InertiaMPDD
A world-model-inspired framework for multimodal depression detection that uses a latent predictor to learn the dynamics of healthy affective states and constructs three types of descriptors—reference deviation, residual decay rate and Jacobian spectrum—to quantify deviations in affective dynamics.
Timeline (last 90 days)
InertiaMPDD is a multimodal depression detection framework motivated by emotional inertia that uses a world-model-inspired latent predictor to learn next-state patterns under healthy states.
The core of InertiaMPDD is a world-model-inspired latent predictor that uses pretrained multimodal representations to predict the next state, with healthy-state patterns as the learning objective.
The paper evaluates on five cohorts from four datasets.
Multimodal fusion reaches 98.41% F1 on the D-Vlog dataset.
On the MPDD dataset, the individual descriptors outperform fixed additive fusion.
On the AVEC2014 dataset, the descriptors have weaker discriminative ability.
The paper's limitations include the weak discriminative ability of the descriptors on AVEC2014, generalization depending on data and scenario, and the lack of annotation for fine-grained dynamics under recording-level supervision.
All Facts (7)
InertiaMPDD is a multimodal depression detection framework motivated by emotional inertia that uses a world-model-inspired latent predictor to learn next-state patterns under healthy states.
50%UnverifiedThe core of InertiaMPDD is a world-model-inspired latent predictor that uses pretrained multimodal representations to predict the next state, with healthy-state patterns as the learning objective.
50%UnverifiedThe paper evaluates on five cohorts from four datasets.
50%UnverifiedMultimodal fusion reaches 98.41% F1 on the D-Vlog dataset.
50%UnverifiedOn the MPDD dataset, the individual descriptors outperform fixed additive fusion.
50%UnverifiedOn the AVEC2014 dataset, the descriptors have weaker discriminative ability.
50%UnverifiedThe paper's limitations include the weak discriminative ability of the descriptors on AVEC2014, generalization depending on data and scenario, and the lack of annotation for fine-grained dynamics under recording-level supervision.
50%