New Feature Construction - Department of Software and Information

advertisement
FROM TINNITUS DATA TO CLASSIFIERS
CONSTRUCTION:
Building Decision Support System
for Diagnosis and Treatment
of Tinnitus
Zbigniew W. Ras1 & Paul Jastreboff2 & Pamela Thompson1
1)
2)
University of North Carolina at Charlotte
College of Computing and Informatics
Tinnitus and Hyperacusis Center
Emory University School of Medicine
1
In collaboration with Jan Rauch
Department of Computer Science
University of Economics, Prague, Czech Republic
Research partially supported by the Project ME913
of the Ministry of Education, Youth, and Sports
of the Czech Republic
2
Methodology
◦ Domain Knowledge
◦ Data Collection
◦ Data Preparation
 New Feature Construction
 Tolerance Relation Based Clustering & New Temporal Features
 Classifiers Construction –
[for Total Score or Difference in Total Score]
 Action Rules Discovery [hints how to treat tinnitus]
 Future Research
From Music to Emotions and Tinnitus Treatment
4
Neil Young, Barbra Streisand, Pete
Townshend, William Shatner, David
Letterman, Paul Schaffer, Steve Martin,
Ronald Reagan, Neve Campbell, Jeff
Beck, Burt Reynolds, Sting, Eric
Clapton, Thomas Edison, Peter
Jennings, Dwight D. Eisenhower, Cher,
Phil Collins, Vincent Van Gogh, Ludwig
Van Beethoven, Charles Darwin, . . .
Introduction
5
Introduction
6

TRT includes
 DIAGNOSIS
◦ Preliminary medical examination
◦ Completion of initial interview questionnaire
◦ Audiological testing
◦ TREATMENT
◦ Counseling
◦ Sound Habituation Therapy
◦ Exposure to a different stimulus to reduce emotional reaction
◦ Visit questionnaire (THI)
◦ Secondary questionnaire (TFI) in new dataset
◦ Instrument tracking (instruments can be table top or in ear,
different manufacturers)
◦ Continued audiological tests
Methodology: Domain Knowledge
7

Original Dataset
◦ 555 patients
◦ Relational
◦ 11 tables

New Dataset
◦ 758 patients
◦ Relational
◦ Secondary questionnaire  Tinnitus Functional Index (TFI)
8
Initial Interview form provides basis for initial patient classification.
Category - 0 to 4 (stored in Questionnaires tables)
0 – low tinnitus only: counseling
1 – high tinnitus: sound generators set at mixing point
2 – high tinnitus w/hearing loss (subjective): hearing aid
3 – Hyperacusis: sound generators set above threshold
of hearing
4 – persistent hyperacusis: sound generators set at
the threshold; very slow increase of sound level
Methodology: Database Features
9
9
Tinnitus Functional Index
New cognitive and emotional questions
Scale of 0 to 10 and some %
Includes questions related to
Anxious/worried
Bothered/upset
Depressed
This new set of features is mapped to “arousal-valence
emotion plane” used for construction of emotion-based
classifiers in music information retrieval domain
(personalization aspects are considered as well).
Methodology: Database Features
10
10
10
Arousal-valence emotion plane - used in Automatic Indexing
of Music by emotions
11
11
New Features Based on the TFI and emotions
Table 2: Tinnitus Functional Index (scale of 0 to 10)
Category of Question
Q1
% of time aware
Awareness
Q2
loud
HEARING
Q3
in control
E11
E-V Scale
Q4
% of time annoyed
Annoyance
Q5
cope
E11
E1
Q6
ignore
E21
E2
Q7
concentrate
THINKING CONCENTRATION
Q8
think clearly
THINKING CONCENTRATION
Q9
focus attention
THINKING CONCENTRATION
Q10
fall/stay asleep
E33
E3
Q11
as much sleep
E33
E3
Q12
sleeping deeply
E33
E3
Q13
hear clearly
HEARING
Q14
understand people
HEARING
Q15
follow conversation
HEARING
Q16
quite, resting activities
E41
E4
Q17
relax
E43
E4
Q18
peace and quiet
E42
E4
Q19
social activities
SOCIAL
Q20
enjoyment of life
E11
Q21
relationships
SOCIAL
Q22
work on other tasks
SOCIAL
Q23
anxious, worried
E23
E2
Q24
bothered upset
E22
E2
Q25
depressed
E31
E3
E1
E1
Sum of values represents E1 Energetic Positive, E2 Energetic Negative, E3 Calm Negative, E4 Calm Positive
New Feature Construction: TFI and Emotions
12

Tinnitus Handicap Inventory
◦
◦
◦
◦
Questionnaire, forms Neumann-Q Table
Function, Emotion, Catastrophic Scores
Total Score (sum)
THI
 0 to 16: slight severity
 18 to 36: mild
 38 to 56: moderate
 58 to 76: severe
 78 to 100: catastrophic
Methodology: Database Features
13
New 8 decision attributes based on different discretizations
of the difference in Total Score (between first and last visit)
Total Score
Difference
Description
Discretization
(score a represents the highest T Score in all cases)
TSa
a= {s: s>0}, b= {0} , c = {s: s < 0}
TSb
a={ s: s>30}, b ={s: 10 < s  30}, c={s: -10 < s  10},
d={s: -40 < s  -10}, e – remaining scores
a={s : s > 28}, b={s: 0 < s  28}, c ={s: -1 < s  0},
d ={s: -15 < s  -1} , e – remaining scores
a={s: s > 40}, b={s: 10 < s  40}, c={s: -10 < s  10},
d={s: -40 < s  -10}, e – remaining scores
a={s: s > 50}, b={s: 0< s  50}, c={s: -50< s  0}, d – remaining scores
TSc
TSd
TSe
TSg
a={s: s > 80}, b={s: 60< s  80}, c={s: 40<s  60}, d={s: 20 < s  40},
e ={s: 0< s  20}, f={s: -20 < s  0}, g={s: -40< s  -20},
h={s: -60 < s  -40}, i – remaining scores
a={s: s > 28}, b={s: 0 < s  28}, c={s: -12 < s  0}, d – remaining scores
TSh
a ={s: s> 10}, b={s: -10  s  10}, c – remaining scores
TSf
New Feature Construction: Decision Feature
14

Data Transformation – ORIGINAL DATABASE
◦ Flattened File (by Patient) From original database,
one tuple per patient with addition of features
◦ Discovered from Text Data
◦ Statistical (standard deviations, averages, ..)
◦ Temporal (sound level centroid, sound level spread,
recovery rate)
◦ Decision Feature – discretized Difference in Total Score
from THI

Data Transformation – NEW DATABASE
 Clustered patient-driven datasets (by similar visit
patterns) with addition of features
 Coefficients, angles
15
16

Text Mining
◦ Text fields
 Demographic, Miscellaneous, Medication tables
 Categories may show cause of tinnitus for patient
 Stress, Noise, Medical:
New Boolean Features Stress, Noise, and Medical Based on
Text Mining of Terms
Stress
stress, depression, emotion, work, marriage, wedding
Noise
accident, noise, concert, loud, music, shooting, blast
Medical
surgery, infection, medicine, depression, hospital
New Feature Construction: Text Features
17

New Temporal Features
◦ Sound Level Centroid
T = Total number of Visits per patient (3)
V is some sound level feature (ex. LDL measurement)
measured at each visit V(1), V(2), V(3)
1/3*V(1) + 2/3 * V(2) + 3/3 * V(3)
V(1) + V(2) + V(3)
New Feature Construction: Temporal Features
18

New Temporal Features
◦ Sound Level Spread
SQRT V(1) * (1/3-C)2 + V(2) * (2/3-C)2 + V(3) * (3/3 – C)2
V(1) + V(2) + V(3)
New Feature Construction: Temporal Features
19

New Temporal Features
◦ Recovery Rate
V0  Vk
T k  T0
, k  min
V i , i  [ 0 , N ]
V = Total Score from THI
Vo = first score (should be less than Vk)
Vk is the best or min score in the vector
Tk is the date of best score
New Feature Construction: Temporal Features
20

In Search for Optimal Classifiers
describing Total Score or changes in Total
Score [new decision attributes]
◦ WEKA
◦ J48 (C4.5 Decision Tree Learner)
◦ Random Forest
◦ Multilayer Perceptron
Data Mining: Unclustered Data
21

Experiments and Results
 1) Original data with Standard Deviations and Averages from
Audiological features
 2) Original data with Standard Deviations, Averages, Sound level
centroid and sound level spread (Sound) only
 3) Original data with Standard Deviations, Averages, and Text
 4) Original Data Standard Deviations, Averages, Text and Sound
 5) Original Data with Text
 6) Original Data with Sound
 7) Original Data with Sound, Text, and Recovery Rate
 8) Original Data with Sound, and Recovery Rate /the winner/
 9) ……………………………………….
Data Mining: Unclustered Data
22
Top Classification Results for all 8 decision variables
Original Data with Sound Level Centroid,
Sound Level Spread, Recovery Rate
0.9
0.8
0.7
Tsa
0.6
TSb
0.5
TSc
0.4
TSd
0.3
Tse
0.2
TSf
0.1
TSg
TSh
0
J48
RF
Precision
MP
J48
RF
Recall
MP
J48
RF
MP
Fmeasure
Data Mining: Unclustered Data
23

Continuing the Search for Optimal
Classifiers
◦ Transformation to Visit Structure
◦ Creating Tolerance-Relation based Datasets
◦ Adding New Features
Two groups of databases: three and
four visit centered sets were constructed.
Data Mining: Clustered Data
24
25
Coefficients and Angles Feature Construction for Dp
where p is a patient with 4 visits:
Clustering Techniques for Temporal Feature Extraction
26
27
Quadratic Equation Based New Features
Clustering Techniques
28
Clustering Techniques
29
Eight new decision attributes based on different
discretizations of Differences in Total Score
Total Score
Difference
Description
Discretization
(score a represents the highest T Score in all cases)
TSa
a= {s: s>0}, b= {0} , c = {s: s < 0}
TSb
a={ s: s>30}, b ={s: 10 < s  30}, c={s: -10 < s  10},
d={s: -40 < s  -10}, e – remaining scores
a={s : s > 28}, b={s: 0 < s  28}, c ={s: -1 < s  0},
d ={s: -15 < s  -1} , e – remaining scores
a={s: s > 40}, b={s: 10 < s  40}, c={s: -10 < s  10},
d={s: -40 < s  -10}, e – remaining scores
a={s: s > 50}, b={s: 0< s  50}, c={s: -50< s  0}, d – remaining scores
TSc
TSd
TSe
TSg
a={s: s > 80}, b={s: 60< s  80}, c={s: 40<s  60}, d={s: 20 < s  40},
e ={s: 0< s  20}, f={s: -20 < s  0}, g={s: -40< s  -20},
h={s: -60 < s  -40}, i – remaining scores
a={s: s > 28}, b={s: 0 < s  28}, c={s: -12 < s  0}, d – remaining scores
TSh
a ={s: s> 10}, b={s: -10  s  10}, c – remaining scores
TSf
New Feature Construction: Decision Feature
30
Classifiers Construction [learning differences in total score]
for clustered data:
J48, Random Forest, and Multilayer Perceptron
(Neural Network) have been tested on the cluster-based
original datasets with:







1)
2)
3)
4)
5)
6)
7)
standard deviations and averages,
coefficients and text,
coefficients and angles,
coefficients only,
angles only,
angles and text,
angles, coefficients and text /the winner/.
Data Mining: Clustered Data
31
Data Mining: Clustered Data
32

Results are quite encouraging
◦ Top precision is .884
◦ This represents an improvement over the classification
precision of .751 with J48 classification on the original dataset
and features Sound Level Centroid, Sound Level Spread and
Recovery Rate being present
Summary Data Mining: Clustered Data
33
Action Rules
34
Action rule is defined as a term
[(ω) ∧ (α → β)] →(ϕ→ψ)
conjunction of fixed condition
features shared by both groups
Information System
A
B
D
proposed changes in values
of flexible features
a1 b2 d1
desired effect of the action
a2 b2
a2 b2 d2
Action Rules
35

New Decision Feature
◦ Boolean features + or – related to a feature
such as Total Score improving or getting worse
 Calculated from score on next visit
 Stored as + or – on visit related tuple
New Feature Construction: Decision Features
showing change over time
36
Rules using LISpMiner
ACTION RULES: EXPERIMENT AND RESULTS
37
Analysis:
Before confidence: 9/9+0
After confidence: 9/ [9+20]
Low confidence but shows promise
ACTION RULES: EXPERIMENT AND RESULTS
38
Summary
39
Continue Action Rule Study
 Develop GUI for patient data entry
 Use knowledge gained from rules to
develop decision support system for
treatment support for tinnitus sufferers
 Continue research with music, emotions,
and tinnitus treatment

Future Research
40
Download