LFVDNet: Low-frequency variable-driven network for medical time series.

Abstract

Medical time series, a type of multivariate time series with missing values, is widely used to predict time series analysis, the "impute first, then predict" end-to-end architecture is used to address this issue. However, existing methods are likely to lead to the loss of uniqueness and key information of low-frequency sampled variables (LFSVs) when dealing with them. In this paper, we aim to develop a method that effectively handles LFSVs, preserving their distinctive characteristics and essential information throughout the modeling process. We propose a novel end-to-end method named Low-Frequency Variable-Driven network (LFVDNet) for medical time series analysis. Specifically, the Time-Aware Imputer (TA) module encodes the observed values and critical time information, and uses the attention mechanism to establish an association between the observed values and the missing values. TA adopts channel-independent strategy to prevent interference from high-frequency sampled variables (HFSVs) on LFSVs, thereby preserving the unique information contained in LFSVs. The Offset-Selection Module (OS) independently selects data points for each variable through offsets, avoiding the natural disadvantages of LFSVs in selection-based imputation, thus solving the problem of the loss of key information of LFSVs. LFVDNet is the first method for analyzing multivariate time series with missing values that emphasizes the effective utilization of LFSVs. We carried out the experiments on four public datasets and the experimental results indicate that LFVDNet has better robustness and performance. All code is available at https://github.com/dxqllp/LFVDNet. This study proposes a novel method for medical time series analysis, namely LFVDNet, which aims to effectively utilize LFSVs. Specifically, we have designed the TA module, which performs imputation through temporal correlations. The OS module, on the other hand, performs selective imputation based on a data point selection strategy. We have verified the effectiveness of this method on four datasets constructed from PhysioNet 2012 and MIMIC-IV.