HiTF: A Hippocampal-Thalamic Function-Inspired Framework for Robust Multimodal Sentiment Analysis with Incomplete Data

Multimodal sentiment analysis with incomplete data faces significant challenges, particularly in the form of random frame-level missingness, where fragmented emotional cues and heterogeneous data quality impede effective understanding. Existing completion methods predominantly rely on cross-modal consistency, often neglecting residual modality specificity and failing to assess cross-modal reliability, which leads to redundancy. Human cognitive systems exhibit remarkable robustness to incomplete sensory inputs, driven by two functional mechanisms: hippocampal memory systems that reconstruct missing content via pattern completion, and thalamic perceptual regulation that dynamically modulates multisensory integration by filtering unreliable signals. Inspired by these brain functions, we propose the Hippocampal-Thalamic Function-Inspired Framework (HiTF). This framework integrates two complementary streams. The hippocampal-inspired intra-modal enhancement stream employs memory retrieval and sparse activation mechanisms to mine modality-specific semantics and reconstruct missing information. Concurrently, the thalamic-inspired inter-modal regulation stream evaluates the reliability of each modality to adaptively integrate high-quality cross-modal information while suppressing redundant interference. Comprehensive experiments on MOSI, MOSEI, and SIMS demonstrate that HiTF achieves superior performance with 1.5%–-2.5% average accuracy improvements over state-of-the-art methods across all missing rates. Notably, even under an extreme 90% missing condition on MOSI, our framework achieves substantial improvements in F1 score over strong baselines, validating the effectiveness of the proposed brain function-inspired framework.