Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TRACE: Intra-visit Clinical Event Nowcasting via Effective Patient Trajectory Encoding

Yuyang Liang, Yankai Chen, Yixiang Fang, Laks V. S. Lakshmanan, Chenhao Ma

TL;DR

This work defines laboratory measurement nowcasting within a hospital visit, addressing the gap in intra-visit predictions for timely clinical insights. It introduces TRACE, a Transformer-based architecture with a time-aware timestamp embedding that combines decay and periodic components and a smoothed denoising attention mask to handle noisy event sequences. Empirical results on MIMIC-III and MIMIC-IV show TRACE achieving state-of-the-art performance across PR-AUC, F1, Precision@k, and NDCG@k, validating the value of intra-visit information for real-time clinical decision-making. The study points to future directions in knowledge-augmented reasoning and advanced temporal trajectory learning to further improve robustness and applicability in EHR settings.

Abstract

Electronic Health Records (EHR) have become a valuable resource for a wide range of predictive tasks in healthcare. However, existing approaches have largely focused on inter-visit event predictions, overlooking the importance of intra-visit nowcasting, which provides prompt clinical insights during an ongoing patient visit. To address this gap, we introduce the task of laboratory measurement prediction within a hospital visit. We study the laboratory data that, however, remained underexplored in previous work. We propose TRACE, a Transformer-based model designed for clinical event nowcasting by encoding patient trajectories. TRACE effectively handles long sequences and captures temporal dependencies through a novel timestamp embedding that integrates decay properties and periodic patterns of data. Additionally, we introduce a smoothed mask for denoising, improving the robustness of the model. Experiments on two large-scale electronic health record datasets demonstrate that the proposed model significantly outperforms previous methods, highlighting its potential for improving patient care through more accurate laboratory measurement nowcasting. The code is available at https://github.com/Amehi/TRACE.

TRACE: Intra-visit Clinical Event Nowcasting via Effective Patient Trajectory Encoding

TL;DR

This work defines laboratory measurement nowcasting within a hospital visit, addressing the gap in intra-visit predictions for timely clinical insights. It introduces TRACE, a Transformer-based architecture with a time-aware timestamp embedding that combines decay and periodic components and a smoothed denoising attention mask to handle noisy event sequences. Empirical results on MIMIC-III and MIMIC-IV show TRACE achieving state-of-the-art performance across PR-AUC, F1, Precision@k, and NDCG@k, validating the value of intra-visit information for real-time clinical decision-making. The study points to future directions in knowledge-augmented reasoning and advanced temporal trajectory learning to further improve robustness and applicability in EHR settings.

Abstract

Electronic Health Records (EHR) have become a valuable resource for a wide range of predictive tasks in healthcare. However, existing approaches have largely focused on inter-visit event predictions, overlooking the importance of intra-visit nowcasting, which provides prompt clinical insights during an ongoing patient visit. To address this gap, we introduce the task of laboratory measurement prediction within a hospital visit. We study the laboratory data that, however, remained underexplored in previous work. We propose TRACE, a Transformer-based model designed for clinical event nowcasting by encoding patient trajectories. TRACE effectively handles long sequences and captures temporal dependencies through a novel timestamp embedding that integrates decay properties and periodic patterns of data. Additionally, we introduce a smoothed mask for denoising, improving the robustness of the model. Experiments on two large-scale electronic health record datasets demonstrate that the proposed model significantly outperforms previous methods, highlighting its potential for improving patient care through more accurate laboratory measurement nowcasting. The code is available at https://github.com/Amehi/TRACE.

Paper Structure

This paper contains 12 sections, 10 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Task Descriptions
  3. TRACE
  4. Medical Event Encoder
  5. Time-aware Transformer
  6. Masked Representation Prediction
  7. Experiment
  8. Experiment Setup
  9. Experimental Results
  10. Abalation Study
  11. Case Study
  12. Conclusion and Future Work

Figures (3)

  • Figure 1: Investigation on MIMIC-IV dataset. (a) Jaccard similarity of diagnosis codes between different visits of patients. (b) The proportion of patients with only one visit and patients with multiple visits in the MIMIC-IV dataset.
  • Figure 2: The TRACE model workflow.
  • Figure 3: A sample case of a patient with kidney dysfunction. For simplicity, the ordered patient trajectory only contains several important medical events. The wrong predictions are colored with red and the correct ones are colored with green.