Temporal robustness of biomarker-based classification algorithms for sepsis.

Summary

In two ICU cohorts with serial 8-hourly biomarker sampling, three immune profiles at admission frequently changed over hours to days. Poor intra-class cohesion (median Rand Index ~65%) indicates that single-timepoint endotyping is unstable, limiting its utility for patient stratification and targeted therapy in sepsis.

Key Findings

• Three immune profiles at ICU admission: adaptive activation (A, 43%), hyperinflammatory (B, 17%), and broadly attenuated (C, 39%).
• By 48 hours, prevalence shifted toward the attenuated profile (C increased to 56%).
• Frequent reclassification with poor intra-class cohesion (median Rand Index ~65%), indicating instability of biomarker-derived endotypes.

Clinical Implications

Avoid relying on single-timepoint immune endotypes for trial enrichment or bedside decisions. Consider repeated biomarker assessments and dynamic models when stratifying patients or timing immunomodulatory therapies.

Limitations

• Moderate sample size (n=345) and potential cohort-specific biomarker effects
• Biomarker panel limited to 30 immune markers; external validation in broader settings needed

Future Directions

Develop time-aware, dynamic endotyping frameworks and adaptive trial designs that account for rapid immunophenotypic transitions; evaluate whether trajectory-based stratification improves treatment effects.