Modeling and Machine Learning

Cedalion supports two complementary analysis strategies:

Model-driven analysis uses the General Linear Model (GLM) to estimate haemodynamic response functions (HRFs) from known stimulus timings. A design matrix is constructed from convolutions of stimulus boxcar functions with HRF basis functions (Gamma, GaussianKernels, DiracDelta). The GLM is fitted per-channel, and beta coefficients quantify the amplitude of the response to each condition. This approach is well-suited when the experimental design is block- or event-based and stimuli are accurately recorded.

Data-driven analysis uses decomposition methods (ICA, CCA, SPoC, mSPoC) to identify shared structure across channels or modalities without relying on a stimulus model. This is particularly useful for naturalistic recordings, EEG–fNIRS co-recordings, or when the neural signal of interest cannot be described by a simple HRF.

Both approaches integrate with scikit-learn: Cedalion DataArrays behave as numpy arrays to sklearn estimators, while the mlutils module provides cross-validation utilities designed for the temporal structure of fNIRS data.

Models

cedalion.models.glm	Tools for describing fNIRS data with general linear models.
cedalion.math.ar_model

Decomposition Methods

cedalion.sigdecomp.unimodal
cedalion.sigdecomp.multimodal	Multimodal signal decomposition methods.

Examples

Basic single trial fNIRS finger tapping classification

Example for ICA Source Extraction

Constrained Independent Component Analysis (ICA)

GLM Basis Functions

GLM Fingertapping Example

GLM Illustrative Example

Statsmodels Methods Overview

Fitting a GLM with Gaussian Kernels