🏥 EHR Predictive Modeling: Mortality & Length-of-Stay

📌 Overview

This repository presents a comprehensive end-to-end pipeline for predicting in-hospital mortality and length of hospital stay (LOS) using Electronic Health Record (EHR) data.

The project systematically compares:

• Traditional Machine Learning models
• Sequential Deep Learning models (LSTM, Transformer)
• Clinical Large Language Models (LLMs)

The dataset is a synthetic EHR dataset inspired by MIMIC-III/IV, consisting of 5,000 patients across multiple relational tables.

---

📂 Repository Structure

├── data/
│   ├── patients_table.csv
│   ├── patient_admissions.csv
│   └── disease_diagnosis_code.csv
│
├── Notebooks/
│   ├── Preprocessing/
│   │   └── EHR_featureEngineering.ipynb
│   │
│   ├── Traditional_ML/
│   │   └── EHR_modelDevelopment.ipynb
│   │
│   ├── LSTM_Transformer/
│   │   └── LSTM_transformer.ipynb
│   │
│   └── LLM/
│       ├── llm_internist_7b.ipynb
│       └── LLM_Bio_ClinicalBERT_Modified.ipynb
│
├── scripts/
│   ├── ehr_featureengineering.py
│   └── ehr_modeldevelopment.py 
│   └── llm_internist_7b.py
│   └── llm_bio_clinicalbert_modified.py
│
└── README.md

---

📊 Dataset Description

The dataset consists of three relational tables:

1. Patients Table

• Demographics: age, gender, year group
• Death information

2. Admissions Table

• Admission/discharge timestamps
• Admission type, insurance, race
• Target variable: hospital_expire_flag

3. Diagnosis Table

• ICD-9/ICD-10 diagnosis codes
• Sequential clinical history per admission

---

🎯 Objectives

• Predict in-hospital mortality (binary classification)
• Predict length of stay (LOS) (regression + classification)
• Evaluate model performance across ethnic groups
• Compare performance across different model architectures

---

🧭 Workflow Overview

The project follows a structured pipeline from raw EHR data to advanced predictive modeling:

1. 🔧 Data Preprocessing

Notebook: Notebooks/Preprocessing/EHR_featureEngineering.ipynb Script: Scripts/EHR_featureEngineering.py

• Data integration across relational tables using subject_id and hadm_id
• ICD-9 → ICD-10 harmonization
• Handling missing values and leakage prevention

🔹 Feature Engineering

• Temporal Features

  • Length of Stay (LOS) from admission and discharge time
  • Admission year, month, and time-based patterns

• Demographic Features

  • Age at admission (derived from anchor variables)
  • Gender encoding
  • Race and ethnicity encoding

• Diagnosis-Based Features

  • ICD code normalization and mapping
  • Grouping using Clinical Classification Software (CCSR)
  • Multi-label binarization of diagnosis categories
  • Number of diagnosis categories per admission

• Categorical Encoding

  • Admission type
  • Insurance type
  • Admission and discharge locations

---

2. 📊 Traditional Machine Learning

Notebook: Notebooks/Traditional_ML/EHR_modelDevelopment.ipynb
Script: Scripts/EHR_modelDevelopment.py

• Models: Logistic Regression, Random Forest, XGBoost, LightGBM
• Tabular feature-based modeling
• Baseline performance benchmarking. Best for structured tabular data. Providing best performance for clinical EHR data outperforming even clinical LLM models.

---

3. 🔄 Sequential Deep Learning

Notebook: Notebooks/LSTM_Transformer/LSTM_transformer.ipynb

• Patient trajectory modeling using sequential data
• Models: LSTM, Transformer
• Capturing temporal dependencies in admissions.

---

4. 🧠 Clinical Large Language Models (LLMs)

Notebooks:

• Notebooks/LLM/llm_internist_7b.ipynb
• Notebooks/LLM/LLM_Bio_ClinicalBERT_Modified.ipynb

Scripts:

• Scripts/llm_internist_7b.py

• Scripts/LLM_Bio_ClinicalBERT_Modified.py

• Conversion of structured EHR data into clinical text

• Models: BioClinicalBERT, Internist-7B

• Fine-tuning and multitask learning (mortality + LOS).

---

📈 Evaluation Metrics

Classification (Mortality)

• F1-score
• AUROC
• AUPRC
• Precision & Recall
• 95% Confidence Interval (CI)

Regression (LOS)

• MAE
• RMSE
• R²

Additional

• Calibration (ECE, Brier Score)
• Fairness Analysis: Subgroup performance across race/ethnicity

5. 📈 Evaluation & Analysis

• Metrics: AUROC, F1-score, Precision, Recall
• Regression: MAE, RMSE, R²
• Calibration: ECE, Brier Score
• Fairness: Subgroup performance across race/ethnicity

---

🧪 Key Findings

• Traditional ML (XGBoost) achieved strongest overall performance
• Sequential models captured temporal signals but did not outperform tabular models consistently
• Clinical LLMs showed strong discrimination but calibration issues
• Model performance varied across demographic subgroups

---

🚀 How to Run

1. Clone repository

git clone https://github.com/your-username/repo-name.git
cd repo-name

2. Install dependencies

pip install -r requirements.txt

3. Run notebooks

• Feature Engineering → EHR_featureEngineering.ipynb
• ML Models → EHR_modelDevelopment.ipynb
• Deep Learning → LSTM_Transformer.ipynb
• LLM Models → llm_internist_7b.ipynb, LLM_Bio_ClinicalBERT_Modified.ipynb

4. Run scripts (Alternative of "Run notebooks")

• Feature Engineering → EHR_featureEngineering.py
• ML Models → EHR_modelDevelopment.py
• LLM Models → llm_internist_7b.py, LLM_Bio_ClinicalBERT_Modified.py

---

🧠 Technologies Used

• Python (Pandas, NumPy, Scikit-learn)
• PyTorch
• HuggingFace Transformers
• XGBoost / LightGBM
• PEFT (LoRA, QLoRA)

---

⚠️ Disclaimer

This project uses a synthetic dataset inspired by clinical data (MIMIC-III/IV).
It is intended for research and educational purposes only and should not be used for real clinical decision-making.

---

📌 Future Improvements

• Incorporate real-world EHR datasets
• Improve LLM calibration
• Advanced fairness-aware modeling
• Multimodal integration (labs, imaging)

---

👩‍🔬 Authors

• Shalini Majumder, MSc Bioinformatics FT, University of Birmingham
• Ekarsi Lodh, MSc Bioinformatics FT, University of Birmingham
• Abhijith Sajeev, MSc Bioinformatics FT, University of Birmingham

---

📜 License

This project is for academic and research purposes only.