigens and are administered to the patients to stimu-

late cellular immune response against the tumor. Pro-

teins extracted from surgically removed tumors, syn-

thetic peptides or recombinant proteins can be used

as tumor antigens for loading into DCs (Yannelli and

Wroblewski, 2004). Hybrids obtained by fusion of

DCs and tumor cells (dendritomas) may be used as a

vaccine as well (Wei et al., 2006).

The necessity to isolate and process each patients

own cells in vitro makes DC vaccines quite expensive

and hard to standardize. It also makes it difﬁcult to

analyze the effectiveness of vaccination, because its

efﬁcacy is inﬂuenced by each patients individual fea-

tures. It is thus essential to reveal patients characteris-

tics critical for the success of vaccine application and

to categorize the patients accordingly. Here we sug-

270

Yu. Lupatov A., I. Panov A., E. Suvorov R., V. Shvets A., N. Yarygin K. and D. Volkova G..

Assessment of Dendritic Cell Therapy Effectiveness Based on the Feature Extraction from Scientiﬁc Publications.

DOI: 10.5220/0005248802700276

In Proceedings of the International Conference on Pattern Recognition Applications and Methods (ICPRAM-2015), pages 270-276

ISBN: 978-989-758-077-2

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

gest to solve this problem by extracting numerical and

symbolic information from available sources includ-

ing journal articles, theses, patients hospital records

and clinical trial reports describing the use of DC vac-

cines and its analysis by machine learning methods.

2 RELATED WORK

Most published studies assessing the efﬁcacy of DC-

based vaccination are focused on one type of vaccine

and one tumor form. It is true even for publications

reviewing meta-analysis data.

There is quite an informative publication analyz-

ing clinical trials of DC vaccines giving a systematic

review and meta-analysis of clinical studies of 17 vac-

cines against prostate cancer and 12 vaccines against

renal cancer (906 patients all together) (Draube et al.,

2011). Such factors as the type of DCs used, anti-

gen loading method, administration route, dose, adju-

vants, toxicity and clinical response were evaluated.

Additional parameters considered were patients age,

sex, stage of disease, previous and concomitant treat-

ment. Unfortunately, this paper summarizes only the

results of using DC vaccines to treat two types of uro-

logical neoplasms.

Some papers analyze the dependence of the DC

vaccines effect from the initial selection criteria ap-

plied to patient selection (Figdor et al., 2004; Murthy

et al., 2009). Those criteria can include age, sex, stage

of the cancer process, method of therapy, comorbidi-

ties, biochemical and hematological parameters of the

blood, etc. Unfortunately, some selection criteria for

a particular study look subjective and are not reason-

ably explained by the authors.

There are studies devoted to the information re-

trieval from medical texts (Aggarwal and Zhai, 2012)

dealing with named entities extraction (drugs, dis-

eases, etc.); ﬁnding of the connection between entities

(genes and diseases, protein interactions, etc.); extrac-

tion of the correlation between time and cause. For

text processing methods of syntactic, semantic and

discourse analysis, co-reference resolution (Aggarwal

and Zhai, 2012; Gaizauskas et al., 2003) are used, as

well as regular expressions; contextual rules and tem-

plates compiled manually; classiﬁers (support vector

machines, etc.). Currently, Deep Learning (Collobert

et al., 2011) develops rapidly. Multilayer neural net-

works are the main tool for text analysis.

3 META-ANALYSIS

METHODOLOGY

As pointed out above, it is important to separate pa-

tients into groups according to their characteristics es-

sential for the vaccination success. It is obvious that

the group formation should take into account nosol-

ogy, because different malignant tumors have differ-

ent degree of immunogenicity and different sensitiv-

ity to immunotherapy. The disadvantage of this ap-

proach is lack of data for some tumors. To avoid

loss of data, we propose to use a two-level method

for group forming. At the bottom level the classiﬁca-

tion into groups is based on the type of disease, for

example “colorectal adenocarcinoma”, “prostate can-

cer”, “glioblastoma”. At the top level malignant tu-

mors are divided into ﬁve groups according to their

origin and pathogenesis. Group I – “carcinoma and

melanoma” includes tumors originating from epithe-

lium or skin pigment cells. Group II – “sarcoma”

embraces tumors from solid tissues such as osteosar-

coma, rhabdomyosarcoma, etc. Malignant tumors

originating from neural tissue such as neuroblastoma,

glioblastoma and others constitute group III. Group

IV – “hematological malignancies and solid tumors of

the immune system” is represented by the diseases of

hematopoietic and immune system such as leukemia,

lymphoma, thymoma, etc. Group V includes tumors

whose origin, etiology or pathogenesis differ signiﬁ-

cantly from the others.

Evaluation of the impact of DC vaccine prepara-

tion procedures demonstrated that there is at least one

vaccine characteristic that might signiﬁcantly affect

the results of immunotherapy. This is DC vaccine va-

lency, i. e. the number of antigenic determinants the

vaccine was designed against. DC vaccine valency is

determined by the substance that is loaded as antigen

during vaccine preparation.

For the purposes of this study, we chose the

clinical trials registry (http://www.clinicaltrials.gov)

and information related to biomedicine retrieval

machine http://www.ncbi.nlm.nih.gov/pubmed(Med-

line) as Internet sources of information about the re-

sults of clinical studies on the use of DC vaccines

against cancer. These resources cover most of the data

required worldwide.

We used keywords and word combinations “can-

cer”, “tumor” and “dendritic cells” to retrieve the nec-

essary data. When using Medline the restriction “clin-

ical trial” was inputted. Sources of results of clinical

trials were chosen with the help of text analysis. Data

of randomized and other types of studies, regardless

of their phase, were collected. Only studies with at

least three participants with one of tumor types men-

AssessmentofDendriticCellTherapyEffectivenessBasedontheFeatureExtractionfromScientificPublications

271

tioned above were taken. Studies applying allogeneic

or other types of non-autologous DCs were excluded.

Analysis of the registry of clinical trials revealed

387 registered trials of cancer treatment methods us-

ing DC vaccines. Unfortunately, only 20 of them pro-

vided information about the end results. The Med-

line search system produced 587 relevant sources of

data related to malignancies treatment using DC vac-

cines. Of those 310 papers refered to the treatment

of the ﬁrst group tumors (carcinoma 148, melanoma

162); 17 articles described the treatment of the second

group tumors (sarcomas); 38 articles were devoted to

the treatment of brain tumors (III group); 103 articles

concerned hematopoietic and immune system tumors

(IV group) and two articles were about vaccination of

patients with the V group tumors.

Besides general individual characteristics of the

patient such as gender, age, race, that can always be

found in the records of clinical studies, the charac-

teristics of the tumor process and immunological sta-

tus of the patient are of great interest. Conventional

nomenclatures such as the International clinical clas-

siﬁcation of tumors by TNM (American Joint Com-

mittee on Cancer) allow characterizing the process

of tumor development only within a nosological unit

which it is not appropriate for this study. Therefore,

we devised our own classiﬁcation of tumor progres-

sion stages. The ﬁrst stage is the absence of detectable

tumor site. This stage is seen in patients who under-

went radical surgery and were assigned immunother-

apy using DCs vaccine to prevent relapse. The sec-

ond stage is seen in patients with a primary tumor

which can be combined with a limited numberof local

metastases. The third stage is a generalized neoplastic

process with multiple distant metastases.

We ranked the immune status as “positive” and

“negative”. “Positive” immune status corresponds to

the patient with the evidence of cell immune response.

“Negative” immune status is characterized by nega-

tive immunoregulation such as increased number of

regulatory T-cells or expression of immunosuppres-

sive cytokines.

To evaluate the results of DC vaccine application

it is necessary to formulate criteria of successful treat-

ment. We concluded that the lifetime increase com-

pared to a median of patient survival time without DC

vaccine application, can be a valid criterion. This is

the only possible criterion for the patients with the

ﬁrst-stage tumor (without visible tumor). Moreover,

it allows not only qualitative, but also quantitative es-

timation of treatment efﬁcacy. Unfortunately, the ap-

plication of this criterion implies long-term follow-up

of vaccinated patients. Consequently, the number of

sources providing related information is limited.

We chose positive clinical response as the second

criterion. That is complete or partial regression of tu-

mor, as well as stabilization of the disease.

4 TECHNICAL MEANS OF

META-ANALYSIS

To facilitate the proposed meta-analysis methodology,

we are developing an informational analytical sys-

tem. The system integrates modules that aim to au-

tomate all the essential steps of the proposed method-

ology: information gathering, information extraction

and analysis of cause-and-effect relations.

4.1 Information Gathering

Information gathering is organized as thesauri-aided

meta-search: user types a search query with the help

of domain-speciﬁc vocabularies (e. g. MeSH, ICD

etc.) and the system redirects this query to such public

libraries of scientiﬁc texts as PubMed, Cochrane, etc.

This leads to ﬁlling of a local publication database.

After an expert concludes that the local database

is full enough they can launch information extraction

module. The aim of this module is to convert unstruc-

tured or semi-structured information about patients

and treatments presented in the documents found into

the vector space representation.

The proposed module for information extraction

implements a rather standard text mining pipeline:

ﬁrst, the documents are converted to plain text, then

segmentation and tokenization, morphological analy-

sis, syntax parsing, semantic role labeling, third-party

information extraction algorithms and our informa-

tion extraction methods follow. The engine is im-

plemented using UIMA (Ferrucci and Lally, 2004),

thus each processing step reads source data from and

writes results to a uniﬁed graph-like data structure

called CAS. CAS contains text and annotations. An-

notations are program objects that are labelled spans

of text. Annotations have attributes (properties) of

various types and may refer to other annotations.

One of the peculiarities of our pipeline is that it in-

corporates Tabula library to extract tables from PDF

documents (Nurminen, 2013). It’s important because

many papers contain information about patients writ-

ten as tables. After Tabula extracts tables, the text

in cells is aligned against the full text and the cor-

responding annotations are added to the CAS. Other

incorporated tool is cTAKES (Savova et al., 2010).

To reduce the necessary manual labor amount we

use on-line active machine learning paradigm: the

overall process of analyzing the documents is split

ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods

272

into iterations, on each the system outputs the ex-

tracted information and asks the user to provide some

feedback, i.e. answers to questions like “Is it a posi-

tive example of X?” and/or “Is this a positive/negative

example of X because of Y”. To efﬁciently analyze

expert’s feedback and generate the questions, we in-

dex all the CASes using a graph database. It’s pre-

ferred, but not necessary to use an incremental learn-

ing algorithm. SVM and inductive rule builders sup-

port this (Cauwenberghs and Poggio, 2000).

The problem of information extraction is treated

as a combination of binary classiﬁcation problems:

ﬁrst, a classiﬁer must determine, whether a token rep-

resents a particular characteristics; second, it must

ﬁnd mentions of patient groups (similarly to the ﬁrst

subproblem); last, it must link characteristics to these

groups. Most of existing classiﬁcation algorithms

work with vector-space object representation, thus it’s

important to choose appropriate feature selection pro-

cedure. Currently, we experiment with information

gain, Euclid distance between positive and negative

examples and various graph traversal strategies for

features extraction and selection. Classiﬁers being

tested are VM, HMM, CRF, ProbLog, decision trees,

neural networks (Collobert et al., 2011).

4.2 Feature Extraction

The information extraction step results in a summary

table in which rows correspond to groups of patients

and columns – to their features. Table 1 is an ex-

ample of such a table. In most cases, it contains

columns that identify patient groups (paper identi-

ﬁer/number of the group), describe diseases (via ICD-

10 codes or verbally), present results of various anal-

yses (e.g. tumor antigens), demographic information

(gender, sex), type of vaccine used, loading technique

and measures of outcomes (averagesurvival time with

and without treatment). Usually the summary table

contains a few hundreds of columns. Additionally, to

increase separability of classes, the binarization tech-

nique may be applied to this table. According to bina-

rization, features with K possible values (K > 2) are

replaced with K new features with only two possible

values, indicating presence or absence of the corre-

sponding value of the source feature.

To identify the set of signiﬁcant features, the pre-

viously mentioned table is analyzed using inductive

machine learning methods. To determ the effective-

ness of speciﬁc DC vaccine types we choose JSM

method (Anshakov et al., 1991). JSM method is an

algorithm that allows discovering of cause-and-effect

relations. Although it was extensively used in vari-

ous areas, its applications are limited to only small-

scale problems (only about a few tens of features). To

overcome this issue, we propose to preliminary select

features using a more lightweight method, e.g. AQ

(Wojtusiak et al., 2006). The idea of the proposed

feature pre-ﬁltering is that only those features that AQ

(or other preprocessing method) chooses to build up

the decision rules are forwarded to JSM method.

For this purpose, the table is represented as a

matrix of feature values A = {a

}. In this matrix

columns correspond features and rows - to theirs val-

ues:

→ (a

, a

, . . . , a

), (1)

and each group of patients (object) corresponds to it

description

→ (p

= a

, p

= a

, . . . , p

= a

), (2)

where p

= a

is a characteristic of an object.

Continuous features are discretized as follows.

The whole interval of feature values is divided into

three subintervals: w

– low values, w

– medium

values, and w

– high values. The system divides

all patients into classes according to successfulness

of the vaccination. As a result of learning, all classes

are described using a set of characteristics (AQ-rules

(Michalski, 1973; Wojtusiak et al., 2006)). Charac-

teristic h

is a disjunction of feature value intervals:

We propose to treat the process of rules genera-

tion as an optimization problem that consists in ﬁnd-

ing a possibly optimal set of rules. However, clas-

sical optimization procedures cannot be applied, be-

cause of the large number of features and their values.

So it is reasonable to use a genetic algorithm (GA).

GA have been extensively applied to solving complex

optimization problems with non-standard algorithmic

assignment of functions, complex conﬁgurationof the

admitted region, with multi-extremal functions, large

number of variables, etc. (Goldberg, 1989).

We use a recently developed modiﬁcation of the

well-known GA - co-evolutionary asymptotic genetic

algorithm (CAGA)(Sergienko and Semenkin, 2013).

It has fewer parameters than the standard GA. It rep-

resents several asymptotic probabilistic GA that work

in parallel and compete for a common resource - a

number of individuals in the population, and share the

best found solutions with each other. Base algorithms

have an adaptive mutation operator and differ from

each other by the selection criteria. Such a combi-

nation of algorithms make it unnecessary to choose

the selection, recombination and mutation operators,

which are individual for each discrete task.

Thus, we suggest to run an iterative process that

uses CAGA to ﬁnd the best rule that covers the maxi-

mum number of positive examples and uses the mini-

mum number of characteristics. Examples that satisfy

AssessmentofDendriticCellTherapyEffectivenessBasedontheFeatureExtractionfromScientificPublications

273

Table 1: Example of structured information about patient groups extracted from scientiﬁc publications about DC trials.

Group Disease HLA Vaccine Load T cell Ir. SR. Before SR. After

1/1 Melanoma IV 1 MAGE A3 (168-176) I - 1 - -

2/1 NSCLC III/IV - MAGE A3 (112-120) I Indir 10 1 52 309

2/2 NSCLC III/IV - MAGE A3 (112-120) I Dir 10 1 280 655

3/1 Glioma 3/4 - OK-432 diff. - 1 400 480

the found decision rule are not excluded and they are

taken into account in subsequent steps. Within each

iteration at least one new object should be covered.

If two rules have the same length and coverage, the

one that covers a greater number of examples of the

original set is chosen.

An individual represents some rule which is en-

coded in a binary string as a sequence of characteris-

tics. Since each feature has three values, a feature is

encoded by three numbers (0 or 1), which are inter-

preted as the presence or absence of the correspond-

ing feature value in the rule. If all feature values are

equal to one, then the feature is insigniﬁcant and it

should be excluded from rules during decoding.

Extracted descriptions of groups of patients are

sparse and often miss some features. To deal with

missing data, we assume that a rule covers an object

if they don’t contradict to each other: the rule must

contain only values, which the object description ei-

ther contains or misses. When covering positive ob-

jects there is one more condition: a rule has to include

at least one the non-missing feature value.

Due to the data sparsity there can be a situation

when a positive example is similar to a set of negative

ones and there is no rule that separates them. This

means that such set of objects has conﬂicts and there-

fore we ﬁrst delete such conﬂicting negativeexamples

from initial set before the rules learning.

We selected the ﬁtness function as a weighted

sum:

αN

init

+ βN

cur

− γN

gen

→ max, α ≫ β ≫ γ > 0, (3)

where N

cur

is a number of covered positive examples

of the current set, N

init

is a number of covered pos-

itive examples of the initial set, N

gen

is a number of

properties involved in the rule. The weights are de-

ﬁned once before running the algorithm so that the

ﬁrst summand makes the greatest contribution to the

sum and the last summand – the smallest contribution.

Such type of function allows to identify the rules that

cover the largest number of examples to choose the

smallest rule.

We evaluated our method on medical dataset

MIMICII (Massachusetts Institute of technology,

2014). We chose two groups of patients with two dif-

ferent diseases (65 and 60 patients correspondingly)

and automatically constructed rules that allowed dis-

tinguishing these patients according to their features.

We also compared our rules with rules obtained by

original AQ-method. The results of the comparison

are present in the table 2. The last column P shows

which part of rules covers a large part of objects (at

least one tenth of all objects).

Both methods were able to construct rules that

cover all positive objects. Number of feature values

in rules is almost the same in both cases. The fol-

lowing differences of methods should be highlighted.

AQ-method generates less number of rules. However,

only small part of these rules covers a large amount

of objects. Thus, there are three rules in average (one

tenth of all rules) that cover more than one tenth of

objects. At the same time our method constructs a set

of rules a bit larger, but there are more than twenty

rules in average (more than a half of all rules), which

cover more than one tenth of objects. Another dif-

ference is that rules generated by our method are sig-

niﬁcantly differ from each other, when rules of AQ-

method are similar to each other and do not allow con-

sidering all valuable features, which are required for

JSM-method.

4.3 Analysis of Cause-and-effect

Relations

As a result of feature extraction, each group of pa-

tients is described with a set of rules containing only

signiﬁcant features:

= {R

[

)}. (4)

The set of features used to search cause-and-effect

relations consists of all the features used in all the

rules generated at the previous step. These cause-

and-effect relations form the so-called fact base that

explains why a certain case of DC vaccination is suc-

cessful or not. Generation of causal hypotheses is

performed using the JSM-method (Anshakov et al.,

1991). Besides, objects of all classes are considered

as the entire set of objects (positive and negative ex-

amples within the fact base).

JSM-method relies on the following assumptions:

ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods

274

Table 2: Comparison of rule sets obtained by our method and the original AQ-method.

Number of rules Length of rules Covered objects (%) P (%)

AQ-method 21–23 9–11 100 10–15

Our method 32–36 14-16 100 50–60

• similarity hypothesis: if descriptions of all objects

with observable feature have only one common

part, then this part is a cause of the feature;

• differences hypothesis: if descriptions of two ob-

jects are similar, except one part, and this part is

present in the case where the feature appears, then

this part is considered as a cause of the feature;

• abduction hypothesis: if a set of parts of the de-

scription explains a set of hypotheses, then these

hypotheses are plausible.

JSM-method generates hypotheses about the

cause of the target feature value in the form of con-

junction of values for the class c

. A set of hypothe-

ses generated for characteristic h

is reduced by the

length and nesting. If a complex cause contains more

than three properties, it is considered as not signif-

icant. Causes which are an extension of other ones

due to the conjunction of the characteristics are also

excluded from the set of hypotheses.

This method for cause-and-effect relations mining

is used to analyze medical and psychological data and

recommended by experts in these ﬁelds due to trans-

parent and intuitive results (Blinova et al., 2003; Finn

et al., 1996).

Finally, we can deﬁne a classiﬁcation problem:

given the features of a patient, disease and a DC vac-

cine, we need to determine whether this treatment will

be successful or not. Let h

(the class label) be the

feature that correspondsto the estimation of efﬁciency

of a DC vaccine. The input feature set consists of

properties identiﬁed by the JSM-method as causes of

a certain value of h

feature.

5 CONCLUSIONS

Classiﬁcation of patients with cancer on the base of

numerical and symbolic information, extracted from

scientiﬁc publications and other sources, is proposed

in this article to increase the effectiveness of DCs vac-

cines application. Two-level division of patients into

groups according to the type of disease and type of

tumor is recommended. To justify such division we

analyzed the registry of clinical trials and the Med-

line search system, found reliable information for dif-

ferent groups of patients and conducted an initial se-

lection of parameters that have a core role in patient

classiﬁcation.

For further analysis of the retrieved information

a combined method for machine learning is proposed.

It combines the advantages of statistical and inductive

analysis and has already proved its applicability for

the analysis of clinical and psychological data.

ACKNOWLEDGEMENTS

This work was supported by the the Russian Foun-

dation for Basic Research (project no. 13-07-12127-

oﬁ

m).

REFERENCES

Aggarwal, C. C. and Zhai, C., editors (2012). Mining Text

Data. Springer.

Anshakov, O. M., Skvortcov, D. P., and Finn, V. K. (1991).

On logical construction of JSM-method of automated

hypotheses generation. Doklady Akademii nauk SSSR,

320(6):1331–1336.

Blinova, V. G., Dobrynin, D. A., Finn, V. K., Kuznetsov,

S. O., and Pankratova, E. S. (2003). Toxicology anal-

ysis by means of the JSM-method. Bioinformatics,

19(10):1201–1207.

Cauwenberghs, G. and Poggio, T. (2000). Incremental

and decremental support vector machine learning. In

Advances in Neural Information Processing Systems,

volume 13.

Collobert, R., Weston, J., Bottou, L., Karlen, M.,

Kavukcuoglu, K., and Kuksa, P. (2011). Natural lan-

guage processing (almost) from scratch. The Journal

of Machine Learning Research, 12:2493–2537.

Draube, A., Klein-Gonz´alez, N., Mattheus, S., Brillant, C.,

Hellmich, M., Engert, A., and von Bergwelt-Baildon,

M. (2011). Dendritic cell based tumor vaccination in

prostate and renal cell cancer: a systematic review and

meta-analysis. PLoS One, 6(4):1–11.

Ferrucci, D. and Lally, A. (2004). Uima: an architec-

tural approach to unstructured information processing

in the corporate research environment. Natural Lan-

guage Engineering, 10(3-4):327–348.

Figdor, C., de Vries, I., Lesterhuis, W., and Melief, C.

(2004). Dendritic cell immunotherapy: mapping the

way. Nat Med., 10(5):475–480.

Finn, V. K., Gergely, T., and Kuznetsov, S. O. (1996). Plau-

sible reasoning for open problem domains. In Pro-

ceedings of the 1996 IEEE International Symposium

on Intelligent Control, pages 242–247, Dearborn, MI.

AssessmentofDendriticCellTherapyEffectivenessBasedontheFeatureExtractionfromScientificPublications

275

Gaizauskas, R., Hepple, M., Davis, N., Guo, Y., Harkema,

H., Roberts, A., and Roberts, I. (2003). AMBIT: Ac-

quiring medical and biological information from text.

In All Hands Meeting., pages 370–373.

Goldberg, D., editor (1989). Genetic Algorithms in Search,

Optimization and Machine Learning. Addison-

Wesley Professional.

Massachusetts Institute of technology (2014). MIMIC II.

Michalski, R. S. (1973). AQVAL/1-computer implementa-

tion of variable-valued logic system VL1 and exam-

ples of its application to pattern recognition. In Proc.

Of the First Int. Joint Conf. on Pattern Recognition,

pages 3–17, Washington, DS.

Murthy, V., Moiyadi, A., Sawant, R., and Sarin, R. (2009).

Clinical considerations in developing dendritic cell

vaccine based immunotherapy protocols in cancer.

Curr Mol Med., 9(6):725–731.

Nurminen, A. (2013). Algorithmic extraction of data in ta-

bles in pdf documents. Master’s thesis, Tampere Uni-

versity of Technology.

Savova, G. K., Masanz, J. J., Ogren, P. V., Zheng, J., Sohn,

S., Kipper-Schuler, K. C., and Chute, C. G. (2010).

Mayo clinical text analysis and knowledge extraction

system (cTAKES): architecture, component evalua-

tion and applications. Journal of the American Medi-

cal Informatics Association, 17(5):507–513.

Sergienko, R. and Semenkin, E. (2013). Michigan and pitts-

burgh methods combination for fuzzy classiﬁer design

with coevolutionary algorithm. In Proc. of 2013 IEEE

Congress on Evolutionary Computation, pages 3252–

3259, Cancun, Mexico.

Shortman, K. and Caux, C. (1997). Dendritic cell develop-

ment: multiple pathways to nature’s adjuvants. Stem

Cells, 15:409–419.

Wei, Y., Sticca, R., Holmes, L., Burgin, K., Li, J.,

Williamson, J., Evans, L., Smith, S., Stephenson,

J., and Wagner, T. (2006). Dendritoma vaccination

combined with low dose interleukin-2 in metastatic

melanoma patients induced immunological and clin-

ical responses. Int J Oncol., 28(3):585–593.

Welzen-Coppens, J., van Helden-Meeuwsen, C., Drexhage,

H., and Versnel, M. (2012). Abnormalities of dendritic

cell precursors in the pancreas of the nod mouse model

of diabetes. Eur J Immunol., 42(1):186–194.

Wojtusiak, J., Michalski, R. S., Kaufman, K. A., and

Pietrzykowski, J. (2006). The AQ21 natural induc-

tion program for pattern discovery: Initial version and

its novel features. In 18th IEEE International Con-

ference on Tools with Artiﬁcial Intelligence (ICTAI

2006), pages 523–526, Arlington, Virginia.

Yannelli, J. R. and Wroblewski, J. M. (2004). On the road to

a tumor cell vaccine: 20 years of cellular immunother-

apy. Vaccine, 23:97–113.

Zhou, L. and Tedder, T. (1996). CD14

blood monocytes

can differentiate into functionally matureCD83

den-

dritic cells. Proc. Natl. Acad. Sci. USA, 93:2588–

2592.

ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods

276