yingqiuz/predict-task-individual-variability

An ensemble model to predict individual variability of task-fMRI using resting-state fMRI

predict-task-individual-variability

This is the repository for https://www.biorxiv.org/content/10.1101/2021.08.19.456783v1.

the baseline model

baseline.py contains the baseline model. For a given task contrast, it seeks to find a linear combination of functional modes to reconstruct the task contrast map for each individual.

example

from baseline import baseline
baseline_predictions = baseline(train_x, train_y, test_x, test_y)

parameters and returns

• train_x and test_x should be lists of ndarrays, e.g., if you have 50 functional modes, then train_x is a list of 50 (n_train, n_voxels) ndarrays.
• train_y and test_y are individual task maps (of a given contrast), e.g., train_y is a (n_train, n_voxels) array.
• It returns a dictionary of predicted task maps {"train": pred_train, "test": pred_test}, where pred_train and pred_test each is a n_subject x n_voxels array.

the sparse model

sparse.py contains the sparse model. It introduces more spatial complexity to the baseline mode.

example

from sparse import sparse
sparse_predictions = sparse(train_x, train_y, test_x, test_y, rest_ic=800, task_ic=800, n_jobs=1, alpha=1)

parameters and returns

• train_x, train_y, test_x, and test_y have the same types/formats as in baseline.py.
• rest_ic can either be an Int, specifying the number of independent compoenents that each array in train_x should be reduced to, or be a dictionary {"train": train_ic, "test": test_ic}, where train_ic and test_ic are precomputed (and concatenated) mixing matrices.
• task_ic can either be an Int, specifying the number of independent compoenents that train_y should be reduced to, or be a dictionary {"train": train_ic, "test": test_ic}, where train_ic and test_ic are precomputed mixing matrix of the given task contrast.
• It returns a dictionary of predicted task maps {"train": pred_train, "test": pred_test}, where pred_train and pred_test each is a n_subject x n_voxels array.

the ensemble model

ensemble.py runs the baseline model and the sparse model and combines them to give a single set of predictions.

examples

from ensemble import ensemble
ensemble_predictions = ensemble(train_x, train_y, test_x, test_y, rest_ic=800, task_ic=800, residualise=True, n_jobs=-1)

parameters and returns

• train_x and test_x should be lists of ndarrays, e.g., if you have 50 functional modes, then train_x is a list of 50 (n_train, n_voxels) ndarrays.
• train_y and test_y are individual task maps (of a given contrast), e.g., train_y is a (n_train, n_voxels) array.
• rest_ic can either be an Int, specifying the number of independent compoenents that each array in train_x should be reduced to, or be a dictionary {"train": train_ic, "test": test_ic}, where train_ic and test_ic are precomputed (and concatenated) mixing matrices.
• task_ic can either be an Int, specifying the number of independent compoenents that train_y should be reduced to, or be a dictionary {"train": train_ic, "test": test_ic}, where train_ic and test_ic are precomputed mixing matrix of the given task contrast.
• residualise specifies whether to run the model on residuals (and add the group-averaged back in at the end) or on the original data.
• It returns a dictionary of predicted task maps {"train": pred_train, "test": pred_test}, where pred_train and pred_test each is a n_subject x n_voxels array.