# Introduction

(rationale)=
## Rationale

Functional near-infrared spectroscopy (fNIRS) aims at measuring and localizing neural 
evoked oxygenation changes in the brain. The recorded signal is inherently confounded by
absorption changes in the cerebral and extracerebral compartments which are not of 
neural origin but stem from systemic physiological activity. These confounding 
components can reduce contrast, sensitivity, and specificity of the hemodynamic response
in fNIRS. 

Three trends in fNIRS methodology are observable.

1. Wearable High-Density DOT: Acquisition hardware improvements allow for denser optode 
arrays in a wider range of usage scenarios that increasingly include wearable and 
naturalistic environments. The increased density allows to cover head areas with 
multiple source-detector pairs at varying distances which helps to discriminate signal
contributions from different tissue depths and brain regions and improves contrast and 
lateral specificity.

2) Simultaneous measurements with complementary neuroimaging modalities like EEG as well
as physiological sensors provide additional information to discriminate confounding 
components from those of interest.

3) The portability and unobtrusiveness of the fNIRS modality makes it suitable for 
studying subjects outside the lab in more naturalistic environments. Unstructured 
protocols and more variable systemic interference (e.g. from participant movements) make
these experimental setups particularly challenging. New methods need to be developed
and provided to the community to use information from additional sensors like 
accelerometers, eye movement trackers and GPS to exploit additional information about 
the state and context in which a fNIRS recording was conducted.

Hence, we envision current and future fNIRS experiments to be increasingly concerned 
with multivariate, compounded time series with data from different neuroimaging 
modalities and sensor types and in which relations between time series must be modeled. 
We assume that machine learning techniques will play a key role in modeling
the complex interplay between these time series.

## Development goals

This development project will tap into the rich Python ecosystem of machine learning and
data science tools. The aim is to provide 
[user-extensible data structures](data_structures/index.md) and functionality that allow for 
easy data exchange with the tensor data types and data frames provided by popular 
frameworks like PyTorch and Pandas. Making this exchange easy will simplify the 
[integration of machine learning workflows](examples/new_conference_example2.ipynb) 
and conventional fNIRS data processing streams. 

Also, we recognize that for each neuroimaging modality versatile and well-tested 
analysis toolboxes with specific preprocessing methods exist. We want to support the 
construction of [workflows](workflows.md) that chain functionality of these toolboxes 
together. [Support for standardized file formats](api/cedalion.io.rst) like SNIRF and 
BIDS will be central to facilitating the data exchange between toolboxes.