NSCI 270 Content Review

What is structural MRI? What is the difference between T1 and T2 weighted images, and how is MRI’s resolution (spatial vs. temporal)? What things can be measured using MRI?

• Magnetic Resonance Imaging

• Structural MRI provides detailed images of brain anatomy using magnetic fields and radio waves (non-invasive)

• Static, does not measure brain activity — used to examine brain structure, anatomy, and pathology

• Resolution:

  • Spatial: Very high (sub-millimeter possible)

  • Temporal: N/A, doesn’t capture time-varying signals

• Weighted Images:

  • T1: highlight gray vs. white matter

  • T2: show fluid differences

What is fMRI? How does it work and what does it actually measure? What is it’s resolution (spatial vs. temporal)?

• Functional Magnetic Resonance Imaging

• How does it work?

  • Nerve cells communicate with each other in the brain, neurons in active parts of the brain require oxygen to communicate

  • Hemoglobin is a protein found in blood that transports oxygen to different parts of the body

  • Signal recorded by the MRI scanner = Blood Oxygen Level-Dependent (BOLD) Signal → reflects changes in blood oxygenation that occur with neural activity

  • Hemodynamic Response (HDR): Process by which blood flow increases to active brain regions, causing changes in oxygen levels that generate the BOLD signals

• Measures brain activity indirectly via BOLD changes, increased blood flow to active areas

• Resolution:

  • Spatial: high (depending on # of voxels, ~ millimeter)

  • Temporal: low, due to slow HDR

What is DTI? What does it measure? What is its resolution and how can we use it in neuroscience?

• Diffusion Tensor Imaging

• MRI-based structural connectivity technique that measures how water diffuses in brain tissue, particularly along white matter fibers (myelinated axons)

• Resolution:

  • Spatial: moderate to high, depends on scan parameters (~ 1-2 mm)

  • Temporal: N/A, static structural measure

• Example of use: studying corpus callosum connectivity in multiple sclerosis or mapping major tracts before neurosurgery

What is tractography and how is it modeled with DTI? What measurements are necessary?

• Tractography: color-coded brain tractography image showing major white matter pathways

• This technique models anisotropy (directional preference of water movement) to map the orientation of fiber tracts

• Anisotropy vs. Isotropy:

  • Anisotropy: Restricted diffusion of water moving along a preferred direction

  • Isotropy: Random, equal diffusion of water everywhere

• Measurements:

  • Fractional Anisotropy (FA):

    • Summary measure of microstructural integrity

    • Directional diffusion of water

    • Highly sensitive to microstructural changes

    • Less specific measure: reduction in FA does not indicate with biological process is driving change

    • [0,1]

  • Mean Diffusivity (MD):

    • Inverse measure of the membrane density

    • Represents the average overall magnitude of water diffusion

    • Very similar for both gray and white matter, and higher for CSF

    • MD is sensitive to cellularity, edema, and necrosis

  • Axial Diffusivity (AD):

    • Tends to be variable in WM changes and pathology

    • Measure diffusion parallel (axial) to the main directional vector

    • AD decreases in axonal injury

    • The ADs of WM tracts have been reported to increase with brain maturation

  • Radial Diffusivity (RD):

    • Increases in white matter with demyelination or dysmyelination

    • Measure diffusion perpendicular (radial) to the main directional vector

    • Changes in the axonal diameter or density may influence RD

What is PET and what does it detect? How is it measure and what is its resolution? How may we use it in research?

• Positron Emission Tomography

• Measures molecular activity by detecting gamma rays from injected radiotracers

  • Radiotracer: a safe, injectable radioactive chemical the scanner detects — diseased cells absorb large amounts of the radiotracer

• Reflects metabolism, receptor binding, protein aggregation, etc.

• Resolution:

  • Spatial: moderate

  • Temporal: low (minutes) — depends on tracer kinetics

• Minimally invasive — requires radiotracer injection but gives molecular specificity you can’t get from fMRI or EEG

• Example of usage: mapping hypometabolism in neurodegenerative disorders or using receptor-specific tracers to study psychiatric disorders

What is EEG? What is it’s resolution and how is pertinent neural activity extracted? What’s the difference between wet and dry EEG?

• Electroencephalography

• Measures electrical activity from the scalp, generated by large populations of neurons

  • Wet EEG: gel on scalp, placed electrodes

  • Dry EEG: head set

• Signal Measurement:

  • Event-Related Potential (ERP): derived from EEG data by extracting and averaging the brain responses time-locked to specific events

  • Averaging process reduced background noise and highlights the neural activity related to the event of interest

• Resolution:

  • Spatial: low (cm), due to signal distortion by skull and scalp

  • Temporal: extremely high (milliseconds) — great for studying the timing of brain processes

What is MEG? What is it’s resolution and how is it used? What are the pros and cons of it?

• Magnetoencephalography

• Non-invasive technique detecting tiny magnetic fields generated by synchronized neural currents

• Resolution:

  • Spatial: Better than EEG

  • Temporal: High (ms)

• Usage:

  • Magnetic fields are less distorted by the skull and scalp than electric potentials → cleaner timing and localization

  • Requires magnetically shielded room, which is expensive and deeper sources are still harder to detect

  • Uses: timing of language/memory processes, mapping epileptic foci pre-surgery, studying sensory processing dynamics

  • Often combined with MRI (for anatomical constraints) or EEG (to leverage complementary strengths)

What are some modalities (beyond neuroimaging) we can use to study the brain?

• Electrophysiology:

  • Invasive recordings (Single-Unit, Local Field Potential, ECoG) offer high temporal and spatial resolution

  • Capture neural activity from individual neurons to local populations

  • Used in animals and clinical contexts to study sensory processing, motor planning, and disease

• Genetic and Transcriptomic Data:

  • Provides insights into gene expression across brain regions and cell types

  • Tools like Allen Brain Atlas help connect molecular biology with brain structure and function

  • Crucial for understanding developmental trajectories and neurological disorders

• Behavioral and Cognitive Data:

  • Assesses task performance, decision-making, and reaction times

  • Often used in combination with neurophysiology for integrative brain-behavior analysis

  • Techniques include eye-tracking, Stroop tasks, and other cognitive paradigms

Why is organization needed in any dataset? What are some common methods used to organize data?

• Standardization: ease of interpretation

• Modes of Organization

  • Brain Imaging Data Structure (BIDS) format

    • Contains folders for each participant, separated by modality and tasks

    • MRI suffix = nii.gz

  • OpenNeuro Specific Organization: BIDS format with additional constraints for sharing and validation, includes metadata (acquisition parameters and task information)

  • NWB: Standardizes how raw signals, stimulus events, and metadata are stored

  • XCEDE: Older format for metadata

  • COINS: Database system for multi-site studies

  • DataLad: Data management tools for large datasets

  • HDF5-Based Formats: Hierarchal data format for large datasets

Why does open science matter?

1. Reproducibility: Allow other researchers to validate findings by repeating analyses

   • Reduces errors, builds confidence in results

2. Collaboration: Fosters collaboration across labs, disciplines, and countries

   • Accelerates discovery and enables interdisciplinary work

3. Transparency: Lets everyone see exactly how conclusions were drawn, combatting questionable research practices and promoting ethical science

4. Accessibility/Equity: Lowers barriers for researchers laving funding for expensive data collection

   • Democratizes research and fosters diversity in ideas and approaches

5. Reuse/Efficiency: Allows researchers to ask new questions without costly or time-consuming data collection

   • Maximizes the scientific value of data, saves resources, and speeds up innovation

6. Real-World Impact: Allows for large-scale projects like the Human Connectome Project and Open Neuro

   • Encourage open practices and prepare students to contribute to research

What is Metadata?

• Provides important information that describes, explains, or gives context to the main data you are working with

• Helps researchers understand, interpret, and properly use the data

• Makes sharing and reusing datasets possible because others can know exactly what the data represents and how it was collected

• Ex:

  • When/how MRI scans were collected (scanner type and settings)

  • Participant info (age, gender, clinical diagnosis)

  • Experimental conditions or task descriptions

  • File formats and data organization details

What are some pros and cons with using existing data, as opposed to collecting new data?

• Using existing open access data sources:

  • Pros:

    • Fast and efficient

    • Can build off existing work

    • Transparency and replicability

  • Cons:

    • Data collection practices may be unethical or irresponsible

    • May not directly answer question

• Collecting new data in a lab:

  • Pros:

    • Can best align data with question

    • Can ensure ethical and responsible data collection practices

  • Cons:

    • Can be resource-intensive (costly/time-consuming)

    • Requires considerable expertise

What is Big Data? How is it described by leaders in the field? What techniques are required for its analysis? How is it evolving in our field?

• Observational study data, large datasets, technology-generated outcomes

• Machine-learning algorithms

• Extremely large datasets to be analyzed computationally to reveal patterns, trends, and associations

• Application of mathematical techniques to large datasets to infer probabilities for prediction and find novel techniques

• Require specific technologies and analytical methods

• High volume, velocity, and variety

• Large, more varied, and complex structure with challenges in storing, analyzing, and visualizing it

• More availability of data, new technologies to collect data

What are three types of data we have in neuroscience?

• Cross Sectional: Collects data from a specific group of individuals at a single point in time

  • Static (snapshot)

  • Multiple subjects at one time point

  • Identifying differences between groups

  • Limitation: cannot track changes over time

• Longitudinal: Collects data from the same group of subjects over an extended period of time

  • Same group of people, tracking progression

  • Limitations: Very large, complex data sets

• Multimodel: Integrates different types of observations or imaging (ex: neuroimaging, ultrasound, etc.)

What are the 5 V’s of Big Data? What are some opportunities and challenges that each present?

1. Volume

   1. Scale enables discovery (rare event detection, multi-cohort/site data)

   2. Challenges: Storage and processing

      1. High-volume data leads to significant costs, which requires careful forecasting of data growth and computational needs (budget planning)

      2. May need institutional support ($$) and high-capacity clouds

      3. Collaboration and shared resources can mitigate costs

2. Velocity

   1. Real-time neuroscience (rapid data acquisition)

   2. Challenges: Computational complexity

      1. More data = harder to clean (removing data - inherently losing signal)

3. Variety

   1. Multi-model insight

   2. Challenges: Computational complexity

      1. High dimensionality problems

      2. Many modes of imaging data → more complex analysis

4. Veracity

   1. Trusted data from reliable science

   2. Challenges: Cleaning and pre-processing

      1. Inconsistent data can bias our results and lead to misinterpretation

      2. Data harmonization between projects

      3. Common templates (like the MNI) can align data from a particular voxel/time-series

5. Value

   1. Impact (important, clinically applicable research)

   2. Challenges: Reproducible science

      1. Allows for new abilities in psychiatry, use of AI-supported predictive modeling

      2. Ex: Diagnostic models, prognostic models and disease progression measurements, patient stratification

What is preprocessing and why is it necessary? What does it do?

• Pre-Processing: Series of automated and manual steps used to clean, correct, and standardize raw brain data before running statistical analysis or creating maps of brain function

• Why?

  • Raw data from scanners/sensory is noisy and highly dimensional (1000-millions of measurements per subject)

  • Data is always contaminated by scanner instability, physical flaws and head movements

    • Ex: magnetic/electrical interference, scanner drift, head movements, etc.

• What does preprocessing do?

  • Removes noise — eliminates scanner instability, thermal static, and environmental interference

  • Corrects artifacts — fixes image distortions caused by body physiology (heartbeats, breathing) and patient movement

  • Standardizes space — aligns individual brain scans to a universal anatomical template (MNI/Talarich) so data can be compared across groups (data harmonization)

• Pre-processing improves signal-to-noise ratio (SNR), measuring the physiological/metabolic brain signal relative to the unwanted background fluctuations, determining the clarity and reliability of the data

• Ensures comparability between subjects and meaningful statistical analysis

What are some common artifacts and what are their impacts on data? How are movement artifacts mitigated?

• Movement Artifacts:

  • Blurring: Movement spreads signals across pixels, reducing spatial resolution and clarity

  • Ghosting: Displaced data creates faint, repeated “shadow” images across the scan matrix

  • Ring: Sudden jerks cause bright concentric ripples around brain boundaries

  • Cardiac Pulsation: Leads to periodic ghosting streaks or focal blur around major blood vessels, the brainstem, and ventricles

  • Breathing: Impacts global brain activity

  • In EEG specifically: Eye blinks, respiration, lateral eye movement, and jaw clenching

• Equipment Drift and Environmental Interference: Gradual signal change over time (scanner drift), power line interference

• Impacts:

  • False Connectivity: Can falsely look like neural synchronization

  • Motion-Related False Positive Activation: Signal in a region interpreted as a response to a task, but may be related to physiology (ex: vascular/CSF/cardiac pulsations, eye motion)

• Corrected by algorithms post-imaging and by physical head restraints/real-time prospective motion correction during imaging

What are the 3 core goals of pre-processing?

1. Standardization/Normalization: Harmonizes the scanner hardware and imaging protocols/Mathematically rescales the resulting data’s spatial dimensions and pixel intensities for consistent analysis

2. Denoising: Suppresses physiological and scanner artifacts to reveal the true underlying neural signal

3. Feature Extraction: Converts raw structural and functional neuroimaging data into quantifiable biomarkers for statistical and clinical analysis

   1. Ex: cortical thickness, tract density, functional connectivity metrics

   2. Pull out areas of interest

What are the differences between data normalization, transformation, and standardization in pre-processing?

• Data Normalization: Ensuring features are on a consistent scale across subjects and measurements

  • Rescale data to a fixed range

  • Adjust voxel intensities and sensory readings to a shared scale (ex: 0 to 1, z-scores, or %s)

  • This enables signals to be compared accurately across different tasks and subjects

  • Spatial normalization warps uniquely shaped brains onto a standardized 3D template to ensure anatomical structures along across everyone in a study (common reference space)

• Data Transformation: Mathematically adjusting raw data to meet statistical assumptions for reliable group analyses

  • Reduces skewness, stabilizes variance, and improves the validity of downstream statistical tests or models

• Data Standardization: Transforms data so it’s in the same numerical space (ex: mean = 0, standard deviation = 1)

  • Allows for group-level comparisons

How are outliers detected/is quality controlled in neuroimaging?

• Removing outliers → removing bias and data skew

• Ensures data is accurate, reliable, and free from major artifacts or errors

• Identification of substantial deviations:

  • Visual Inspection (gold standard) — manual removal

  • Statistical Methods:

    • Feature extraction

    • Thresholds, z-scores > 5

    • Outliers flagged

  • Machine Learning Approaches:

    • Isolation forest algorithm detects abnormal connectivity and patterns deviating from cohort distribution → abnormality score

• Things to check for in QC:

  • DTI: Motion artifacts, signal dropout, distortions

  • MRI: Image quality, contrast, coverage

  • fMRI: Head motion, low SNR, ghosting

  • EEG: Noisy/flat channels, muscle/eye-blink artifacts, amplifier saturation, power spectrum consistency, correlation across electrodes

What is dimensionality? Why do we need to do dimensionality reduction?

• Dimensionality:

  • Refers to the number of data points that make up a single brain scan

  • 1 scan: > 100k to 1 million voxels

• Dimensionality Reduction:

  • Re-representing massive, multi-dimensional brain data (complex 3D data sets)

  • Into a lower-dimensional space

  • Has fewer features

  • Still preserves most critical, meaningful structural or functional properties

  • Why?

    • Beat the curse of dimensionality — reduction in the number of features prevents overfitting (finding trends that aren’t there)

    • Fixes multiple comparisons — minimizes false positives (Type I Error), reduces the # of independent statistical tests used

    • Eliminates spatial redundancy — resolves collinearity by compressing correlated neighboring voxels into a single unified network signal

    • Boosts computational speed

What are some common dimensionality reduction techniques?

• Region of Interest (ROI) Analysis:

  • Instead of analyzing 1,000,000 individual brain voxels, researchers use an anatomical atlas to segment the brain into 100 or 200 distinct regions

  • The voxels within each region are averaged together, instantly shrinking the dataset’s dimensions by several orders of magnitude

• Principal Component Analysis (PCA):

  • A mathematical approach that finds orthogonal (uncorrelated) linear combinations of voxels that explain the maximum amount variance across the subjects

  • It condenses the entire brain’s variance into a handful of principal components

• Independent Component Analysis (ICA):

  • Widely used in functional (MRI), ICA separates a massive, mixed spatiotemporal dataset into distinct, independent brain networks

  • Turns thousands of voxel time-courses into a few meaningful network signals

Compare and contrast PCA and ICA in terms of goal, components, focus, use case, and assumptions.

• PCA (components that explain the most variance):

  • Goal: Reduce dimensionality while preserving maximum variance

  • Components: Orthogonal (uncorrelated)

  • Focus: Captures variance in the data

  • Use Case: Noise reduction, feature compression, identifying major patterns (Dimensionality reduction, visualization)

  • Assumption: Data can be represented by uncorrelated components (linear combination)

• ICA (components that are statistically independent even if variance is small):

  • Goal: Separate mixed signals into statistically independent sources

  • Components: Statistically independent (not necessary orthogonal)

  • Focus: Captures independent signals in the data

  • Use Case: Artifact removal, source/signal separation, identifying functional networks

  • Assumption: Observed signals are linear mixtures of independent sources (Linear combinations, non-Gaussian sources)

What is feature extraction and why is it important? How do you choose a feature?

• Feature Extraction: Process of transforming raw neuroimaging data, often high-dimensional and noisy, into meaningful summary measures that can be used for analysis, modeling, or ML

• Extracted features summarize key patterns, and making downstream statistical or ML analyses feasible and interpretable

• Neuroimaging produces enormous, noisy arrays (voxels x time for fMRI, channels x time for EEG/MEG)

• Directly feeding raw arrays into statistical tests or ML models is feasible because of:

  • Computational cost and storage

  • High-dimensional → overfitting with limited subjects

  • Noise and artifact sensitivity (motion, physiological artifacts, eye blinks)

  • Difficult interpretability of model weights in raw space

• Feature extraction compresses the signal into smaller sets of interpretable, task-relevant summaries (ex: time-series, band powers, connectivity edges, spatial maps) so models are faster, more robust, and results are more interpretable

What is feature selection and what are some methods we can use?

1. Filter Methods: Sort and select features using quick statistical tests before any machine learning begins

   • Instantly screen out millions of useless background brain voxels to save computing power

   • Select features based on statistics (independence of model)

   • Pros: Fast, simple

   • Cons: Ignores feature interactions

   • Role of ML: None

   • How it works: Scores each brain region individually and drops low scores

   • Key Examples: ANOVA, correlation, t-tests

   • Main Advantage: Extremely fast, handles massive voxel data instantly

2. Wrapper Methods: Use trial-and-error to test different combinations of features directly inside the AI model

   • Excellent for detecting how different brain regions interact together as complex functional networks

   • Use a predictive model to evaluate feature subsets

   • Pros: Considers interactions, can be highly predictive

   • Cons: Computationally intensive

   • Role of ML: Used as an external judge to grade feature combinations

   • How it works: Trains the model repeatedly using trial-and-error subsets

   • Key Examples: Recursive Feature Elimination (RFE) with classifier feedback

   • Main Advantage: Tests how different brain regions interact together

3. Embedded Methods:

   • Automatically choose the most important features while the AI model is actively training

   • Naturally reward key brain networks and ignore noise during learning, balancing speed and accuracy

   • Happens during model training via regularization

   • Pros: Efficient, integrates selection with modeling

   • Cons: Requires parameter tuning

   • Role of ML: Built directly into the training algorithm itself

   • How it works: Trains the model exactly once, picking features automatically

   • Key Examples: LASSO, ElasticNet regression

   • Main Advantage: Very efficient and highly resistant to overfitting

What is Multi-Modal Feature Integration? What information can be obtained from different modalities?

• Complementary information: Different modalities capture distinct aspects of the brain

  • fMRI: Functional connectivity, network dynamics

  • DTI: Water matter integrity and tract-specific metrics

  • Structural MRI: Cortical thickness, gray matter volumes

  • EEG: Oscillatory activity and temporal dynamics

  • PET: Molecular and metabolic activity

• Enhanced prediction: Combining features can improve accuracy in classification, disease detection, or cognitive outcome prediction

• Biomarker discovery: Multi-modal integration helps identify robust, multimodal biomarkers that may be more sensitive and specific than single-modality features

What are the key steps of Multi-Modal Feature Integration?

1. Feature extraction per modality: Compute modality-specific features (ex: fMRI connectivity matrix, DTI FA, PET SUVR)

2. Dimensionality Reduction: Apply PCA, ICA, or other methods to reduce feature dimensionality within each modality — prevents model overload and overfitting due to high-dim data

3. Normalization/Scaling: Bring all features onto comparable scales

4. Feature Integration: Concatenate or fuse reduced feature sets into a single multiple-modal dataset

5. Analysis/Modeling: Apply ML or statistical modeling for prediction, classification, or biomarker discovery

Why do we visualize data, and how do we decide how to visualize? What dimensions does neuroscience data have?

• Why? — Answering research question in a clear and simple manner

• Question dictates type of visualization

• Dimensions:

  • Time — When does activity happen?

  • Space — Where does activity occur?

  • Participants — How do people differ?

  • Connections — Which regions are related?

  • Frequency — What rhythms are present?

  • Condition — Control vs. patient, task A vs. B

  • Timepoint — Baseline vs. follow-up

What is being visualized in EEG? What are some key visualization questions we have? What are the 5 frequency broad bands in EEG?

• EEG measures: electrical potential, measured at electrodes with very high temporal resolution (=/= individual neurons)

• Key Visualization Questions:

  • When did activity occur?

  • What does the signal look like?

  • Where on the scalp did it occur?

  • What frequencies are present?

• Frequency Broad Bands:

  • Delta (0.3-4 Hz): Dreamless sleep, subconscious, loss of bodily awareness

  • Theta (4-8 Hz): Inner peace, REM sleep, deep meditation

  • Alpha (8-13 Hz): Creativity, flow state, focus

  • Beta (13-30 Hz): Concentration, arousal, alertness

  • Gamma (30 Hz +): Multi-processing

What are some common visualizations used in EEG? What are FFS, PSD, and wavelet?

• Visualizations:

  • Topographic Maps:

    • A 2D representation of EEG activity across the scalp

    • Each electrode has a location

    • Color represents voltage/activity (intensity)

    • Used when visualizing spatial distribution of EEG activity, event-related potentials, frequency band power, differences between conditions, changes over time (not a time-series), etc.

  • Time-Series Visualization:

    • Wave form visualization

    • Oscillations: Repeated peaks and troughs

    • Amplitude: How large are the fluctuations?

    • Frequency: How quickly does the signal oscillate → brain state

    • Artifacts: Large or unusual changes may arise from eye blinks, muscle activity, movement, and electrical interference

    • Multiple time-series → superimposing several different channels (comparative view)

  • Average Signal/Variability:

    • Overall activity across the brain → mean

    • Is variation resulting as a result of an event or is it an outlier → use standard deviation

  • Group Comparisons

• Frequency Visualization:

  • Time domain: what happens over time? Freq. domain: what frequencies make up the signal? Freq. spectrum: can show peaks

  • Fourier Frequency Spectrum (FFS): Shows the individual frequency components (amplitude and phase) that make up a signal

  • Power Spectral Density (PSD): Shows how the total power of a signal is distributed across those frequencies

  • Wavelet: Maps how brainwave frequencies change over time

What does fMRI visualize?

• BOLD signal (=/= direct measurement of neural firing)

• High spatial resolution

• Relatively slow temporal dynamics

• Data is inherently 3D + time

• Key Question: Where in the brain is the signal changing?

What is correlation? How are correlation heat maps used?

• Correlation asks: Do two variables tend to change together?

• Correlation DOES NOT TELL US that one variable causes the other (Correlation =/= Causation) — requires stats to back it up

• Correlation Coefficient:

  • Used to find how strong a relationship is between data

  • Value between -1 and 1

    • 1 indicates a strong positive relationship

    • -1 indicates a strong negative relationship

    • 0 indicate no relationship at all

• Correlation Heat Maps:

  • A heatmap represents numbers using color → instead of looking at hundreds of individual lines, we can see many patterns across many measurements simultaneously

  • EEG: rows → channels, columns → time, color → voltage

  • fMRI: rows → participants, columns → voxels/regions, color → activation

  • DTI: rows → participants, columns → connections, color → connectivity strength

  • How to read a heatmap:

    • Read the axes — what does each row/column represent?

    • Read the colorbar — what does dark/light or blue/red mean?

    • Look for patterns — horizontal patterns, vertical patterns, blocks/clusters and isolated extreme values

    • Biological meaning — do several channels show a similar pattern at the same time? (synchrony/correlation)

What are some common visualizations used with fMRI?

• Brain Networks:

  • Correlation networks

  • A network is only as meaningful as the definition of the nodes. edges, and threshold

    • Lower threshold = more edges = potentially more noise

    • Higher threshold = fewer edges = potentially hides meaningful relationships

    • Changing the threshold changes the network

    • Adjust threshold based on literature

  • Node: A brain region or channel (EEG electrode, fMRI region, DTI brain region)

  • Edge: A relationship between two nodes

  • Edge weight: A relationship between two nodes

  • Hub: A highly connected node

• Brain Maps/fMRI Slice Montage:

  • A brain map shows where an effect occurs

  • Typical visualization: anatomical brain image, statistical/activation overlay, color represents the direction or magnitude of the effect

  • Important questions: Which anatomical region? Which direction? How strong is the effect? What statistically threshold was used? What comparison produced the map?

What are we visualizing in DTI? How is directionality displayed in DTI?

• Measures the directionality of water diffusion

• Used to estimate: White matter organization, fiber orientation, structural connectivity, tract integrity and properties

• Key Question: How is white matter organized and connected?

• Directionality:

  • Depending on the visualization: color can encode diffusion direction, brightness can encode diffusion properties

  • Color is an encoding of a variable

What does DTI tractography represent?

• Lines represent estimated white-matter pathways

• ESTIMATED using mathematical properties

• Based on diffusion measurements (diffusion properties of water)

• They are not photographs of individual axons

What are some common ways to visualize data distributions? Why do we visualize outliers?

• Boxplot (box and whiskers):

  • Middle Line: Median

  • Box: Middle 50% of observations

  • Whiskers: Spread according to the plotting convention

  • Points beyond whiskers: Potential outlier

• Scatterplots:

  • Are two variables related?

  • Upwards/downwards trends, obvious relationships or outliers, different groups

• Histograms:

  • Visualizes data distributions using discrete, blocky bars grouped into intervals (bins)

  • Density Plot: Uses Kernel Density Estimation (KDE) to smooth those points into a continuous, fluid probability curve — estimate the shape of distribution

• Violin Plots

• Clustering/Dendrograms:

  • Clustering group observations based on similarity

  • Tree-like visualization showing how observations are progressively grouped

• Outlier Visualization:

  • Outlier: Observation that is unusually different from the rest of the dataset

  • Causes: Measurement error, artifact, data-entry problem, unusual participant, genuine biological variation

  • Not automatically bad data — but do we choose to keep or neglect?

How do we compare measurements longitudinally?

• Longitudinal Visualization: Comparing measurements at different time points

  • Baseline → follow-up, before → after treatment, younger → older, task A → B

• Visualization options

  • Connected dot plots

  • Line plots

  • Mean ± error

  • Change/delta plots

  • Change heat maps

    • Delta = Follow-up - Baseline

      • Positive = value increased

      • Negative = value decreased

      • 0 = No change

    • Makes patterns more noticeable (participant/connection changes, consistency, outliers)

What is a Scree Plot and a Cumulative Variance Plot?

• Scree plot: shows how much variance each PCA component explains

  • PC1 explains the most variance, PC2 explains the next largest, etc.

• Elbow: Point where adding more components explain progressively smaller amounts of variance

What are normality and variance and why are they important? How are parametric and non-parametric tests different?

• Normality: Whether the data are approximately normally distributed

• Variance: How much the data vary within each group (groups have similar variance = equal variance)

• Violations of normality or unequal variance can affect statistical interference, so they should be checked when choosing and evaluating an appropriate model

• Parametric vs. Non-Parametric:

  • Parametric tests assume data is normality distributed (bell-shaped) and variances are roughly equal across groups

  • Non-parametric tests make fewer assumptions and are safer when assumptions are violated

What are features and labels? What is the difference between labeled and. unlabeled data?

• Data = information organized in rows and columns

• Each row = one observation

• Each column = a feature that describes the data

• The label = the feature we want to predict

  • Labeled data: specifically named inputs and outputs

  • Unlabeled data: not distinctly labelled (may be used to find patterns)

  • Labels can represent numbers or categories/classes

What are the four types of machine learning?

1. Supervised: The algorithm learns to make predictions or decisions based on labeled data with input-output pairs, and the goal is to minimize the error between predicted and actual outputs

2. Unsupervised: The algorithm learns to identify patterns and relationships in unlabeled data, with the goal of discovering hidden structure and using it to make predictions or decisions

3. Reinforcement: The agent learns to make decisions based on feedback received from the environment in the form of rewards or punishments, with the goal of maximizing cumulative reward over time

4. Semi-supervised: The algorithm learns from a combination of labeled and unlabeled data, which can improve accuracy and reduce the need for labeled data

How does supervised machine learning work? What models can predict labels?

• The algorithm learns from labeled data which includes both inputs (features) and outputs (labels)

• The goal is to learn a mapping function that connects inputs to outputs by minimizing the error between predicted and actual labels

• Models learn from examples so it can predict the label of new, unseen data

• Steps:

  1. The model first remembers the dataset

  2. The model formulates a model or a rule

  3. When a new image comes in, the model makes a prediction about what the label of the new image is

• Regression and Classification Models:

  • Regression Models: Predict a number (continuous output which can take any value within a range)

  • Classification Models: Predict a category or state, handling more than two classes (multi-class classification)

How does unsupervised machine learning work? What are some common algorithms they employ and tasks they complete?

• Algorithm learns from unlabeled data (only features, no labels) with the goal of discovering hidden patterns or structure

• Algorithms:

  • K-means: Groups points around random centers, adjusts iteratively

    • Ex: Segment voxels in fMRI data into regions with similar activation patterns

  • Hierarchical: Groups closest points first, builds a tree/cluster step by step

    • Ex: Build a hierarchy of brain regions based on connectivity strength

  • DBSCAN: Groups dense areas, ignores isolated points as noise/sparse points

    • Ex: Detects clusters of high activity fMRI while ignoring noisy voxels

  • Gaussian Mixture: Assigns probabilities of belonging to multiple clusters

    • Ex: Identify overlapping functional networks where a voxel may belong partially to multiple networks

• Tasks:

  • Clustering: grouping similar data points

    • Grouping data based on similarity

    • Algorithm learns relationship

  • Dimensionality Reduction: simplifying data while keeping essential information

    • Simplifying our data and describing it with fewer features without losing much generality

    • Preprocessing data: for supervised learning

How does reinforcement machine learning work?

• The agent has a goal or set of goals

• Using this type of machine learning, the agent learns to make decisions based on the feedback (rewards or punishments) it receives from its environment

• The goal is to learn a policy that maximizes the cumulative reward over time

• There is no labeled dataset, the agent learns by trial and error

What are training and testing datasets? What is overfitting and how can we prevent it?

• Data is usually split (~70-80% training, ~20-30% testing) to ensure the model generalizes well, not just memorizing the training data

• Training data: Used to teach the model, the algorithm learns patterns and relationships from this data

  • Algorithm studies the hidden patterns and insights from observations and learns from them

• Testing data: Used to evaluate the models performance on unseen data

  • Should not contain any training dataset data

  • Predict unknown results

  • Checks the performance, accuracy, and precision of the model

• Overfitting:

  • Model learns the training data too well, including the noise

  • Failure to generalize beyond what it’s seen

  • Performs very well on training data, but poorly on new data

  • Prevention:

    • Use a validation dataset during training

    • Apply cross-validation techniques

    • Simplify the model or gather more data

What is validation data? What are some basic techniques for splitting training data?

• Validation data: A subset of training data used to tune/optimize the model (adjust parameters, choose features, prevent overfitting)

• Techniques:

  • Holdout Method: Dataset split once into separate training, validation, and testing sets to evaluate model performance on unseen data

  • Random Sampling: Data is shuffled and split into training/validation/test sets based on a user-defined ratio (defined in literature)

    • Works best on balanced datasets, may create bias if imbalanced

  • Stratified Sampling: Ensures each split preserves the class distribution of the dataset

    • Ideal for imbalanced datasets to maintain representative samples

What is K-Folds cross-validation? What is Leave-One-Out Cross Validation?

• K-Folds Cross Validation:

  • Dataset is split into K equal-sized folds (subsets)

  • One-fold is used for validation, and the remaining K-1 folds are used for training — the process repeats K times, each fold serves once as the validation set

  • Validation accuracy is recorded for each round

    • Final model accuracy = average

  • Common choice: K = 10 (balances bias and variance)

  • More representative of the dataset

• Leave-One-Out Cross Validation:

  • Algorithm loops n times for n total data points, with one data point being reserved to test the model

  • Final score = errors from all n tests averaged together

  • Pros: low bias, same result every time

  • Cons: high computational demand (time and money)

How are Machine Learning models evaluated using accuracy and confusion matrices?

• Accuracy: Measures the proportion of correctly predicted samples (# correct/total predictions)

• Confusion Matrix: Table showing how well the model performs across classes

  • Rows = actual classes, columns = predicted classes

  • Identifies types of errors

• Derived Metrics from Confusion Matrix:

  • Precision = TP / (TP + FP) — out of all predicted positives, how many were actually positive?

  • Recall/Sensitivity = TP / (TP + FN) — out of all actual positives, how many did the model correctly find?

  • F1-Score = 2 (precision x recall) / (precision + recall) — harmonic mean of precision and recall