🧬 Survival Analysis & Risk Prediction using Gene Expression (TCGA-BRCA)

🌐 Live Demo

👉 Try the app here: https://survival-analysis-9gze2w98kxusgx6eclcj7x.streamlit.app/

---

📌 Project Overview

This project predicts high-risk vs low-risk breast cancer patients using gene expression data from TCGA-BRCA (Breast Cancer subset of TCGA).

The Cancer Genome Atlas is a large-scale cancer genomics database containing multiple cancer types. In this project, we specifically use TCGA-BRCA to ensure consistent disease biology and meaningful modeling.

The pipeline integrates:

• Survival Analysis (Kaplan–Meier, Cox Model)
• Feature Selection (LASSO)
• Machine Learning (XGBoost, Random Forest, Neural Network)
• Deployment using Streamlit

---

🎯 Objectives

• Analyze survival patterns in breast cancer patients
• Identify key genes associated with survival
• Build ML models to classify high-risk vs low-risk patients
• Deploy a clinical-style application for real-time prediction

---

📂 Dataset

Source: TCGA-BRCA

Data Used:

• Clinical data (survival time, event)
• Gene expression data (RNA-seq)

Final Dataset:

• 41 patients

• 20 selected genes

• Target: risk_label

  • 1 → High risk
  • 0 → Low risk

---

⚙️ Methodology

1️⃣ Data Preprocessing

• Cleaned missing values ("--" → NaN)

• Converted survival columns to numeric

• Created:

  • time (survival duration)
  • event (death indicator)

---

2️⃣ Survival Analysis

• Kaplan–Meier estimation
• Cox Proportional Hazards model
• L1 regularization (LASSO) for feature selection

---

3️⃣ Feature Selection

~60,000 genes → Top 500 → LASSO → Top 20 genes

---

4️⃣ Risk Label Creation

risk_label = (time < median(time)).astype(int)

---

5️⃣ Machine Learning Models

• XGBoost (ROC-AUC: 0.79)
• Neural Network (0.76)
• Random Forest (0.71)

Evaluation:

• Cross-validation (cv=5)
• ROC-AUC metric

---

🚀 Deployment (Streamlit App)

A Streamlit web application was developed to simulate a clinical decision-support tool.

Features:

• Upload patient gene expression data (CSV)
• Predict survival risk (High / Low)
• Display prediction probability
• Provide model-based insights

👉 Live App: https://survival-analysis-9gze2w98kxusgx6eclcj7x.streamlit.app/

---

📊 Results

Model	ROC-AUC
XGBoost	0.79
Neural Network	0.76
Random Forest	0.71

---

🧠 Key Insights

• Gene expression data can predict patient survival risk
• Multiple genes contribute to cancer progression (not just BRCA1/2)
• XGBoost achieved the best performance
• Combining survival analysis with ML improves interpretability

---

⚠️ Limitations

• Small sample size (41 patients)
• Risk label simplifies survival (does not fully account for censoring)

---

🚀 Future Work

• Use larger datasets
• Apply deep learning survival models (e.g., DeepSurv)
• Integrate multi-omics data
• External validation on independent cohorts

---

🛠️ Tech Stack

• Python
• Pandas, NumPy
• Scikit-learn
• XGBoost
• Lifelines
• Streamlit

---

📌 Conclusion

This project demonstrates an end-to-end pipeline from genomic data to deployment. It highlights how machine learning and survival analysis can support clinical decision-making in breast cancer.

---

👨‍💻 Author

Abraham Pulickakudiyil Master’s Student – Life Science Informatics

---

⭐ If you like this project, consider giving it a star!