TFNet: Time and Frequency Modeling for Irregular Multivariate Medical Time Series

Abstract

Time Series Classification (TSC) involves assigning a target label based on features involving time series data. TSC arises in a variety of domains, like healthcare, finance, process control, weather pattern prediction, etc. This work is focused on exploiting both frequency and time domains of a time series. Inspired by the TimesNet proposed in [27], which learns multi-periodic variations, we proposed Time-Frequency Network (TFNet), a novel Deep Learning model, and applied it to irregular medical time series data. Earlier methods used either only features captured in the time domain or in the frequency domain. It is di"cult to learn both temporal dependencies and understand cyclic or seasonality patterns when analyzed in a single domain. To tackle these limitations, we extend the TimesNet model to perform time domain analysis. Our proposed TFNet achieves an improved performance when applied to in-hospital mortality (IHM) prediction based on 48 hours of ICU stay, on a dataset extracted from Medical Information Mart for Intensive Care (MIMIC-III)