Exploring Unconscious Mental Abilities Using fMRI and Artificial Intelligence: A Multimodal Neuroimaging Study

Abstract

The unconscious mind remains one of the most elusive frontiers in cognitive neuroscience, shaping perception, emotion, and behavior through neural processes that evade explicit awareness (Marcel, 1983; Goodale & Milner, 1992; Bargh & Chartrand, 1999; Baars, 1988; Kahneman, 2011). Despite substantial behavioral evidence for unconscious cognitive processing (Bechara et al., 1997; Naccache & Dehaene, 2001; Soon et al., 2008; Dijksterhuis et al., 2006), the neural mechanisms underlying these phenomena remain incompletely characterized. This study develops an integrated framework combining functional magnetic resonance imaging (fMRI), advanced machine learning algorithms (Hussein et al., 2026; Yousif et al., 2026; Islam et al.., 2024; Abuzreda et al., 2025; Yousif et al., 2026; Shomata et al., 2025; Abuzreda et al., 2026), and information-theoretic analyses (Williams & Beer, 2010; Luppi et al., 2023) to decode the neural signatures of unconscious mental processes. Thirty healthy participants (15 male, 15 female; mean age 28.4 ± 4.2 years) underwent task based and resting state fMRI while performing a subliminal priming paradigm with emotionally valenced facial expressions presented below the threshold of conscious awareness (16 ms) (Tottenham et al., 2009; Levitt, 1971). Data were analyzed using the General Linear Model (GLM) (Friston et al., 1994), Independent Component Analysis (ICA) (Beckmann & Smith, 2004), Convolutional Neural Networks (CNNs), Support Vector Machines (SVM) (Norman et al., 2006), Artificial Neural Networks (ANN), and Partial Information Decomposition (PID) (Williams & Beer, 2010). Results demonstrated significant unconscious amygdala activation in response to masked fearful faces (Whalen et al., 1998) (p < 0.001), with AI models achieving 94.7% accuracy in stimulus classification and 68.2% accuracy in detecting unconscious emotional processing—substantially outperforming traditional statistical approaches (R² = 0.79 vs. 0.58 for linear regression) (Dehaene et al., 2001; Naccache & Dehaene, 2001). Critically, we observed increased structure function coupling during unconscious states and a breakdown of synergistic information integration (Luppi et al., 2023; Luppi et al., 2024; Luppi et al., 2023), consistent with the hypothesis that consciousness depends on the brain’s capacity for synergistic information processing (Tononi, 2006; Dehaene & Naccache, 2001). These findings establish fMRI AI integration as a powerful paradigm for investigating the neural correlates of unconscious cognition, with significant implications for the clinical assessment of disorders of consciousness (Bodien et al., 2024) and the development of objective diagnostic biomarkers (Mei et al., 2022; Mei & Soto, 2025; Qiu et al., 2022; Qureshi & Stevens, 2022; Gomez et al., 2024; Castro et al., 2024).