Explainable Machine Learning for Sepsis Outcome Prediction Using a Novel Romanian Electronic Health Record Dataset

1. Problem Statement

Sepsis remains a leading cause of global mortality, with significant regional disparities in outcomes. While machine learning (ML) models for sepsis prediction exist, they are predominantly trained on North American (e.g., MIMIC-III/IV, eICU-CRD) or Western European datasets. These models often fail to generalize to Eastern European populations due to differences in healthcare infrastructure, treatment protocols, and patient demographics. Furthermore, existing clinical scoring systems (SOFA, qSOFA, APACHE) rely on linear assumptions and limited variable sets, failing to capture complex, non-linear interactions between patient factors. There is a critical need for explainable, region-specific ML models that utilize routine laboratory data to predict sepsis outcomes with high accuracy, particularly in resource-constrained settings.

2. Methodology

Dataset and Cohort

• Source: Retrospective data from the University Emergency Hospital of Bucharest (UEHB), Romania's largest emergency hospital.
• Scale: 12,286 hospitalizations involving 12,089 adult patients diagnosed with sepsis between 2007 and 2024 (an 18-year span).
• Data Scope: Includes demographics, ICD-10 diagnostic codes, and 600 types of laboratory tests.
• Ethical Approval: Granted by the Carol Davila University of Medicine and Pharmacy and UEHB; informed consent was waived due to the retrospective, anonymized nature of the study.

Data Preprocessing & Feature Engineering

• Handling Sparsity: To address the extreme sparsity of 600 lab tests, the authors created subsets of the Top-10, 20, 30, 40, and 50 most frequent laboratory tests. This allowed for a trade-off analysis between feature richness and patient coverage (e.g., Top-10 covered 92.59% of patients, while Top-50 covered only 8.68%).
• Temporal Aggregation: Due to data sparsity in time-series, temporal variations were not modeled. Instead, the mean value of laboratory results over the hospitalization period was used.
• Diagnostic Encoding: ICD-10 codes (approx. 17,000 unique) were mapped into 14 clinically relevant comorbidity categories (e.g., Infectious Diseases, Cardiovascular, Nervous System) using a simplified encoding mechanism to avoid excessive sparsity.
• Class Imbalance: Random undersampling was applied to balance classes for specific binary tasks (e.g., matching "Deceased" vs. "Recovered" counts).

Classification Tasks

The study formulated three distinct binary classification scenarios:

1. Deceased vs. Discharged (Alive): Predicting survival vs. mortality (groups all non-deceased statuses together).
2. Deceased vs. Recovered: Distinguishing between mortality and full recovery.
3. Recovered vs. Ameliorated: Distinguishing between full recovery and partial improvement among survivors.

Machine Learning Models

Five algorithms were trained and compared:

• Logistic Regression (LR)
• Support Vector Classifier (SVC)
• Random Forest (RF)
• Gradient Boosting (GB)
• Histogram-based Gradient Boosting (HistGB)

Experimental Design:

• Variants: Models were tested in two configurations:
  • Base: Demographics + Lab results.
  • Extended: Demographics + Lab results + Comorbidities (ICD-10 derived).
• Evaluation Metrics: Accuracy, Precision, Recall, F1-score, and Area Under the Curve (AUC).
• Explainability: SHAP (SHapley Additive exPlanations) was used to interpret feature importance and ensure clinical transparency.

3. Key Results

Performance Metrics

The Histogram-based Gradient Boosting (HistGB) model consistently outperformed other algorithms, particularly when combined with diagnostic features.

• Task 1 (Deceased vs. Discharged):

  • Best Model: HistGB with Top-40 labs + Diagnostics.
  • Performance: AUC = 0.920, Accuracy = 0.843.
  • Significance: Demonstrates robust mortality prediction even without full recovery granularity.

• Task 2 (Deceased vs. Recovered):

  • Best Model: HistGB with Top-30 labs + Diagnostics.
  • Performance: AUC = 0.983, Accuracy = 0.930.
  • Significance: This is the highest-performing task, surpassing many state-of-the-art results on MIMIC/eICU datasets (typically 0.94–0.96 AUC).

• Task 3 (Recovered vs. Ameliorated):

  • Best Model: HistGB with Top-20 labs + Diagnostics.
  • Performance: AUC = 0.865.
  • Significance: Successfully identifies subtle differences in functional recovery among survivors.

Feature Importance (SHAP Analysis)

The SHAP analysis revealed both expected and novel predictors:

• Top Predictors: Cardiovascular comorbidities, Urea levels, Aspartate Aminotransferase (AST), Platelet count, and Eosinophil percentage.
• Key Discovery (Eosinopenia): The study identified Eosinopenia (low eosinophil count) as a top predictor across all tasks. Despite being the 26th most frequent test, it was a stronger predictor than many established markers. The model confirmed that lower eosinophil percentages correlate inversely with survival (higher mortality risk).
• Other Markers: Elevated Urea (kidney dysfunction/catabolism), low Platelets (coagulopathy), and liver enzymes (AST/ALT) were critical for mortality prediction.

4. Key Contributions

1. First Eastern European Sepsis ML Study: This is the first study to apply ML for sepsis outcome prediction using a large-scale, long-term (18-year) dataset from an Eastern European country, addressing a significant geographic gap in medical AI literature.
2. Novel Dataset: Introduction of a curated dataset of ~12,000 sepsis cases with 600 lab types and detailed ICD-10 mappings, offering a robust alternative to Western-centric datasets.
3. Superior Performance: Achieved state-of-the-art results (AUC 0.983 for Deceased vs. Recovered), outperforming previous benchmarks on MIMIC and eICU datasets.
4. Clinical Interpretability: The use of SHAP not only validated known clinical markers (urea, platelets) but also highlighted Eosinopenia as a high-value, underutilized marker for early sepsis risk stratification.
5. Resource Efficiency: Demonstrated that high-performance models can be built using routine laboratory tests (Top-20 to Top-40) and basic demographics, making the solution deployable in hospitals with limited advanced monitoring capabilities.

5. Significance and Implications

• Clinical Decision Support: The models can assist clinicians in triage, helping to identify high-risk patients earlier than traditional scores (SOFA/qSOFA) and potentially reducing mortality through earlier intervention.
• Resource Allocation: By distinguishing between patients likely to fully recover versus those with only partial improvement, the models can aid in optimizing ICU bed usage and transfer decisions, a critical factor in resource-strained healthcare systems.
• Generalizability: The study suggests that ML models trained on local, diverse data are superior to those trained on foreign datasets, emphasizing the need for region-specific AI development.
• New Biomarker Validation: The findings provide strong statistical evidence for the inclusion of eosinophil counts in sepsis assessment protocols, potentially updating current clinical guidelines.

Limitations

• Data Sparsity: The study relied on mean values rather than temporal trends due to missing time-series data.
• Subjectivity: Discharge status for survivors was recorded subjectively by treating physicians.
• Subset Constraints: Only patients with complete data for specific lab subsets were included, which may introduce selection bias.

Conclusion

The paper successfully demonstrates that explainable machine learning, specifically Histogram-based Gradient Boosting, can achieve exceptional accuracy in predicting sepsis outcomes using routine Romanian hospital data. The identification of Eosinopenia as a critical predictor and the high performance of the models suggest a viable path toward integrating AI-driven decision support into clinical workflows in Eastern Europe and similar healthcare environments.