Challenges for Analyzing Slow Rhythms in MEG DataJonathan Z. SimonUniversity of MarylandDepartments of Biology and Electrical & Computer EngineeringSupported by NIHR01 AG 027573R01 DC 007657R01 DC 008342R01 DC 005660Auditory Rhythms in Speech•Amplitude Modulation (AM)•Frequency Modulation (FM)•Other (e.g. binaural)Slow Neural Rhythms2 – 20 Hz!! ! Critical rates for speech! generating, or competing with2 – 7 Hz! ! ! Delta, Theta8 – 12 Hz!! ! Alpha12 – 30 Hz !! Beta 30+ Hz ! ! ! GammaSlow Rates ⇒ Speech ProcessingCortical Modulation FilterbanksCortical encoding of multiple modulations analogous to cochlear encoding of multiple frequencies?Juanjuan Xiang (in preparation)Slow Rate = Noisy BackgroundYadong Wang, Nai Ding (in preparation)Denoising is CriticalFocus: Neural background noiseExample: Data driven spatial filtering, e.g. Denoising Source Separation (DSS)Generates spatial filters & their outputs (“components”)ordered by reproducibility1st component “maximally reproducible” = stimulus drivenTemporal DSS ExamplesMost reproducible filter & componentOptimally filters out trial-to-trial-variable signal = neural noiseFilter can be applied to other signals, e.g. single trialsde Cheveigné & Simon, J. Neurosci. Methods 2008Spectral DSS ExamplesFrequency Spectrum before DSSFrequency Spectrum after DSSDing & Simon, J. Neurophysiol 2009Phase DSS ExamplesPhase Spread before DSSPhase Spread after DSSAlpha Phase ParameterAlpha Phase Parameterby subjectby subjectDing & Simon, J. Neurophysiol 2009Challenges OvercomeWith careful attention paid to noise and variability ⇒ denoising, cross-validation, etc.! Faint, variable signals can be made ! robust & reliableApplication: Natural SpeechLengthy natural speech stimuli (2 minutes of “The Legend of Sleepy Hollow”).MEG response cleaned with DSS and reverse correlated with stimulusRobust MEG representation of Speech, even after a single presentation: STRF25050010002000400012550 100 150 200 2500250500100020004000125TimeSpectrumSpectrumTimeRateRate Rate RateRateTime (ms)Spectral Response Fields,evolving in timeImpulse Responses,parametrized by spectral band50 100 150 200 2500Time (ms)50 100 150 200 2500Time (ms)50 100 150 200 2500Time (ms)250500100020004000125Interpreting STRFsCross-section InterpretationsFrequency (Hz)Frequency (Hz)Frequency (Hz)Neuronal STRFsSimon et al. Neural Computation (2007)Time (ms)Frequency (Hz)Speech STRF (Left Hemisphere) 0 100 200 300 400 200 400 600110019003200ï0.100.1MEG Representation of SpeechRobustly represented by Spectro-Temporal Response Function (STRF)Stimulus representation dominated by frequencies from 500 to 1200 HzResponse dominated by slow frequencies < 8 Hz1 7 15 3100.050.10.150.2Right HemisphereFrequency (Hz)CorrelationNai Ding (in preparation)ï0.1 0 0.1 0.2 0.3 0.4Time (s)Multiple Speakers No ProblemSpeech impulse responses to speech from multiple simultaneous speakersSpeech representation strongly modulated by attentionSingle SpeakerAttended SpeechUnattended SpeechNai Ding (in preparation)SummarySlow rhythms critical to neural representations of speechStimulus generated slow rhythms are easily masked by intrinsic rhythms and other neural background Neural representation of speech and other speechlike sounds can be made visible and robustSignificance by Frequency051015202530354000.1250.250.3750.50.6250.750.87510 10 20 30 40 5005101520253035400 10 20 30 40 5000.1250.250.3750.50.6250.750.87510 10 20 30 40 50Fig. 1Modulation Frequency (Hz)PowerRatio of significant