Early Prediction of Sepsis

Competition: PhysioNet Challenge 2019
Research Aim: Develop a model to predict sepsis at an early stage, before clinical diagnosis.

Challenges

• Limited Raw Features: Only 40 features are available in the dataset.
• High Missing Rates: Most features have missing rates above 90%.
• Variable-Length Inputs: The number of available records varies at each hour.
• Label Imbalance: Only 7.26% of patients develop sepsis (1.8% of total records).

Contributions

• Address dataset imbalance and variable-length time series issues
• Improve clinical outcome prediction
• Enhance model interpretability

Methodology

Pipeline

Input → Feature Engineering → Missing Value Imputation (LOOF) → Sampling (Down/Up) → Modeling (XGBoost)

Feature Engineering

• Feature Selection: Remove non-informative features (Bilirubin_direct, TroponinI, Fibrinogen)
• Missingness Indicators: Add features reflecting missing data patterns
• Sliding-Window Statistics: Compute mean, variance, and other time-windowed statistics
• Empirical Scores: Incorporate clinical scores such as SOFA and ∆SOFA
• Textual Representation: Convert [column name] [value] into embeddings using BioClinicalBERT

Evaluation

Evaluation Metric

Evaluation Strategy

Addressing Key Challenges

• High Missing Rate: Missing-value imputation + XGBoost
• Limited Features & Variable-Length Inputs: Feature engineering and BioClinicalBERT representations
• Imbalanced Dataset: Explore up-sampling, down-sampling, and weighted loss approaches

Results

Model Performance Comparison

Effectiveness of Class Imbalance Strategies

Impact of Different Feature Engineering Techniques

Contribution of Text-Based Features

Effect of Adding Text Features to a Deep Learning Model (MLP)

Comparing Deep Learning Models (FT-Transformer) vs. XGBoost

Feature Importance Analysis