TL;DR
This paper introduces a novel time-aware soft clustering algorithm to identify distinct sepsis sub-phenotypes from electronic health records (EHRs). By incorporating temporal dynamics and clinical context, they discover six novel hybrid sub-phenotypes that offer more nuanced insights into sepsis progression and can be used to build more accurate early-warning prediction models.
What Problem Does This Paper Address?
Sepsis is a life-threatening condition with high mortality, characterized by a highly variable presentation and progression. This heterogeneity makes it challenging to develop effective, targeted treatments and accurately predict patient outcomes. While previous studies have attempted to identify sepsis sub-phenotypes using EHR data, they often failed to capture the temporal evolution of patient data or made rigid assumptions about sub-phenotype relationships. This paper aims to address these limitations by developing a more sophisticated phenotyping approach that accounts for the dynamic nature of sepsis and the complex interplay of clinical factors over time.
Methods
The core of this work is a time-aware soft clustering algorithm. This algorithm is a hybrid of harmonic K-Means and overlapping K-Means, designed to be less sensitive to initial centroid placement and to allow patients to belong to multiple sub-phenotypes simultaneously. Key innovations include:
- Time-aware Clustering: The algorithm explicitly incorporates temporal information from EHR data, allowing it to capture the evolving physiological and laboratory values of patients.
- Clinical Context Integration: The clustering process is guided by clinical context, specifically organ dysfunction types (liver, kidney, lung), using semi-supervised learning to calibrate cluster centroids. This ensures the identified sub-phenotypes are medically plausible.
- Soft Clustering: Unlike hard clustering, soft clustering assigns a degree of membership to each sub-phenotype, reflecting the complex and sometimes overlapping nature of sepsis progression.
- Post-soft Clustering Analysis: After the initial soft clustering, a K-Medoids clustering step is applied to group the soft clustering results, leading to the identification of six distinct "hybrid sub-phenotypes."
- Early-Warning Prediction Model: The identified hybrid sub-phenotypes are then used as ground truth labels to train a logistic regression-based early-warning prediction model.
Datasets
The study utilizes two large, de-identified EHR datasets:
- MIMIC-IV Database: Contains data from over 40,000 patients admitted to the ICU at Beth Israel Deaconess Medical Center (BIDMC) from 2008 to 2019.
- eICU Collaborative Research Database: Contains records from over 200,000 patients admitted to ICUs across the continental US in 2014 and 2015.
The authors selected patients based on specific Diagnosis Related Group (DRG) codes indicative of sepsis and focused on the first 120 hours of ICU stay.
Key Insights and Ideas
- Temporal Dynamics Matter: Capturing the time-series nature of clinical data is crucial for accurately phenotyping complex diseases like sepsis, as static snapshots miss critical information about disease progression.
- Hybrid Sub-phenotypes are Medically Plausible: The identified six hybrid sub-phenotypes show distinct patterns in clinical features and correlate with mortality rates, suggesting they represent meaningful clinical distinctions.
- Soft Clustering Enhances Granularity: Allowing patients to belong to multiple sub-phenotypes provides a more realistic representation of sepsis, which can affect multiple organ systems concurrently or sequentially.
- ABM as a Severity Indicator: The "ABM" (a measure of distance from cluster centroids) appears to correlate with disease severity and mortality, offering a data-driven way to assess patient condition.
Key Results
- Six Novel Hybrid Sub-phenotypes: The algorithm successfully identified six distinct sepsis hybrid sub-phenotypes from both MIMIC-IV and eICU datasets.
- Mortality Correlation: Hybrid sub-phenotype 1 consistently showed the highest mortality across both datasets, while other sub-phenotypes exhibited varying mortality rates that generally aligned with the ABM indicator.
- Early-Warning Prediction: The early-warning prediction model achieved promising results, with the best accuracy reaching 0.668 and AUPRC of 0.671 using 24 hours of ICU data on the MIMIC-IV dataset. The model's performance was generally better with earlier data (12-48 hours) compared to the full 120 hours.
Limitations and Future Work
- Cardiovascular Dysfunction Exclusion: The current model did not explicitly include cardiovascular dysfunction due to the complexity of incorporating vasopressor use. Future work aims to integrate this component.
- Generalizability: While tested on two large datasets, further validation on diverse databases and private hospital data is needed to confirm generalizability.
- Sub-phenotype Evolution: The study did not track how patients' assigned sub-phenotypes might change over time during their ICU stay. This dynamic aspect is a target for future research.
- Non-ICU Patients: The current study focused on ICU patients. Future work will explore phenotyping for non-ICU patients, which presents challenges due to wider time ranges and sparser data.
- Data-driven vs. Clinical Interpretation: While the identified sub-phenotypes show clinical plausibility, further investigation is needed to fully understand the clinical implications and potential treatment strategies for each.
Summary updated