Predictive AI Models for the Personalized Medicine

Luigi Lella

, Ignazio Licata

, Gianfranco Minati

, Christian Pristipino

, Antonio Giulio De Belvis

and Roberta Pastorino

ASUR, Regional Health Agency of Marche, AN, Italy

ISEM, Inst. for Scientific Methodology, PA, Italy

AIRS, Italian Association for Systems Research, MI, Italy

ASSIMSS, Italian Association for Systems Medicine & Healthcare, Rome, Italy

Section of Hygiene, Inst. of Public Health, Università Cattolica del Sacro Cuore, Rome, Italy

Keywords: e-Health, e-Health Applications, Pattern Recognition and Machine Learning, Decision Support Systems.

Abstract: Innovative information systems which enable personalized medicine are presented. The designed decision

support systems are expected to infer with an excellent level of accuracy the outcome of a therapeutic

intervention through the analysis of biometric, genetic and environmental data. They are also capable to

motivate their predictions according to a dynamic knowledge base, which is kept updated with new

analysed cases. These systems can be used by researchers to identify useful correlations between biometric,

genetic and environmental data with potential risks and benefits of certain therapeutic choices. They can

also be used by the patients to choose the most appropriate therapeutic intervention according to their needs

and expectations. In other words the presented decision support tools can realize the vision of the predictive,

preventive, personalized and participatory (P4) medicine pursued by the systemic medicine.

1 INTRODUCTION

As reported in (Personalized Medicine, 2013),

personalized medicine has opened a new rapidly

growing market in the European industry, also

creating new job opportunities.

The purpose of personalized medicine is

essentially to contain healthcare expenditure at a

time when the cost of healthcare delivery is growing

throughout Europe along with the prevalence of

chronic diseases and disorders, and more than 6% of

readmission cases hospital due to acute conditions

are caused by serious adverse drug reactions.

Research on the correlations between biological

mechanisms, environmental interactions and the

development or evolution of certain diseases and

disorders will have a significant impact throughout

the health care chain, from the research world to the

provision of health care services (Saqui M. et al.

2016).

Despite the development of some personalized

medicine approaches, we are still in one of the first

stages of implementation of this intervention

strategy (Nimmersgern E., 2017). According to a

recent review of the personalized medicine literature

presented by (Di Paolo A. et al. 2017), focused on

research carried out within the European Union,

there would not seem to be even sufficient

consensus on the definition and conception of

personalized medicine itself.

Some articles correlate its definition to the

concept of stratification or subdivision of patients

into subgroups, depending on the probability of

receiving benefits from the adoption of a specific

pharmaceutical therapy or clinical treatment. Others

instead frame it as the assignment of a tailored

therapy to patients on the basis of new individual

and dynamical classifications of diseases based on

their molecular basis and networking characteristics

rather than only on clinical grounds.

As pointed out by the authors, the initial state of

the patients is almost always evaluated considering

mainly their genetic data and their biological

markers together with the outcome of some

specialized examinations. Instead, other factors such

as the clinical evolution over time, as well as the

needs and preferences of the patient should be

considered as also required by a recent European

recommendation (Personalised Medicine 2010)

(Sagner M. et al 2017). Also according to (Di Paolo

396

Lella, L., Licata, I., Minati, G., Pristipino, C., De Belvis, A. and Pastorino, R.

Predictive AI Models for the Personalized Medicine.

DOI: 10.5220/0007472203960401

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 396-401

ISBN: 978-989-758-353-7

A. et al. 2017), further research work is aimed both

at predicting the individual outcomes of certain

treatments and the probability of incurring collateral

effects (Baumbach J. et al. 2018).

Regarding the technologies used in these deep

learning tasks, literature seems to converge in recent

years on the use of recurrent neural networks, in

particular those models based on the Long Short-

Term Memory (LSTM) paradigm for the analysis of

genetic data (Vohradsky J. 2009; Xu R. et al. 2007)

or the analysis of data contained in electronic

medical records (Lipton Z.C. et al. 2017; Pham T. et

al.2017).

This paper presents some solutions based on

machine learning systems able to infer the outcome

of a given treatment together with any side effects

on the basis of patients status (genetic data,

biometric data, environmental data), the chosen

therapies, their needs and preferences.

In agreement with (Di Paolo A. et al. 2017) we

believe that just by the adoption of a holistic

approach, that does not consider only genetic data

and biological markers but also the environment and

the needs of the patients, it is possible to effectively

deal with the problem of personalized medicine,

adapting it exactly to the profile of patient. Through

the development of outcome measures co-developed

between researchers, patients and subject experts we

will cover what really matters to patients, embracing

the cognitive, self-cognitive, psychological,

symbolic, social, ecological and environmental

dimensions.

Most machine learning techniques are oriented

towards a kind of structural representation of

knowledge. This can be symbolic or subsymbolic.

Sub-symbolic models can achieve the best results in

problems that are difficult to solve if a static

knowledge base consisting of simple logical

production rules is adopted. Sub-symbolic models

can be further subdivided into classification learning

algorithms (Kohonen T., 1988; Rumelhart D.E. and

McClelland J. L., 1986), association learning

algorithms (Kohonen T., 1989) and clustering

learning algorithms (Van Hulle M. M., 2012;

Kohonen T. 1990; Fritzke B., 1994; Licata I. and

Lella L., 2007).

In classification learning, the system is trained to

provide a given output (a class) from a set of

classified examples. This type of model, to which

LSTMs belong, is only effective if the correlations

between non-class attributes and all the possible

classes are known in advance. This model does not

therefore adapt to the case of the predictions of

therapeutic choices in personalized medicine, since

it can be very complicated to define the rules of

association between individual profiles of patients

and possible therapeutic interventions.

In association learning there are no specific

classes, the system only tries to find an interesting

scheme or a correlation between the data.

Association rules can be used to predict attributes of

any kind, not just class ones. Since we are interested

in predicting the therapeutic choice, the duration of

therapy, the risks and the results that can be

achieved, association learning models are not suited

to solve the problem.

Finally clustering algorithms are unsupervised,

meaning that there is no set of classified examples

that can be used to train the system. If we choose the

duration of therapy, the achievable results and

possible side effects as class attributes, the system

can extrapolate several clusters related to class

attributes. In this way it is possible to avoid the

presence of human experts making this solution

more interesting and easy to implement.

Among the algorithms belonging to this last

family the SOM (Kohonen T., 1989) have been

widely used in healthcare, but we believe that the

best results can be obtained using more adaptive

models. In this type of unsupervised learning

activity there is no clear correlation between class

attributes and the other ones. In other words, the

exact topology of the input space is unknown.

B. Fritzke in one of his articles showed that his

network model called GNG (Fritzke B., 1994) is

able to identify exactly the local dimension of the

input space, i.e. a GNG can find how many

attributes in the defined input space are needed to

accurately predict the considered class attributes.

As a further model to be compared with the self-

organizing neural networks and the LSTMs, we will

test the self-organizing symbolic model of the Non

Organized Turing Machine (A-Type) (Turing A.,

1948) consisting essentially of a network of NAND

gates by which it is possible to construct a sort of

knowledge base modelling the problem. This

network will evolve through the use of various

algorithms that encode the network configuration by

means of fixed-length bit sequences. In particular we

will consider Genetic Algorithms (Eiben A. E. and

Smith J. E., 2015; Mitchell A.E., 2015) and Swarm

Intelligence algorithms (Praveena S., 2018).

A Non-Organized Turing Machine is a symbolic

model from which we expect a lower performance in

terms of prediction accuracy than the considered

subsymbolical models, but an A-type may be able to

justify the inferred answers by resorting to a

dynamic logical formalism.

Predictive AI Models for the Personalized Medicine

397

The trained machine learning models will allow

professionals and assistants to select the most

suitable therapeutic regimen to treat the clinical case

taken care of. The patient will have the opportunity

to evaluate the outcome of a pharmacological or

specialist therapy by selecting it from the list of

those already used to treat similar cases.

This will lead not only to the patient's

empowerment, but it will also lead to the realization

of the long-awaited therapeutic alliance between

caregiver and doctor who has taken care of him,

limiting inappropriate interventions.

To achieve this result it is important to define

diseases more precisely and to stratify patients into

subgroups, based on their likelihood of responding

to a given treatment, and also to stratify healthy

citizens according to their risk of disease.

The classic approach of diagnosis and treatment

must be overcome through specific omics data

acquisition, the individual profile of the subject is

assessed, enabling the choice of a specific

therapeutic strategy. It is thus possible to minimize

the "toxic cost" of the therapy, improving the

patient's quality of life and optimizing the

management of the available economic resources.

The described models will be tested at the the

University Polyclinic Foundation Agostino Gemelli

Hospital Center with the help of ad hoc resources

collecting information which is not already collected

routinarily. The expected pathology specific clinical,

economic, quality and humanistic outcomes will be

suggested by the involved multidisciplinary team.

2 METHODS

As input data to encode patients status, a binary

vector will be assembled that encodes the genetic

information, the molecular fingerprints (e.g. -omics),

the biometric information, the clinical data, the

therapeutic choice, the exposome, and the needs and

the psychological dimensions of the patient and of

his/her social networks.

As far as genetic information is concerned, a

selection could be made, at least during the test

phase of the developed decision support system, of

all the possible about 30,000 human genes,

considering only those that research considers useful

for predicting the onset of disorders or diseases. For

cancer alone, for example, large-scale studies (Hill

S., 2018) have confirmed that there are about 450

"key genes" to be considered in the prediction of the

onset or evolution of different forms of cancer.

In order to reduce the training and processing

times of the chosen machine learning models, the

considered cases could be limited only to a set of

tumor forms that are particularly incident on the

territory.

The encoding of such data will be accompanied

by the codification of the outcome of some related

specialist examinations. For example, the key gene

for breast cancer called HER2 (Perez E.A. et al.,

2014) is associated with the IHC (Immuno Histo

Chemistry) exam that identifies the percentage of

HER2 proteins in tumor cells, and with FISH

(Fluorescence In Situ Hybridization), SPoT-Light

HER2 CISH test and Inform HER2 Dual ISH test to

identify if there are too many pairs of HER2 genes

in tumor cells. The outcome of all these specialized

examinations must be appropriately coded using a

simple binary coding in the case of results that can

be simply positive or negative or a "one-hot" coding,

having as many bits as all the possible outcomes,

and with only one of these coded as 1. For the IHC

test of the HER2 gene, for example, the code "1000"

can be used for the "negative" result, the code

"0100" for the result "also negative", the code

"0010" for the result "borderline" and the code

"0001" for the "positive" result. A one-hot code

should also be used to codify the choice of treatment

regimen, the status and the needs of the patient.

The indicators to be taken into consideration to

define the patient's status and needs will be taken

from the information systems for the measurement

of the outcomes reported by the patients as the one

developed within the PROMIS project (Cella D. et

al., 2010) or other outcome measures reported from

the patients studied in literature (Black N., 2013)

(Donabedian A., 1988). The outcome measures

reported by the patients (PROM) are measures of

functionality and well-being in the sphere of the

patient's physical, mental and social health (Black

N., 2013).

To codify the output of the chosen forecasting

models, a vector with the one-hot coding of the

duration of the therapy will be assembled (duration

divided into classes or periods, for example: 0-6

months, 6-12 months> 1 year), together with a one-

hot vector with the possible pathology specific

outcomes (also in this case we will adopt the

PROMIS coding system), and a sequence of binary

codes (present or not present) associated with

possible side effects.

To improve the learning process of the chosen

self-organizing networks (SOM and GNG) as well

as the Non-Organized Turing Machine, we will

adopt the methodology suggested by Kohonen

HEALTHINF 2019 - 12th International Conference on Health Informatics

398

(Kohonen T., 1990). The input vector of the chosen

models will be constructed by concatenating a

contextual part that represents the class attributes of

the instance and a symbolic part composed of the

other attributes. The part of the symbol and the part

of the context will therefore be represented by two

orthogonal vectors such that the norm of the second

is larger than that of the first. In this way, in the

subsimbolic prediction models taken into account

the symbols can be coded in a topological order

(connection between neural units) that reflects the

logical analogies.

The implemented evolutionary algorithms

(genetic and swarm intelligence) will instead be able

to make more accurate predictions by selecting them

from the considered space of the solutions.

The data set will be divided into a part equal to

the 66% of the samples used as a training set, and a

part equal to the remaining 34% of the samples used

as a test set to evaluate the predictive accuracy of the

model. All of these models have already been

successfully tested in computationally similar

contexts like the length of hospital stay prediction on

the basis of the data contained in the patients

admission forms (Lella L. and Licata I.,2017; Lella

L. and Licata I., 2018).

Finally, it has to be noticed that these models

must be trained with a large number of data, or

rather, following the definition of big data provided

by (Anderson C., 2008; Mayer-Schonberger V. and

Cukier K., 2017; Godsey B., 2018), automatically

collecting, storing and analysing all the clinical data,

managing them as soon as they become available.

As expressed by (Naimi A.I. and Westreich D. J.,

2014) we will not consider the automatic analysis of

all the data as the best adoptable scientific approach.

According to the book review, we believe that all the

available data will never be completely free of bias

and in any case it will be necessary to adopt

preprocessing techniques including resampling.

Instead, it will be fundamental to monitor in real

time all the patients available data in order to follow

the evolution of their clinical picture, suggesting

possible prevention and treatment pathways.

In a future in which the personal, health-related

and environmental information of each individual

will be contained within a "personal data cloud" it

will be possible to analyse in real time all this

amount of data in order to provide people with

useful coaching suggestions on how to improve their

health preventing chronic disorders.

It will be possible, for example, to suggest to an

individual, who has a genetic variant associated with

a high predisposition to type 2 diabetes and a rapid

increase in blood glucose level, to undergo a series

of tests and to adopt certain dietary regimens and

levels of physical activity to avoid the devastating

effects of this disorder.

By activating the participatory component of

medicine, patients will be more involved by making

them aware of the possible consequences of their

behaviour. This will reduce the onset of chronic

disorders through self-monitoring and self-

assessment leading to improved quality of life for

patients and their caregivers.

3 CONCLUSIONS

Data mining and knowledge discovery processes do

not follow precise rules. There is no model or

method capable of producing useful results in any

context of use.

In the case of the prediction of the duration of the

therapy, of the outcome and the side effects of a

personalized medicine case, it may be useful to use

models such as GNG that perform the so-called

dimensionality reduction. These models can find a

sub dimensional space that contains most of all input

data. The GNG model has the potential to adapt

effectively to the input space, but it must be trained

through the use of appropriate preprocessing

techniques. We believe that the GNG model will

perform better than other considered self-organizing

networks, achieving a greater prediction accuracy.

We will also test a second symbolic model that

implements the Non-Organized Turing Machines

that is able to justify its predictions and to

autonomously evolve its knowledge base over time.

The development of these systems is perfectly in

line with two of the objectives specified in the

European Union report on personalized medicine

(Personalized Medicine, 2013), which are primarily

to reduce the number of unnecessary interventions

and adverse events by maximizing the added value

perceived by patients, but also to favour a

containment of welfare costs.

The use of artificial intelligence models in

forecasting the outcomes of therapeutic choices can

contribute to implement the predictive, preventive,

personalized and participatory (P4) vision predicted

and desired by some pioneers of systems medicine

(Flores M. et al. 2013; Auffray C. et al. 2017).

Decision support systems supported by AI

models, such as those presented in this work, will

also make it possible to improve the effectiveness of

medical decisions by moving from symptom-

focused medicine to medicine focused on causes,

Predictive AI Models for the Personalized Medicine

399

highlighting the therapies with the highest

probability of success with the lower level of risk for

each individual wherein the participation of the

patient remains pivotal (Leyens L. et al. 2014).

REFERENCES

Anderson C., 2008. The end of theory: the data deluge

makes the scientific method obsolete, Wired.

Auffray C., Sagner M., Abdelhak S., Adcock I., 2017.

Viva Europa, a Land of Excellence in Research and

Innovation for Health and Wellbeing. Progress Prev

Med 2017; 2(3):e006. doi: 10.1097/pp9.000000000

0000006

Baumbach J., Schmidt H., 2018. The End of Medicine as

We Know It: Introduction to the New Journal,

Systems Medicine. Syst Med 2018; 1: https://

doi.org/https://doi.org/10.1089/sysm.2017. 28999.jba

Black N., 2013. Patient reported outcome measures could

help transform healthcare, BMJ n.346, p.167.

Cella D., Yount S., Rothrock N., Gershon R., Cook K.,

Reeve B., Ader D., Fries J. F., Bruce B., Rose M.,

2010. The patient-reported outcomes measurement

information system (PROMIS), PMC.

Di Paolo A., Sarkozy F., Ryll B., Siebert U., 2017.

Personalized medicine in Europe: not yet personal

enough?, BMC Health Services Research, vol.17, Issue

289.

Donabedian A., 1988. The quality of care. How it can be

assessed?, JAMA n.260, pp.1743-1748.

Eiben A. E., 2015. Smith J. E. Introduction to

Evolutionary Computing, Springer: Natural

Computing Series, 2nd ed.

Flores M., Glusman G., Brogaard K., Price N.D., Hood L,

2013. P4 medicine: how systems medicine will

transform the healthcare sector and society,

Personalized Medicine vol.10, issue 6, pp. 565-576.

Fritzke B., 1994. A Growing Neural Gas Network Learns

Topologies. Part of: Advances in Neural Information

Processing Systems 7, NIPS.

Godsey B., 2018. Think like a data scientist: tackle the

data science process step by step, Manning

Publications Co.

Hill S., 2018. Introducing genomics into cancer care, BJS

n.105, pp. e14-e15.

Kohonen T., 1988. An introduction to neural computing,

Neural Networks, vol.1, pp.3-16.

Kohonen T., 1989. Self-Organization and Associative

Memory, Berlin: Springer-Verlag.

Kohonen T., 1990. The Self Organizing Map, Proc. of the

IEEE, vol.78, n.9.

Lella L., Licata I., 2017. Prediction of length of hospital

stay using a growing neural gas model, Proc. of

IMCIC 2017.

Lella L., Licata I., 2018. Length of hospital stay prediction

through unorganised Turing machines, Proc. of

BIOSTEC 2018, vol.5: HEALTHINF, pp. 402-407.

Leyens L., Hackenitz E., Horgan D., Richer E., Brand A.,

Bubhoff U., Ballensiefen W., 2014. CSA Permed:

Europe’s commitment to personalised medicine,

Eurohealth.

Licata I., Lella L., 2007. Evolutionary Neural Gas (ENG):

A model of self-organizing network from input

categorization”, EJTP, vol.4, n.14.

Lipton Z.C., Kale D.C., Elkan C., Wetzel R., 2017.

Learning to Diagnose with LSTM Recurrent Neural

Networks, Available online: arXiv:1511.03677.

Mayer-Schonberger V., Cukier K., 2017. Big Data: A

revolution that will transform how we live, work and

think, John Murray Publishers.

Mitchell M., 1996. An Introduction to Genetic

Algorithms, Cambridge, MA: MIT Press.

Naimi A. I., Westreich D. J., 2014. Big Data: A revolution

that will transform how we live, work and think –

book review, American Journal of Epidemiology,

Available online: https://academic.oup.com/aje/article

/179/9/1143/2739247.

Nimmesgern E., Benediktsson I., Norstedt I., 2017.

Personalized Medicine in Europe, Clin.Transl Sci.

n.10, pp.61-63.

Perez E. A., Cortes J., 2014. Gonzalez-Angulo A. M.,

Bartlett J. M. S. HER2 testing: current status and

future directions, Cancer Treatment Reviews n.40, pp.

276-284.

Personalised Medicine – opportunities and challenges for

European healthcare, 2010. Available online:

http://ec.europa.eu/research/health/pdf/13th-european-

health-forum-workshop-report_en.pdf.

Personalized Medicine, 2013. Available online: https://ec.

europa.eu/research/health/index.cfm?pg=policy&polic

yname=personalised

Pham T., Tran T., Phung D., Venkatesh S., 2017.

DeepCare: A Deep Dynamic Memory Model for

Predictive Medicine, Available online:

arXiv:1602.00357.

Praveena S., 2018. Review on Swarm Intelligence

Algorithms, International Journal of Scientific

Research in Computer Science, Engineering and

Information Technology, Vol.3, Issue 4.

Rumelhart D. E., McClelland J. L., Eds., 1986. Parallel

Distributed Processing, vol.1, Cambridge, MA: MIT

Press.

Sagner M., McNeil A., Puska P., Auffray C., Price N. D.,

Hood L., Lavie C. J., Han Z.G., Chen Z., Brahmachari

S. K., McEwen B. S., Soares M. B., Balling R., Epel

E., Arena R., 2017. The P4 Health Spectrum - A

Predictive, Preventive, Personalized and Participatory

Continuum for Promoting Healthspan. Prog

Cardiovasc Dis.2017;59(5):506-521.

Saqi M., Pellet J., Roznovat I., Mazein A., Ballereau S.,

De Meulder B., Auffray C., 2016. Systems Medicine:

The Future of Medical Genomics, Healthcare, and

Wellness. Methods Mol Biol.2016;1386:43-60.

Turing A., 1948. Intelligent Machinery, in Collected

Works of A.M.Turing:Mechanical Intelligence. Edited

by D.C.Ince. Elsevier Science Publishers, 1992.

Van Hulle M.M., 1989. Self Organizing Maps, Handbook

of Natural Computing, pp. 585-622.

HEALTHINF 2019 - 12th International Conference on Health Informatics

400

Vohradsky J., 2009. Neural network model of gene

expression, the FASEB Journal, vol.19, no.2, pp. 320-

329.

Xu R., Wunsch II D., Frank D., 2007. Inference of genetic

regulatory networks with recurrent neural network

models using particle swarm optimization, IEEE/ACM

Transactions on Computational Biology and

Bioinformatics (TCBB), vol. 4, n.4, pp. 681-692.

Predictive AI Models for the Personalized Medicine

401