CNN-Based Modeling Reveals Temporal Brain Dynamics of Auditory Intensity Processing

Understanding how the human brain encodes auditory intensity remains a fundamental challenge, particularly at the level of single-trial hemodynamic responses. Traditional fNIRS analysis methods rely on fixed hemodynamic response models that collapse all trials into a single amplitude estimate, potentially overlooking rich temporal information. In this study, we introduce a Siamese convolutional neural network designed to decode intensity levels directly from full HbO trial waveforms, enabling a dynamic, waveform-level characterization of cortical responses. fNIRS data were collected from eighteen normal-hearing adults across two stimulus cohorts with three or four intensity levels plus a silence baseline. All analyses were performed using leave-one-subject-out validation to ensure cross-participant generalizability. Across participants, the network accurately discriminated intensity-level pairs ( $\gt {85}\%$ accuracy) and reconstructed full intensity rankings with robust performance in both cohorts. Training augmentation using cross-subject trial pairs further improved discrimination, revealing that intensity-related hemodynamic patterns are remarkably consistent across individuals with normal hearing. In contrast, disrupting temporal structure-either by randomizing trial order or by artificially jittering the alignment between stimulus onset and the extracted hemodynamic response resulted in substantial reductions in decoding accuracy, highlighting the critical role of time-locked waveform features in cortical intensity encoding. Together, these findings provide neural evidence that intensity perception is supported by highly consistent and temporally structured cortical responses in normal-hearing listeners. The results also demonstrate the value of dynamic, waveform-based deep learning approaches for fNIRS analysis and suggest a path toward more sensitive and physiologically grounded models of auditory processing.