Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Swathi KiranSpeech & Hearing Sciences; Boston University
Department of Neurology; Massachusetts General HospitalCommunication Sciences & Disorders; University of Texas at
Austin
A computational account of bilingual aphasia rehabilitation
Funding support from NIH/NIDCD: R21 DC009446; ASHF-Clinical Research Grant, ASHF New Investigator Grant
MASSACHUSETTSGENERAL HOPSITAL
What is aphasia? Aphasia is characterized by language deficits such as problems
speaking, understanding people, reading and writing
Approximately 80,000 people incur aphasia each year
It is estimated that 60% of the world is bi/multi-lingual
Bilingual Aphasia
Schematic of treatment for each participant
Session 1: Training
Session 2: Testing & Training
Session 1: Training
Session 2: Testing & Training
Session 1: Training
Session 2: Testing & Training
Pre –treatment assessment: Western Aphasia Battery, BNT, Bilingual Aphasia Test
Treatment on 1 set of examples in 1 language
Session 1: Training
Session 2: Testing & Training
Session 1: Training
Session 2: Testing & TrainingUntil 80% accuracy achieved on items trained
Week 1
Week 2
Week 3
Week 4
Post –treatment assessment:
Standardized language tests
. . . . . .
Baselines: Naming across consecutive sessions & languages
Edmonds & Kiran, (2006) JSLHR
0102030405060708090
100
1 3 5 7 9 11 13 15 17 19 21 23 25
Perc
ent a
ccur
acy
Weeks
Treatment in English English Set 1
Spanish Set 1
• Obviously, this translates to an increase in clinical need to address bilingual aphasia rehabilitation, but no clear guidelines on how to do so…
• No consistent results on rehabilitation of bilingual aphasia (Lorenzen & Murray, 2008; Faroqi-Shah et al., 2010)
• Few systematic studies that have examined and observed the extent of cross language transfer but results vary• (Croft et al., 2011; Edmonds & Kiran, 2006; Miertsch et al., 2009,
Kiran & Roberts, 2009)
Bilingual Aphasia Rehabilitation
Stroke
L1 language exposure L1 Post stroke impairment
Schema for Bilingual Aphasia
L1 AoA L2 language exposure L 2 Post stroke impairmentL2 AoA
Pre-stroke proficiency
Is there another way to understand the nature of bilingual aphasia
rehabilitation?
Develop a computational simulation of bilingual aphasic naming deficits and rehabilitation of bilingual aphasia.
Similar to predicting rehabilitation of naming deficits (Plaut, 1996) which has led to the complexity account of treatment deficits for naming deficits (Kiran, 2007)
Computational Modeling: SOM
Self Organizing Maps (Kohonen, 1995) operate in two modes Training -builds the map using input examples Mapping- classifies a new input vector
SOMs have been used to understand bilingual language learning (Li,
Zhao & McWhinney, 2007) and biological/psychiatric conditions (Hamalainen,1994; Hoffman, Grasemann, & Miikkulainen, 2011)
Semantic map
English phonetic map Spanish phonetic map
The Bilingual DISLEX Model
Grasemann et al., 2011; Mikkulainen & Kiran, 2009
Semantic representations260 hand‐coded binary features
Phonetic representations•Based on English and Spanish IPA transcriptions•Numerical representations of phonemes
The Bilingual DISLEX Model
Semantic map
English phonetic map Spanish phonetic map
Grasemann et al., 2011; Mikkulainen & Kiran, 2009
Naming Task in Bilingual DISLEX Model
Semantic map
“Dog”English phonetic map Spanish phonetic map
Grasemann et al., 2011; Mikkulainen & Kiran, 2009
Model of Bilingual Lexical Access
Semantics
SpanishEnglish
Asymmetrical Model(Kroll & Stewart, 1994
Kroll et al., 2010)
Step 1
• Model pre‐stroke/normal bilingual language performance• Use AoA and exposure as training parameters• DISLEX should be able to match pre‐stroke English and Spanish performance
Step 2
• Simulate damage to the lexicon• Distort associative connections with noise• DISLEX should be able to model impairment in patients
Step 3• Use the model to predict treatment outcomes• Examine improvements in trained language and cross language transfer
Develop a computational simulation of bilingual aphasic naming deficits and rehabilitation of bilingual aphasia.
Step 1
• Model pre‐stroke/normal bilingual language performance• Use AoA and exposure as training parameters• DISLEX should be able to match pre‐stroke English and Spanish performance
Grasemann et al., 2011; Mikkulainen & Kiran, 2009
Approach
L1 language exposure
L1 AoA
L2 language exposureL2 AoA
Pre-stroke proficiency
English performance: 80- 96%
Spanish range: 65% - 92%
Information about AoA, Language exposure, proficiency obtained from a language use question – Kiran et al.( 2010, submitted)
AoAEarly: < 10 years
ExposureHigh > 60%
Results of simulation of normal bilingual individuals
(Grasemann et al., 2010; Grasemann et al., 2011)
Step 2
• Simulate damage to the lexicon• Distort associative connections with noise• DISLEX should be able to model impairment in patients
Simulation of bilingual aphasia-DISLEX Model Lesion was applied to the
connections from the semantic map to the phonetic maps
Adding Gaussian noise with μ = 0 to all these connections.
The amount of damage (the “lesion strength”) in each case was adjusted by changing the \sigma (σ) of the noise between 0 and 1.0 in steps of 0.01.
Grasemann et al., 2011; Kiran et al., 2010
Semantic map
English phonetic map Spanish phonetic map
More damage Pre‐stroke
Results from DISLEX Model – Modeling Impairment in one patient
Grasemann et al., 2011; Kiran et al., 2010
Results DISLEX Model– Modeling Impairment in one patient
More damage Pre‐stroke
Grasemann et al., 2011; Kiran et al., 2010
Patient BNT scoresEnglish = 35%Spanish = 1.7%
DISLEX Model: Modeling impairment for 16 patients with aphasia
L1 language exposure
L1 AoA
L2 language exposureL2 AoA
Pre-stroke proficiency
StrokeL1 Post stroke impairment
L 2 Post stroke impairment
Lesion damage
Accounting for pre-stroke proficiency and lesion damage adequately simulates naming impairment in bilingual aphasia
Step 3• Use the model to predict treatment outcomes• Examine improvements in trained language and cross language transfer
Patient Study 3: (N = 17)AOA Lifetime Exposure Treatment Effect Size
English Spanish English Spanish English Spanish
UTBA01 Native native high low 12.70 0.58
UTBA02 late native low high 4.95 11.08
UTBA07 native native moderate moderate 3.11 12.41
UTBA09 early native moderate moderate 2.07 10.97
UTBA11 late native moderate high 14.90 1.15
UTBA16 native native high low 6.82 0.83
UTBA17 early native high low 5.32 1.19
UTBA18 late native moderate moderate 1.73 15.17
BUBA01 late native low high 4.92 1.42
BUBA04 early native high low 2.61 16.50
BUBA07 late native low high 2.89 4.08
UTBA19 late native low high 1.44 4.90UTBA20 late native low high 0 0UTBA21 early native 0 0UTBA22 late native low high 0.13 12.73UTBA23 early native low high 10.68 13.84BUBA12 late native low high 8.16 0
Schematic of treatment for each participant
Session 1: Training
Session 2: Testing & Training
Session 1: Training
Session 2: Testing & Training
Session 1: Training
Session 2: Testing & Training
Pre –treatment assessment: Western Aphasia Battery, BNT, Bilingual Aphasia Test
Treatment on 1 set of examples in 1 language
Session 1: Training
Session 2: Testing & Training
Session 1: Training
Session 2: Testing & TrainingUntil 80% accuracy achieved on items trained
Week 1
Week 2
Week 3
Week 4
Post –treatment assessment:
Standardized language tests
. . . . . .
Baselines: Naming across consecutive sessions & languages
Edmonds & Kiran, (2006) JSLHR
0102030405060708090
100
1 3 5 7 9 11 13 15 17 19 21 23 25
Perc
ent a
ccur
acy
Weeks
Treatment in English English Set 1
Spanish Set 1
Treatment protocol in Behavioral studies
1. Name picture2. If incorrect, told correct name 3. Choose 6 correct features from 12
cards4. Answer 15 yes/no questions about
the item 5. Named item again with feedback
Treatment always provided only in one language (either English/Spanish) and amount of improvement examined
Generalization (cross language transfer) examined to untrained language
L2L1
“Celery” “Apio”
TREATMENT
Long and green.
Vegetable
Crunchy
Found in produce section
Eaten Fresh
Nutritious
Edmonds & Kiran, 2006; Kiran & Roberts, 2009
Rehabilitation in the DISLEX Model The starting point was set to either a severe
impairment in naming (30% or less accuracy) or mild impairment (70% or high naming accuracy).
Model retrained trained with different number and schedule of presentations of words in one language
Treatment always provided only in one language (either English/Spanish) and amount of improvement examined
Generalization (cross language transfer) examined to untrained language
L2L1
“Celery” “Apio”
TREATMENT
Long and green.
Vegetable
Crunchy
Found in produce section
Eaten Fresh
Nutritious
Edmonds & Kiran, 2006; Kiran & Roberts, 2009
In order to evaluate the modelMatch the patient and model’s parameters on AoA,
exposure and damage parameters and see if the model’s predictions match the actual data obtained.
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
UTBA01 UTBA02 UTBA07 UTBA09 UTBA11 UTBA16 UTBA17 UTBA18 UTBA19 UTBA22 UTBA23 BUBA01 BUBA04 BUBA07 BUBA12
Cross Correlation r with Model Trained Language
Cross Correlation r with Model Untrained Language
Cross-correlation coefficient > .6 significant; > 2 SD errors
Patient and computational results:Both languages high damage
0
10
20
30
40
50
60
70
80
90
100
1 3 5 7 9 11 13 15 17 19 21 23 25
Perc
ent a
ccur
acy
Treatment in EnglishEnglish Set 1Spanish Set 1
UTBA01
UTBA 01: English: EarlySpanish: Native
English: High exposureSpanish: Low exposure
Spanish ES: .58English ES: 12.7
0
10
20
30
40
50
60
70
80
90
100
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Perc
ent a
ccur
acy
Probes
Treatment in English English Set 1Spanish Set 1UTBA16
UTBA16: English: EarlySpanish: Native
English: Moderate exposureSpanish: Moderate exposure
Spanish ES: .83English ES: 6.8
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Perc
ent a
ccur
acy
Probes
Training in Spanish English Set 1
Spanish Set 1-TrainedBUBA04
BUBA04: English: EarlySpanish: Native
English: High exposureSpanish: Low exposure
Spanish ES: 16.5English ES: 2.52
Patient and computational results:Both languages low damage
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Perc
ent a
ccur
acy
Probes
Treatment in EnglishEnglish Set 1-TrainedSpanish Set 1BUBA01
BUBA01 English: LateSpanish: Native
English: Low exposureSpanish: High exposure
Spanish ES: 1.42English ES: 4.92
Patient and computational results:Both languages differential
damage
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Perc
ent a
ccur
acy
Probes
Training in English English Set 1(trained)Spanish Set 1
UTBA 17:
English: EarlySpanish: Native
English: Moderate exposureSpanish: Moderate exposure
Spanish ES: 5.32English ES: 1.19
UTBA 22: English: LateSpanish: Native
English: Low exposureSpanish: High exposure
Spanish ES: 12.7English ES: 1.89
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Perc
ent a
ccur
acy
Probes
Treatment in Spanish English Set 1
Spanish Set 1UTBA 22
The model also does not always predict correct performance
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
Perc
ent a
ccur
acy
Probes
Treatment in English English Set 1
Spanish Set 1BUBA07
BUBA07English: LateSpanish: Native
English: Low exposureSpanish: High exposure
Spanish ES: 4.08English ES: 2.8
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
80.00
90.00
100.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Perc
ent a
ccur
acy
Probes
Treatment in EnglishEnglish Set1
UTBA11English: LateSpanish: Native
English: High exposureSpanish: High exposure
Spanish ES: 1.15English ES: 14.9
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Perc
ent a
ccur
acy
Probes
Treatment in Spanish English Set 1
Spanish Set 1 -trained
UTBA20
UTBA 09:English: LateSpanish: Native
English: Low exposureSpanish: High exposure
Spanish ES: 0English ES: 0
The model can also predicts what treatment outcome may have been if the other language was trained
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Perc
ent a
ccur
acy
Probes
Training in Spanish EnglishSet 1
UTBA09UTBA 09:English: EarlySpanish: Native
English: Moderate exposureSpanish: Moderate exposure
Spanish ES: 10.97English ES: 2.07
Summary Model can predict rehabilitation outcomes Of the 17 patients, good fit for 12 patients, For patients that do not have a good fit, model overestimates
outcomes Education/literacy issues in patients
Severe phonological output deficits
Severity of language/cognitive issues
Provides a starting point for understanding why patient did not improve
Model can also predict what treatment outcome may have been if treatment plan was different that what was followed…
Conclusions These results highlight the important interaction between
language proficiency, stroke impairment and language recovery No individual factor can independently predict the amount of
treatment recovery.
Training always improves the trained language but cross language transfer depends on AoA, amount of language exposure pre-stroke and extent of nature of stroke impairment e.g., in individuals with late AoA, low exposure and high damage, cross-language
transfer is less likely to occur
e.g., in individual with early AoA and moderate-high exposure, cross language transfer may be expected.
Conclusions.. While preliminary, the combination of computational
modeling and behavioral treatment provide a promising approach to examining the important issue of recovery of language in bilingual aphasia
Uli GrasemannUT-Austin
Risto MiikkulainenUT-Austin
Chaleece SandbergBoston University
Acknowledgements
UT Austin Anne Alvarez Ellen Kester Rajani Sebastian
Boston University Danielle Tsibulsky Fabiana Cabral Lauren Liria