Leakage-Free Multimodal Sepsis Prediction (LF-MSP)

Overview

This project implements a leakage-free multimodal sepsis early warning system that integrates structured Electronic Health Record (EHR) time-series data with clinical text. The system combines a two-layer LSTM, Bio_ClinicalBERT embeddings, and a logistic regression meta-model to produce early sepsis risk predictions with unified explainability.

The model predicts sepsis up to six hours prior to onset while enforcing strict temporal constraints to prevent data leakage, a common issue in retrospective clinical machine learning systems .

---

Objectives

• Develop a leakage-free multimodal prediction pipeline
• Integrate structured EHR and unstructured clinical notes
• Provide unified, clinician-centered explanations
• Evaluate performance under strict temporal validation

---

System Architecture

The system consists of four primary components:

1. Temporal Preprocessing

• Windowed EHR sequences (18-hour observation window)
• Strict masking of post-onset data
• Timestamp-based filtering of clinical notes

2. Modality Encoding

• LSTM for structured time-series data
• Bio_ClinicalBERT for clinical text representation

3. Fusion Layer

• Logistic regression meta-model combining modality outputs
• Robust handling of missing modalities

4. Explainability

• Gradient-based attribution for LSTM inputs
• TF-IDF-based term importance for text
• Fusion-level contribution analysis

---

Performance

Evaluation on a held-out test set shows:

• AUROC: 0.9770
• AUPRC: 0.6071
• Sensitivity: 0.8047
• Negative Predictive Value (NPV): 0.9945

These results were obtained under strict leakage control and demonstrate strong predictive capability relative to baseline methods .

---

Technology Stack

• Backend: FastAPI, Uvicorn
• Machine Learning: PyTorch, scikit-learn
• NLP: Hugging Face Transformers (Bio_ClinicalBERT)
• Data Processing: pandas, NumPy
• Visualization: matplotlib
• Deployment: Docker

---

Project Structure

sepsis_app/
├── model_artifacts/      # Trained models and preprocessing artifacts
├── static/               # Frontend assets
├── main.py               # FastAPI application
├── requirements.txt      # Python dependencies
├── Dockerfile            # Container configuration
├── setup.sh              # Environment setup script
├── run                   # Application run script

---

Local Setup

./setup.sh
source venv/bin/activate
./run

---

Docker Deployment

docker build -t sepsis-app .
docker run -p 8000:8000 sepsis-app

---

API

• Swagger UI: http://localhost:8000/docs
• Endpoint: POST /predict

The API accepts ICU patient data in CSV format and returns:

• Sepsis risk probability
• Model explanations
• Supporting visualizations

---

Dataset

This project is based on the MIMIC-IV v3.1 dataset, accessed via PhysioNet under appropriate data use agreements.

---

Notebooks

Full pipeline and experiments are available via Google Colab:

Data Extraction and Observation Window Sensitivity https://colab.research.google.com/drive/19ZVd51yaJM-8iih4zjh_e_nqjWK246He?usp=sharing

Model Training (LSTM and Meta-Fusion) https://colab.research.google.com/drive/1rtiGnGp7_e-l7BT6FYaMwxFye780y5kG?usp=sharing

---

Limitations

• Evaluation limited to a single dataset (MIMIC-IV)
• No external validation across institutions
• Transformer-based inference introduces latency

---

Future Work

• External validation on multi-site datasets
• Real-time clinical integration
• Model optimization for lower latency
• Advanced multimodal fusion techniques

---

Author

Dihas Sathnindu BSc Computer Science Informatics Institute of Technology (IIT) University of Westminster

---

License

This project is intended for academic and research purposes.