Clinically Contextualized State Spaces Provide Continuous Representations of Sepsis

Technical Summary: Clinically Contextualized State Spaces Provide Continuous Representations of Sepsis

Problem Statement
Precision medicine in sepsis is hindered by the syndrome's heterogeneity and the limitations of current phenotyping approaches. Conventional methods typically rely on cross-sectional, outcome-agnostic clustering, which groups patients based solely on biological or physiological similarity. The authors argue that this approach fails to capture clinically relevant heterogeneity because biologically similar states may follow divergent clinical trajectories, while biologically distinct states may share similar outcomes. Furthermore, existing longitudinal clustering often discovers temporal patterns without conditioning on specific clinical outcomes, and supervised dimensionality reduction (e.g., PLS-DA) is often limited to single outcomes. Sepsis, however, involves multiple interrelated trajectories (e.g., recovery with/without ventilation, mortality with/without renal replacement), necessitating a representation that integrates biological characteristics with these complex, multi-dimensional clinical contexts.

Methodology: CERL Framework
The authors introduce Contextual Endotype and State Representation Learning (CERL), a framework designed to generate continuous, clinically contextualized state representations.

• Core Architecture: CERL leverages TabPFN (Tabular Prior-data Fitted Network), a foundation model pretrained on over 100 million synthetic datasets, as its backbone. TabPFN is noted for its native robustness to noise and missingness.
• Learning Mechanism: Unlike traditional unsupervised clustering or single-outcome supervised methods, CERL learns a nonlinear latent representation of input features (routine laboratory measurements) conditioned on multiple, partially overlapping clinical outcomes (e.g., mortality, invasive mechanical ventilation, renal replacement therapy).
• State Space Construction: The framework produces a continuous manifold rather than discrete clusters. In this space, observations with similar biological profiles but different clinical contexts occupy distinct positions, while distinct biological profiles sharing a clinical context may occupy related regions.
• Discrete State Derivation: While the output is a continuous space, discrete clinical-state groups can be derived from the CERL embeddings using standard unsupervised clustering (e.g., K-means) for interpretability.
• Validation Strategy:
  • Synthetic Data: Experiments used simulated datasets with predefined endotypes and nonlinear generation mechanisms to test recovery of latent structures using the Adjusted Rand Index (ARI).
  • Real-World Data: The framework was applied to the eICU Collaborative Research Database (eICU-CRD), specifically a cohort of 2,127 patients meeting Sepsis-3 criteria.
  • Reproducibility: Validation was performed using institution-held-out splits to prevent data leakage from institution-specific practices. Reproducibility was assessed via cosine similarity of embeddings, Pearson correlation of laboratory profiles, and Mean Absolute Error (MAE) of clinical trajectory distributions.
  • Benchmarking: CERL was compared against conventional K-means clustering and Bayesian Gaussian Mixture (BGM) models applied directly to raw laboratory data.

Key Results

1. Synthetic Recovery: CERL consistently outperformed conventional unsupervised approaches (K-means, BGM, agglomerative clustering) in recovering predefined latent endotypes under nonlinear data-generating mechanisms. In a specific scenario (1,500 subjects, 3 endotypes, nonlinear mechanism), CERL achieved a mean ARI of 0.539, compared to 0.058 for K-means.
2. Reproducibility: In the eICU cohort, CERL demonstrated high reproducibility across institution-held-out validation sets.
   • Embedding Similarity: Cosine similarity between corresponding states in derivation and validation cohorts ranged from 0.888 to 0.974.
   • Biological Consistency: Pearson correlations of state-specific laboratory profiles ranged from 0.852 to 0.983.
   • Clinical Consistency: The distribution of clinical trajectories (e.g., acute recovery, late death) within derived states showed low error (MAE of 0.051–0.075) between derivation and validation cohorts.
3. Clinical Trajectory Informativeness: When benchmarked against K-means and BGM, CERL-derived states retained greater or comparable information regarding clinical trajectories. Specifically, CERL showed higher Area Under the Precision-Recall Curve (AUPRC) for acute recovery, late recovery, and invasive mechanical ventilation.
4. Real-World Application: Applied to the full eICU sepsis cohort, CERL identified three reproducible states (S0, S1, S2) visualized via PHATE:
   • S0: Enriched for acute death and invasive mechanical ventilation without renal replacement.
   • S1: Characterized by acute recovery.
   • S2: Enriched for renal replacement therapy, thrombocytopenia, and elevated BUN/creatinine.
   • The continuous space revealed outcome-specific temporal redistribution, where recovery trajectories concentrated in specific regions of the manifold over time, while mortality trajectories distributed more broadly.

Significance and Claims
The paper positions CERL not merely as a phenotyping tool but as a foundational framework for longitudinal pathological progression modeling and future digital twin approaches in critical care.

• Contextual Integration: The primary contribution is the construction of a state space where clinical trajectories actively contribute to the organization of physiologic states, rather than being evaluated post-hoc.
• Continuous Representation: By providing a continuous manifold, CERL preserves biological heterogeneity while organizing states according to clinically relevant trajectories, enabling the analysis of state transitions and dynamics that discrete clustering might obscure.
• Foundation for Future Work: The authors claim that this framework reduces reliance on post-hoc associations between subgroups and outcomes, offering a more robust basis for causal modeling and simulating disease progression.

Limitations Acknowledged
The authors note that CERL generates representations for individual observations and does not inherently model within-patient dependence; downstream analyses must account for repeated measures. Additionally, the representations are conditional on the specific outcomes defined a priori, meaning the inclusion of new outcomes requires reconstructing the representation rather than retroactively interpreting existing embeddings. Finally, the real-world evaluation was limited to a single database, and heterogeneity across healthcare systems remains a factor for future investigation.