Prediction-guided clustering for sepsis phenotyping: a retrospective cohort analysis

TL;DR

This study introduces a machine learning method to identify distinct sub-phenotypes of sepsis patients, which could lead to better personalized treatment strategies.

Contribution

A novel prediction-guided clustering approach that integrates deep learning with clinical outcomes to identify interpretable sepsis sub-phenotypes.

Findings

Six distinct sepsis sub-phenotypes with varying risk profiles and clinical presentations were identified.

The sub-phenotypes showed robust generalizability across different ICU datasets.

Reinforcement learning revealed different optimal treatment strategies for each sub-phenotype.

Abstract

Sepsis is a major cause of morbidity and mortality worldwide, with its heterogeneous and dynamically evolving clinical presentation complicating diagnosis, treatment, and prognosis. The identification of clinically meaningful sub-phenotypes within the sepsis population could help tailor interventions and improve outcomes. However, existing phenotyping studies have yielded inconsistent results with limited clinical utility. In this study, we propose a novel, guided machine-learning approach to identify clinically relevant sub-phenotypes within the sepsis condition by integrating deep representation learning with prediction-guided clustering to capture temporal disease trajectories.

We trained a recurrent neural network-based encoder to generate compact, predictive representations of sepsis patients over time. During training, the encoder is guided by four auxiliary prediction objectives (i.e., 90-day mortality, remaining length of stay, need for mechanical ventilation, and need for renal replacement therapy), which encourage the model to create representations that are relevant with respect to patient-centred outcomes. After training, patient representations were clustered using the K-means algorithm. The identified sub-phenotypes were compared across two large ICU data sets (AmsterdamUMCdb and MIMIC-IV) and interpreted using Integrated Gradients-based attribution maps. Practical and clinical utility of the phenotypes was evaluated using a reinforcement learning framework to evaluate optimal treatment strategies within each sepsis sub-phenotype.

Through our approach, we identified six clinically distinct sub-phenotypes with varying risk profiles and presentations. The learned representations demonstrated robust generalisability across the different data sets, and the reinforcement learning results indicated that the different sub-phenotypes were associated with different optimal treatment strategies, highlighting the potential for phenotype-informed decision-making.

This study introduces a flexible and effective framework for the identification of robust and clinically meaningful sub-phenotypes within the population of sepsis patients. Moreover, the identified sub-phenotypes are clinically interpretable, and the proposed trajectory-aware phenotyping approach may support the future development of personalised and precision medicine strategies.

The online version contains supplementary material available at 10.1186/s40635-026-00882-9.

Figures