3.0 Methods

Outline 3.0 Methods

• 3.1 Data Acquisition & Preprocessing
  • Human connectome project
  • participants
  • scanning parameters (fMRI, 7T?)
  • Preprocessing (FDR, Denoising?)
• 3.2 Selection of Regions of Interest (ROIs)
  • 3.3.1 Seed definition
    • FEF
    • IFJa
  • 3.3.2 Target Definition
    • what targets (liste)
    • where targets (liste)
• 3.3 Matrix Construction
  • Toolbox erklären
  • 3.4.1 functional connectivity
    • EC vs RSFC erklären
  • 3.4.2 Partial/full Correlation
  • 3.4.3 Correction (FDR) Erklärung
• 3.4 Brain Behavior Correlation
  • wie machen wir Predicitie Modelling (WM/Language_Story)

3.0 Methods

3.1 Data Acquisition & Preprocessing

3.1.1 Dataset and Participants

This study utilized resting-state functional magnetic resonance imaging (rs-fMRI) data from the S1200 release of the Human Connectome Project (HCP). To ensure high data quality and completeness, a strictly defined sub-group of $N=812$ healthy young adult subjects was selected. This specific cohort was chosen because these subjects completed all four rs-fMRI runs, and their data were reconstructed using the latest r227 reconstruction algorithm, which corrects for specific phase-encoding artifacts present in earlier releases.

3.1.2 Data Acquisition

Imaging data were acquired using a customized Siemens 3T Connectome Skyra scanner. The rs-fMRI data were collected utilizing a gradient-echo EPI sequence with the following standard HCP parameters: a repetition time (TR) of $720\text{ ms}$, an echo time (TE) of $33.1\text{ ms}$, a flip angle of $52^{\circ }$, and an isotropic spatial resolution of $2.0\text{ mm}$. Each subject completed two sessions on separate days. Each session consisted of two 15-minute runs with opposing phase-encoding directions (Right-to-Left and Left-to-Right), resulting in 1200 frames per run and a total of 4800 time points per subject. During the acquisition, subjects were instructed to keep their eyes open and maintain fixation on a red crosshair on a dark background.

3.1.3 Preprocessing

The raw fMRI data underwent preprocessing via the standardized HCP Minimal Preprocessing Pipelines. This first pipeline performs spatial distortion correction, motion correction, and registration to the structural final brain mask by reducing the bias field and normalization. Subsequently, the data were transformed from native mesh to fs_LR registered 32k mesh (2mm average vertex spacing).

To account for noise and motion artifacts, the data were denoised using ICA-FIX (Independent Component Analysis-based X-noiseifier). This automated method decomposes the data into independent components and regresses out those identified as structured noise (e.g., cardiac, respiratory, and head movement artifacts) without applying aggressive global signal regression.

---

Link zum Original

3.2 Selection of Regions of Interest (ROIs)

To test the hypothesis of supramodal prefrontal control over auditory processing, a comprehensive set of 36 target ROIs and 2 primary seed regions (plus one control seed) per hemisphere was defined. All regions were identified utilizing the multimodal cortical parcellation (HCP-MMP1.0) by Glasser et al. (2016) - Nature. This atlas provides superior neuroanatomical precision by integrating structural, functional, and connectivity data.

The selected ROIs are categorized into functional networks reflecting the auditory dual-stream architecture and their respective prefrontal controllers. The large sample size (N=812) provides sufficient statistical power to detect moderate connectivity effects across all ROI pairs after multiple comparison correction

3.2.1 The HCP-MMP1 Atlas

Before defining specific regions, the choice of the underlying atlas must be justified. Unlike traditional parcellations based solely on cytoarchitecture (e.g., Brodmann areas), the HCP-MMP1.0 atlas integrates four distinct modalities:

1. Cortical Myelin Content: Identified via T1w/T2w ratios.
2. Cortical Thickness: Measuring structural differences.
3. Task-fMRI Activation: Pinpointing functional hubs.
4. Resting-State Functional Connectivity: Mapping intrinsic networks.

This multimodal approach is essential for this study because it allows for the differentiation of functionally distinct areas that appear homogenous in classical maps. For instance, it enables the isolation of the IFJa from the surrounding prefrontal cortex, which is important for the auditory dual-stream architecture.

3.2.2 Prefrontal Seed Regions (The Conductors)

We selected distinct prefrontal control hubs based on the functional dissociation previously established in the visual system by Bedini & Baldauf (2021). This selection tests whether these visual control architectures map onto auditory processing.

• Frontal Eye Field (FEF): A core node of the Dorsal Attention Network (DAN), hypothesized to exert top-down control over the spatial auditory “Where”-stream.
• Anterior Inferior Frontal Junction (IFJa): Part of the Frontoparietal Network (FPN), hypothesized to mediate feature-based attention and semantic processing within the auditory “What”-stream Bedini & Baldauf (2021).
• Posterior Inferior Frontal Junction (IFJp) - Control Seed: Included strictly as a methodological baseline. IFJp is associated with the Multiple-Demand (MD) system for general-purpose executive tasks Bedini & Baldauf (2021). Factoring out its variance allows for a double dissociation, ensuring that the observed control over the auditory network is highly specific to the IFJa.

3.2.3 The Dorsal Pathways (Where & How)

The dorsal pathways process spatial localization (Where) and sensorimotor integration (How) (Hickok & Poeppel 2007 - Nature). ROI selection relies on the connectivity profiles outlined by Rolls et al. (2023) - Cerebral Cortex and recent functional data by Dureux (2024).

3.2.3.1 The Spatial-Orienting Network (Where-Stream)

This network tracks moving auditory objects and directs spatial focus.

• 7AL, 7Am, 7PC: Superior parietal areas receiving direct input from early auditory regions, crucial for auditory spatial attention and motion tracking Rolls et al. (2023) - Cerebral Cortex.
• PF, PFop: These areas correspond to the Inferior Perietal Lobule (Baker (2018)) and according to Rauschecker & Scott (2009) - Nature Neuroscience the IPL plays an important role in the dorsal auditory pathway (Glasser et al. (2016) - Nature)
• PFcm: Defined by Glasser et al. (2016) - Nature as a heavily myelinated inferior parietal region with strong somatosensory properties along with OP1-4 and FOP1. Although recent evidence indicates a lack of direct auditory responsiveness (Dureux et al., 2024), it was strategically included due to its anatomical position bridging the auditory-motor interface (OP4) and multimodal integration hubs (PSL).
• MT, MST: Classical supramodal motion hubs. They receive effective connectivity from auditory areas (A4/A5) to enable the tracking of auditory movement in space Rolls et al. (2023) - Cerebral Cortex.

3.2.3.2 The Motor Interface (How-Stream)

This interface translates auditory representations into articulatory motor plans (Sound-to-Motor mapping, Hickok & Poeppel (2004) - Cognition, Hickok & Poeppel 2007 - Nature)

• 55b: A premotor hub exhibiting responses to vocal stimuli being assigned to a language area. Interesting because 55b is a neighbor of the prefrontal seed region FEF and Glasser et al. (2016) - Nature; Dureux (2024).
• 44 (Broca’s pars opercularis): According to Rolls et al. (2023) - Cerebral Cortex Area 44 might be forming a dorsal auditory pathway with PBelt, A4, A5 and motion areas MT and MST.
• OP4, FOP1, FOP2, FOP3, 43: According to Frühholz the frontal operculum is connected to posterior STG and anterior IFG via dorsal pathways Frühholz (2015) - NeuroImage. OP4 specifically demonstrates exclusive sensitivity to human vocalizations Dureux (2024). Area 43 was included due to its anatomical position between the FOP and premotor cortex, potentially serving as an interface along the auditory-motor-pathway.
• SCEF (Supplementary and Cingulate Eye Field): A medial frontal cingulate area anatomically adjacent to the FEF. Despite responding almost exclusively to vocalizations (Dureux (2024)), its medial frontal anatomy and oculomotor system affiliation place it within the dorsal network. Dureux (2024) further groups SCEF functionally with premotor area 55b, OP4, and Broca’s areas 44 and 45 in the same vocalization-selective cluster.

3.2.4 The Ventral Pathways (What)

The ventral stream decodes auditory object identity, semantics, and speech perception.

3.2.3.1 Ventral Semantic and Identity Network

• STGa, STSda, STSdp, STSva, STSvp, TA2: The semantic core of the ventral pathway processing complex auditory objects. The STS complex integrates vocal inputs with facial motor representations to decode identity and message (Glasser et al. (2016) - Nature, Rolls (2022) - NeuroImage, Rolls et al. (2023) - Cerebral Cortex)
• AVI: The Anterior Ventral Insula shows activations to auditory stimuli alongside inferior frontal regions 55b, IFJp, IFJa, IFSp, 44, 45, and OP4. AVI exhibits highly specific capabilities in distinguishing vocalizations from noise, extending the semantic network into insular evaluation regions Dureux (2024).
• 45, 47l, IFSp: The frontal termini of the ventral meaning pathway. Area 45 (Broca’s pars triangularis) and 47l handle semantic selection, while IFSp shows responses to vocalization along with 44, 45, IFJp, OP4 and 55b ; Rolls et al. (2023) - Cerebral Cortex, Dureux (2024)

3.2.4.2 Anterior Temporal Semantic Regions

• TE1a, TGd, TGv (Temporal Pole): Following the effective connectivity analysis of Rolls (2022) - NeuroImage, these areas represent the semantic integration hubs of the temporal lobe. We intentionally differentiated between these regions to test the specificity of prefrontal control:
  • TE1a & TGd (Group 1): Associated with an inferior, visual-semantic system. Their lack of effective connectivity with the IFJa makes them ideal candidates to evaluate the boundaries of the IFJa-controlled auditory network.
  • TGv (Group 2): Integrated into a frontal system involving speech production and syntax. Its robust connectivity with the IFJa, FEF, and Area 55b makes it an interesting target for top-down modulation during linguistic and executive tasks.

3.2.5 Hierarchical Gateways and Connectors

These regions serve as routing hubs and convergence zones between early acoustic analysis and higher-order integration.

• PBelt: A4, A5: The primary “routers” exiting the auditory core. PBelt and A4 show effective connectivity to parietal regions 7AL, 8AM, 7PC. Rolls et al. (2023) - Cerebral Cortex suggests that PBelt, A4 and A5 might form a language-related dorsal pathway, adding them to the auditory where-stream. Glasser et al. (2016) - Nature parcellation on the other hand places A4 and A5 together with STSdp, STSda, STSvp, STSva, STGa, and TA2 in a region naming them auditory association cortex. This area rather belongs to the ventral what-stream, making a clear classification difficult. Ambiguities will be resolved and discussed in chapters 4.4 and 5. (wo auch immer ich das mache)   
• PGi: Effective connectivity to semantic areas as STS, TGv, TGd and TE1a place PGi in the auditory what-pathway. Rolls (2022) - NeuroImage places PGi in Group 1 along with STSva, STSvp and TE1a and TGd.
• PSL, STV, TPOJ1: Multimodal convergence zones (Group 3 networks) bridging auditory semantics with visual and somatosensory inputs (Rolls (2022) - NeuroImage, Rolls et al. (2023) - Cerebral Cortex) showing effective connectivity with PBelt, A4 and A5. While PSL is anatomically integrated into the semantic network, functional models suggest it acts as an abstract linguistic interface rather than a primary auditory responder Dureux (2024). That’s why we placed these regions in the semantic what-pathway for the following analyses. TPOJ1 shows weak responses to non-vocal sounds according to Dureux (2024).

3.2.6 Tables

Table 1: Prefrontal Seed Regions

Kürzel	Voller Name	Location (Stream)	Quelle
FEF	Frontal Eye Field	Prefrontal (Dorsal Attention)	Bedini & Baldauf (2021); Salmi et al. (2009)
IFJa	Anterior Inferior Frontal Junction	Prefrontal (Frontoparietal)	Bedini & Baldauf (2021)
IFJp	Posterior Inferior Frontal Junction	Prefrontal (Multiple-Demand)	Bedini & Baldauf (2021)

Table 2: Where & How Stream (Dorsal)

Kürzel	Voller Name	Location (Stream)	Quelle
7AL	Area 7 Anterior Lateral	Dorsal (Where - Spatial)	Rolls et al. (2023)
7Am	Area 7 Anterior Medial	Dorsal (Where - Spatial)	Rolls et al. (2023)
7PC	Area 7 Posterior Capsular	Dorsal (Where - Spatial)	Rolls et al. (2023)
PF	Area PF (Inferior Parietal)	Dorsal (Where - Spatial)	Baker (2018)
PFop	Area PF Opercular	Dorsal (Where - Spatial)	Rauschecker & Scott (2009)
PFcm	Area PF Complex Medial	Dorsal (Where/Somatosensory)	Glasser (2016)
MT	Middle Temporal Area	Dorsal (Where - Motion)	Rolls et al. (2023)
MST	Medial Superior Temporal Area	Dorsal (Where - Motion)	Rolls et al. (2023)
55b	Area 55b	Dorsal (How - Motor Relay)	Dureux (2024)
44	Area 44 (Pars Opercularis)	Dorsal (How - Motor)	Rolls et al. (2023)
OP4	Frontal Opercular Area 4	Dorsal (How - Motor interface)	Dureux (2024)
FOP1-3	Frontal Operculum 1, 2, 3	Dorsal (How - Motor planning)	Frühholz (2015)
43	Area 43	Dorsal (How - Motor planning)	Frühholz (2015)
SCEF	Supp. & Cingulate Eye Field	Dorsal (How - Cingulate Attn.)	Dureux (2024)
A4	Auditory Area 4	Gateway (Sensory Router)	Rolls et al. (2023)
PBelt	Parabelt Complex	Gateway (Sensory Router)	Rolls et al. (2023)

Table 3: What pathway (ventral)

Kürzel	Voller Name	Location (Stream)	Quelle
A5	Auditory Area 5	Ventral (What - Semantic Gateway)	Rolls et al. (2022)
STGa	Superior Temporal Gyrus Ant.	Ventral (What - Identity)	Glasser (2016)
STSda/dp	STS Dorsal Ant. / Post.	Ventral (What - Semantic)	Rolls et al. (2023)
STSva/vp	STS Ventral Ant. / Post.	Ventral (What - Semantic)	Glasser (2016)
TA2	Area TA2	Ventral (What - Semantic)	Glasser (2016)
STV	Superior Temporal Visual Area	Ventral (What - Multimodal)	Rolls (2023)
TPOJ1	Temp.-Par.-Occ. Junction 1	Ventral (What - Convergence)	Rolls (2023)
PGi	Area PGi (Inferior Parietal)	Ventral (What - Visual Semantic)	Rolls (2022)
AVI	Anterior Ventral Insula	Ventral (What - Evaluation)	Dureux (2024)
45	Area 45 (Pars Triangularis)	Ventral (What - Semantic)	Rolls et al. (2023)
47l	Area 47 Lateral	Ventral (What - Semantic)	Rolls et al. (2023)
IFSp	Inferior Frontal Sulcus Post.	Ventral (What - Semantic)	Dureux (2024)
TE1a	Area TE1 Anterior	Ventral (What - Periphery)	Rolls (2022)
TGd / TGv	Temporal Gyrus Dor. / Ven.	Ventral (What - Periphery)	Rolls (2022)
PSL	Perisylvian Language Area	Connector (Linguistic Interface)	Rolls (2023)

Link zum Original

3.3 Matrix Construction

3.3.1 Functional Connectivity

Resting-state functional connectivity (RSFC) is defined as the temporal correlation between neurophysiological events in spatially distinct brain regions Friston (1994). In this study, RSFC is operationalized as the statistical dependency between BOLD (Blood Oxygen Level Dependent) signal time series, reflecting the intrinsic functional architecture of the brain during resting-state Biswal (1995). Unlike EC, which models directed causal influences, RSFC is a symmetric, undirected measure that captures the degree to which two regions show correlated activity over time, independent of any explicit task or stimulus.

3.3.2 Time Series Extraction

We utilized the preprocessed dense CIFTI timeseries from the HCP S1200 release. For each subject, the BOLD signal was spatially averaged across all vertices within each of the 360 parcels (180 per hemisphere) of the HCP MMP1 atlas (Glasser et al. (2016) - Nature). Mean BOLD time series were extracted for all seed regions (FEF, IFJa, IFJp) [WELCHE noch? 55b, 44, 45?], as well as all auditory target regions defined in section 3.2.

3.3.3 Connectivity Matrix Construction

Pairwise RSFC values were computed as Pearson correlation coefficients between the extracted BOLD time series of each seed region and all parcels of the HCP-MMP1 atlas - consistent with the HCP’s own connectivity analyses (Glasser et al. (2016) - Nature). This led to a connectivity vector for each subject representing the full RSFC profile for each seed across the cortex. All matrix construction steps were implemented using a custom MATLAB toolbox developed within the Baldauf Research Group. [TOOLBOX Ref - Daniel fragen]

3.3.4 Single Seed vs. Contrast Analysis

We used two complementary analysis modes. In the single-seed analysis, the functional connectivity of a given seed region (FEF or IFJa) was assessed against the mean whole-brain connectivity of that subject, detecting parcels that are significantly more correlated with the seed region than the average cortical region. In the contrast analysis, the connectivity profiles of two seed regions were directly compared, revealing parcels with preferential connectivity to one seed over the other.

3.3.5 Partial vs Full Correlation

When characterizing functional brain networks, a critical distinction needs to be made between full and partial correlation. Full Correlation calculates the pairwise statistical relationship between two regions A and B without controlling for other variables. Because any two regions that both correlate with a third region C will appear correlated with each other - therefore displaying indirect correlation without true connectivity. Partial Correlation on the other hand, estimates the relationship between A and B after regressing out the shared variance to all other simultaneously measured regions (e.g. region C, Marrelec 2006 - NeuroImage, Smith (2011)). As a result, the partial correlation connectome is considerably sparser than its full correlation counterpart, while retaining the most direct functional connections (Glasser et al. (2016) - Nature).

In this study, both measures are applied: full correlation provides a global view of each seed’s connectivity landscape, while partial correlation isolates direct connections after filtering out contributions from the other regions included in the analysis.

3.3.6 Statistical Testing: Wilcoxon Signed-Rank Test & FDR Correction

Statistical significance of RSFC estimates was assessed using the paired Wilcoxon signed-rank test, an alternative to the Gaussian-based paired t-test. This test was chosen because RSFC values derived from Pearson correlations do not necessarily follow a Gaussian distribution across subjects (Soyuhos, O., & Baldauf, D. (2022)) For the single-seed analyses, a one-tailed test was applied, evaluating whether connectivity with the seed region exceeded the mean whole-brain connectivity. For contrast analyses, a two-tailed test was used to determine whether connectivity was significantly higher for one seed than for the other.

Each seed region (FEF or IFJa) was tested independently; their connectivity profiles were not averaged before statistical testing, preserving each region’s functional fingerprint.

To control for the elevated risk of false positives arising from testing all 360 cortical parcels simultaneously, all p-values were corrected for multiple comparisons using the False Discovery Rate (FDR) procedure (Benjamini, Y., & Hochberg, Y. (1995)). In contrast to the more stringent Bonferroni correction, FDR correction maintains higher statistical power by controlling the expected proportion of false positives among all rejected null hypotheses, rather than the probability of any single false positives. A significance threshold of q < 0.05 (FDR-corrected) was applied in all analyses.

For visualization, significant p-values were converted to z-scores. The z-scores were summed across all seeds for which a given parcel reached significance, and then divided by the total number of seeds showing significance for that parcel. The normalization generates a score that weights connectivity strength by consistency of the effect across seeds.

Link zum Original

3.4 Brain-Behavior Correlation

To validate our RSFC patterns performed in Section 4, we assessed whether individual differences in RSFC could predict performance on behavioral tasks that specifically test the auditory ‘what’ and ‘where’ pathways. This approach links the network architecture observed during rest to cognitive function during task engagement, contributing validity for the supramodal organization hypothesis.

3.4.1 Behavioral Subject Subsample

For the connectome-based predictive modeling, we used behavioral data of N = 371 participants from the HCP S1200 release. While the full resting-state cohort counts $N=812$ participants, restricting the behavioral analyses to the smaller subset guarantees complete data across all evaluated tasks. For each task, subjects were additionally excluded if they exceeded a head motion threshold of 0.15 mm framewise displacement or had missing behavioral scores, resulting in slightly varying effective sample sizes per analysis (see sections 3.4.2.1 and 3.4.2.2).

3.4.2 Behavioral fMRI Tasks

Three fMRI tasks were selected to evaluate the auditory dual-stream framework, defined in our study. The tasks were selected to enable a functional dissociation between semantic processing ‘what’, auditory-spatial decoding ‘where’, and low-level acoustic filtering, while acknowledging the limitations of the HCP task set.

3.4.2.1 Language Task (Story Comprehension)

In the story condition of the HCP Language Task, participants listened to short auditory stories adapted from Aesop’s fables, followed by a two-alternative forced-choice (2-AFC) question testing the semantic content of the stories (Binder (2011), Barch et al. (2013)). Two performance measures were extracted: Story Accuracy measures the proportion of correct responses, and since the vocabulary of the stories is simple, we observe a ceiling effect in which the sustained attention is measured rather than semantic access. However, Story Median Reaction Time captures the speed of semantic retrieval and therefore is used as the primary measurement for ventral ‘what’-stream efficiency. For this analysis, 20 subjects were excluded from the N = 371 subsample (18 due to head motion > 0.15 mm, 2 due to missing scores), yielding an effective N = 351. The Spearman correlation between behavioral scores and head motion was $r=0.00$ ($p=0.939$), confirming that motion artifacts did not confound the results.

3.4.2.2 Working Memory Task (2-Back Place)

The HCP Working Memory Task consists of a visual N-Back design in which participants track sequential images and respond whether the current image matches the one presented two trials before (Barch et al. (2013)). The place subcondition, in which stimuli consisted of landscape and scene photographs, was extracted as a measure for visuo-spatial working memory (Barch et al. (2013)).
This task was chosen as the closest available test of spatial processing capacity, linked to the FEF’s dorsal stream. Because the HCP dataset does not include an auditory spatial task, this visual task tests the supramodal hypothesis. Participants might solve these tasks via object recognition rather than spatial navigation, which might reduce observed effects in the dorsal network. Additionally, since IFJa has been shown to be involved in working memory (Bedini & Baldauf (2021)), its connectivity may contribute to Place task predictions independently of any semantic strategy, making it difficult to cleanly attribute ventral model effects to a single mechanism. For this analysis, 19 subjects were excluded from the N = 371 subsample (18 due to head motion > 0.15 mm, 1 due to missing scores), yielding an effective N = 352. However, the Spearman correlation between behavioral scores and head motion was $r=-0.16$ ($p=0.002$), indicating a significant association between motion and Place Acc performance. This constitutes a potential confound and is discussed as a limitation in section 5.7.

3.4.2.3 Words-in-Noise (NIH Toolbox Noise Comparison)

The NIH Toolbox Words-in-Noise test assesses the ability to recognize spoken words in background noise by adaptively decreasing the signal-to-noise ratio across trials (Zecker (2013)). The task is a valid measure of low-level acoustic signal separation rather than high-level semantic processing. This test was included as a control paradigm to determine whether IFJa network connectivity generalizes across auditory tasks or whether acoustic signal separation relies on a functionally different network.

3.4.3 Connectivity-Based Behavioral Prediction

The connectivity values were derived from partial correlation matrices computed across the N = 371 subsample, reflecting direct functional coupling between ROIs with shared variance removed. The prediction model itself was fitted using linear correlation between the connectivity predictor matrix and the behavioural outcome variable. Separate models were computed for the ventral ROIs subset and the dorsal ROIs subset, enabling predictions specific to the streams.

We evaluated model performance using K-fold cross-validation (K = 371, leave-one-out, as the primary analysis; K = 10 for comparison), providing an R value (correlation) as the indicator of predictive accuracy. A significance threshold of p < 0.05 was applied for all tasks.

Link zum Original

Notes & Scrapbook

Hier Dinge abladen, die noch keinen Platz im Text haben, damit der Schreibfluss nicht stoppt.