Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Temporal Sepsis Modeling: a Fully Interpretable Relational Way

Vincent Lemaire, Nédra Meloulli, Pierre Jaquet

TL;DR

This work tackles the challenge of early sepsis detection amid heterogeneous temporal trajectories by casting patient histories as relational data and applying a Minimal Description Length (MDL) based propositionalization to produce interpretable aggregates. The flattened data are then analyzed with a selective Bayesian classifier (Fractional Naive Bayes) that preserves probabilistic interpretability, aided by a two-stage MDL-driven variable selection process. The framework demonstrates strong predictive performance on the MIMIC-III dataset (around 3940 patients with a $12$-hour window and horizons $h=3$ and $h=6$), while providing multilevel explanations: univariate, global, local, and counterfactual trajectories. This combination of horizon-flexible modeling and intrinsic interpretability offers a practical pathway toward clinically actionable early sepsis prediction and broader deployment across diverse patient populations.

Abstract

Sepsis remains one of the most complex and heterogeneous syndromes in intensive care, characterized by diverse physiological trajectories and variable responses to treatment. While deep learning models perform well in the early prediction of sepsis, they often lack interpretability and ignore latent patient sub-phenotypes. In this work, we propose a machine learning framework by opening up a new avenue for addressing this issue: a relational approach. Temporal data from electronic medical records (EMRs) are viewed as multivariate patient logs and represented in a relational data schema. Then, a propositionalisation technique (based on classic aggregation/selection functions from the field of relational data) is applied to construct interpretable features to "flatten" the data. Finally, the flattened data is classified using a selective naive Bayesian classifier. Experimental validation demonstrates the relevance of the suggested approach as well as its extreme interpretability. The interpretation is fourfold: univariate, global, local, and counterfactual.

Temporal Sepsis Modeling: a Fully Interpretable Relational Way

TL;DR

This work tackles the challenge of early sepsis detection amid heterogeneous temporal trajectories by casting patient histories as relational data and applying a Minimal Description Length (MDL) based propositionalization to produce interpretable aggregates. The flattened data are then analyzed with a selective Bayesian classifier (Fractional Naive Bayes) that preserves probabilistic interpretability, aided by a two-stage MDL-driven variable selection process. The framework demonstrates strong predictive performance on the MIMIC-III dataset (around 3940 patients with a -hour window and horizons and ), while providing multilevel explanations: univariate, global, local, and counterfactual trajectories. This combination of horizon-flexible modeling and intrinsic interpretability offers a practical pathway toward clinically actionable early sepsis prediction and broader deployment across diverse patient populations.

Abstract

Sepsis remains one of the most complex and heterogeneous syndromes in intensive care, characterized by diverse physiological trajectories and variable responses to treatment. While deep learning models perform well in the early prediction of sepsis, they often lack interpretability and ignore latent patient sub-phenotypes. In this work, we propose a machine learning framework by opening up a new avenue for addressing this issue: a relational approach. Temporal data from electronic medical records (EMRs) are viewed as multivariate patient logs and represented in a relational data schema. Then, a propositionalisation technique (based on classic aggregation/selection functions from the field of relational data) is applied to construct interpretable features to "flatten" the data. Finally, the flattened data is classified using a selective naive Bayesian classifier. Experimental validation demonstrates the relevance of the suggested approach as well as its extreme interpretability. The interpretation is fourfold: univariate, global, local, and counterfactual.
Paper Structure (23 sections, 1 equation, 3 figures, 3 tables)

This paper contains 23 sections, 1 equation, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Context and Related Works
  3. Time-series and deep learning approaches
  4. Feature-based and window-aggregation approaches
  5. MDL-based and Bayesian explainable models
  6. Prediction horizons in early sepsis modeling
  7. Modeling: From temporal Data to Flat Data
  8. Dataset used
  9. Viewing the data as relational data
  10. Flatening the data
  11. Variable selection and learned classifier
  12. Filtering uninformative variables
  13. Finding the most informative subset of variables:
  14. Learning classifier
  15. Results
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Time series relational data flattening
  • Figure 2: Temporal relationship between the fixed observation window ($p=4,6$ hours) and different prediction horizons $h=3,1$.
  • Figure 3: Screenshot of Khiops Visualization after analyzing the sepsis database and constructing 10000 aggregates