Early Sepsis Detection Using Machine Learning and Deep Learning Models

This project explores machine learning (ML) and deep learning (DL) approaches for early detection of sepsis, a life-threatening condition requiring timely intervention. Using clinical data from PhysioNet Sepsis Challenge 2019, MIMIC-IV Clinical Demo, and MIMIC-IV-ED Demo, we evaluate and compare models including Logistic Regression, Random Forest, XGBoost, LSTM, and a Stacking Ensemble (RF + LSTM)

Datasets Used

PhysioNet Sepsis Challenge 2019 – ICU time-series data for sepsis prediction.

MIMIC-IV Clinical Demo – Structured ICU patient data.

MIMIC-IV-ED Demo – Emergency Department data for early-stage identification.

All datasets are publicly available through PhysioNet under appropriate credentialed access.

Methodology Overview

Exploratory Data Analysis (EDA): Handling missing values, class imbalance, feature distribution, correlation analysis.

Preprocessing & Feature Engineering: Rate of change features, time-window aggregation, normalization.

Model Training & Evaluation:

Traditional ML models: Logistic Regression, Random Forest, XGBoost

Deep learning: LSTM (with class weights)

Ensemble: Stacking Classifier (RF + LSTM)

Metrics Used

Accuracy

Recall (TPR)

False Positive Rate (FPR)

ROC AUC

Key Findings

Stacking Classifier (RF + LSTM) achieved the highest recall (87.69%) — crucial for medical diagnosis.

LSTM (with class weights) provided the best accuracy-recall balance.

Models trained on imbalanced data showed inflated accuracy but failed to detect true sepsis cases (low recall).