Unsupervised Clustering Reveals Clinically Distinct Sepsis Phenotypes and Racial Disparities in ICU Outcome Classification: A Retrospective Cohort Study Using MIMIC-IV and eICU

Technical Summary: Unsupervised Clustering Reveals Clinically Distinct Sepsis Phenotypes and Racial Disparities in ICU Outcome Classification

Problem Statement
Sepsis is a heterogeneous syndrome responsible for approximately 11 million deaths annually, yet it is often managed as a single clinical entity. While unsupervised machine learning has been proposed to identify distinct sepsis phenotypes from electronic health records (EHRs) to enable precision medicine, existing studies have largely overlooked whether the clinical variables used for clustering introduce systematic demographic bias. Specifically, the study investigates whether serum creatinine—a standard marker of renal function included in severity scores like SOFA—disproportionately influences phenotype assignment in Black patients. This is critical because serum creatinine levels are known to vary systematically across racial groups due to differences in average muscle mass rather than kidney function alone, a factor that has led to the removal of race from estimated glomerular filtration rate (eGFR) calculations. The authors hypothesize that unsupervised clustering models relying on creatinine may inadvertently misclassify Black patients into more severe phenotypes, creating a source of racial measurement bias in clinical decision support.

Methodology
The study employed a retrospective cohort design using two large, de-identified ICU databases:

1. Discovery Cohort: MIMIC-IV (version 3.1), containing data from Beth Israel Deaconess Medical Center (2008–2022).
2. External Validation Cohort: eICU Collaborative Research Database (2014–2015), covering 208 US hospitals.

Data Processing and Feature Extraction:

• Population: Adult patients (≥18 years) with sepsis (ICD-10 codes A41%, R65%) and an ICU stay ≥24 hours.
• Features: 24 clinical variables extracted from the first 24 hours of ICU admission, including vital signs (heart rate, MAP, SpO2, temperature, respiratory rate) and laboratory values (creatinine, lactate, WBC, bilirubin, platelets).
• Preprocessing: Missing values were imputed using median imputation; continuous variables were standardized to zero mean and unit variance.

Clustering and Modeling:

• Phenotype Discovery: K-Means clustering was applied to the MIMIC-IV cohort. The optimal number of clusters ($k$) was determined using the silhouette coefficient, selecting $k=4$. The algorithm was initialized 20 times with different seeds to ensure stability.
• Classification: A Random Forest classifier was trained on the MIMIC-IV phenotype labels (60% training, 20% validation, 20% testing) and applied directly to the eICU cohort without retraining to assess generalizability.
• Bias Analysis:
  • Creatinine Sensitivity: The clustering procedure was repeated after removing all creatinine-related variables. The proportion of patients whose phenotype assignment changed was compared between Black and White patients using the Adjusted Rand Index (ARI).
  • Health Equity: Race-stratified analyses assessed phenotype distribution and mortality outcomes. Multivariable logistic regression was used to determine if race remained an independent predictor of high-mortality phenotype assignment after adjusting for creatinine and other severity markers.
  • Interpretability: SHapley Additive exPlanations (SHAP) values were calculated to identify feature importance.

Key Results

• Phenotype Identification: Four distinct phenotypes were identified in the MIMIC-IV cohort ($N=5,888$):
  • Phenotype A (3.4%): Hepatic failure dominant (high bilirubin), 63% 28-day mortality.
  • Phenotype B (46.6%): Hemodynamically stable, lowest mortality (29%).
  • Phenotype C (42.3%): Hyperdynamic/febrile, intermediate mortality (31%).
  • Phenotype D (7.8%): Metabolic crisis/high lactate (8.02 mmol/L), highest mortality (68%).
• External Validation: The Random Forest classifier achieved 89.5% test accuracy on MIMIC-IV and successfully reproduced the biological profiles and mortality ranking (D > A > C > B) in the eICU cohort ($N=19,609$) without retraining, yielding an AUC of 0.721.
• Racial Disparities and Creatinine Bias:
  • Race significantly predicted phenotype assignment ($\chi^2=34.0, p=0.0007$).
  • Black patients had significantly higher median serum creatinine than White patients (1.50 vs. 1.30 mg/dL, $p<0.001$).
  • Sensitivity Analysis: Upon removing creatinine, 49.4% of Black patients were reassigned to a different phenotype compared to 46.3% of White patients. Notably, nearly 90% of the reassigned Black patients were originally in Phenotype C.
  • Mortality Paradox: Within the highest-risk Phenotype D, Black patients exhibited paradoxically lower mortality (57.4%) compared to White patients (68.8%), suggesting that Black patients in this group may have been misclassified into a higher severity category due to creatinine levels.
  • Mediation: Multivariable logistic regression showed that while race was initially associated with high-mortality phenotype assignment, this association disappeared after adjusting for serum creatinine and other severity markers (Black: OR = 0.72, $p=0.110$), indicating creatinine mediates the disparity.
• Feature Importance: While SHAP analysis ranked creatinine only 15th/16th in global importance, its removal disproportionately altered assignments for Black patients, demonstrating that global feature importance metrics are insufficient for detecting demographic bias.

Significance and Claims
The paper claims to provide the first evidence that unsupervised sepsis phenotyping models can introduce racial measurement bias through creatinine-based features. The authors argue that:

1. Measurement Bias, Not Algorithmic Bias: The observed disparity originates from the measurement itself (creatinine's physiological variation by race) rather than the machine learning algorithm.
2. Upward Severity Misclassification: The findings suggest that Black patients may be systematically assigned to higher-severity phenotypes (and thus potentially receive more aggressive interventions) based on demographic differences in creatinine rather than true physiological severity.
3. Limitations of Global Metrics: Conventional feature importance measures (like SHAP) may fail to identify demographic bias if a feature has low global importance but high influence within specific subpopulations.
4. Clinical Implication: Before deploying sepsis phenotyping systems for clinical decision support, they must be evaluated for racial measurement bias. The study concludes that fairness evaluation should become a routine component of developing and validating clinical phenotyping systems to ensure precision medicine improves outcomes equitably across diverse populations.

The authors maintain a modest tone, acknowledging limitations such as the single-center origin of the discovery cohort, self-reported race data, and the assumption that removing creatinine yields a "fairer" solution (which warrants further investigation with alternative biomarkers like cystatin C). They do not propose specific new algorithms but rather call for rigorous bias evaluation in existing and future models.