Handwriting - marker for Parkinson’s Disease
P. Drotár et al.
Signal Processing LabDepartment of Telecommunications
Brno University of Technology
3rd SPLab Workshop, 2013
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 1 / 28
Motivation
Outline
1 MotivationParkinson’s Disease
2 Data
3 Handwriting features
4 Classification
5 Additional handwriting analysis
6 Conclusions
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 2 / 28
Motivation Parkinson’s Disease
Parkinson’s Disease
neurodegenerative disorder affecting mainly elderlyhigh prevalence rates (4.1 mil) - second most commonneurodegenerative disorder
difficult diagnosis
causes are unknowncauses under research include:genetics, age, toxins
syndroms: tremor , rigidity,bradykinesia
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 3 / 28
Motivation Parkinson’s Disease
Handwriting - marker for PD
bradykinesia ⇒ micrographia - characterized by decreased lettersize and by changes in kinematic aspects of handwriting
kinematic variables have been shown to be sensitive measures foralterations of handwriting movements
decreased letter size, decreased velocities of writing, increasedmovement time, . . .complete picture of the extent to which any one measurement orset of measurements is useful in predict is still missing
handwriting - sensitive measure of micrographia in PD
advantages over other approaches include:avoid usual pickles of speech signals acquisition and processingno need for special sensors (wearable sensors)
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 4 / 28
Data
Outline
1 MotivationParkinson’s Disease
2 Data
3 Handwriting features
4 Classification
5 Additional handwriting analysis
6 Conclusions
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 5 / 28
Data
Data acquisition
Recorded signals:x-y coordinatestime stamppressurebutton statusazimuthaltitude
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 6 / 28
Data
Template
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 7 / 28
Data
Subjects
Altogether, 75 subjects were enrolled at the First Department ofNeurology, St Annes University Hospital in Brno.
37 Parkinsonian patients (19 men/18 women) + 38 healthy controls (20men/18 women)
Table: Parkinson’s handwriting dataset characteristics
Age UPDRS (part V) Years since diag.mean std mean std mean std
PD 69.3 10.9 2.27 0.84 8.37 4.8H 62.4 11.3 - - - -
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 8 / 28
Handwriting features
Outline
1 MotivationParkinson’s Disease
2 Data
3 Handwriting features
4 Classification
5 Additional handwriting analysis
6 Conclusions
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 9 / 28
Handwriting features
In-air movement
Hand movement during handwriting = on-surface movement + in-airmovement
Hypothesis: In-air trajectory during handwriting contain informationreflecting syndromes of PD.
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 10 / 28
Handwriting features
In-air movement
Hand movement during handwriting = on-surface movement + in-airmovement
Hypothesis: In-air trajectory during handwriting contain informationreflecting syndromes of PD.
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 10 / 28
Handwriting features
In-air movement
Hand movement during handwriting = on-surface movement + in-airmovement
Hypothesis: In-air trajectory during handwriting contain informationreflecting syndromes of PD.
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 10 / 28
Handwriting features
In-air movement
Hand movement during handwriting = on-surface movement + in-airmovement
Hypothesis: In-air trajectory during handwriting contain informationreflecting syndromes of PD.
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 10 / 28
Handwriting features
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 11 / 28
Handwriting features
Handwrittig sample
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 12 / 28
Handwriting features
Sample recordings
1000 2000 3000 4000 5000 60001200
1300
1400
1500
1600
In−air movement
1000 2000 3000 4000 5000 60001000
1200
1400
1600
1800
On−surface movement
Figure: Handwriting sample of healthy control
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 13 / 28
Handwriting features
Sample recordings
4500 5000 5500 6000 6500 7000 7500 80001600
1700
1800
1900
2000
In−air movement
4500 5000 5500 6000 6500 7000 7500 80001400
1600
1800
2000
On−surface movement
Figure: Handwriting sample of PD patient
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 14 / 28
Handwriting features
Handwriting features
Feature Description
stroke speed trajectory during stroke divided by stroke durationspeed trajectory during handwriting divided by handwriting durationvelocity rate at which the position of a pen changes with timeacceleration rate at which the velocity of a pen changes with timejerk rate at which the acceleration of a pen changes with timehorizontal velocity/acceleration/jerk velocity/acceleration/jerk in horizontal directionvertical velocity/acceleration/jerk velocity/acceleration/jerk in vertical directionnumber of changes in velocity direction (NCV) the mean number of local extrema of velocitynumber of changes in acceleration direction (NCA) the mean number of local extrema of accelerationrelative NCV NCV relative to writing durationrelative NCA NCA relative to writing durationin-air time time spent in-air during writingon-surface time time spent on-surface during writingnormalised in-air time time spent in-air during writing normalised by whole writing duration
normalised on-surface timetime spent on-surface during writing normalised by wholewriting duration
in-air/on-surface ration ratio of time spent in-air/on-surface
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 15 / 28
Handwriting features
FeatureMutualInformation
CorrelationCoefficient
stroke speed(on surface, standard dev.)
6.09 -0.388
velocity(in air, standard dev.)
5.94 -0.387
vert. jerk(in air, min.)
5.7 0.383
acceleration(in air, standard dev.)
5.92 -0.38
horz. jerk(in air, range)
5.72 -0.379
jerk(in air, standard dev.)
5.96 -0.389
horz. acceleration(in air, range)
5.81 -0.375
horz. velocity(in air, range)
5.87 -0.371
horz. velocity(on surface, quantile 75%)
4.46 -0.37
vert. acceleration(in air, min.)
5.74 -0.369
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 16 / 28
Classification
Outline
1 MotivationParkinson’s Disease
2 Data
3 Handwriting features
4 Classification
5 Additional handwriting analysis
6 Conclusions
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 17 / 28
Classification
Support Vector Machines
To learn non-linearly separable functions, the data are implicitlymapped to a higher dimensional space by means of a kernel function,where a separating hyperplane is found. New samples are classifiedaccording to the side of the hyperplane they belong to.
RapidMiner
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 18 / 28
Classification
Numerical results
Feature selectionRelief algorithm
SVMRadial basis function kernelThe parameters kernel gamma γ, penalty parameter C andconvergence epsilon ε were optimized using grid search ofpossible values. Specifically, we searched over the grid (C, γ, ε)defined by the product of the sets C = [10−5,10−4, . . . ,103,104],γ = [10−5,10−4, . . . , , 102,103] and ε = [10−5,10−4, . . . ,102,103]
Classifier validation was conducted using a leave-one-outapproachThe process was repeated a total of 50 times, where in eachrepetition the original dataset was randomly permuted prior tosplitting into training and testing subsets
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 19 / 28
Classification
Relief FS; SVM classification
4 8 12 16 20 24 28 32 36 4055
60
65
70
75
80
85
Number of features used for classification
Cla
ssific
atio
n a
ccu
racy (
%)
in−air + on−surface
in−air
on−surface
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 20 / 28
Classification
Relief(50) + Sequential forward FS; SVM classification
Relief select subset of 50 featuresSFFS
evaluating all feature subsets which consist of only one inputattributeselects only the best k subsetscopies of the attribute set are made and exactly one of thepreviously unused attributes is add to the attribute setiteration algorithm continues and next unused feature is added
Results:Sequential forwardfeature selection
all features
in-air+on-surface 85.61% 68.83%in-air 84.43% 68.83%
on-surface 78.16% 71.31%
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 21 / 28
Additional handwriting analysis
Outline
1 MotivationParkinson’s Disease
2 Data
3 Handwriting features
4 Classification
5 Additional handwriting analysis
6 Conclusions
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 22 / 28
Additional handwriting analysis
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 23 / 28
Additional handwriting analysis
Feature analysis
FeatureSVM predictionaccuracy [%]
CorrelationCoefficient
stroke speed(task 8, standard dev.)
70.6 -0.39
stroke width(task 3, percentile 1st)
68.3 -0.40
horz. velocity(task 8, percentile 99th)
68.2 0.33
stroke width(task 3, mean)
66.7 -0.24
stroke length(task 3, percentile 1st)
65.3 -0.24
stroke length(task 3, standard dev.)
64 0.32
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 24 / 28
Additional handwriting analysis
evaluation of individual tasks (only on-surface)
TaskAccuracy
task 1 65.4task 2 70.0task 3 72.3task 4 65.4task 5 66.7task 6 67.7task 7 67.1task 8 78.7overall 79.4
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 25 / 28
Conclusions
Outline
1 MotivationParkinson’s Disease
2 Data
3 Handwriting features
4 Classification
5 Additional handwriting analysis
6 Conclusions
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 26 / 28
Conclusions
1 Handwriting as an tool for monitoring and diagnosis of PD
2 In-air movement - new modality PD evaluation
3 In-air + on-surface movement = clinically relevant classificationaccuracy (> 85%)
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 27 / 28
Appendix
I
ER Dorsey et alProjected number of people with Parkinson disease in the mostpopulous nations, 2005 through 2030..Neurology, 2007.
P. Drotár et al. (Brno University of Technology) Handwriting - marker for PD SPLab Workshop 2013 28 / 28