Prediction of human subjective time from functional magnetic resonance imaging

• Ano de defesa: 2024
• Autor(a) principal: Souza, Erick Almeida de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Cerebral cortex; Córtex cerebral; Functional MRI; Human time perception; IRM funcional; Machine learning; Neurociência; Neuroscience; Percepção do tempo humana
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/59/59135/tde-22012025-121936/

Time perception is one of the essential components of human perception and can be biased by different features present in daily stimuli, causing substantial variations in relation to real (clock) time. In this study, we sought to investigate the neural bases of the timing of durations and validate the hypothesis that subjective time is determined by the accumulation of salient changes in the perceptual processing. To this end, we executed five approaches based on data obtained from an experiment in which healthy human participants watched and estimated the duration of silent videos of varying lengths (a few seconds) while functional magnetic resonance imaging (fMRI) scans were acquired. In the first approach, we aimed to reproduce the findings reported in the original study using equivalent methodology. We used three brain parcellation schemes based on perceptual hierarchies (visual, auditory, and somatosensory) to predict the subjective time metric (normalized duration bias) from salient events accumulated in the fMRI signal. We confirmed the original finding that the visual hierarchy has the strongest association with duration bias, given that the stimulus is purely visual. However, contrary to the original study, we found that the association observed in the somatosensory hierarchy is also substantial and should be further evaluated. In the second approach, we applied a methodology similar to the previous one but with a parcellation based on functional networks composed of 360 cortical regions. In this case, we showed that regions in networks other than the visual network also exhibited significant associations with duration bias: the dorsal attention network, the cingulo-opercular network, and the somatomotor network. In the third approach, we used a recurrent (LSTM) deep learning model to capture features present in the fMRI signal across 360 cortical regions that might be associated with duration bias. We evaluated the model\'s performance across a wide range of hyperparameter combinations. With this model, it was not possible to predict duration bias. In the fourth approach, we used the same 360-region atlas to evaluate whether functional connectivity (FC) between these regions was associated with duration biases, since in the second approach we showed that different functional networks were associated with this metric. Using a regularized linear regression model, we were unable to predict duration bias from inter-individual variations in FC. In the fifth approach, we applied independent component analysis (ICA) to the fMRI data to assess the existence of spatial components significantly associated with the experimental design and thus the time perception task. None of the components obtained showed consistent correlation across participants. The findings of this study highlight that the salient event approach requires further validation, as the level of association obtained with duration bias was low. However, we demonstrated that regions within additional functional networks likely play important roles in the timing of durations in a visual stimulation context. Future studies should focus on a more precise evaluation of which regions within these networks are most predictive of subjective time.