Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Large Language Models For Irregular Time Series Classification In Critical Care

Feixiang Zheng, Yu Wu, Cecilia Mascolo, Ting Dang

TL;DR

This work tackles the challenge of irregular ICU time series by evaluating LLM-based time-series models, emphasizing two core components: time series encoders and multimodal alignment strategies. Through a systematic benchmark of four SOTA LLM-based methods against supervised and self-supervised baselines on ICU mortality tasks, the study finds that irregular-aware encoders (notably mTAND) drive large performance gains (approximately $12.8\%$ in AUPRC over a vanilla Transformer), while alignment strategies yield smaller improvements (up to $2.9\%$). However, the computational cost of LLMs remains prohibitive, with training times about $10\times$ longer and limited benefits in few-shot settings, and Warpformer still often outperforms LLM-based approaches. These results highlight both the potential and current limitations of LLMs for irregular time series in critical care, pointing to the need for more efficient architectures that balance accuracy and practicality. The code is available at the project page.

Abstract

Time series data from the Intensive Care Unit (ICU) provides critical information for patient monitoring. While recent advancements in applying Large Language Models (LLMs) to time series modeling (TSM) have shown great promise, their effectiveness on the irregular ICU data, characterized by particularly high rates of missing values, remains largely unexplored. This work investigates two key components underlying the success of LLMs for TSM: the time series encoder and the multimodal alignment strategy. To this end, we establish a systematic testbed to evaluate their impact across various state-of-the-art LLM-based methods on benchmark ICU datasets against strong supervised and self-supervised baselines. Results reveal that the encoder design is more critical than the alignment strategy. Encoders that explicitly model irregularity achieve substantial performance gains, yielding an average AUPRC increase of $12.8\%$ over the vanilla Transformer. While less impactful, the alignment strategy is also noteworthy, with the best-performing semantically rich, fusion-based strategy achieving a modest $2.9\%$ improvement over cross-attention. However, LLM-based methods require at least 10$\times$ longer training than the best-performing irregular supervised models, while delivering only comparable performance. They also underperform in data-scarce few-shot learning settings. These findings highlight both the promise and current limitations of LLMs for irregular ICU time series. The code is available at https://github.com/mHealthUnimelb/LLMTS.

Rethinking Large Language Models For Irregular Time Series Classification In Critical Care

TL;DR

This work tackles the challenge of irregular ICU time series by evaluating LLM-based time-series models, emphasizing two core components: time series encoders and multimodal alignment strategies. Through a systematic benchmark of four SOTA LLM-based methods against supervised and self-supervised baselines on ICU mortality tasks, the study finds that irregular-aware encoders (notably mTAND) drive large performance gains (approximately in AUPRC over a vanilla Transformer), while alignment strategies yield smaller improvements (up to ). However, the computational cost of LLMs remains prohibitive, with training times about longer and limited benefits in few-shot settings, and Warpformer still often outperforms LLM-based approaches. These results highlight both the potential and current limitations of LLMs for irregular time series in critical care, pointing to the need for more efficient architectures that balance accuracy and practicality. The code is available at the project page.

Abstract

Time series data from the Intensive Care Unit (ICU) provides critical information for patient monitoring. While recent advancements in applying Large Language Models (LLMs) to time series modeling (TSM) have shown great promise, their effectiveness on the irregular ICU data, characterized by particularly high rates of missing values, remains largely unexplored. This work investigates two key components underlying the success of LLMs for TSM: the time series encoder and the multimodal alignment strategy. To this end, we establish a systematic testbed to evaluate their impact across various state-of-the-art LLM-based methods on benchmark ICU datasets against strong supervised and self-supervised baselines. Results reveal that the encoder design is more critical than the alignment strategy. Encoders that explicitly model irregularity achieve substantial performance gains, yielding an average AUPRC increase of over the vanilla Transformer. While less impactful, the alignment strategy is also noteworthy, with the best-performing semantically rich, fusion-based strategy achieving a modest improvement over cross-attention. However, LLM-based methods require at least 10 longer training than the best-performing irregular supervised models, while delivering only comparable performance. They also underperform in data-scarce few-shot learning settings. These findings highlight both the promise and current limitations of LLMs for irregular ICU time series. The code is available at https://github.com/mHealthUnimelb/LLMTS.
Paper Structure (9 sections, 3 figures, 4 tables)

This paper contains 9 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Experimental Setup
  4. Results and Findings
  5. RQ1. To what extent can LLMs developed for regular time series effectively generalize to irregular ICU healthcare data?
  6. RQ2. Which components of LLMs for time series have the greatest impact on irregular time series modeling? Is it encoder or alignment between textual and temporal representations?
  7. RQ3. What are trade-offs between computational overhead introduced by LLMs and the gains they offer for irregular time series modeling, compared to traditional methods?
  8. RQ4. How effective are time series LLMs at the few-shot learning for irregular time series?
  9. Conclusion

Figures (3)

  • Figure 1: LLM architectures for time series (TS) modeling.
  • Figure 2: Workflow of LLMs. Irregular ICU time series is encoded and aligned with textual embeddings for classification.
  • Figure 3: AUPRC of LLMs on the semi-synthetic irregular ECG dataset as the proportion of missing values increases.