A cross-sectional analysis of research waste in randomized controlled trials on postoperative cognitive dysfunction.

Functional magnetic resonance imaging (fMRI), with its many advantages, including high spatial resolution, non-invasive nature, and radiation-free acquisition, has been a widely used neuroimaging modality in recent times. By capturing dynamic fluctuations in the blood oxygenation level-dependent (BOLD) signal, fMRI enables examination of both localized and distributed neural processes. Based on their acquisition, the fMRI is broadly categorized into task-based fMRI (t-fMRI) and resting-state fMRI (rs-fMRI). t-fMRI records neural responses during explicit tasks and is typically employed to map functional specialization across cortical and subcortical regions. However, it suffers from limitations such as dependence on subject compliance, variability in task performance, and relatively low test–retest reliability, which constrain its utility in clinical and developmental populations. In contrast, rs-fMRI captures spontaneous low-frequency BOLD fluctuations in the absence of task engagement, providing a potential alternative in scenarios where task performance is difficult or unreliable, particularly benefiting non-compliant populations, such as the elderly, infants, or people with disabilities. To this end, recent studies have explored the feasibility of predicting the t-fMRI activation maps using the rs-fMRI data only. In a seminal work, Tavor et al. proposed a linear predictive framework for estimating task-induced fMRI activation maps using functional connectivity (FC) maps obtained from rs-fMRI. The preliminary analysis was conducted using group principal component analysis (PCA), followed by group independent component analysis (ICA), which generated 40 spatial FC maps for each hemisphere. Based on the similarity within each hemisphere, 33 cortical FC maps were selected for further analysis. Similarly, for the subcortex region, separate clustering methods were used to generate 32 subcortical features. These FC features, together with a few structural features, resulted in a total of 98 features that were provided to generalized linear regression models (GLM) to obtain the parcel-wise predictions. The evaluation across seven task contrasts in the Human Connectome Project (HCP) dataset, yielded mean correlation values ranging from to . Following this, Teterva et al. extended the framework where similar FC maps were used as features in flat and stacked prediction models. However, unlike the simple GLM, they employed sophisticated regressors like elastic net, random forest, XGBoost, and support vector regression (SVR), obtaining a test-retest correlation of 0.7. Cohen et al.also proposed a similar rest-to-task activation prediction framework, which modeled the FC features using non-linear models, viz., bagging and neural networks, resulting in correlations between 0.1 to 0.8. Tik et.al. tested the generalizability of the FC and GLM-based model across datasets and populations by training the model on one dataset and evaluating it on different datasets. The mean correlations obtained for this cross-dataset generalization were between and 0.4. Vaibhav et al. used a different strategy for rest-to-task activation map predictions, mapping the motor contrast in accordance with the pre-surgical planning for the brain tumor. Unlike the aforementioned methods, they modeled using connectome fingerprinting on different parcellations, viz., the multimodal parcellation, the Schaefer parcellation, and a comparison of the Schaefer and multimodal parcellations. A correlation in the range of 0.45 to 0.63 was obtained for all the copes in motor contrasts. Extending to modality-specific predictions, Zhou et al. demonstrated an accurate prediction of visual cortex activations by introducing Rest2Visual, a volumetric encoder-decoder design that takes 3D features from rs-fMRI as input, modulates them as image embeddings to obtain task activation predictions. This study used a natural scenes dataset, achieving between predicted and actual visual task maps. More recently, with the advent of the field of deep learning, many deep learning-based models have been tried on these predictive frameworks. Ngo et al. developed a surface-based deep learning framework, brain-surf CNN, that works with a brain’s cortical sheet representation. The resulting predictions, with a dice coefficient between 0.50 and 0.64 and an AUC between 0.20 and 0.30, were on par with the target-repeat reliability of the measured contrast maps. Kwon et al. used 3D fMRI data as input to SwiFUN, a transformer-based method that used shifted-window attention and contrastive learning to understand complex patterns in sequential data. It demonstrated 27% higher performance in the ABCD dataset compared to the existing methods. Patel et al., in, employed and compared the performance of graph-based, kernel-based, and transformer-based models in the task activation map prediction. The transformer-based models achieved the highest performance, with AUC = 0.82 and correlation