Making sense of muscle activity in sensorimotor deficits Lena H. Ting, PhD If two interventions are shown to have equivalent clinical outcomes: • Are they redundant? – e.g. can choose either to same effect? • Are they complementary? – e.g. doing both is better • Is one more effective for some people than others? • How can the interventions be improved to optimize outcomes for an individual? • Why do they work? – E.g. what neural deficit are improved and what changes or neuroplasticity is induced? Progress in rehabilitation depends upon mechanistic metrics of motor function that can: • Identify neural mechanisms underlying motor deficit or improvement – Why do successful rehabilitation interventions work? Levin et al. 2009, Coote et, al 2009 – Which subpopulations stand to benefit most? Kwakkel 2009, Dobkin 2011 – Are improvements in function due to compensation or restoration (or vicariation) of function? Metz et al 2005, Raghavhan 2010 • Predict motor performance in the real world – Do clinical score predict mobility and falls risk? Bowden et al. 2010 EMG is noninvasive, neural organization & biomechanics…and is difficult to interpret cortex brainstem spinal cord Sensory receptors muscles skeleton tissues environment Tresch commentary on Lockhart and Ting Nature Neuroscience 2007 How can we get the information out? Framework Behavioral experiments Functional data analysis Modeling and simulations • Theoretical framework with key components for movement • Experiments that probe specific neural mechanisms • Computational methods for understanding complex signals • Predictive models of movement in experiments and disease Clinical movement disorders Neurophysiology Sensorimotor feedback Biomechanics Sensorimotor feedback framework for understanding muscle activity • Hierarchal relationship of temporal and spatial structure underlying muscle activation – Temporal structure related to task-level goals – Spatial structure related to multi-joint coordination • Using models + experiments + computation in – Peripheral neuropathy balance – Parkinson’s disease balance – Stroke walking • Advantages of sparse analysis methods – Wavelet functional ANOVA for temporal structure – NMF based muscle synergies for spatial structure Sensorimotor feedback framework for understanding muscle activity • Hierarchal relationship of temporal and spatial structure underlying muscle activation – Temporal structure related to task-level goals – Spatial structure related to multi-joint coordination • Using models + experiments + computation in – Peripheral neuropathy balance – Parkinson’s disease balance – Stroke walking • Advantages of sparse analysis methods – Wavelet functional ANOVA for temporal structure – NMF based muscle synergies for spatial structure Framework Temporal recruitment of muscle synergies based on task-level goals Sensorimotor transformation Sensory input Motor output EMG Temporal & spatial complexity Ting and McKay, Current Opinion in Neurobiology 2007; Chiel et al. J Neuorsci 2009 ;Ting et al. IJNMBME 2012 Postural responses to perturbation: Sensorimotor transformations in action quiet standing passive response & sensory encoding EMG cortex brainstem short-latency (spinal / local variables) long-latency (brainstem / task variables) spinal cord step transition Sensory “decision” receptors muscles Force skeleton CoM tissues 50 100 200 ms environment perturbation Functional data analysis Spatiotemporal patterns related to a few parameters related to CoM motion Ting and McKay, Curr Op in Neurobiol 2007; Chiel et al. J Neuorsci 2009 ;Ting et al. Int J Numer Method Biomed Eng 2012 Functional data analysis Time course of muscle activity reflects sensed body motion CoM backward CoM motion backward d(t) CoM motion time v support surface translation feedback gains time delay delayed sum scaled CoM a kd kv v (t a time ka recorded TA EMG Sensory feedback Sensory estimates of center of mass (CoM) motion threshold reconstructed TA timing and magnitude d d tibialis anterior (TA) brainstem & spinal cord time Sensorimotor Response Model (SRM) Motor responses measured with electromyogram (EMG) Functional data analysis Agonist (TA) muscles activity reflects sensed backward CoM motion CoM backward CoM motion backward d(t) CoM motion time v support surface translation a Sensory encoding Sensory estimates of center of mass (CoM) motion feedback gains time delay delayed sum scaled CoM threshold reconstructed TA timing and magnitude d d tibialis anterior (TA) brainstem & spinal cord kd kv v (t time a ka Conduction & Synaptic & circuit gain processing delay Summation interneuron or motor neuron Sensorimotor Response Model (SRM) recorded TA EMG time Motor responses measured with electromyogram (EMG) Functional data analysis Antagonist (MG) muscles reflect sensed forward CoM motion forward CoM forward motion CoM motion CoM d(t) kv delayed sum scaled CoM threshold Reconstructed MG timing and magnitude v (t a ka a Sensory backward encoding CoM motion backward CoM motion d time delay kd v support surface translation feedback gains d d medial gastrocnemius (MG) brainstem & spinal cord Summation basal& interneuron Synaptic & Conduction inhibit or inappropriate ganglia program circuit gain processing delay motor neuron brainstem & spinal cord feedback gains time delay delayed sum scaled CoM threshold • Agonist muscle (TA)d is inhibited and antagonist k muscle (MG) is excited at the end of the v (t k perturbation a d v a v ka Coactivation with threat: Carpenter MG, Frank JS, Adkin AL, Paton A, Allum JH. J Neurophysiol 92: 3255-65, 2004 Behavioral experiments Common sensory signals drive reciprocal agonist/antagonist muscle activity Safavynia and Ting, J Neurophysiol 2013 Behavioral experiments CoM kinematics predict EMG better than joint kinematics in humans Safavynia, and Ting J Neurophysiology 2013 Delayed, task-level error feedback predicts temporal recruitment of muscle synergies Ting and McKay, Curr Op in Neurobiol 2007; Chiel et al. J Neuorsci 2009 ;Ting et al. Int J Numer Method Biomed Eng 2012 Each muscle synergy has a preferred direction of recruitment Safavynia and Ting, J Neurophysiol 2013 Time course of muscle synergy recruitment in preferred direction is described by CoM feedback Safavynia and Ting, J Neurophysiol 2013 Parameters from preferred direction predict activity in other directions • Based on projection of CoM kinematic error along each muscle synergy’s preferred direction Safavynia and Ting, J Neurophysiol 2013 Each functional muscle synergy produces an consistent endpoint force direction Activation level! Muscle synergy recruitment, ci Muscle weights, wi Synergy vector! Ting and Macpherson, J Neurophysiology 2005; Torres-Oviedo, Macpherson, Ting, J Neurophysiology 2006 Muscle synergies quantify variability across subjects using a common framework Torres-Oviedo, Macpherson, Ting, J Neurophysiology 2006 Muscle synergies underlie trial by trial variability • Ankle and hip strategy are implemented by different synergies • Muscle synergy activations vary from trial to trial Torres-Oviedo and Ting, J Neurophysiology 2007 • Muscle activity variations are not random but coupled in muscle Averaging synergy patterns Sensorimotor feedback framework for understanding muscle activity • Hierarchal relationship of temporal and spatial structure underlying muscle activation – Temporal structure related to task-level goals – Spatial structure related to multi-joint coordination • Using models + experiments + computation in – Peripheral neuropathy balance – Parkinson’s disease balance – Stroke walking • Advantages of sparse analysis methods – Wavelet functional ANOVA for temporal structure – NMF based muscle synergies for spatial structure Modeling and simulations Neuromechanical model can predict optimal solutions (data-free) in cat Musculoskeletal system Muscle activation Sensory feedback Nervous system Lockhart and Ting, Nature Neuroscience 2007, Welch and Ting, J Neurophysiology 2008, 2009, Safavynia and Ting 2011, 2013 Detailed features of muscle activity can be predicted using by only 3 parameters Lockhart and Ting, Nature Neuroscience 2007 “Bounce” in muscle activity reflects non-ideal perturbation acceleration, not neural processing Lockhart and Ting, Nature Neuroscience 2007 Pyridoxine (B6) overdose causes permanent sensory neuropathy • Affects sensory afferents with cell bodies in dorsal root ganglia; largest neurons first • Dosed to target Group I (10-20 um) afferents – Muscle spindles – Cutaneous afferents – Golgi Tendon organs • IP injections on 2 consecutive days • Ataxia, followed by motor recovery over 2-3 weeks Stapley, Ting, Hulliger, Macpherson, J Neuroscience 2002 Muscle activity reflects loss of acceleration sensory information • Optimal muscle pattern in sensory loss differs from intact • Kinematics are quantitatively but not qualitatively different Lockhart and Ting, Nature Neuroscience 2007 Sensorimotor feedback framework for understanding muscle activity • Hierarchal relationship of temporal and spatial structure underlying muscle activation – Temporal structure related to task-level goals – Spatial structure related to multi-joint coordination • Using models + experiments + computation in – Peripheral neuropathy balance – Parkinson’s disease balance – Stroke walking • Advantages of sparse analysis methods – Wavelet functional ANOVA for temporal structure – NMF based muscle synergies for spatial structure Parkinson’s disease impairs balance and mobility leading to falls and injury • PD is a neurodegenerative disorder affecting the basal gangia – Loss of dopamine production in substantia nigra • • • • 3-month fall rate ≈ 46%1 6-month fall rate ≈ 5× vs. age-matched adults2 Risk of a hip fracture ≈ 3x vs. age-matched adults3 Among least responsive to dopaminergic pharmacotherapy2,4 1Pickering RM et al. Mov Disord 22: 1892-1900, 2007; 2Bloem BR et al. J Neurol 248: 950-958-958, 2001" 3Melton LJ, 3rd et al. Mov Disord 21: 1361-1367, 2006; 4Klawans HL. Mov Disord 1: 187-192, 1986." Dynamic co-contraction of TA and MG in PD Pre-AT Subject 3 H&Y 2.5 kd kv ka Kp=0.024, Kv=0.054, Ka=0.661,Td=0.100 Kp=0.000, Kv=0.000, Ka=0.000,Td=0.100 CoM bias=0.093,R2=0.553, VAF=0.759 d(t) EMG (mean) EMG (trials) SRM fit 0.5 nu TA Appopriate SRM Component Inappropriate SRM Component kd kv ka Kp=0.000, Kv=0.000, Ka=0.000,Td=0.100 Kp=0.035, Kv=0.036, Ka=0.507,Td=0.100 MG bias=0.369,R2=0.559, VAF=0.925 0.5 nu support surface translation McKay, Hackney, Ting, in progress 0.0 5 cm time (sec) 1.0 • TA and MG both excited in response to backward CoM motion and inhibited in response to forward CoM motion • MG not excited in response to forward CoM Antagonist may also co-activate with agonists for stability, or with threat forward CoM forward motion CoM motion CoM d(t) kv backward CoM motion backward CoM motion v a delayed sum scaled CoM threshold v (t a ka a d time delay kd v support surface translation feedback gains d d medial gastrocnemius (MG) brainstem & spinal cord basal ganglia brainstem & spinal cord feedback gains time delay delayed sum scaled CoM d kd kv v (t a inhibit inappropriate program threshold Biphasic response due to summation of two neural pathways ka Coactivation with threat: Carpenter et al. J Neurophysiology 2004 Coactivation with threat: Carpenter MG, Frank JS, Adkin AL, Paton A, Allum JH. J Neurophysiol 92: 3255-65, 2004 Adapted tango (AT) neurorehabilitation improves balance and mobility in individuals with PD • Promotes adherence → counters excessive attrition in PD • Adapted for PD → safety, dynamic balance, internal + external movement cues • Effective → gains on Berg Balance Scale, gait speed, and other clinical measures1-4 – Retained for 1 month1 1Hackney ME, and Earhart GM. Neurorehabil Neural Repair 24: 384-392, 2010; 2Am J Dance Ther 31: 41-45, 2010; 3J Rehabil Med 41: 475-481, 2009; 4Duncan RP, and Earhart GM. Neurorehabil Neural Repair 26: 132-143, 2012 Intensive AT Rehabilitation improved clinical measures of static and dynamic balance Pre Post P-value (paired t-test) Unified Parkinson s Disease Rating Scale 30±5! 28±4! 0.189*! Berg Balance Scale 50±7! 54±4! 0.028*! Fullerton Advanced Balance Scale 27±8! 31±6! 0.004*! Functional Reach (cm) 30±8! 33±6! 0.061! Dynamic Gait Index 19±4! 21±3! 0.090*! 6-Minute Walk Test (m) 406.1±100.7! 419.6±103.8! 0.721*! Decreased forward CoM motion in perturbations McKay, Hackney, Ting, in prep Muscle activity + computational analysis + multiple experiments reveals: • A common sensorimotor transformation governing healthy and impaired movements • A specific channel of lost sensory information not evident in kinematics • Impaired and improved motor program selection in PD Supports (Mink 1996) • Improved involuntary brainstem sensorimotor response after exercise-based neurorehabilitation Sensorimotor feedback framework for understanding muscle activity • Hierarchal relationship of temporal and spatial structure underlying muscle activation – Temporal structure related to task-level goals – Spatial structure related to multi-joint coordination • Using models + experiments + computation in – Peripheral neuropathy balance – Parkinson’s disease balance – Stroke walking • Advantages of sparse analysis methods – Wavelet functional ANOVA for temporal structure – NMF based muscle synergies for spatial structure If you don’t have a model, data discovery is also important but curves are hard to compare Statistics based on user-selected features, e.g. peak or time bins are biased • A lot of information is not used, particularly concerning the shape of the curves • Data from Welch and Ting J Neurophysiology 2009 Functional data analysis Timepoint analysis lack statistical power • ANOVA on individual timepoints only identifies discrete regions with large differences • Reduces power due to multiple comparisons • But observations are not independent McKay, Welch,Vidakovic, Ting J Neurophysiology 2013 ANOVA in wavelet domain produces smooth difference curves over time; increased power McKay, Welch,Vidakovic, Ting J Neurophysiology 2013 Wavelets produce sparse representations using a few time-localized functions McKay, Welch,Vidakovic, Ting J Neurophysiology 2013 Difference curves for 3 factor functional ANOVA allow for data discovery subject x acceleration x velocity contrast •wfANOVA contrasts ignore high variability areas • Sensorimotor feedback framework for understanding muscle activity • Hierarchal relationship of temporal and spatial structure underlying muscle activation – Temporal structure related to task-level goals – Spatial structure related to multi-joint coordination • Using models + experiments + computation in – Peripheral neuropathy balance – Parkinson’s disease balance – Stroke walking • Advantages of sparse analysis methods – Wavelet functional ANOVA for temporal structure – NMF based muscle synergies for spatial structure Muscle synergies can account for cycle-bycycle variations in walking Clark, Ting, Neptune, Zajac, Kautz J Neurophysiology, 2010 McGowan, Neptune, Clark, Kautz, J Biomechanics , 2010 NMF produces sparse representations that represent features or events in the data Sparse = A few muscles in each synergy A few synergies at a time A few synergies contribute to a muscle PCA uses all elements all the time All muscles in each synergy All synergies at a time Most synergies contribute to a muscle Chronic post-stroke hemiplegia NMF identifies muscle synergies that relate to motor function Merged proximal and distal extensors, PA=0.27±0.16 Independent proximal and distal extensors, PA=0.52±0.32 • Preserved spinal structure, reduced corticospinal drive? Clark, Ting, Neptune, Zajac, Kautz J Neurophysiology, 2010 PCA is descriptive, NMF is prescriptive, e.g. constrains the set of possible patterns Ting and Chvatal 2010, Motor Control, chapter and tutorial available for download PCA is descriptive, NMF is prescriptive, e.g. constrains the set of possible patterns Ting and Chvatal 2010, Motor Control, chapter and tutorial available for download Fewer muscle synergies in the paretic leg in chronic post-stroke hemiplegia Low Moderate High Level of locomotor complexity • These were what they used, but not necessarily what they could use (learned non-use?) àneed experiment • We don’t know if these are feedforward or feedback mediated à need experiments Clark, Ting, Neptune, Zajac, Kautz J Neurophysiology, 2010 Number of muscle synergies predicts functional locomotor performance • High complexity subjects could walk faster than moderate, but chose not to à need to find constraints • Subjects with the same number of synergies as healthy controls walked slower à need temporal analysis Clark, Ting, Neptune, Zajac, Kautz J Neurophysiology, 2010 Number of muscle synergies predicts motor function similar better than Fugl-Myer score • Balance and propulsive scores are predicted by number of muscle synergies à generalized motor capacity? Bowden, Clark, Kautz, Neurorehabilitation and Neural Repair, 2010 Muscle synergies are both a useful tool and a theory of the brain, but use wisely Framework Behavioral experiments Functional data analysis Modeling and simulations • Theoretical framework with key components for movement • Experiments that probe specific neural mechanisms • Computational methods for understanding complex signals • Predictive models of movement in experiments and disease Clinical movement disorders Neurophysiology Sensorimotor feedback Biomechanics Can structure and organization of EMG output during movement: • Differentiate individuals across motor function? – Can it do as well as a clinical test? – How does muscle activity differ in impairment? • Identify mechanisms of deficit and recovery? – Can it predict more than a clinical test? – Differentiate compensatory vs restorative mechanisms? – Do they change after successful rehabilitation? – What are “good” and “bad” patterns of muscle activity? • Predict behavior important in the real world? – Does the organization of activity generalize across behaviors? Sensorimotor feedback framework for understanding muscle activity • Hierarchal relationship of temporal and spatial structure underlying muscle activation Delis et al J Neurophysiology 2013 McCrea &Rybak Brain Research Reviews 2007, Progress in Brain Research 2007 • Using models + experiments + computation to – Identify neural substrates; feedforward vs feedback; restoration vs substitution – Predict current and potential motor ability Allen, Kautz, Neptune Clinical Biomechanics 2013; Walter … Fregly Journal of Biomechanical Engineering 2014 • Advantages of sparse analysis methods – Sparse coding predicts sensory representations and could be a general principle for sensing and movement Olshausen and Field Current Opinion in Neurobiology 2004; Ting and McKay Current Opinion in Neurobiology 2007; Poggio and Bizzi Nature 2004 Ting Neuromechanics Lab Current: Jessica Allen, PhD Harrison Bartlett Kyle Blum, BS Kim Lang, BS Lauren Levey, BS Lucas McKay, PhD Andrew Sawers, PhD, MSPO Hongchul Sohn, MS NIH R01 HD046922, NS053822, NIH R21 HD075612 NSF EFRI 1137229 Some Alumni: Jeff Bingham, PhD Stacie Chvatal, PhD Nate Bunderson, PhD Julia Choi, PhD Gelsy Torres-Oviedo, PhD Seyed Safavynia, PhD Torrence Welch PhD Some Collaborators: Tom Burkholder Young-Hui Chang Charlie Kemp Madeleine Hackney Karen Liu Jane Macpherson Richard Nichols Randy Trumbower