1. Machine Learning for Alzheimer Disease
Prediction
Oleksandr Kazakov.
Big Data Dev @Wajam
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 1 / 41
2. Alzheimer Quick Facts
More than 5 million people in US and 35 million worldwide are living
with Alzheimer’s
Every 66 seconds someone in the US develops disease
Kills more than breast and prostate cancer combined
6th leading cause of death in the US and Canada
No current cure
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 2 / 41
3. Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 3 / 41
6. Why Raman Spectroscopy?
It is cheap!
It can provide fingerprinting-type information on the total biochemical
state of blood or cerebrospinal fluid.
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 6 / 41
7. Why Raman Spectroscopy?
It is cheap!
It can provide fingerprinting-type information on the total biochemical
state of blood or cerebrospinal fluid.
It is super cheap and...portable(true if number of handles<=2).
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 6 / 41
17. Biological Neuron
yi = f(
d
i
wixi + w0)
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 16 / 41
18. Feed Forward Neural Network with Backpropagation
Learning
Activity of ith neuron:
xi = f( i)
Where:
i = woi +
j⊂Γi
wijxj
Applying chain rule to compute each weight in the network for any
arbitrary error function E
∂E
∂wij
=
∂E
∂ i
∂ i
∂neti
∂neti
∂wij
Once we know partial derivative of each weight the minimization is done
via gradient descent:
wij(t + 1) = wij − λ
∂E
∂wij
(t)
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 17 / 41
19. Typical three layers feed-forward neural network
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 18 / 41
22. f(x) =
N
i
wih(x)i
Most Popular choices of radial basis function are:
Gaussian: e
−( v2
β2 )
Multiquadric: (v2 + β2)
1
2
Inverse Multiquadric: 1
(v2+β2)
1
2
Where v = x − c and represents distance from the input vector to the
radial basis function centre, and β is a width of a radial basis function.
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 20 / 41
25. Some Theory
Classifier for SVM:
F(x) = sign(
n
j=1
ωjxj − ω0) = sign( ω, x − w0) (1)
Where x = (x1, ...., xn)- evidential description of an object X; vector
ω = (ω1, ...., ωn) and value of ω0 are some parameters of the algorithm.
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 23 / 41
26. Some Theory
Classifier for SVM:
F(x) = sign(
n
j=1
ωjxj − ω0) = sign( ω, x − w0) (1)
Where x = (x1, ...., xn)- evidential description of an object X; vector
ω = (ω1, ...., ωn) and value of ω0 are some parameters of the algorithm.
If set is linear separable then error function can be described as following:
E(ω, ω0) =
l
i=1
[yi( ω, xi − ω0) < 0] (2)
Such condition gives a numerous sets of hyperplanes but we want the one
that will be a maximum distance from the separable set.
−1 < ω, x − ω0 < 1
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 23 / 41
28. Width of The Separation Plane
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 25 / 41
29. Non-linear SVM Classification
The idea is to map given feature space X to the new H using transition
ψ : X → H and if H has enough dimensionality we can hope that set will
be separated. Kernel will be defined as K(x, x ) = ψ(x), ψ(x ) and
classifier can be rewritten as:
F(x) = sign(
l
j=1
λiyiK(xi, x) − w0)
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 26 / 41
30. Non-linear SVM Classification
The idea is to map given feature space X to the new H using transition
ψ : X → H and if H has enough dimensionality we can hope that set will
be separated. Kernel will be defined as K(x, x ) = ψ(x), ψ(x ) and
classifier can be rewritten as:
F(x) = sign(
l
j=1
λiyiK(xi, x) − w0)
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 26 / 41
31. Types of Kernel
Polynomial: K(x, x ) = ( x, x + const)d Example
Radial Basis Function: K(x, x ) = e−γ x−x 2
, γ > 0
Gaussian Radial basis function: K(x, x ) = e−0.5 x−x 2(σ−2)
Sigmoid: K(x, x ) = tanh(k x, x + c), k > 0, c > 0
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 27 / 41
32. Similarity with Neural Networks
SVM structure:
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 28 / 41
34. Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 30 / 41
35. GA Basics
Randomly choose initial population P of N individuals.
Determine the fitness of each individual.
Assign probability of reproduction pi to each individual.
Create a new population selecting individuals according to pi. Then
selected individuals will generate offsprings via crossovers or
mutations.
Repeat if needed.
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 31 / 41
37. RBF and MLP result for Serum
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 33 / 41
38. SVM for AD vs HC
Class Sensitivity(%) Specificity(%)
Healthy Control 87 90
Alzheimer dementia 84 87
Healthy Control 81 93
Mild Alzheimer dementias 67 90
Moderate Alzheimer dementias 74 94
Healthy Control 86 88
Mild Alzheimer dementia 88 87
Healthy Control 82 83
Moderate Alzheimer dementia 81 82
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 34 / 41
39. MLP and SVM for AD vs HC for CSF
For MLP:
Class Sensitivity(%) Specificity(%)
Healthy Control 81 83
Mild Alzheimer dementia 82 81
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 35 / 41
40. MLP and SVM for AD vs HC for CSF
For MLP:
Class Sensitivity(%) Specificity(%)
Healthy Control 81 83
Mild Alzheimer dementia 82 81
And even below 74% Sensitivity and Specificity for SVM.
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 35 / 41
41. MLP for AD vs OD for Serum
Class Sensitivity(%) Specificity(%)
Other dementias 95 98
Alzheimer dementia 96 95
Other dementias 92 97
Mild Alzheimer dementias 95 92
Moderate Alzheimer dementias 93 99
Other dementias 91 94
Mild Alzheimer dementia 84 82
Other dementias 96 95
Moderate Alzheimer dementia 92 99
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 36 / 41
42. SVM for AD vs OD for Serum
Class Sensitivity(%) Specificity(%)
Other dementias 90 95
Alzheimer dementia 95 90
Other demetias 94 87
Mild Alzheimer demetias 89 94
Moderate Alzheimer demetias 93 92
Other dementias 92 93
Mild Alzheimer dementia 84 82
Other dementias 92 93
Moderate Alzheimer dementia 93 92
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 37 / 41
43. Selected Networks Architectures
Type of Network Network Structure Sensitivity(%) Specificity(%)
MLP 5-20-20-1 97 95
MLP 5-50-10-1 96 96
MLP 5-100-1 96 95
MLP 5-200-20-1 96 94
RBF 5-20-20-1 95 94
RBF 5-50-20-1 96 95
RBF 5-100-1 94 94
RBF 5-200-20-1 94 94
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 38 / 41
44. Genetic Algorithm Results for Serum
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 39 / 41
45. Genetic Algorithm Results for CSF
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 40 / 41
46. Conclusions
We demonstrated power of Raman Spectroscopy as diagnostic tool
for AD detection.
We demon- strated the high probability for single subject to be
correctly assigned to the particular diagnostic category.
Our results are consistent with current AD biomarker findings.
Oleksandr Kazakov. Big Data Dev @Wajam Machine Learning for Alzheimer Disease Prediction 41 / 41