Functional magnetic resonance imaging (fMRI) captures brain dynamics, offering crucial insights into brain function and disorders. However, its high-dimensional, complex, and noisy nature makes interpretation challenging. Ensuring model interpretability is essential, especially in high-stakes domains like medicine. Addressing this concern requires the development of specialized methods. Another significant challenge is data scarcity, as privacy laws often limit access to clinical data. In such cases, efficient pretraining techniques can be valuable, enabling models to work effectively with limited data while still producing reliable results. To address the challenge of data scarcity, we propose a novel pretraining method called time reversal. Our approach leverages self-supervised learning to train a model on the temporal direction of ICA-preprocessed fMRI data. The pretrained model is then applied to downstream classification tasks for three disorders: schizophrenia, Alzheimer's disease, and autism. Through extensive experiments, we demonstrate that during pretraining, the model effectively learns temporal patterns from a separate dataset. This learned temporal information enhances performance in downstream tasks, as evidenced by improved AUC scores compared to models trained from scratch. Our findings highlight time reversal as a promising approach for capturing essential temporal features and transferring this knowledge to related tasks. To enhance interpretability, we employ model introspection techniques to interpret the proposed pretraining method. We use one of the popular methods (Integrated Gradients) to generate saliency maps that offer post-hoc explanations for pretraining, while Earth Mover’s Distance (EMD) quantifies the temporal dynamics of salient features in the downstream schizophrenia classification task. The saliency maps reveal more concentrated and biologically meaningful salient features along the time axis, aligning with the episodic nature of schizophrenia. We show that, by linking model predictions to meaningful temporal patterns in brain activity, time reversal strengthens the connection between deep learning and clinical relevance. Additionally, it is possible to enhance interpretability by making the intermediate representations of the input more transparent. In most deep learning frameworks, an encoder maps the input to a latent representation, which is then decoded and used for prediction. We develop methods to interpret the latent representations in our self-supervised pretraining task, which focuses on the order of time points. To achieve this, we pretrain a model using time reversal, extract its latent representations, and feed them into a probe (logistic regression) for further analysis. The fMRI data consists of 53 components, which are associated with seven functional brain networks: sensorimotor, visual, sub-cortical, cognitive control, default mode, cerebellar, and auditory. These networks represent connectivity patterns across different brain regions. We first establish a mapping between the fMRI components and the latent features, allowing us to analyze the learned representations in a biologically meaningful way. Using this mapping, we examine the coefficients of the logistic regression probe to determine the contribution of each brain region to schizophrenia classification. This approach provides deeper insights into how specific brain networks influence model predictions, bridging the gap between deep learning and neuroscience.