Abstracts

Rationale: Deep learning based approaches have demonstrated success in analyzing brain connectivity based on functional magnetic resonance imaging (fMRI), but the scarcity and heterogeneity of fMRI data still pose challenges in clinical applications such as predicting the future onset of Post-Traumatic Epilepsy (PTE) from acute data acquired shortly after traumatic brain injury (TBI).
Identification of subjects at high risk of developing PTE can eliminate the need to wait for spontaneous epileptic seizures to occur before starting treatment and enable the mitigation of risks to subjects whose seizures could result in serious injury or death. fMRI plays a vital role in identifying biomarkers for PTE. The presence of lesions in TBI patients can alter resting-state brain dynamics. The neuropathology and disruptions caused by brain injury will be reflected in fMRI data collected after injury, which can therefore provide valuable biomarkers for PTE. However,
TBI datasets are usually characterized by high variability among subjects and limited numbers of subjects, presenting a significant challenge for training of deep learning methods to predict PTE. Foundation models pre-trained on separate large-scale datasets can improve the performance on scarce and heterogeneous datasets.

Methods: We explored a novel strategy which trains the foundation model to generalize from normal/control features to scarce clinical features.
Here we propose Meta Transfer of Self-supervised Knowledge (MeTSK), which harnesses meta-learning to facilitate the generalization of self-supervised features from large-scale control to scarce clinical datasets. The proposed strategy MeTSK is designed to enhance the foundation model's generalization capacity to new and unseen clinical data in challenging downstream applications by leveraging the learned generalization from control features to scarce clinical features during pre-training. We apply our foundation model trained using MeTSK to directly generate features for the downstream fMRI data without any fine-tuning. We then input these features to a linear classifier and perform PTE classification.

Results: We use the Maryland TBI MagNeTs dataset for downstream performance evaluation. All subjects suffered a traumatic brain injury. Of these we used acute-phase (within 10 days of injury) resting-state fMRI from 36 subjects who went on to develop PTE and 36 who did not. We compare the linear probing performance across different foundation models as well as the performance of classifiers trained with functional connectivity features extracted from raw fMRI data. A 5-fold cross-validation was applied and AUCs for PTE v.s. non-PTE classification were computed. Our proposed method achieved the best AUC.

Conclusions: We developed a foundation model tailored for neurological disorders using a novel training strategy, which achieved superior generalization to a small-scale PTE dataset and significantly improved the performance on the challenging PTE prediction task compared to existing state-of-the-art foundation models for fMRIs.

Funding: NIH grants: R01EB026299, R01NS074980; DoD grants: W81XWH181061, HT94252310149