Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ImputeINR: Time Series Imputation via Implicit Neural Representations for Disease Diagnosis with Missing Data

Mengxuan Li, Ke Liu, Jialong Guo, Jiajun Bu, Hongwei Wang, Haishuai Wang

TL;DR

This work introduces ImputeINR, a time-series imputation framework that uses implicit neural representations to learn a continuous function mapping timestamps to multi-variable values. By decomposing the INR into trend, seasonal, and residual components and guiding its parameters with a transformer-predicted token set, along with variable clustering and multi-scale feature extraction, ImputeINR achieves state-of-the-art imputation performance even at very high missing rates. It also demonstrates that imputed healthcare data can meaningfully improve downstream disease diagnosis, validated on eight datasets and multiple baselines, while maintaining computational efficiency. The approach enables fine-grained imputations at arbitrary temporal resolutions, which is particularly valuable for real-world clinical decision support with sparse observations.

Abstract

Healthcare data frequently contain a substantial proportion of missing values, necessitating effective time series imputation to support downstream disease diagnosis tasks. However, existing imputation methods focus on discrete data points and are unable to effectively model sparse data, resulting in particularly poor performance for imputing substantial missing values. In this paper, we propose a novel approach, ImputeINR, for time series imputation by employing implicit neural representations (INR) to learn continuous functions for time series. ImputeINR leverages the merits of INR in that the continuous functions are not coupled to sampling frequency and have infinite sampling frequency, allowing ImputeINR to generate fine-grained imputations even on extremely sparse observed values. Extensive experiments conducted on eight datasets with five ratios of masked values show the superior imputation performance of ImputeINR, especially for high missing ratios in time series data. Furthermore, we validate that applying ImputeINR to impute missing values in healthcare data enhances the performance of downstream disease diagnosis tasks. Codes are available.

ImputeINR: Time Series Imputation via Implicit Neural Representations for Disease Diagnosis with Missing Data

TL;DR

This work introduces ImputeINR, a time-series imputation framework that uses implicit neural representations to learn a continuous function mapping timestamps to multi-variable values. By decomposing the INR into trend, seasonal, and residual components and guiding its parameters with a transformer-predicted token set, along with variable clustering and multi-scale feature extraction, ImputeINR achieves state-of-the-art imputation performance even at very high missing rates. It also demonstrates that imputed healthcare data can meaningfully improve downstream disease diagnosis, validated on eight datasets and multiple baselines, while maintaining computational efficiency. The approach enables fine-grained imputations at arbitrary temporal resolutions, which is particularly valuable for real-world clinical decision support with sparse observations.

Abstract

Healthcare data frequently contain a substantial proportion of missing values, necessitating effective time series imputation to support downstream disease diagnosis tasks. However, existing imputation methods focus on discrete data points and are unable to effectively model sparse data, resulting in particularly poor performance for imputing substantial missing values. In this paper, we propose a novel approach, ImputeINR, for time series imputation by employing implicit neural representations (INR) to learn continuous functions for time series. ImputeINR leverages the merits of INR in that the continuous functions are not coupled to sampling frequency and have infinite sampling frequency, allowing ImputeINR to generate fine-grained imputations even on extremely sparse observed values. Extensive experiments conducted on eight datasets with five ratios of masked values show the superior imputation performance of ImputeINR, especially for high missing ratios in time series data. Furthermore, we validate that applying ImputeINR to impute missing values in healthcare data enhances the performance of downstream disease diagnosis tasks. Codes are available.
Paper Structure (32 sections, 17 equations, 5 figures, 4 tables, 2 algorithms)

This paper contains 32 sections, 17 equations, 5 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Time Series Imputation
  4. Implicit Neural Representations
  5. Implicit Neural Representations on Time Series Data
  6. Methodology
  7. Problem Formulation
  8. Imputation with Missing Data
  9. Disease Diagnosis
  10. Method Overview
  11. Variable Clustering
  12. Multi-scale Feature Extraction
  13. INR Continuous Function
  14. Trend and Seasonal
  15. Residual: Adaptive Group-based Architecture
  16. ...and 17 more sections

Figures (5)

  • Figure 1: The illustration of INR applied to time series data.
  • Figure 2: The overall workflow of the proposed method. The input data with missing values is imputed using the ImputeINR model, and the imputed data is then fed into the disease diagnosis model to obtain the diagnostic results. In ImputeINR, the INR tokens are predicted using a transformer encoder. These tokens serve as the parameters for the INR continuous function, which takes the timestamp $t$ as input.
  • Figure 3: The four architectures we test to evaluate the representation capability of the INR continuous function for the synthetic time series dataset. The synthetic dataset consists of four variables, generated from two distinct distributions. The results prove that the representation capability of INR is strongest when variables from the same distribution are fitted by the same group.
  • Figure 4: A bubble chart reporting running time vs. imputation performance. The size of each bubble refers to the model size.
  • Figure 5: Robustness analysis for mask rates and the number of variables.