A benchmarking framework for improving machine learning with fNIRS neuroimaging data
Functional near-infrared spectroscopy (fNIRS) is a non-invasive neuroimaging technology that presents attractive advantages such as a good compromise between temporal resolution, spatial resolution, and portability. Thanks to its properties and relative robustness to motion, it is often used in natu...
| Main Author: | |
|---|---|
| Format: | Thesis (University of Nottingham only) |
| Language: | English |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://eprints.nottingham.ac.uk/78005/ |
| _version_ | 1848801044522336256 |
|---|---|
| author | Benerradi, Johann |
| author_facet | Benerradi, Johann |
| author_sort | Benerradi, Johann |
| building | Nottingham Research Data Repository |
| collection | Online Access |
| description | Functional near-infrared spectroscopy (fNIRS) is a non-invasive neuroimaging technology that presents attractive advantages such as a good compromise between temporal resolution, spatial resolution, and portability. Thanks to its properties and relative robustness to motion, it is often used in naturalistic settings and one of its key domains of application is the study of working memory and mental workload. However fNIRS is less popular and tested than other technologies such as electroencephalography (EEG) or functional magnetic resonance imaging (fMRI) which can nowadays be used in a clinical context. As such, there is no real consensus yet in the research community as to how fNIRS data should be used and analysed. While efforts to establish best practices with fNIRS have been published, there are still no community standards for using machine learning with fNIRS data. Moreover, the lack of open source benchmarks and standard expectations for reporting means that published works often claim high generalisation capabilities, but with poor methodology practices or reporting with missing details. These issues make it hard to evaluate the performance of models when it comes to choosing them for brain-computer interface (BCI) applications.
In this thesis, we present the creation of an open-source benchmarking framework called BenchNIRS to establish a best practice machine learning methodology to develop, evaluate, and compare models for classification from fNIRS data, using open-access datasets from the literature. This framework makes the implementation of a robust machine learning methodology for fNIRS much simpler and less time-consuming.
We demonstrate the utility of the framework by presenting a benchmarking of 6 baseline machine learning models (linear discriminant analysis (LDA), support vector machine (SVM), k-nearest neighbours (kNN), artificial neural network (ANN), convolutional neural network (CNN) and long short-term memory (LSTM)) on 5 open-access datasets and investigate the influence of different factors on the classification performance. Those benchmarks set a solid basis for future comparisons of machine learning approaches for fNIRS classification and reveal that most models have lower performances than expected when evaluating them in an unbiased way.
We then go on to use the framework in a specific challenging use case, the classification of mental workload, and demonstrate how it is used to develop machine learning models tailored to a specific task or dataset. We show that models using as inputs longer durations of fNIRS recordings did not necessarily predict n-back levels better and that the classification from fNIRS data on this task is relatively challenging, despite the tailoring of deep learning models to specific device configurations being promising.
Finally, we go beyond supervised learning and extend the framework to study how unlabelled segments of the data can be used to improve classification. This leads us to perform transfer learning with a self-supervised representation learning pretext task and study to what extent it can be useful to fNIRS data classification. We explore the use of opposite hemoglobin type reconstruction as a pretext task and extend the framework to support the exploration of more pretext tasks for transfer learning in the future. |
| first_indexed | 2025-11-14T21:01:12Z |
| format | Thesis (University of Nottingham only) |
| id | nottingham-78005 |
| institution | University of Nottingham Malaysia Campus |
| institution_category | Local University |
| language | English |
| last_indexed | 2025-11-14T21:01:12Z |
| publishDate | 2024 |
| recordtype | eprints |
| repository_type | Digital Repository |
| spelling | nottingham-780052024-07-23T04:40:28Z https://eprints.nottingham.ac.uk/78005/ A benchmarking framework for improving machine learning with fNIRS neuroimaging data Benerradi, Johann Functional near-infrared spectroscopy (fNIRS) is a non-invasive neuroimaging technology that presents attractive advantages such as a good compromise between temporal resolution, spatial resolution, and portability. Thanks to its properties and relative robustness to motion, it is often used in naturalistic settings and one of its key domains of application is the study of working memory and mental workload. However fNIRS is less popular and tested than other technologies such as electroencephalography (EEG) or functional magnetic resonance imaging (fMRI) which can nowadays be used in a clinical context. As such, there is no real consensus yet in the research community as to how fNIRS data should be used and analysed. While efforts to establish best practices with fNIRS have been published, there are still no community standards for using machine learning with fNIRS data. Moreover, the lack of open source benchmarks and standard expectations for reporting means that published works often claim high generalisation capabilities, but with poor methodology practices or reporting with missing details. These issues make it hard to evaluate the performance of models when it comes to choosing them for brain-computer interface (BCI) applications. In this thesis, we present the creation of an open-source benchmarking framework called BenchNIRS to establish a best practice machine learning methodology to develop, evaluate, and compare models for classification from fNIRS data, using open-access datasets from the literature. This framework makes the implementation of a robust machine learning methodology for fNIRS much simpler and less time-consuming. We demonstrate the utility of the framework by presenting a benchmarking of 6 baseline machine learning models (linear discriminant analysis (LDA), support vector machine (SVM), k-nearest neighbours (kNN), artificial neural network (ANN), convolutional neural network (CNN) and long short-term memory (LSTM)) on 5 open-access datasets and investigate the influence of different factors on the classification performance. Those benchmarks set a solid basis for future comparisons of machine learning approaches for fNIRS classification and reveal that most models have lower performances than expected when evaluating them in an unbiased way. We then go on to use the framework in a specific challenging use case, the classification of mental workload, and demonstrate how it is used to develop machine learning models tailored to a specific task or dataset. We show that models using as inputs longer durations of fNIRS recordings did not necessarily predict n-back levels better and that the classification from fNIRS data on this task is relatively challenging, despite the tailoring of deep learning models to specific device configurations being promising. Finally, we go beyond supervised learning and extend the framework to study how unlabelled segments of the data can be used to improve classification. This leads us to perform transfer learning with a self-supervised representation learning pretext task and study to what extent it can be useful to fNIRS data classification. We explore the use of opposite hemoglobin type reconstruction as a pretext task and extend the framework to support the exploration of more pretext tasks for transfer learning in the future. 2024-07-23 Thesis (University of Nottingham only) NonPeerReviewed application/pdf en cc_by https://eprints.nottingham.ac.uk/78005/1/PhD_Thesis_Johann.pdf Benerradi, Johann (2024) A benchmarking framework for improving machine learning with fNIRS neuroimaging data. PhD thesis, University of Nottingham. fNIRS machine learning mental workload |
| spellingShingle | fNIRS machine learning mental workload Benerradi, Johann A benchmarking framework for improving machine learning with fNIRS neuroimaging data |
| title | A benchmarking framework for improving machine learning with fNIRS neuroimaging data |
| title_full | A benchmarking framework for improving machine learning with fNIRS neuroimaging data |
| title_fullStr | A benchmarking framework for improving machine learning with fNIRS neuroimaging data |
| title_full_unstemmed | A benchmarking framework for improving machine learning with fNIRS neuroimaging data |
| title_short | A benchmarking framework for improving machine learning with fNIRS neuroimaging data |
| title_sort | benchmarking framework for improving machine learning with fnirs neuroimaging data |
| topic | fNIRS machine learning mental workload |
| url | https://eprints.nottingham.ac.uk/78005/ |