Identification of Diagnostic

1
Identification of Diagnostic Evoked Response
Potential Segments in Alzheimers Disease
Benvenuto, Jin, Casale, Lynch, Granger, Experiment
al Neurology 176, 269-276 (2002)
2
Background

• EEG coherence is thought to be a measure of
  regional cortical synchronization and possibly
  the functional status of intracortical
  communication.
• Since the early, and especially in the late 90s,
  a number of quantitative EEG (mostly coherence)
  measures have been used in attempt to identify
  physiological correlates of the cognitive changes
  found on early stages Alzheimers disease
• Slowing of spectral EEG predicts the rate of
  subsequent cognitive and functional decline in
  patients with AD (multiple linear regression
  analysis).
• Patients with AD had significantly lower intra-
  and interhemispheric coherence than controls in
  the alpha and beta frequency bands. AD patients,
  particularly those with severe cognitive
  impairments, have reduced alpha band coherence in
  temporo-parieto-occipital areas.
• Further evidence linking coherence to the
  evolution of AD comes from results suggesting
  that patients homozygous for the Apo-E epsilon4
  allele, a predisposing condition for sporadic
  Alzheimers, have particularly reduced bilateral
  coherence in select cortical fields.
• Other EEG indices (such as the EEG complexity)
  are reported to discriminate AD from normal
  controls or other types of dementia, as well as
  to differentiate subgroups among AD patients.

3
The novelty
The evolved method uses projection pursuit
algorithms to search for differentially
diagnostic segments within the time locked
signals, with correlated co-occurrences of
segments used as composite features in
classification. Because time-locked signals are
required, evoked response potentials (ERPs) to
photic driving were used in the studies instead
of free running EEG. The results indicate that
application of iterative projection pursuit to
ERPs can be used to recognize AD with a high
degree of accuracy.
4
Methods

• Patient Population
• 15 AD patients unhospitalized, unmedicated,
  aged 76.2 5.7 years. The severity of illness
  was limited to mild or moderate, based on a
  screening tests.
• Normal Controls
• 17 normal subjects matched for age, gender, and
  education level were recruited from the
  community. Each subject was interviewed by a
  research psychiatrist rule out AD and other
  psychiatric diagnoses.
• Exclusion Criteria
• Severe or unstable disease other than AD
• Medical or psychiatric disorders that might
  complicate the assessment of dementia
• A disability that may prevent the subject from
  completing all study requirements (e.g.,
  blindness, deafness, language difficulty)
• Recent intake of an investigational drug, drug
  known to cause major organ system toxicity, any
  CNS-active medication, or any recreational drug.

5
Methods
Apparatus Electrode cap, amplifier system, PC,
bright light source. ERPs were collected from 19
sites on the skull through scalp electrodes
embedded in a tight-fitting meshwork cap.
Procedures Subjects were acclimated to the
apparatus for 5 min during which time the quality
of each of the 19 leads was checked. Once normal
voltage EEG was recorded from all sites, stimuli
(visual light flashes) were presented at 60 per
minute (1 Hz) for a session of 5 min in duration,
and continuous ERPs were collected.
Illustration Anonymous ERP subject.
6
Methods

• Vectorizing subjects
• For each subject, for each electrode channel
• Split the 300 sec of measurements into segments
  of 1 sec each

1. Average the 1 sec segments

Average

1. Bin the resulting average into 5-ms segments
   (resulting in 200 values)

Binning
(v1, v2, , v200)
The values from the 19 channels conflated into 1
vector (of 3800 elements)
(u1, u2, , u200)
(u1, , u200, v1, , v200, w1, , w200) (e1,
e2, , e3800)
Conflating
(v1, v2, , v200)
(w1, w2, , w200)
7
Methods

• Analyses
• After a single vector was constructed for each
  subject, jackknife analyses were run, in which
  all the subject records but one were used as
  matching data, and the remaining subject was
  tested to see which category (Alzheimers or
  control) that subject would be placed in. Three
  classification algorithms with three parameter
  settings (k1,3,5) were used
• k-nearest-neighbors analysis.
• Projection pursuit analysis.
• Extended projection pursuit analysis.

8
Methods
k-nearest-neighbors
k nearest neighbors of the subject under
investigation are being chosen. The diagnosis is
set according to the neighbors majority
type. The distance metric is based on a
Mahalanobis distance (norm)
C
C
C
C
A
C
C
C
A
?
A
A
C
A
(S the covariance matrix.)
A
A
The analysis was performed for k1,3,5.
Illustration The k here is 3. The subject under
investigation will be classified as Control. In
reality, the number of dimensions is 3800, not 2.
9
Methods
Projection pursuit
Subsets of the subject-vectors were randomly
generated, focusing on a few voltage values out
of 3800. The subsets are equivalent to a
projection on a lower-dimension space. In each
subspace, k nearest neighbor was performed as
before. Votes for Alzheimers versus Control
were tallied across all subspaces, and the
resulting majority classification was used as a
diagnosis. The analysis was performed for
k1,3,5.
Illustration Two projections of the same 3D data
set.
10
Methods
Extended projection pursuit
The previous projection pursuit procedure was
performed. Based on its findings, the most
predictive subspaces were selected and the
process performed again this iterative
compilation of subspaces continued until all
subspaces chosen were more predictive than a
preselected threshold amount. The resulting
majority classification was used as a
diagnosis. The analysis was performed for
k1,3,5.
subject vector
PPKNN
Selection
PPKNN
Selection
Illustration Iterative PP, KNN and sub-space
selections.
11
Results
Discrimination of AD from control The sensitivity
statistic gives attention to the rate of the AD
patients diagnosed

• For k nearest neighbors (k1,3,5), the
  sensitivity does not exceed 25. The maximum
  false positive rate is 12.
• For projection pursuit methods, the best
  sensitivity is 75, with corresponding 29 false
  positive rate.
• For all extended projection pursuit methods, the
  sensitivity is 100, with a false positive rate
  of 6.1 (one subject).

12
Results
Plot of Sensitivity vs. False Positive Rate
13
Results
Temporal Location of Predictive Features in the
ERP The predictiveness is defined as the
percentage of times the segment is used in
correct prediction. The following typical table
represents the predictiveness of 100-ms segments
originating from C4 electrode.
14
Results
Spatial and Frequency Location of Predictive
Features The following figure shows the relative
power in each of four frequency bands (delta,
theta, alpha, beta) for averages of Alzheimers
and matched controls, plotted across the 19
electrodes of the headset apparatus.
15
Authors' Summary

• Extended pursuit projection identifies
  correlates of AD in ERPs elicited by simple
  visual stimuli (sensitivity of 100, false
  positive rate of 6).
• The most distinctive features occurred within
  two temporal segments (200 and 400 ms and from
  800 to 1000 ms) and arose from fronto-parietal
  recording sites.
• There is prior evidence indicating that
  correlates of mild AD are found within these
  spatio-temporal coordinates. Although there was
  evidence that simple learning contributed to the
  observed differentiation of the AD group, the
  unstructured stimuli used in the study have the
  disadvantage of not activating cognitive
  activities thought to be impaired by AD.
• Between-group differences could be enhanced, and
  probably markedly so, with paradigms that engage
  attention to novelty or working memory. On the
  other hand, unstructured cues have the important
  advantages of test simplicity and applicability
  across patient populations.

16
Conclusions and criticism

• The number of subjects is small, the number of
  methods tried is at least 9.
• The projection pursuit procedure is unclear.
  What does it mean randomly generated subspaces,
  few voltage values?
• The presented projection pursuit method
  resembles deduction based on bagging (or
  majority voting) rule. There is no attempt at
  projection pursuit optimization via
  index/objection functions (such as the ones
  suggested by Friedman-Tukey Jones-Sibson
  Intrator-Cooper, Hebb-Oja). The only article
  quoted on the subject of projection pursuit is
  from 1985.
• It is implied that the same significant spaces
  are shared by and calculated across the different
  jackknifed subjects, which means that the test
  subjects are influencing the results! If that is
  the case, this involves a risk of circular
  reasoning.
• And yet the potential seems to be there (the
  results may be ok, even if their derivation was
  problematic.). A more elaborate projection
  pursuit, or other clustering methods carefully
  applied might yield more founded results.

17
Time for Questions?!
18
Thank You.