A new method for assessing sepsis severity has been developed, moving beyond the static, decades-old scoring systems currently in use. Researchers created a sepsis index that learns directly from patient data over a 72-hour treatment window, using mortality as a ranking signal rather than fixed, hour-by-hour targets. This approach, detailed in a study published on arXiv, was validated across two hospital systems, involving a total of 36,807 adult patients meeting Sepsis-3 criteria.
Traditional sepsis severity indices, such as the Sequential Organ Failure Assessment (SOFA) score, rely on variables and weights established many years ago. These older metrics are often coarsely discretized and calibrated to patient populations that no longer accurately reflect contemporary critical care practices. The newly developed index aims to address these limitations by learning from patient trajectories.
The study utilized 43 routinely charted variables, including physiological, laboratory, and treatment data, to develop the sepsis index. Unlike previous methods that might target specific states or outcomes at fixed time points, this new approach uses mortality as a treatment-level ranking signal. This allows the model to redistribute "credit" or severity non-uniformly across timesteps, better reflecting the dynamic progression of the illness. The researchers employed an inverse reinforcement learning method called Trajectory-ranked Reward Extrapolation (T-REX) for this purpose.
Evaluation of the new index was conducted on a permanent 20% test holdout. The study found that non-survivors scored between 1.19 and 1.64 points higher than survivors on a 0-10 scale, even when stratifying patients by baseline SOFA scores. Similar results were observed when stratifying by lactate, mean arterial pressure (MAP), and creatinine levels. Furthermore, changes within a patient's index score correlated with changes in their lactate levels, with a Spearman correlation of 0.39 for 1,854 patients. Weaker correlations were found for MAP and creatinine.
Cross-institutional agreement between models trained on data from different hospital systems demonstrated a high degree of consistency, achieving 70-77% of the agreement seen within a single site. External within-patient correlations were reported at 0.54 and 0.59, against ceilings of 0.92 and 0.90. The new index also showed correlation with established indices, while control models with null data remained near zero. The researchers noted that their index demonstrated hourly prognostic information that meaningfully separates patient outcomes and aligns with clinical expectations, suggesting its potential as a decision support tool to complement clinical judgment.
This work builds upon previous research in developing data-driven severity scores for sepsis. A 2014 study developed a Sepsis Severity Score using logistic regression on data from over 23,000 patients, which accurately estimated hospital mortality. More recently, studies have explored machine learning frameworks for learning severity scores directly from clinical comparisons, showing improved accuracy over traditional scores like APACHE-II and SOFA in distinguishing sepsis stages. However, these often relied on specific clinical comparisons or fixed variables.
The current study's innovation lies in its use of mortality as a trajectory-level ranking signal, moving away from per-hour supervision. This allows the model to capture the dynamic nature of sepsis progression more effectively. The researchers suggest that a data-driven, continuously updated sepsis severity score derived directly from patient trajectories could substantially improve triage and resource allocation decisions in critical care settings. While the index demonstrated strong performance, the study acknowledges the need for local recalibration for clinical deployment.