Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: The canonical HMM approach can be applied in MEG sensor space. This allows networks to be approximately inferred when source reconstruction is not feasible. (A) Topographic maps of the relative state power (1–45 Hz) at each MEG sensors viewed from above and PSDs for the 10-state canonical HMM networks at the sensor level, in the same order of states as for the parcel-space canonical HMM shown in . The solid line shows the state PSD averaged over regions (parcels), and the dashed line shows the PSD averaged across states. (B) Illustrative state probabilities inferred from the first 8 seconds of the first session in Cam-CAN when applying the canonical HMM networks to the sensor-level data (top) or the (prepared) parcellated data (bottom). (C) Normalized confusion matrix: average accuracy (across participants) for the state assignment at each time point between the sensor-level and source-level canonical models.

Article Snippet: The sensor-level canonical HMMs were inferred using the MaxFiltered Cam-CAN data (Elekta, N = 621 , rest and task).

Techniques:

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: Overview of HMM approaches. A conventional HMM approach involves training a new model on the boutique dataset (usually multiple times to ensure results are robust), whereas the canonical HMM approach uses an HMM pre-trained on a large independent dataset. The HMM is used to infer the state probabilities at each time point, and dual estimation is used to infer individualized network properties for each state.

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques:

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: Ten state parcel-level canonical HMM networks inferred from Cam-CAN MEG data ( N = 621, rest and task). (A) Relative state power maps, coherence networks (top 3%), and PSDs showing activity in the frequency range 1–45 Hz. The solid line shows the state PSD averaged over regions (parcels), and the dashed line shows the PSD averaged across states. (B) Mean coherence vs power for each parcel averaged over subjects. (C) State transition probability matrix. (D) State summary statistics for dynamics: fractional occupancy; mean lifetime; mean interval; and switching rate.

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques: Activity Assay

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: Applying the canonical HMM approach provides a common set of networks that can be compared across studies and reduces the computational resources needed. A conventional (A) and canonical (B) HMM modeling approach was applied to the BioFIND dataset. Relative state power maps (1–45 Hz), coherence networks (1–45 Hz, top 3%), and summary statistics for dynamics are shown. State activations from each approach show good agreement (normalized confusion matrix shown in C) and the same differences between mild cognitive impairment (MCI) and healthy control groups (D). Group differences were calculated using a General Linear Model (GLM) with age and sex as confounds. Row-shuffle permutations were used for statistical significance testing with the maximum t -statistic across all features and states being used to control for multiple comparisons. The asterisk ( * ) indicates a p -value < 0.05 .

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques: Control

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: The canonical HMM networks can describe brain activity related to higher-order cognitive processes. A conventional (A) and canonical (B) HMM modeling approach was applied to the MEGUK dataset. Relative power maps (1–45 Hz), coherence networks (1–45 Hz, top 3%), and summary statistics for dynamics are shown. State activations from each approach show good agreement (normalized confusion matrix shown in C), and the same network response for the 2-back minus 0-back condition is observed (D). The dashed vertical line is the presentation of the stimulus. Horizontal bars indicate time points with p -value < 0.05 , which were calculated using sign-flip GLM permutations taking the maximum statistic (COPE) across time and states to control for multiple comparisons.

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques: Activity Assay, Control

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: MEG-like networks can be inferred from parcellated EEG data. A conventional (A) and canonical (B) HMM modeling approach was applied to the BioFIND dataset. Relative power maps (1–45 Hz), coherence networks (1–45 Hz, top 3%), and summary statistics for dynamics are shown. Fractional verlap between state activations from each modelling approach (i.e., normalized confusion matrix) is shown in (C). The same differences in state fractional occupancy for an eyes open vs closed contrast are found in both modeling approaches (D). Contrasts were calculated using a GLM with row-shuffle permutations for statistical significance testing using the maximum t -statistic across states to control for multiple comparisons. The asterisk ( * ) indicates a p -value < 0.05 .

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques: Control

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: The canonical HMM approach can be applied in MEG sensor space. This allows networks to be approximately inferred when source reconstruction is not feasible. (A) Topographic maps of the relative state power (1–45 Hz) at each MEG sensors viewed from above and PSDs for the 10-state canonical HMM networks at the sensor level, in the same order of states as for the parcel-space canonical HMM shown in . The solid line shows the state PSD averaged over regions (parcels), and the dashed line shows the PSD averaged across states. (B) Illustrative state probabilities inferred from the first 8 seconds of the first session in Cam-CAN when applying the canonical HMM networks to the sensor-level data (top) or the (prepared) parcellated data (bottom). (C) Normalized confusion matrix: average accuracy (across participants) for the state assignment at each time point between the sensor-level and source-level canonical models.

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques:

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: The canonical HMM infers consistent state time courses for differing data processing choices. Fraction of time points where state assignments agree (i.e., normalized confusion matrix) for data bandpass filtered over different frequency bands using the Glasser52 parcellation (A) and different parcellations using a 0.5–80 Hz bandpass filter (B). State time courses were inferred on the full Cam-CAN dataset.

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques:

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: The canonical HMM networks provide a useful description for a variety of tasks. (A) Network response (using the 10-state canonical HMM networks) to a passive visual task (left), passive audio task (middle), and visual/auditory sensorimotor task (right). See ) and ) for details regarding the tasks. (B) Network response for the higher-order networks (states 1, 2, 9, 10). (C) Network response for the lower-order networks corresponding to primary sense responses (states 3 and 4 for visual, states 7 and 8 for auditory, and states 5 and 6 for sensorimotor). Horizontal bars indicate time points with p -value < 0.05 , which were calculated using sign-flip GLM permutations taking the maximum statistic (COPE) across time and states to control for multiple comparisons.

Article Snippet: Therefore, the user should note the (current) canonical HMM was trained on Elekta data only and should control for scanner differences in subsequent analysis.

Techniques: Control

Journal: Imaging Neuroscience

Article Title: Canonical Hidden Markov Model Networks for studying M/EEG

doi: 10.1162/IMAG.a.1190

Figure Lengend Snippet: The canonical HMM approach can be applied in MEG sensor space. This allows networks to be approximately inferred when source reconstruction is not feasible. (A) Topographic maps of the relative state power (1–45 Hz) at each MEG sensors viewed from above and PSDs for the 10-state canonical HMM networks at the sensor level, in the same order of states as for the parcel-space canonical HMM shown in . The solid line shows the state PSD averaged over regions (parcels), and the dashed line shows the PSD averaged across states. (B) Illustrative state probabilities inferred from the first 8 seconds of the first session in Cam-CAN when applying the canonical HMM networks to the sensor-level data (top) or the (prepared) parcellated data (bottom). (C) Normalized confusion matrix: average accuracy (across participants) for the state assignment at each time point between the sensor-level and source-level canonical models.

Article Snippet: For scenarios where the end user does not wish to adopt the same source reconstruction, we provide a sensor-level canonical HMM for Elekta MEG.

Techniques: