Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Random Missingness: Clinically Rethinking for Healthcare Time Series Imputation

Linglong Qian, Yiyuan Yang, Wenjie Du, Jun Wang, Richard Dobsoni, Zina Ibrahim

TL;DR

This paper argues that random masking inadequately reflects real-world clinical missingness in healthcare time series and systematically evaluates clinically informed masking strategies across eleven imputation methods on the PhysioNet 2012 dataset. It introduces two masking schemes Augmentation RMEO and Overlay RMOD, along with pre-masking vs mini-batch masking and normalization timings NBΜ vs NAM, to study their impact on imputation accuracy and downstream mortality prediction. The findings show masking choices significantly influence both imputation metrics and clinical task performance, with recurrent models like BRITS showing robustness, while some architectures such as TimesNet may underperform in preserving predictive patterns. The work emphasizes the need for clinically grounded evaluation frameworks to ensure reliable deployment of imputation methods in healthcare settings, as high imputation accuracy does not always translate to better downstream decision support.

Abstract

This study investigates the impact of masking strategies on time series imputation models in healthcare settings. While current approaches predominantly rely on random masking for model evaluation, this practice fails to capture the structured nature of missing patterns in clinical data. Using the PhysioNet Challenge 2012 dataset, we analyse how different masking implementations affect both imputation accuracy and downstream clinical predictions across eleven imputation methods. Our results demonstrate that masking choices significantly influence model performance, while recurrent architectures show more consistent performance across strategies. Analysis of downstream mortality prediction reveals that imputation accuracy doesn't necessarily translate to optimal clinical prediction capabilities. Our findings emphasise the need for clinically-informed masking strategies that better reflect real-world missing patterns in healthcare data, suggesting current evaluation frameworks may need reconsideration for reliable clinical deployment.

Beyond Random Missingness: Clinically Rethinking for Healthcare Time Series Imputation

TL;DR

This paper argues that random masking inadequately reflects real-world clinical missingness in healthcare time series and systematically evaluates clinically informed masking strategies across eleven imputation methods on the PhysioNet 2012 dataset. It introduces two masking schemes Augmentation RMEO and Overlay RMOD, along with pre-masking vs mini-batch masking and normalization timings NBΜ vs NAM, to study their impact on imputation accuracy and downstream mortality prediction. The findings show masking choices significantly influence both imputation metrics and clinical task performance, with recurrent models like BRITS showing robustness, while some architectures such as TimesNet may underperform in preserving predictive patterns. The work emphasizes the need for clinically grounded evaluation frameworks to ensure reliable deployment of imputation methods in healthcare settings, as high imputation accuracy does not always translate to better downstream decision support.

Abstract

This study investigates the impact of masking strategies on time series imputation models in healthcare settings. While current approaches predominantly rely on random masking for model evaluation, this practice fails to capture the structured nature of missing patterns in clinical data. Using the PhysioNet Challenge 2012 dataset, we analyse how different masking implementations affect both imputation accuracy and downstream clinical predictions across eleven imputation methods. Our results demonstrate that masking choices significantly influence model performance, while recurrent architectures show more consistent performance across strategies. Analysis of downstream mortality prediction reveals that imputation accuracy doesn't necessarily translate to optimal clinical prediction capabilities. Our findings emphasise the need for clinically-informed masking strategies that better reflect real-world missing patterns in healthcare data, suggesting current evaluation frameworks may need reconsideration for reliable clinical deployment.
Paper Structure (28 sections, 1 figure, 1 table)

This paper contains 28 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep Learning Architectures for Medical Time Series
  4. RNN-based Approaches
  5. Attention-based Models
  6. Probabilistic Approaches
  7. Alignment with Clinical Missing Patterns
  8. Clinical Missing Mechanisms
  9. Current Evaluation Limitations
  10. Recent Developments
  11. Experimental Design
  12. Clinical-Informed Masking Strategies
  13. Timing of Artificial Missingness Introduction
  14. Normalization Procedures
  15. Downstream Clinical Task Evaluation
  16. ...and 13 more sections

Figures (1)

  • Figure 1: Masking techniques and approaches demonstrated over a time-series of five features ($x_1 \sim x_5$) and five-time points ($t_1 \sim t_5$): (a) random masking, (b) temporal masking, (c) spatial masking, (d) block masking. The yellow cells indicate those labelled as missing via masking. In (e) augmentation and (f) overlaying, the blue cells indicate cells that are missing within the original data. In (e), the masked (yellow) cells have no overlap with the original missingness in the data. Green: masked data coming from both the original missingness and artificial missingness. In (f), overlaying masks cells from either the original missingness or simulates artificial missingness from non-missing data.