ORAHS2025
Abstract Submission

222. Sequential Selection of Biomarkers in Sepsis Risk Scores Using Reinforcement Learning

Invited abstract in session TD-1: Blood and sepsis, stream Sessions.

Tuesday, 15:30-17:00
Room: NTNU, Realfagbygget R5

Authors (first author is the speaker)

1. Anandakrishnan Nandakumar
Mathematics, Amrita School of Physical Sciences
2. Dipu T. S.
Infectious Disease, Amrita Institute of Medical Sciences
3. Georg Gutjahr
Department of Health Science Research, Amrita Institute of Medical Sciences and Research Center, Kochi, Kerala, India

Abstract

Sepsis is a life-threatening condition caused by a dysregulated host response to infection. Early recognition and timely treatment are crucial, as delayed diagnosis can rapidly lead to organ damage, tissue injury, and mortality. Sepsis risk scores help quantify disease severity, predict disease onset and progression, and guide treatment strategies. High-precision risk scores often rely on costly biomarkers, including cytokines, procalcitonin, and immunophenotyping markers. In low-and-middle-income countries, a key question is whether a combination of surrogate markers can initially substitute for the costly markers, reserving the latter for high-risk patients identified during the initial screening phase. To address this question, we consider the sequential decision-making problem of selecting an initial set of biomarkers and subsequently adding more costly markers over the following days in a data-dependent way. We formulate the problem as a Markov decision process and use reinforcement learning to find a policy that minimizes cost while maintaining a high predictive performance, measured by the Net Reclassification Improvement (NRI) metric compared to full risk scores. The model is developed using data from the Indian OASIS cohort and the US MIMIC cohort and validated with a training-test split. We identify the potential for cost savings with an acceptable reduction in predictive accuracy.

Keywords

Status: accepted


Back to the list of papers