Identifying Sepsis Subphenotypes via Time-Aware Multi-Modal Auto-Encoder

Changchang Yin, Ruoqi Liu, Dongdong Zhang, Ping Zhang

Abstract

Sepsis is a heterogeneous clinical syndrome that is the leading cause of mortality in hospital intensive care units (ICUs). Identification of sepsis subphenotypes may allow for more precise treatments and lead to more targeted clinical interventions. Recently, sepsis subtyping on electronic health records (EHRs) has attracted interest from healthcare researchers. However, most sepsis subtyping studies ignore the temporality of EHR data and suffer from missing values. In this paper, we propose a new sepsis subtyping framework to address the two issues. Our subtyping framework consists of a novel Time-Aware Multi-modal auto-Encoder (TAME) model which introduces time-aware attention mechanism and incorporates multi-modal inputs (e.g., demographics, diagnoses, medications, lab tests and vital signs) to impute missing values, a dynamic time wrapping (DTW) method to measure patients’ temporal similarity based on the imputed EHR data, and a weighted k-means algorithm to cluster patients. Comprehensive experiments on real-world datasets show TAME outperforms the baselines on imputation accuracy. After analyzing TAME-imputed EHR data, we identify four novel subphenotypes of sepsis patients, paving the way for improved personalization of sepsis management.

Introduction

Sepsis remains a primary driver of hospital mortality, largely due to its extreme clinical heterogeneity. As a syndrome rather than a single disease, patients exhibit vastly different physiological responses and outcomes. Identifying distinct subphenotypes is critical for personalizing treatment, yet Electronic Health Records (EHRs) present significant challenges: pervasive missing data and irregular measurement intervals. Standard models often fail to capture the temporal trajectories of a patient's health, either by ignoring time gaps or using simplistic imputation methods that mask clinical nuances.

Methodology

The researchers developed a three-stage framework to extract meaningful patterns from noisy EHR data.

  1. TAME (Time-Aware Multi-modal auto-Encoder): This novel imputer processes multi-modal data, including demographics, diagnoses, medications, and lab results. It utilizes a bidirectional LSTM network with a time-aware attention mechanism. By treating the intervals between measurements as informative features rather than noise, TAME can more accurately predict missing values.
  2. Dynamic Time Warping (DTW): To compare patient trajectories, the framework employs DTW. Unlike static distance metrics, DTW aligns sequences that may progress at different speeds, allowing for a comparison of the "shape" of clinical recovery or decline over time.
  3. Weighted K-means: Finally, a robust clustering algorithm groups patients based on their DTW-aligned trajectories while minimizing the impact of outliers.
The TAME framework for sepsis subtyping, integrating imputation, similarity measurement, and clustering.

Experimental Results

The framework was validated on two major datasets: DACMI (8,267 patients) and MIMIC-III (11,715 adult sepsis patients). TAME demonstrated superior imputation accuracy compared to state-of-the-art baselines like BRITS and 3DMICE, achieving the lowest normalized Root Mean Square Error (nRMSE) across nearly all variables.

The analysis identified four distinct subphenotypes with unique "organ failure signatures" and mortality profiles:

  • Minimal Organ Dysfunction: The largest cohort (4,342 patients) with a low 4.1% mortality rate.
  • Renal Dysfunction: Characterized by significant kidney-related failure (8.7% mortality).
  • CNS & Respiration Dysfunction: High neurological and respiratory failure (15.9% mortality).
  • Maximum Organ Dysfunction: The highest-risk group (1,799 patients) with a 42.4% mortality rate.
Mortality rates over 40 days for the four identified sepsis subphenotypes.

Discussion and Conclusion

By leveraging multi-modal data and accounting for temporal irregularities, the TAME framework provides a more accurate lens into sepsis heterogeneity. The identification of these four subphenotypes suggests that the first 24 hours of an ICU stay contain sufficient information to predict long-term patient trajectories.

Future work should expand beyond the initial 24-hour window to capture subphenotypes that emerge later in the clinical course. Additionally, while the clusters are clinically distinct, translating these findings into specific treatment protocols—such as tailored fluid management for the renal dysfunction group—remains the next frontier for precision sepsis management.

Summary updated

BibTeX
@inproceedings{yin2020identifying,
  author = {Changchang Yin and Ruoqi Liu and Dongdong Zhang and Ping Zhang},
  title = {{Identifying Sepsis Subphenotypes via Time-Aware Multi-Modal Auto-Encoder}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2020},
  doi = {10.1145/3394486.3403129},
}
Figure