Using FT Transformer for Sepsis Prediction

This project aims to assist our partner hospital's Emergency Department (ED) in developing a sepsis early prediction system using electronic health records (EHRs) and a state-of-the-art FT Transformer model.

Sepsis

Sepsis is a life-threatening condition that arises when the body's response to infection causes injury to its own tissues and organs

Challenges in Sepsis Diagnosis:

• Sepsis is commonly misdiagnosed and mistreated because deterioration with organ failure is also common in other diseases.
• The diagnosis of sepsis is often equivocal due to the varied nature of infection sources and wide-ranging patients’ responses.
• The diversity in age, gender, and comorbidities affect the symptoms and outcome of septic patients.

Why Machine Learning?

• Extract complex patterns from structured (e.g., vitals, labs) and unstructured (e.g., notes) EHR data.

• Provide real-time clinical support for sepsis diagnosis in high-pressure ED environments.

• Improve diagnostic accuracy and reduce treatment delays.

FT Transformer (Feature Tokenizer Transformer)

The FT Transformer is a powerful model architecture designed to handle tabular data efficiently, offering:

• A Feature Tokenizer that converts heterogeneous input features into embeddings.

• A Transformer encoder that models feature interactions.

• Superior performance in many tabular data tasks compared to traditional models.

Reference Paper:
Gorishniy et al., "Revisiting Deep Learning Models for Tabular Data," Link to paper

1. data_preprocessing demo

A step-by-step demo of how raw EHR data is cleaned, encoded, and transformed into a format suitable for the FT Transformer model.

data_preprocessing_demo.ipynb

2. model training demo

A notebook showcasing how to train the FT Transformer model for sepsis prediction using processed EHR data.

model_train_demo.ipynb