<p class="MsoNormal" style="margin-bottom: 12.0pt; text-align: left; mso-line-height-alt: 14.0pt; layout-grid-mode: char; mso-layout-grid-align: none;" align="left">Physiological Phenotypes of Sepsis Reveal a Normotensive High-Risk “Cryptic Shock” State: An Unsupervised Machine Learning Analysis of a Multicentre ICU Cohort and External Validation in Mimic-IV

Background: Sepsis is a heterogeneous syndrome, yet current definitions and management guidelines emphasize standardized, blood pressure–driven interventions. This "one-size-fits-all" approach often obscures physiologically unstable patients who lack overt hypotension. We hypothesized that distinct, reproducible physiological phenotypes of sepsis exist and that some high-risk states remain under-recognized by traditional assessment. Methods: We conducted a retrospective observational study using the PhysioNet 2019 Challenge dataset, comprising high-frequency ICU data from two hospital systems. To ensure independence, we reconstructed unique patient trajectories, yielding 40,217 unique patients. Features included vital signs and markers of perfusion (lactate, creatinine), with engineered variables such as Shock Index (HR/SBP). Unsupervised K-Means clustering was applied to a derivation cohort (70%) to identify phenotypes, with internal validation in the remaining cohort. We assessed the prevalence of algorithmic Sepsis-3 criteria across phenotypes and benchmarked supervised models (logistic regression vs. gradient boosting) to test physiological non-linearity. Findings: Four stable phenotypes were identified: Stable (38.6%, low risk); Hyperdynamic (28.7%, hypertensive/wide pulse pressure); Renal Dysfunction (3.7%, organ-dominant failure); and Cryptic Shock (29.0%). The Cryptic Shock phenotype was characterized by normotension (mean MAP ~75 mmHg) but significant metabolic distress (tachycardia, elevated lactate, high Shock Index). This group demonstrated a significantly higher prevalence of meeting algorithmic sepsis criteria compared with the Stable phenotype (10.2% vs 5.8%; p &lt; 0.001). Gradient boosting significantly outperformed logistic regression (AUC 0.744 vs 0.637), confirming that sepsis risk is driven by non-linear interactions poorly captured by linear scores. Interpretation: We identified a prevalent "Cryptic Shock" phenotype characterized by normotensive metabolic distress that is poorly detected by blood pressure–centric protocols. These findings have direct implications for early risk stratification and trial design, suggesting that blood pressure targets alone are insufficient endpoints for resuscitation in this subgroup.