Cogan Lab@CoganLab

Cogan Lab at Duke University: Investigating speech, language, and cognition using invasive neural human electrophysiology

 Durham NC
coganlab.org
Joined July 2019
4/5
Two plots testing how neural-interface sampling affects cross-patient speech decoding. Decoding improves as electrode spacing becomes denser and as cortical grid coverage becomes broader. Aligned cross-patient models outperform patient-specific models at the highest electrode densities and broadest coverage, showing that both dense and broad sampling help recover shared neural structure.

ALT Two plots testing how neural-interface sampling affects cross-patient speech decoding. Decoding improves as electrode spacing becomes denser and as cortical grid coverage becomes broader. Aligned cross-patient models outperform patient-specific models at the highest electrode densities and broadest coverage, showing that both dense and broad sampling help recover shared neural structure.

1
32
3/5
Plot of phoneme decoding accuracy as the amount of data available from a new patient is reduced. Across the full range, the aligned cross-patient model in purple performs better than the patient-specific model in blue, with an advantage still present when only 5 percent of the new patient’s data is available.

ALT Plot of phoneme decoding accuracy as the amount of data available from a new patient is reduced. Across the full range, the aligned cross-patient model in purple performs better than the patient-specific model in blue, with an advantage still present when only 5 percent of the new patient’s data is available.

1
50
2/5
Schematic from the paper showing neural state-space trajectories from two participants before and after alignment. Before alignment, the blue and red latent trajectories differ substantially. Canonical correlation analysis transforms the patient-specific latent dynamics so that the trajectories become more similar in a common neural space.

ALT Schematic from the paper showing neural state-space trajectories from two participants before and after alignment. Before alignment, the blue and red latent trajectories differ substantially. Canonical correlation analysis transforms the patient-specific latent dynamics so that the trajectories become more similar in a common neural space.

1
1
101
4/5
Intracranial decoding results showing that the correct serial order of phonemes emerges during articulation and monitoring rather than being fully ordered during planning. Sensorimotor cortex also represents transitions between successive phonemes, suggesting that discrete speech units are dynamically transformed into continuous motor sequences.

ALT Intracranial decoding results showing that the correct serial order of phonemes emerges during articulation and monitoring rather than being fully ordered during planning. Sensorimotor cortex also represents transitions between successive phonemes, suggesting that discrete speech units are dynamically transformed into continuous motor sequences.

1
1
43
3/5
Time-resolved speech decoding shows that syllable representations emerge before phoneme representations, but the separation shrinks as speech approaches execution. The syllable-to-phoneme lag is about 250–350 milliseconds during planning, 110–130 milliseconds during articulation, and only 17–25 milliseconds during monitoring.

ALT Time-resolved speech decoding shows that syllable representations emerge before phoneme representations, but the separation shrinks as speech approaches execution. The syllable-to-phoneme lag is about 250–350 milliseconds during planning, 110–130 milliseconds during articulation, and only 17–25 milliseconds during monitoring.

1
1
90
2/5
“Planning happens first.” Data from intracranial recordings show a temporal cascade through the speech network. Prefrontal planning activity precedes sensorimotor articulation by a median 180 milliseconds, and articulation precedes auditory monitoring by 398 milliseconds. Participant-level functional connectivity shows the same ordering.

ALT “Planning happens first.” Data from intracranial recordings show a temporal cascade through the speech network. Prefrontal planning activity precedes sensorimotor articulation by a median 180 milliseconds, and articulation precedes auditory monitoring by 398 milliseconds. Participant-level functional connectivity shows the same ordering.

1
116
Two new papers from the lab in Nature Human Behaviour and Nature Communications, spanning the neural mechanisms of speech production and new approaches to speech neurotechnology! 🧠 Speech plans → fluent motor sequences 🗣️ Speech BCIs across people Links below 👇 1/3
Graphic announcing two new Cogan Lab papers. The Nature Human Behaviour paper asks how the brain turns a speech plan into fluent movement, illustrated as planning → articulation → monitoring. The Nature Communications paper asks whether a speech BCI can learn from other people, illustrated by neural activity from two individuals converging on a shared latent representation.

ALT Graphic announcing two new Cogan Lab papers. The Nature Human Behaviour paper asks how the brain turns a speech plan into fluent movement, illustrated as planning → articulation → monitoring. The Nature Communications paper asks whether a speech BCI can learn from other people, illustrated by neural activity from two individuals converging on a shared latent representation.

1
15
55
5,275
From Cogan Lab Journal Club with @zspald11: these decomposition acronyms are getting out of hand!
1
6
336
We’re happy to present @zspald11 's work on shared neural representations of speech production across individuals! We find that patient-specific data can be aligned to a shared space that preserves speech information, enabling cross-patient speech BCIs. biorxiv.org/content/10.1101/…
1
42
111
9,186
In addition to sequences of discrete phonemes, we show that the motor cortex also tracks the transitions between phonemes (phonotactics), suggesting that speech execution combines both discrete and continuous articulatory properties. 8/10
1
1
106
During speech execution, we show evidence for motor sequencing by extracting sequential patterns of phonemes and find that sequencing only occurs in execution regions during speech production. 7/10
1
1
84
Using neural decoding models, we characterize the hierarchical relationship between syllables and phonemes and demonstrate that this relationship is temporally distinct in planning vs. execution. 6/10
1
1
103
Using high-resolution cortical recordings, we also show that this temporally distinct syllabic activation follows an anatomical spatial gradient from pars opercularis to pre/motor cortex that transitions in time from planning to articulation. 5/10
1
1
108
We identified distinct coding for syllable frames and found that this code first occurred during planning in the left-hemispheric pre-frontal cortex and was sustained during speech motor execution. 4/10
1
1
116
We organized speech neural activations in the high-gamma band (HG) into distinct anatomical networks that were specific to planning and execution (articulation and monitoring). These networks were active sequentially prior to and during speech production. 3/10
1
1
124
We performed intracranial recordings on 52 patients while they articulated pseudowords in a delayed speech repetition task. Constructed pseudowords were either monosyllabic or disyllabic and contained a fixed set of phonemes at each position within the syllable frame. 2/10
1
1
182
Coming to Chicago for #SfN24? Interested in intracranial EEG and speech and cognition? Come see the lab’s posters!
1
4
20
2,419
Last week, Nicole Liddle (@nicoleliddlee, CRS, Sr.) presented Zilio, Gomez-Pilar and colleagues’ 2021 paper on how intrinsic neural timescales (INTs) relate to sensory vs. motor processing in abnormal states. This 🧵 explores her thoughts (🤍 & ❔) sciencedirect.com/science/ar…
1
2
4
560
Congratulations to Dr. Duraivel for passing his dissertation defense with flying colors! It'll be sad to see him go, but he's off to @MIT to do a postdoc with @ev_fedorenko and @MarkRichardson, so he'll be in great hands! @DukeBrain @DukeEngineering @Duke_Neurology @Dukeneurosurg
1
23
5,379
Our recent lab paper led by @dsuseendar was one of the top 25 most read Nature Communications articles in health sciences in 2023! Big thanks to the whole lab and our collaborators for all of their hard work! #NCOMTop25 nature.com/collections/dbigc…
2
9
1,213
Coming to #SfN2023? @SfNtweets Check out posters from the Cogan Lab on studying speech using intracranial neural recordings!
1
3
20
3,230
µECoG revealed temporal sequencing of phonemes in speech motor cortex. A non-linear recurrent model that captured the specific spatio-temporal neural patterns resulted in better decoding performance than linear techniques.
1
1
169
Higher spatial resolution was also required to achieve better decoding as more unique phonemes were included in the analysis.
1
1
130
We decoded spoken phonemes from micro-scale HG activations and found that µECoG achieved up to 57% accuracy in predicting spoken phonemes with total spoken duration of <2.5 minutes. This accurate decoding outperformed standard IEEG by 35%.
1
1
163
µECoG captured fine-scale spatio-temporal patterns in SMC. These spatio-temporal patterns revealed separation of both speech articulators and individual phonemes. This clustering was dependent on high-resolution sampling.
1
1
178
Speech neural activations in the high-gamma band (HG) were spatially specific, and our high-definition recordings demonstrated spatially discriminant neural signals at < 2 mm spacing. This micro-scale neural activations achieved 48% higher SNR compared to macro-ECoG and SEEG.
1
1
192
high-density ECoG arrays (4 mm). 4 patients performed a speech repetition task during their awake neurosurgery.
1
1
220
Do you have (or about to have) a PhD and are you interested in speech and intracranial recordings (seeg, ecog, and micro-ecog)? The Cogan Lab at Duke is looking for a postdoc! International applicants are welcome. Join us! coganlab.org/postdoc23
10
23
6,360
Some good news: our R01 has been funded! We will study the interactions between speech production and vWM using intracranial recordings. A big thank you to the whole lab and all of our collaborators. @GregoryHickok @Duke_Neurology @Dukeneurosurg @LadNeurosurgery @DukeBrain @DukeU
2
4
23
6,763
The Cogan Lab is looking for a research technician to analyze data from invasive human neural recordings acquired during verbal working memory tasks. The candidate will work closely with the PI @Cogan_G and Neurosurgery Resident Daniel Sexton. Join us! coganlab.org/researchtech23
1
9
21
9,884
The Cogan Lab at Duke University is looking for a Clinical Research Specialist to consent, collect, and process behavioral and neural data from epilepsy patients to help us understand how the brain processes speech, language, and cognition. Come join us! coganlab.org/clinicalresearc…

ALT Clockwork Gears GIF
9
12
6,585
🤍1️⃣: Example visualizations using the models being debated (Levelt in Figures 2/3) provide a helpful side-by-side comparison.
1
83
🤍3️⃣: The motivation of bimanual movement as a nonlinear task in Figure 2 is useful for explicitly showing why nonlinear decoders are necessary, beyond that past literature shows improvements when using them over linear decoders.
1
26
🤍2️⃣: The laterality analysis in Figure 5 provides an important point of interpretability as to how nonlinear neural network models can utilize different aspects of the neural data to improve performance.
1
1
35
❔2️⃣: Is there something about the architecture of Yamnet that leads to worse performance in layers after 12 across all mappings?
1
100
❔2️⃣: In Fig. 4d, it appears that there is a smaller RMS power drop for the 3-syllable NP as compared to 4. Does the reduction in power scale linearly with longer phrases? If so, could this be at least in part, a signature of other cognitive processes (e.g. working memory)?
1
125
❔1️⃣:  What is the explanation behind the peak at 3 Hz in Figure 1c? Is it simply due to harmonics or something else?
1
1
109
❔1️⃣: Would word error rates saturate at some point when adding higher channel counts to a small region? The extrapolation in Figure 4b suggests that the log-linear trend should continue for 1000 electrodes, but why should that trend be true for that many electrodes?
1
179
The Cogan Lab is looking for a research assistant! Ideal candidate has experience in matlab/python and an interest in speech/language and cognitive/systems neuroscience. Position starts ASAP (Late 2019/Early 2020) See website for details: coganlab.org/news
1
7
9
The Cogan Lab is looking for a research assistant! Ideal candidate has experience in matlab/python and an interest in speech/language and cognitive/systems neuroscience. Position starts Fall 2019 and is open until filled. See website for details: coganlab.org/news
11
15
The Cogan Lab is looking for a postdoc to explore the exciting world of speech and language using advanced ECoG! Position open until filled. See lab website for more details: coganlab.org/news Please retweet/forward to anyone you think might be interested.
28
21
Load more