Building therapies on neurobiological principles Word-level therapy for apraxia of speech Rosemary Varley Division of Psychology & Language Sciences University.

Building therapies on neurobiological principlesWord-level therapy for apraxia of speech

Rosemary Varley

Division of Psychology & Language SciencesUniversity College Londonrosemary.varley@ucl.ac.uk

1

CollaboratorsSandra Whiteside & Patricia Cowell (HCS, Sheffield)

Core research team

Lucy Dyson; Lesley Inglis; Abigail Roper

Additional assistance from:

Andrew Harbottle; Jenny Ryder; Vitor Zimmerer

Additional collaboratorsSLTs across South Yorkshire, & in particular Rotherham NHS

Catrin Blank (Clinical Neurology, NHS Sheffield)Tracey Young (ScHaRR, Health Economics, Sheffield)

FundersThe Health Foundation; BUPA Foundation: University of Sheffield/HEIF4

knowledge transfer grants 2

Conflict of Interest Statement

Sword Software

• Software program is commercially available• ‘Inventors’ Varley, Whiteside & Cookmartin receive

share of royalties from sales

3

Post-Stroke Speech Production Impairments

4

AnomiaLexical-semantic

Apraxia (AOS)Phonology-phonetic

DysarthriaPhonetic

Generative-Computational Modelsof Speech & Language

• Minimise storage, maximise computation• ‘Elegant’, ‘Parsimonious’• Phonology: store small number of units

(phonemes, distinctive features e.g. [+ voice]) & large combinatorial mechanism to create syllables

e.g., Shattuck-Hufnagel’s (1979) slots and fillers model

5

Slots & Fillers

• Abstract phonological representation is ‘scanned’ /k æ t/

• Slots (syllable frame determined): _ _ _• Fillers: Mechanism locates segments/phones

corresponding to phonemes [k] [æ] [t]• Fillers inserted into slots [k æ t]

6

Apraxia of Speech (AOS)• Under C-G view, underlying impairment:– Inaccessible segmental plans– Impairment of allocating segment to slot

• Classical apraxia therapy: microstructural (also articulatory-kinematic or sound production therapy)– Rebuilding segmental plans– Practice in generating cohesive syllables through

combination of segmental plans i.e., subcomponents & generative mechanism

7

Example AOS & Microstructural Therapy

• Articulatory errors; Altered durational characteristics; Loss of speech automaticity; non-fluent, effortful, struggle & groping. 8

Evidence-Base for Microstructural Therapy

• Wambaugh, J. et al. (2006). J. Medical Speech-Language Pathology, 14(2), xv-xxxiii Treatment Guidelines for Acquired Apraxia of Speech: A Synthesis and Evaluation of the Evidence

– Majority of research on articulatory therapies– Learning of targeted gesture/syllable– Poor generalisation of learning– Expensive

• Cochrane Review (2009): “No evidence was found for the treatment of AOS.”

• Therapists view as hard-to-treat condition.9

Generative Models Under Attack• At all levels of structural linguistic processing (phonology,

morphology, syntax) generative-computational models under attack– Syntax: I + am + go + ing + to + ____

vs.

– Morphology: un + fortunate + lyvs.

– Phonology y+e+s+t+er+d+ayvs.

• Neurocognitive implausibility– “human memory capacity is quite large” Bybee (2006: 717)

10

I’m going to ____

unfortunately

yesterday

Alternative Model

• Usage/frequency-mediated models of processing– Sequences which are frequently repeated become

stored as complete units (complete words/ phrases/clauses)

11

Shattuck-Hufnagel’s puzzle & Levelt’s and paradox

Shattuck-Hufnagel (1979): “perhaps [the] most puzzling aspect is the question of why a mechanism is proposed for the one-at-a-time serial ordering of phonemes when their order is already specified in the lexicon.”

Levelt (1992) : “Why would a speaker go through the trouble of first generating an empty skeleton for the word, and then filling it with segments? In some way or another both must proceed from a stored phonological representation, the word’s phonological code in the lexicon. Isn’t it wasteful of processing resources to pull these apart first, and then to combine them again (at the risk of creating a slip).” 12

Re-thinking AOS via Frequency-mediated Account

• High frequency constructions stored as complete plans.

• Speech control (Crompton, 1982; Whiteside & Varley, 1998; Varley & Whiteside, 2001): frequently used output stored as complete phonetic plans.

• Biologically more plausible, and capable of delivering fast, cohesive and error-free movements.

13

Building Therapies on Neurobiological Principles (Varley 2011. Int. J. SLP)

1. Frequency-mediated vs. computational– Therapy focuses on whole words

2. Procedural vs. declarative learning3. Interconnectivity of processing systems4. Errorless learning/error-reducing strategies5. Therapy intensity (‘Dose’)

- Illustrate with reference to AOS therapy14

Procedural vs. Declarative Learning‘Doing vs. Talking about doing’

• Some speech/language interventions involve giving patient explicit knowledge of how (we think) speech/language systems operate i.e., metalinguistic knowledge

• Assumption that patient will assimilate this explicit knowledge & ‘re-boot’ the automatic/procedural systems that govern listening & talking

• In the case of AOS, clinician shares explicit knowledge of articulatory phonetics – patient becomes patient becomes a ‘mini-phonetician’.

• But notice, most healthy speakers produce speech fluently without any awareness of phonetics 15

Procedural vs. Declarative Learning‘Doing vs. Talking about doing’

• Some speech/language interventions involve giving patient explicit knowledge of how (we think) speech/language systems operate i.e., metalinguistic knowledge

• Assumption that patient will assimilate this explicit knowledge & ‘re-boot’ the automatic/procedural systems that govern listening & talking

• In the case of AOS, clinician shares explicit knowledge of articulatory phonetics – patient becomes patient becomes a ‘mini-phonetician’.

• But notice, most healthy speakers produce speech fluently without any awareness of phonetics 16

Procedural vs. Declarative Learning‘Doing vs. Talking about doing’

• Wulf et al (2001) Quarterly J Expt. Psych. Internal focus of attention leads to less automaticity in complex motor skill learning.

• Ballard et al (2011) Motor Control. Poorer retention of a novel speech movement in healthy speakers within kinematic feedback, than those without constant kinematic feedback.

• Possible link to learned misuse & constraint therapies: by making patient consciously aware of articulatory movements may result in learned non-use of usual automatic/procedural mechanisms of fluent speech control. 17

Interconnected of sensory-motor systems• AOS therapy often uses nonsense syllables &

pure production therapy (modules/autonomy)• Observing movement results in sensory-

perceptual activation, and ‘mirror’ activation in motor cortex (‘mirror neurons’, e.g. Wilson et al. 2004. NatNeurosci).

• Fridriksson et al (2009, Stroke): therapy consisting of word-picture match + observing video of mouth resulted in improved word production in non-fluent aphasia.

18

Sword (Sheffield Word) http://www.propeller.net/sword.htm

19

20

Therapy Structure• Sensory-perceptual phase: 6 modules– Computer models errorless spoken word-picture

matching– Computer models errorless spoken-written word

matching– Participant performs spoken word-picture

matching task– Participant performs spoken-written word

matching task– Computer models errorless spoken lexical decision

task– Participant performs spoken lexical decision task 21

Errorless Learning/Error-reduction Strategies(Whiteside et al. 2012. JNeuroRehab.)

• Errorful vs. Errorless/error reduction techniques.• Errorless learning may be particularly important

in procedural/motor learning. Errors prevent formation of stable movement memories.

• Therapy designed to minimise errors :– Priming output via sensory-perceptual phase– Imagined movement (Page et al. 2005)– Immediate – Delayed Repetition – Independent

production 22

Therapy Structure (Output: 7 modules)

• Observe video of speaker saying target word• Imagined production of words• Immediate repetition of words ----- delayed

repetition• Repetition with audio-recording & playback• Practise of target words in sentence frames• Production of word in isolation, with cue

support if needed

23

24

Therapy Dose Pulvermüller & Berthier, 2008. Aphasiology.

• Neuronal plasticity & Hebbian learning• Hebb (1949) described how connections

between synapses alter as a result of learning:• “any two cells or systems of cells that are

repeatedly active at the same time will tend to be become ‘associated’, so that activity in one facilitates activity in the other.” (1949, p. 70).

• Computer therapy cost-effective means of achieving necessary ‘dose’.

25

Therapy Study

• Therapy based on neurocognitive principles• Computer therapy, allowing participants to self-

administer intervention with potential to achieve high dose

• Advised to use program regularly for short periods (‘little and often’)

• Program records user interactions• 50 participants with single therapy protocol• Speech intervention contrasted to sham/placebo

computer intervention26

Visual Sham Intervention

27

Participants• 50 participants with AOS (+ aphasia) recruited;– 25 female: 25 male– Mean age = 65 years– >5 months post-LH-stroke (Mean = 22 months)

• 44 participants completed study• Varying levels of computer experience (novice to

expert).• Participants randomly assigned to:– Speech first (speech program – sham program) or,– Sham first (sham program – speech program)– No significant differences at baseline between two groups

28

Study DesignTwo period cross-over design (Cowell, et al. 2010. Frontiers in Human

Neuroscience)

Rest Maintenance

29

Period 1Treatment

16 weeks

Period 2Treatment 2

6 weeks

Sham-First

Speech-First

Rest4 weeks

Maintenance8 weeks

Baselines 1-2

3-4 weeks

SHAM

SPEECH SHAM

SPEECH

Measures of Behaviour

30

• Word Production (35 in each set)– Treated: ‘dog’– Phonetically-matched, untreated: ‘door’– Frequency-matched, untreated: ‘game’– Performance measured in naming & repetition– Also collected spontaneous speech samples at baseline &

maintenance

• Untreated behaviours– Written word-picture matching (PALPA 48, Kay et al., 1992)– Spoken sentence-picture matching (CAT, Swinburn et al.,

2004)

• Health Economic Assessment

Speech Analysis• Word-level– Repetition accuracy (0-7 scale) & word duration

(fluency, cohesiveness of articulatory routines)• 7 = error-free, response latency <2 sec; normal

word duration• 6 = error-free, response latency >2sec or slowed

duration• 4 = one segment error• 2 = two segment error + groping• 0 = no response or off-target

– Naming communicative adequacy (0/1) (would a naive listener understand the intended meaning?)

31

Blinding• ‘Open-label’ trial as not possible to blind clinician or

participant to treatment being administered.• Rater for word-level outcome measures blind to

randomisation to speech-first/sham-first allocation.• Inter-rater reliability check by further rater blinded to

phase & rater 1 measurement (10% data).

32

Measure Correlation

Repetition accuracy Spearman r=.892, p<.001Repetition duration Pearson r=.956, p<.001Naming Spearman r=.912, p<.001

Results

33

Compliance with ‘little and often recommendation’

• Program use (hours:mins in 42 day period)– Speech program: 3:32 – 50:29; M=16:48– Sham program: 0:41 – 50:09; M=14:54– No significant difference between sham/speech-

first groups in level of use of either program – 11 participants completed entire speech program;

33 completed word-level production tasks.

34

Outcomes Summary• Baseline behaviour was stable• Untreated behaviours showed no significant

change over course of study– Spoken sentence-picture match (t(43)=-0.113,

p=.911)– Written word-picture match (t(42)=-1.017, p=.315)

No spontaneous change in behaviour• Sham program had no influence on word

production scores. Any treatment effect was not placebo

35

Results Format

36

Baselines

Post-Tx2

Post-Tx1

Sham-First

Speech-First

Maintenance

37

Naming Communicative Adequacy (Treated Words)*

*

Speech vs. Sham program F(1,39)=14.486, p=0.0001)

Treatment X Sequence interaction approached significance F(1,39)=4.006, p=0.052

Sham-First

Speech-First

Repetition Accuracy - Treated

38

Speech>Sham program

F(1,39)=4.562, p=0.039. No interaction with sequence

*

*

Correct/Fluent scores across word sets (repetition)

39

Phonetically-matched ‘door’

Frequency-matched ‘game’

Treated ‘dog’

0.00

5.00

10.00

15.00

20.00

0.00

5.00

10.00

15.00

20.00

0.00

5.00

10.00

15.00

20.00

*

ns

*

ns

ns

Error/Struggle scores across word sets(repetition)

40

Phonetically-matched ‘door’

Frequency-matched ‘game’

Treated ‘dog’

0.00

4.00

8.00

12.00

0.00

4.00

8.00

12.00

0.00

4.00

8.00

12.00

* ns

nsns

41

High users (over 25 hours)

Low users (under 10 hours)

Delta/change scoresBaseline = 0Post-tx = 7Delta 0 – 7 = -7

Results Summary (1)• Group level: significant improvement in

naming & repetition accuracy of treated words following speech program, & improvements maintained 8 -18 weeks after withdrawal of therapy.

• Evidence of generalisation to phonetically-matched words.

• Pattern of response of speech-first group generally better than that of sham-first.

42

Results Summary (2)• Individual differences in response– Some high users showed improvement on both

treated and untreated words– Other high users responded but little

generalisation– Some lower users showed good response (‘super

learners’)

43

Participants’ Attitudes

• Generally positive regarding model of service delivery, when combined with therapist support.

• Those with family members who could support use of computer were more positive.

• Many found the repetitive stimulation ‘boring’ & likely to be factor in low compliance in some participants

• Positive responses from carers:“I felt I could leave him, knowing he had something useful to get on with.”“I got more gardening done that summer.”

44

Why impairment therapies sometimes don’t work (Varley, 2011. IntJSLP)

• Based on biologically implausible computational models

• Low dose.• Focus on conscious, metalinguistic, declarative

knowledge vs. Implicit/procedural knowledge.• Insufficient practice of ‘getting it right’.• Focus on isolated level (module/level of

representation) & ignore interconnectedness of information processing & neural system.

45

Summary & Conclusions• Intervention study with larger sample of patients,

administered single treatment protocol.• Self-administered, IT-based therapy may represent

cost-effective way of resolving dosage problem.• Word-level therapy for AOS is effective.• Effects most evident on treated word forms, with

maintenance of therapy gains.• Unlike microstructural therapies, evidence of

generalisation to untreated words if phonetically similar.

• Sub-groups may provide insight into those individuals who benefit most from therapy 46

Thank you

r.a.varley@sheffield.ac.uk

47

Macrostructural Therapy for AOS• Macrostructural (whole word/utterance) therapies used in

AOS e.g. Key word therapy (Square-Storer, 1989) & Melodic Intonation Therapy.

• Intervention study: whole words, self-administered computer therapy.

• Self-administered computer therapy allows users to deliver intervention at times/locations convenient to them & without therapist being present.

• Potential to achieve high dose therapy.• Computers ideal for delivering repeated stimulation

necessary to stimulation reorganisation of damaged neural system (Bhogal et al., 2003; Pulvermüller & Berthier, 2008; Varley, 2011). 48

Word Duration (Treated)

49

Segments & Speech Control

• Segmental models built upon evidence of switch errors (Jeremy Hunt, the culture secretary).

• But these errors rare in novice users of speech production mechanism (appearing after 7 yrs Stemberger, 1989), & rare in acquired speech disorders (Varley & Whiteside, 2001).

• Influenced by word frequency, occurring on lower frequency words.

50

Background Assessments

• Screens for:– Severity of aphasia– Dysarthria– Presence of oral

apraxia– Severity of AOS

• Psycholinguistic battery

• Prognostic indicators– Auditory processing– Non-word repetition– Repetition priming– Cloze/isolated

production• Health economic

assessment

51

Data Analysis (Cowell, et al. 2010. Frontiers in Human Neuroscience).

• Cross-over designs typical in drug trials.• Small number in cog-behavioural interventions (e.g. Fillingham,

2005; Raymer et al, 2010).• Avoids ‘resentful demoralisation’ of randomisation to sham in

open-label RCT.• But semi-permanence of successful cog-behav. therapy

creates statistical problem: setting of new baselines across phases.

• Use of delta scores; lambda statistic to determine if possible to join 2 phases of trial.Accuracy score: Test 1 20/35; Test 2 30/35; delta 20-30=-10Word duration: Test 1 400ms; Test 2 300ms; delta 400-300 = +100

52