Transcript Using neural networks for differential diagnosis of

Using neural networks for
differential diagnosis of
Alzheimer Disease and Vascular
Dementia
Author: Elizabeth Gaarcia-Perez, Arturo Violante,
Francisco Cervantes-Perez
Expert Systems with Applications
Introduction
• Several studies have shown that in
people 65 years old or older, the
presence of Alzheimcr Disease (AD)
has increased from 1.3 to 6.2% (Ueda &
Kawano, 1992; Gorelick & Roman, 1993; Joachin et al., 1988)
• the Mexican Society for Alzheimer
has reported that 6% of the people
over 65 years of age have been
diagnosed with Alzheimer (Cummings &
Benson, 1992; Friedland, 1993)
Introduction
• Within the analysis of dementia, the
diagnosis of AD and VD is one of the
main concerns, they represent almost
90% of the illnesses presented by
patients with dementia (O'Brien, 1992; Boiler
et al., 1989).
Introduction
• diagnose VD several techniques have
been developed, like the Hachinski
scale (Hachinski & Lassan, 1974)
• without the possibility of obtaining a
correct differential diagnosis VD
(Villardita, 1993; Gorelick & Roman, 1993; von Reutern,
1991).
Introduction
• Artificial Intelligence, AI
• complex problems in medical diagnosis can
be approached. For example, pattern
recognition in X-ray images (Boone et al., 1990a,b;
Gross et al., 1990; Hallgren & Reynolds, 1992),
biomedical
signals analysis (Gevins & Morgan, 1988; Mamelak et al.,
1991; Alkon et al., 1990; G~ibor & Seyal, 1992; Gfibor et al., 1993)
and prediction and diagnosis problems
(Casselman & Maj, 1990; Poli et al., 1991; Moallemi, 1991;
Baxt, 1991).
Data collection: Training
and Test sets
• To carry out a differential diagnosis of AD
and VD
• Collection data as follow (Bolla et al., 1991; Fisher
et al., 1990; Krall, 1983; Rovner et al., 1989):
– how the problem started (i.e. sudden, or slow
and progressive)
– nature of the initial dysfunction (e.g. loss of
memory, language alterations, problems to
execute motor action, and the incapacity for
recognizing objects, colors or situations)
– Information about changes in personality and
depressive symptoms
Data collection: Training
and Test sets
• In addition, without a unique methodology
to carry out the differential diagnosis of
AD and VD
• Findings generated by:
– (a) different tests (e.g. physical and
neurological exams, as well as blood tests)
– (b) a psychological interview
– (c) nutritional information
– (d) an evaluation of the vascular disease
Data collection: Training
and Test sets
• Demographic
– patient's age, sex, civil state, patient's education,
Occupation
• Antecedents
– smoke, alcoholism, hereditary antecedents, hypertension,
history of depressive states, etc.
• Symptoms and signs
– illness time evolution, if the patient has orientation
problems, changes in personality, problems with
numerical calculus, language problems, or psychotic
symptoms, etc.
Data collection: Training
and Test sets
• Neurological and neuropsychological scales
– patient's clinic history and a clinical exam
– Loeb scale (Loeb, 1988; Cummings, 1985)
• (in both scale was evaluated how the illness started)
– The neuropsychological tests
• (MMSE (Folstein et al., 1975);
• Geriatric Depression Scale (Mattis, 1976; Diaz &
Garcfa de la Cadena, 1993);
• Common Activities Scale (Khachaturain, 1985; Diaz &
Garcfa de la Cadena, 1993).
Data collection: Training
and Test sets
• Electrophysiolog
– EEG
– P300
• Neuroimaging analysis and other
studies
– Tomography(斷層掃描法) and Magnetic
Resonance analyses(核磁共振) are used
to valorize AD pathologies(DeLeon et al.,
1980, 1983; Fox et al., 1975)
Data collection: Training
and Test sets
• 58 paitents
• National Institute of Neurology and
Neurosurgery Manuel Velasco Sudrez
• These cases were organized in three sets:
– Set /----19 subjects diagnosed with VD.
– Set II 16 subjects diagnosed with AD.
– Set 111--23 subjects with diagnosis of
dementia (AD or VD).
Network architecture
and training parameters
Learning rate 0.1
Momentum 0.1
Initial weights 0.3
Error value to stop the training 0.0000002
46 neurons
29 neurons
2 neurons
Results
• a neural network was trained during
65 hours in order to reach the
minimum average error of 0.0000002
• we presented the data corresponding
to the 23 cases of the test set, and
only obtained the correct
classification of 19 cases, that is
an 82.6% efficacy.
Results
• Five networks classify correctly 21
of 23 test cases;
• Five other networks classify
correctly 20 of 23 test cases
• The network trained with data from
demographic records and scales
studies, produces the best results,
22 of 23 test cases were classified
correctly
New Network
• A correct classification was obtained for all 23
cases in the test set, that is, an efficacy of
100%.
conclusions
• In medicine, there are many illnesses
whose diagnosis is a very difficult
task, and people are still searching
for more efficient solutions
• This automata performs quite well:
– It presents a 100% efficacy
– it helps improve the efficiency in the
differential diagnosis of AD and VD
– it helps to reduce costs