Masked features of task states found in individual brain networks

Abstract

Completing complex tasks requires that we flexibly integrate information across brain areas. While studies have shown how functional networks are altered during different tasks, this work has generally focused on a cross-subject approach, emphasizing features that are common across people. Here we used extended sampling "precision" fMRI data to test the extent to which task states generalize across people or are individually specific. We trained classifiers to decode state using functional network data in single-person datasets across 5 diverse task states. Classifiers were then tested on either independent data from the same person or new individuals. Individualized classifiers were able to generalize to new participants. However, classification performance was significantly higher within a person, a pattern consistent across model types, people, tasks, feature subsets, and even for decoding very similar task conditions. Notably, these findings also replicated in a new independent dataset. These results suggest that individual-focused approaches can uncover robust features of brain states, including features obscured in cross-subject analyses. Individual-focused approaches have the potential to deepen our understanding of brain interactions during complex cognition.

Keywords: BOLD fMRI; functional connectivity; individual differences; machine learning; task states.

Fig. 1

Overview of each analysis. Classifiers were trained to distinguish task from rest FC using many sessions of data from a single participant with ridge regression. a) For our first set of models, classifiers were trained to discriminate between all tasks and rest (80 total samples per participant). b) In the second group of models, classifiers were trained to distinguish a single task (vs. rest). c) Finally, multiclass models were trained to discriminate among all task states at once. In all cases, models were then tested on independent session data from either the same person (gray bar) or a different person; performance was contrasted across these tests to identify generalizable or individually specific components of task state.

Fig. 2

Comparison of classifier accuracy within- and between-person. a) Performance of an individualized machine learning classifier for discriminating task from rest states when tested on independent data from either the same individual or different individuals (colored lines = different participants). Initial models contrasted all task states with equal samples of rest (N = 80 samples total per person). b) Secondary analyses trained models to discriminate a single task from rest. On average, all models were able to predict task state significantly better than chance in the same person (P < 0.001 when compared to a permuted null with randomized task/rest labels) and in new people (P < 0.001; red dashed lines = chance). However, model performance was significantly higher when tested on data from the same participant relative to other participants (^***P < 0.001 compared to a permuted null on within/between-person labels; All Tasks vs. Rest: effect size = 0.24; Motor vs. Rest = 0.15; Coherence vs. Rest = 0.30; Memory vs. Rest = 0.09; Semantic vs. Rest = 0.27). Similar results were seen when tested with other machine learning algorithms (Supplementary Fig. S1). For a more detailed breakdown of person-to-person performance, see Supplementary Fig. S2.

Fig. 3

Multiclass performance. a) Here we trained models to discriminate all tasks and rest using a multiclass approach. For a detailed breakdown of performance in each individualized multiclass model, see Supplementary Fig. S6. Multiclass prediction was significant both in the same and new people (red line = chance performance across 5 states), with higher performance in the same individual (^***P < 0.001). In addition, we calculated confusion matrices for classifiers tested on independent data from the (b) same or (c) a different person. Classifiers performed relatively accurately when classifying new data in the same person, exhibiting few biased errors. In contrast, classifiers tested in new people exhibited more errors, especially for the motor and semantic tasks. Classifier performance is averaged across participants (see Supplementary Fig. S6 for confusion matrices from individual participants).

Fig. 4

Feature weight analysis. Average absolute feature weights for each brain region for classifiers built to discriminate between all task states and rest. a) Feature weights for a single person (MCS05) across each fold of the classifier. b) Feature weights from the standard groupwise approach for each fold of the classifier (in this case, training data come from different participants, rather than different sessions of the same participant). c) Average feature weights for each individualized classifier for each participant. d) Average feature weights for the groupwise approach. Note that while feature weights are fairly similar across folds within a person, they show variation across people and compared to the groupwise approach.

Fig. 5

Single network block performance in decoding task state. Here we trained classifiers to decode all tasks from rest using subsets of network to network connections (each block was an independent classifier). We then tested the classifier on new data from either (a) the same person or (b) a new person. Due to low feature numbers prior to training the classifiers, we standardized features (see Methods). Task state information was distributed across many networks of the brain: In within-person analyses, 99% of blocks could decode task state; 95% of blocks could decode task state in between-person analyses. Within-person performance was consistently higher than between-person classification.

Fig. 6

Overview of feature selection. Model classification performance when classifiers were trained on randomly selected features of different numbers or features associated with a given network. Random feature number varied from 10 to 50,000, randomly indexed across all unique FC edges. Networks have been plotted on top at the position equivalent to their feature size. Accuracy is shown both for models tested on the same person (circle markers, blue line) or a different person (square markers, orange line). Error bars on the random feature selection line represent the 5th and 95th percentile across iterations. a) Performance when training and testing to discriminate all tasks from rest. b) Performance when training and testing to discriminate a single task from rest. For a breakdown of participants, see Supplementary Fig. S9.

Fig. 7

Classification performance built on individualized network parcellations. Comparison of performance of individualized classifiers on networks derived from individual parcellations. a) Individual person parcellations and network definitions. FC matrices were calculated at the network level (network × network) using each participant’s network definitions in order to align across people. These FC matrices are then used to train a classifier to decode all task states from rest as in previous analyses. b) Individualized classifiers based on individualized networks were significantly able to decode task from rest in both the same and new participants but performed consistently better for the same person (P < 0.001).