JMIR Publications

ABSTRACT

Background:

The Sequential Organ Failure Assessment (SOFA) score is fundamental for risk stratification in intensive care units (ICUs), yet its expert-defined thresholds may be suboptimal for contemporary populations due to changes in case mix and advances in organ support therapies. While machine learning (ML) models have demonstrated superior predictive performance, most function as black boxes that lack the transparency required for clinical adoption.

Objective:

To develop an interpretable, data driven framework that recalibrates the SOFA score while preserving its additive structure and clinical transparency.

Methods:

We analyzed 14,848 ICU admissions from the eICU Collaborative Research Database (208 United States hospitals, 2014-2015), a multicenter dataset spanning diverse institutional practices and patient populations. The dataset was randomly split into training (60%), validation (20%), and test (20%) sets. The framework jointly learns data-driven discretization thresholds and interval-specific integer weights under explicit structural constraints. Discrimination was assessed using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC); calibration was evaluated graphically.

Results:

The recalibrated score achieved AUROC 0.825 (95% confidence interval [CI]: 0.805–0.844) versus 0.781 (95% CI: 0.760–0.802) for the original SOFA (p<0.001). AUPRC improved from 0.421 (95% CI: 0.383–0.456) to 0.530 (95% CI: 0.488–0.570). Improvements were consistent across length-of-stay strata, with calibration plots demonstrating superior alignment between predicted and observed mortality.

Conclusions:

Data-driven recalibration via constrained optimization and ML components, offers a practical pathway to modernize legacy clinical scores. By learning contemporary cut-points and weights directly from outcome data while preserving explicit additive structure, this approach enables systematic updating of established instruments without sacrificing clinical familiarity, auditability, or bedside usability.