MedTVL: Harnessing Vision and Language for Medical Time Series Classification

Recent research introduces MedTVL, a text-guided dual-pathway architecture for medical time series classification. It integrates time series, vision, and language to mirror clinical diagnostic practices, utilizing a convolution-based temporal pathway. Separately, a new method called Local Reference Geometry (LRG) addresses imbalanced time series classification. LRG identifies training-local geometry failures where minority cases reside in sparse or mixed feature-space neighborhoods, proposing a representation-level solution to ensure local reliability around minority regions despite global class structures.