Understanding Medical Time Series Event Piece by Piece: A Fine-Grained Event Detection Network

Qianru Jia, Xuehui Wang, Xiaojun Ning, Xiyuan Jin, Ziyu Jia, Mengjie Zhai, Youfang Lin, Jing Wang

Keywords: medical time series, fine-grained event detection, human comprehension tasks, multi-task learning

TL;DR: We propose EventCompreNet, a universal network for fine-grained event detection that integrates human-inspired auxiliary tasks, enabling comprehensive understanding of medical time-series events.

Abstract: Event detection in medical time series is fundamental to supporting health monitoring and clinical decision-making. However, most existing methods divide time series into fixed-length segments and perform coarse-grained, segment-level detection, which fails to precisely localize the start and end times of events. This limitation can mislead clinical assessment and obscure the true severity of conditions. To address this, we propose EventCompreNet——a universal network for fine-grained event detection leveraging auxiliary tasks. Inspired by the cognitive processes that human detect events, we introduce four human comprehension tasks to enhance the model’s understanding of each piece of events. Moreover, to overcome the limited knowledge transfer in existing multi-task learning structures, we develop a task-deep-coupling framework that facilitates deeper interaction among tasks. Through these designs, EventCompreNet achieves a comprehensive understanding of the entire event life cycle. Experimental results on four benchmark datasets demonstrate that our model significantly outperforms existing state-of-the-art time series models in fine-grained event detection and exhibits strong event comprehension capabilities.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 12460