Can a brain-connectivity model make better fMRI classifiers?
Classifying people by the connection strengths of a fitted brain model, rather than raw voxel activity, identified aphasia patients almost perfectly and pointed to which connections mattered.
Source
Generative embedding for model-based classification of fMRI data
Study at a glance
- Design
- Computational / modelling — Method comparison on existing fMRI data: subject-wise DCMs of a 6-region auditory network are inverted and their 22 parameters fed to a linear SVM, compared with activation- and correlation-based feature sets and ill-informed control models under leave-one-subject-out cross-validation.
- N
- N=37 · 37 participants: 26 healthy controls and 11 patients with moderate aphasia after stroke, each classified by leave-one-subject-out cross-validation.
- Population
- Adults with post-stroke aphasia and healthy controls performing a speech-listening task in fMRI
- Outcome
- Balanced accuracy of patient-vs-control classification; which model parameters drive it
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
Using fMRI from 26 healthy adults and 11 stroke patients with aphasia listening to normal and reversed speech, the authors fitted a dynamic causal model (a generative model of how six auditory regions influence each other) to each person. The model's 22 connection parameters became the features for a linear support vector machine. They compared this 'generative embedding' with classifiers using voxels chosen by anatomy, contrasts, searchlights or PCA, with correlation-based features, and with deliberately less plausible models, all under leave-one-subject-out cross-validation.
What they found
Generative embedding classified 36 of 37 people correctly (98% balanced accuracy), significantly beating every conventional approach, whose accuracies ranged from 62% to 83%. Less biologically plausible models did worse (77% for a feed-forward model, 81% left hemisphere only, 59.3% right hemisphere only). A sparse classifier kept a consistent set of 9 of the 22 parameters, still reaching 98%, and these were connections between cortical regions, especially right-to-left hemisphere links converging on the left planum temporale.
The limits
What it doesn't show
The sample is small (only 11 patients) and comes from one dataset, so the near-perfect accuracy may not replicate in new cohorts or scanners. Groups also differed in age, which a classifier could exploit. The approach depends on specifying the right brain model in advance, and the paper shows accuracy falls sharply with a poorer model. The extracted text ends partway through the Discussion, so the authors' own stated limitations are not available here.
Key terms
- Generative embedding
- Mapping each subject's data into the parameter space of a fitted generative model and classifying in that space.
- Dynamic causal modelling (DCM)
- A generative model of fMRI in which hidden neural activity in regions influences other regions through directed connections.
- Support vector machine
- A classifier that finds the boundary separating two classes with the largest margin.
- Leave-one-subject-out cross-validation
- Training on all subjects but one, testing on the left-out subject, and repeating for everyone.
- Balanced accuracy
- The average of accuracy within each class, which corrects for unequal group sizes.
Flashcards
0 of 9 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What was used as the feature space in generative embedding?
Common questions
Why is classifying raw voxels hard?
A scan has vastly more voxels than there are subjects and the signal is noisy, so classifiers struggle to find informative voxels without overfitting.
Why use balanced accuracy here?
There were many more controls than patients, so plain accuracy could look high just by predicting 'control' often.
Does the model show what causes aphasia?
No — it shows which model connections distinguish groups, which suggests mechanisms, but differences could be consequences of the stroke rather than causes of symptoms.
More on Generative models