MULTIPLE SCLEROSIS: A DOMAIN FOR THE APPLICATION
OF ADVANCED AI TECHNIQUES
Mauro Gaspari*; Lino P. Marchello†; Cinzia Scandellari†; Sergio Stecchi†
* Dipartimento di Scienze Dell'Informazione, Università di Bologna, Italy
†Centro Sclerosi Multipla, Villa Mazzacorati, A.U.S.L.Bologna, Italy
Multiple Sclerosis is still an unsolved disease and although several advances have been achieved in the last few
years some problems such as the ethiology and the prognostic criteria are unknown or still have not a satisfying
solution. In this paper we give a brief introduction to this disease and we underlying the complexity of its domain
and the knowledge involved. Finally we present a study on a phase of this disease that could be approached ex-
ploiting Artificial Intelligence techniques.
1 Introduction formation retrieval from computerised patients records
in Multiple Sclerosis, for instance ontologies (Falas-
Multiple sclerosis (MS) is an inflammatory disease coni and Stefanelli, 1994), data warehousing (Barquin
characterised by demyelinization of the central nervous and. Edelstein, 1996), and data mining (Agrawal,
system (CNS) (Storch and Lassmann, 1997). It is an 1994).
immunological disease, and even if its pathogenetic
mechanism is well known, some problems such as the In this paper we give a brief introduction to Multiple
ethiology and the prognostic criteria are unknown or Sclerosis and we underline the complexity of its do-
still have not a satisfying solution (Cook, 1996). main and the knowledge involved. Finally we present
Although several advances have been achieved in the a study on a phase of this disease that could be ap-
last few years, predicting with accuracy how and when proached exploiting Artificial Intelligence techniques.
the problem of multiple sclerosis will eventually be
solved is not easy (Compston, 1994).
The efforts of contemporary research are oriented to 2 Multiple Sclerosis
develop complex strategies based on ideas originating
from immunology, neurobiology, brain imaging, and Histopathology of MS is defined by a chronic inflam-
animal modelling. The advent of these increasingly mation and demyelination. Ongoing disease activity is
complex strategies makes clinical trials difficult to con- due to an active inflammatory process, mainly mediat-
trol, in fact, numerical powerful studies on uncertain ed by T lymphocytes and macrophages, and is associat-
markers are needed involving multicenter collaboration ed with blood-barrier damage. B lymphocytes and
for reaching a critical mass of patients. plasma cells are present especially in the lesions that
occur during the late chronic stage of the disease.
The MS research community has recognised these Although all plaques are characterized by demyelina-
problems developing databases to store computerised tion, the patterns of oligodendroglia destruction and of
patient records stored in a standard form: EDMUS damage to other tissue elements, such as axons and as-
(Confavreux, 1994) and MS-COSTAR (Paty et al., trocytes, are variable in different cases. Oligodendro-
1994). Unfortunately, the uncertainty, the incomplete- cytes are less affected by plaques that develop during
ness, and the temporal nature of some of the observed the first bouts of the disease than by those plaques aris-
data make ordinary database query techniques and ing after several years of disease duration.
tools not completely adequate to support an automatic
interpretation of future clinical trials. In our opinion
more complex techniques are needed to manage in-
For MS the best set of diagnostic criteria is still that of perivenular demyelination illness without evi-
the Schumacher Panel (Schumacher et al., 1968). This dence of infection of the CNS.
essentially spells out MS as a white matter disease with - In human immunodeficiency virus-en-
evidence of continued or repeated clinical activity over cephalopathy and myelopathy virus is present
time. These criteria have been amended by the Poser in macrophages and microglia and the myelin
committee (1983) to include laboratory evidence, the abnormalities apparently are caused by solu-
paraclinical evoked responses, and the cerebrospinal ble factors such as viral proteins, cytokines, or
fluid (CSF). neurotoxins.
However the MS diagnosis is essentially a clinical pro- These findings may have implications on how, when
cess based on spatial and temporal dissemination of and were to seek viruses in MS.
symptoms and lesions.
2.4 Classification of MS Forms
McAlpine and associates (1998) have depicted the vari-
There is evidence to support the view that MS is a ation in the natural history of MS in his textbook. Six-
complex trait determined by both genetic and environ- ty percent of MS patients have attacks, particularly at
mental factors. The genetic component is reflected in the beginning of their illness, with quite good recovery
the higher rate of concordance in monozygotic (MZ) vs and minimal deficit. Ultimately MS culminates in a
dizygotic (DZ) twins and in familial recurrence risk progressive course with relatively fewer exacerbations,
data. yet gradually worsening disability.
There is no evidence that any genetic marker, acting Fifteen percent of MS patients have progressive MS
alone or in combination, protect individual from devel- from onset, with no or relatively few attacks, but devel-
opment of the disease. But with the more recent com- op gradually worsening disability. Finally, there is an-
pletion of full genome screens, new regions of potential other group that constitutes 15 or 20% of all MS pa-
genetic interest with respect to multiple sclerosis sus- tients, who have benign MS, having relatively few at-
ceptibility have emerged. tacks early on, without developing any or very little
Nevertheless, there remains a substantial amount of in- permanent disability.
formation gathered from a variety of epidemiological In summary, the course of the disease can be character-
and genetic sources which provides evidence for the ised the following forms (Lublin and Reingold, 1996):
contribution of genetic factors in determining suscepti- - benign forms, with few attacks and no disabil-
bility to multiple sclerosis; what remains uncertain is ity;
the number of genes involved and their location within
the genome, the relative importance of genes and the
- primary-progressive (PP) with an highly dis-
environment, the mechanisms of their interactions and ability degree;
the issue of heterogeneity. - relapsing-remitting (RR) which is character-
ised by exacerbations which normally remits
and sometimes evolves in secondary pro-
2.3 Virology of demyelinating disease gressive forms.
- secondary progressive (SP) having a progres-
Infectious agents have been postulated as causes of sive course with relatively fewer exacerba-
multiple sclerosis for over a century. tions, yet gradually worsening disability.
The possible role of a virus or viruses is supported by
data that a childhood exsposure is involved and “viral”
infections may precipitate exacerbations of disease, ex-
perimental infections in animals and natural infections
3 The Data
in humans can cause diseases with long incubation pe-
riods, remitting and relapsing courses, and demyelina- The principal descriptors of the natural history of MS
tion and patients with MS have abnormal immune re- are attacks (frequency, severity and recovery), remis-
sponses to viruses. sions (frequency and duration), temporal course (re-
The pathogenesis of three human demyelinating dis- lapsing and remitting vs progressive) and disability sta-
eases of known viral etiology is discussed: tus.
- In progressive multifocal leukoencephalopa- - An attack (or relapse) require the appearance
thy, a papovavirus selectively infects oligo- of a new symptom or worsering of an old
dendrocytes and causes focal areas of de- symptom over at least 24 h that could be at-
myelination. tributed to MS activity and is preceded by sta-
- In postmeasles encephalomyelitis, the virus is bility or improvement for at least 30 days
(Shumacher et al., 1968).
lymphotrophic and disrupts immune regula-
tion than can result in an autoimmune
- A remissions is a regression of most of the Evoked potentials are electrical waveforms elicited by
symptoms of the associated attack. A remis- and temporally related to a stimulus, most commonly
sion can be complete, mostly in the early stage an electrical or magnetic stimulus delivered to a sens-
of the disease, or incomplete with permanent ory receptor or nerve (Chiappa, 1990). These tech-
deficits. niques allow the specialist to verify the assessment of
- The temporal course depends on the form of the entire length of the sensory or motor pathways.
SM (PP, RR, SP). Evoked potentials are distinguishable in:
- The disability status is the global deficit - visual evoked potential;
caused by the disease. - brain stem auditory evoked potential;
The instrumental tools for the diagnosis and the follow - somatosensory evoked potential;
up of the MS are: Magnetic Resonance Imaging (MRI), - motor evoked potential.
evoked potentials, liquor examination. For each of these evoked potential there are two signi-
ficant parameters: latency and amplitude of the re-
sponse. Latency is misured in milliseconds and amp-
3.1 Disability Status Scale litude in microV; it can be represented by a floating
One measure of neurologic status is the Disability Sta- Abnormal evoked potentials may have a role in predict-
tus Scale (DSS) (Kurtzke, 1955), which grades clinical ing both the diagnosis of MS and clinical deterioration.
impairment due to MS on a 0 (normal) to 10 (death due
to MS) basis. The expanded DSS (EDSS) (Kurtzke, 3.4 Liquor examination
1983) subdivides each step 1 through 9 into two. Type
and severity of neurological impairment is defined by Cerebrospinal fluid analysis in MS checks for the pres-
graded involvement in the following eight functional ence of particular proteins in the fluid. The most spe-
systems (FS): pyramidal, cerebellar, brainstem, senso- cific abnormality is the oligoclonal banding i.e. the
ry, bowel and bladder, visual, cerebral and other. presence of two or more distinct IgG bands in the
gamma region of the electrophoresis (McApline 1998).
3.2 Magnetic Resonance Imaging This pattern usually remain the same over time and if it
is positive in 85%-95% of clinical defined MS and pos-
Magnetic Resonance Imaging (MRI) is increasingly itive in 40%-50% of suspected MS. The number of
being used as a measure of pathological disease activi- bands can be represented as an integer.
ty in monitoring the efficacy of potential new treat-
ments for MS. A major advantage of MRI over clinical 3.5 Additional Problems
monitoring is that it detects a large amount of sub-clini-
cal disease activity. There several problems which do not depend from
The two main approaches to MRI are detecting active computer science or artificial intelligence aspects that
lesions and measuring total lesion load. In RR and SP make the availability of these data more difficult.
MS gadolinium enhancement increases the number of
detectable active lesions and also probably correlates 3.5.1 Old Patient Records
with pathological activity.
An important limitation of MRI monitoring is that con- Since patient records are the result of several exams
ventional brain MRI abnormalities often show little or starting from the time of the diagnosis until now, it
no relationship with clinical disability. may happen that they do not contain all the significant
One explanation of this fact may be the pathological data. Typically new medical analysis and criteria
heterogeneity of lesions that all look the same on a which often are more significant could not be included
conventional image. Other MR techniques are needed in old patient records.
that specifically identify the pathological features most
likely to result in disability, namely demyelination and 3.5.2 Geographical Distribution of Patient Records
Magnetization transfer imaging and proton MR spec- To realise more significant experiments and trails it is
troscopy are two techniques that show promise in this necessary to consider data which belongs from differ-
regard. ent laboratories in the same country, and also data from
Significant data are the number of lesions, and the fact different countries.
that these lesions are active or not. It is difficult to Thus a distributed solution is necessary which allows
evaluate the influence of the place where the lesions the specialist to integrate data which belong from sev-
appear. eral laboratories. This implies that issues such has dis-
tribution, interoperability and integration of multi-lin-
3.3 Evoked Potentials gual records must also be addressed to obtain more in-
4 Analysis of Disease Progression DEFINITION - marker
A marker is a predicate P such that the following rules
A promising research direction concerns the analysis of hold:
the disease progression with the aim of individuate pro- ∀X. Patient(X) ∧ RRtoSP(X) → P
gression markers. In particular, a transitional phase ∀X. Patient(X) ∧ RR(X) → ¬P
has been observed for some of the patients affected by
a RR form which gradually (in one or two years) The goal of an automatic analysis is to verify if such a
evolves to a SP form (Stecchi et al., 1998). An early predicate exists, note that P is not necessarily an atomic
individuation of this transitional phase is essential for predicate it can be the combination of more observa-
an adequate treatment of the disease which in many tions representing parameters or examinations, and also
cases can prevent the evolution of the disease into it can include a degree of uncertainty.
highly disabling forms. A good progression marker al-
lows the specialist to early discover the occurrence of
this transitional phase and to plan for an adequate treat- 5 Related Work
As far as we know there are no attempts to exploit AI
Although, in the last few years, this transitional phase technology in Multiple Sclerosis which have been de-
has been recognised in the literature, it is still not clear scribed in the literature. Among the other approaches it
if an adequate progression marker exists. Candidate is worth to remember two information systems: ED-
markers are: MUS (Confavreux 1992,1994) and MS-COSTAR
1. Axonal damage. (Paty et al. 1994). These two database have been de-
2. External events, for instance a virus. signed to store clinical information concerning multiple
3. Number of relapses/year. scerosis EDMUS is the result of an European effort,
4. Progression index (EDSS/Duration while MS-COSTAR has been used mainly in USA and
of the disease). Canada. See (Confavreux and Paty, 1995) for an ex-
tensive comparison between EDMUS and MS-COST-
In order to individuate other possible markers and to AR. Other computer based approaches only concen-
verify that a given marker is significant, an analysis of trate on particular aspects of the disease, for instance
clinical, instrumental and laboratory data for patients the elaboration of RMI images (Grimaud et al., 1996;
affected by multiple sclerosis is necessary. Note that Molyneux et al., 1998).
this data are the result of a set of observations on the
patients affected by multiple sclerosis repeated over
time; they represent the history of the evolution (natur-
al history) of the disease for all the patients. This con-
sists of a large set of heterogeneous uncertain data
We strongly believe that a AI approach to multiple
which must be correlated in order to confirm the can-
sclerosis can be useful to medical specialist dealing
didate markers. It is very difficult for the specialist to
with MS in several ways. First a complete knowledge
achieve this goal without the support of an automatic
acquisition procedure can give an order to the medical
tool, and artificial intelligence techniques seems to be
knowledge on this disease which is wide, hill-struc-
the more adequate to deal with this complexity.
tured and characterised by a strong degree of uncer-
tainty. Second several subtasks in the complex domain
In the following we provide a more precise character-
of this disease could benefit of a set of support mechan-
isation of the concept of marker to give an idea of the
isms to automatically verify some of the specialists hy-
problem which must be addressed in term of AI. Note
pothesis on large, possibly distributed, amount of data.
that the following definitions do not intend to be com-
Third, more ambitiously, discovering new results con-
plete and usable, a long and complex knowledge ac-
cerning the diagnosis and treatment of the disease.
quisition phase is needed to reach this status.
We assume the following predicates:
Patient(X): X is a patient. Acknowledgements
RR(X): X is affected by a relapsing-remitting
form. We would like two thank Simon Colton for his valu-
RRtoSP(X): X is affected by a secondary-pro- able comments on a draft of this paper.
gressive form the previously was a
Given this set of predicates a marker can be defined
more formally as follows:
R. Agrawal. Tutorial: Data Mining. Proceedings of the GA. Schumacher, G. Beebe, RF. Kibler, et al Prob-
ACM SIGACT SIGMOD SIGART Symposium on lems of Experimental Trials of Therapy in Mul-
Principles of Database Systems. Minneapolis; tiple Sclerosis: report by the panel on the evalu-
ACM Press, New York, NY 10036, USA ation of experimental trials of therapy in Multiple
13:75--76, 1994. Sclerosis. Ann NY Acad. SCI 122:552-568, 1668.
RB. Altman. Informatics in the care of patients: Ten M. Storch, H. Lassmann. Pathology and Pathogenesis
Notable Challenges. The Western Journal of Medi- of Demyelinating Deseases. Current Opinion in
cine 166(2):118-122, 1996. Neurology 10:186-192 1997.
R. Barquin, S. Edelstein, Planning and Designing the S. Stecchi, C. Scandellari, E. Saba, C. Trocino, LP.
Data Warehouse, Prentice Hall, 1996. Marchello, E. Montanari. Intermediate Transition-
KH. Chiappa. Evoked Potentials in Clinical Medicine al Phase (ITP)of Multiple Sclerosis Patients Re-
II ed. Raven Press, N.Y., USA, 1990. sponsive to Mitoxantrone. Multiple Sclerosis
C. Confavreux, DW. Paty. Current status of computer- 4(4):381, 1998.
ization of multiple sclerosis clinical data for re-
search in Europe and North America: The ED-
MUS/MS-COSTAR connection. Spec Article,
C. Confavreux. Establishment And Use Of Multiple-
Sclerosis Registers - EDMUS. Annals of Neuro-
logy, 36(SS): S 136-S 139. 1994.
C. Confavreux.. EDMUS, a European database for
multiple sclerosis. J Neurol Neurosurg Psychiatry.
1992 Aug; 55(8): 671-676. PMID: 1527537; UI:
SD. Cook. The Epidemiology of Multliple Sclerosis:
clues to the etiology of a mistereous disease.
NeuroScientist 2:172-180, 1996.
WM. Detmer & EH. Shortliffe. Using the Internet to
Improve Knowledge Diffusion in Medicine. Com-
munications of the Association of Computing Ma-
chinery 40(8):101-108, 1997.
S. Falasconi, M. Stefanelli. A library of medical onto-
logies. In Nicolaas J.I. Mars, editor, ECAI94
Workshop - Comparison of Implemented Ontolo-
gies, pages 81-91, Amsterdam, 8-10 August 1994.
J. Grimaud, et al. Quantification of MRI lesion load in
multiple sclerosis: a comparison of three com-
puter-assisted techniques. Magn Reson Imaging.
1996; 14(5):495-505. PMID: 8843362; UI:
FD. Lublin, SC. Reingold. Defining the Clinical
Course of Multiple Sclerosis: results of an Interna-
tional Survey. Neurology 46:907-911 1996.
McAlpine's. Multiple Sclerosis III ed. Churchill Liv-
ingstone, London, UK 1998.
PD. Molyneux, et al. Precision and reliability for
measurement of change in MRI lesion volume in
multiple sclerosis: a comparison of two computer
assisted techniques. J Neurol Neurosurg Psychi-
atry. 1998 Jul; 65(1): 42-47. PMID: 9667559; UI:
D. Paty, D. Studeney, K. Redekop, F. Lublin. MS CO-
STAR: A Computerized Patient Record Adapted
for Clinical Research Purposes. Ann Neurol 1994;
C.M. Poser, D.V. Paty, L. Scheinberg et al. New Dia-
gnostic Criteria for MS: Guidelines for Research
Protocols. Ann. Neurol. 13:227-231, 1983.