🏥 Cardiovascular Disease Analysis: Probabilistic Modeling & Decision Making

Advanced Machine Learning Project: Systematic exploration of probabilistic methods for medical data analysis - from mixture models to reinforcement learning

🎯 Project Overview

This project systematically explores probabilistic modeling techniques applied to cardiovascular disease data, progressing from basic clustering to advanced sequential decision making. Each milestone builds upon previous work, culminating in a reinforcement learning system for treatment optimization.

Core Learning Path

• 📊 M0: Exploratory Data Analysis & Statistical Foundations
• 🎯 M1: Gaussian Mixture Models for Patient Clustering
• 🕸️ M2-M4: Bayesian Networks (Design → Inference → Learning)
• ⏱️ M5: Hidden Markov Models for Disease Progression
• 🤖 M6: Reinforcement Learning for Treatment Policies

Key Features

• 📚 Systematic Methodology: Six milestones covering core probabilistic ML techniques
• 🔬 Hands-on Implementation: Build models from scratch using pgmpy, pygraphiz, hmmlearn, custom RL
• � Real-world Complexity: Handle missing data, discretization, temporal dependencies
• 🎓 Educational Focus: Deep understanding over black-box solutions

🏗️ Project Structure

├── data/
│   ├── raw/                    # 📁 Original datasets (INCLUDED in repo)
│   │   ├── patients.csv        # Static patient demographics & risk factors (3,000 patients)
│   │   ├── encounters.csv      # Longitudinal clinical encounters (24,000 visits)
│   │   └── README.md           # Comprehensive data documentation
│   ├── interim/                # Processed data (imputed, discretized)
│   └── processed/              # Final model-ready datasets
├── milestones/                 # 📋 Six milestone deliverables
│   ├── M0/                     # Exploratory Data Analysis
│   ├── M1/                     # Gaussian Mixture Models
│   ├── M2/                     # Bayesian Network Design
│   ├── M3/                     # BN Inference (Exact & Approximate)
│   ├── M4/                     # BN Learning from Data
│   ├── M5/                     # Hidden Markov Models
│   └── M6/                     # Reinforcement Learning
└── docs/                       # 📚 Documentation & project guidelines

🚀 Quick Start

1. Environment Setup

# Clone repository
git clone <repository-url>
cd Project

# Create virtual environment
python -m venv adv_ml_venv
source adv_ml_venv/bin/activate  # Linux/Mac
# adv_ml_venv\Scripts\activate   # Windows

# Install dependencies
pip install -r requirements.txt

2. Data Overview

import pandas as pd

# Load datasets (included in repo!)
patients = pd.read_csv('data/raw/patients.csv', sep=';')
encounters = pd.read_csv('data/raw/encounters.csv', sep=';')

print(f"📊 {len(patients):,} patients, {len(encounters):,} encounters")
print(f"⏱️  {encounters.groupby('patient_id').size().iloc[0]} time points per patient")

3. Start Exploring

# Run milestone notebooks
jupyter lab milestones/M0/M0.ipynb     # Data exploration
jupyter lab milestones/M1/M1_G14.ipynb # Feature engineering & modeling

🎓 Learning Objectives

This project systematically builds expertise in probabilistic machine learning:

Milestone	Core Technique	Learning Focus	Deliverable
M0	Exploratory Data Analysis	Data understanding, visualization	Statistical summaries & insights
M1	Gaussian Mixture Models	Unsupervised clustering, EM algorithm	Patient phenotype discovery
M2	Bayesian Network Design	Graphical models, conditional independence	Hand-designed probabilistic model
M3	BN Inference	Variable elimination, belief propagation	Clinical query answering
M4	BN Parameter/Structure Learning	Maximum likelihood, score-based search	Data-driven model comparison
M5	Hidden Markov Models	Temporal modeling, Viterbi decoding	Disease progression analysis
M6	Reinforcement Learning	Q-learning, policy optimization	Treatment recommendation system

Key Skills Developed

• 🧮 Probabilistic Reasoning: Understanding uncertainty and conditional dependencies
• 📊 Graphical Models: Designing and interpreting Bayesian networks
• ⏱️ Temporal Modeling: Capturing disease progression with HMMs
• 🎯 Decision Making: Optimizing treatment policies with reinforcement learning
• 🔬 Model Evaluation: Comparing hand-designed vs. learned models

📈 Milestones & Progress

📊 M0: Exploratory Data Analysis

• Dataset overview and baseline characteristics
• Disease state distributions and patient trajectories
• Treatment patterns and outcome analysis
• Missing data assessment and handling strategies

🎯 M1: Gaussian Mixture Models

• Feature selection and preprocessing for clustering
• GMM fitting with optimal cluster selection (AIC/BIC)
• Cluster characterization and clinical interpretation
• Comparison with true disease states

�️ M2: Bayesian Network Design

• Structure design with clinical justification
• Variable discretization for categorical BNs
• CPT estimation from encounter data
• Conditional independence analysis

🧠 M3: Bayesian Network Inference

• Exact inference (Variable Elimination, Belief Propagation)
• Approximate inference with sampling methods
• Clinical query design and interpretation

📚 M4: Learning Bayesian Networks

• Parameter learning with train/test patient splits
• Structure learning with score-based search
• Model comparison: hand-designed vs. learned
• Expert knowledge vs. data-driven trade-offs

⏱️ M5: Hidden Markov Models

• Temporal modeling with troponin + symptom features
• Baum-Welch parameter learning
• Viterbi decoding and state sequence analysis
• Comparison with true disease progression

🤖 M6: Reinforcement Learning

• MDP environment setup with state/action/reward
• Tabular Q-learning for treatment policies
• Policy evaluation vs. random/heuristic baselines
• Clinical interpretation of learned strategies

💡 Key Innovations

1. Multi-task Learning: Simultaneous disease prediction + treatment optimization
2. Temporal Modeling: Account for disease state transitions over time
3. Missing Data Realism: Handle realistic clinical data gaps
4. Utility-based Evaluation: Beyond accuracy - optimize patient outcomes
5. Causal Treatment Effects: Estimate counterfactual treatment scenarios

🔬 Data Highlights

• Domain: Synthetic cardiovascular clinical study
• Design: Longitudinal (8 time points over ~2 years)
• Patients: 3,000 diverse synthetic patients
• Variables: Demographics, risk factors, symptoms, labs, treatments, outcomes
• Missing Data: Realistic patterns (10-20% symptoms, 5-15% labs)
• Ground Truth: Disease states provided for educational purposes

📖 Full data documentation: data/raw/README.md

🤝 Contributing

See CONTRIBUTING.md for development workflow, branching strategy, and code standards.

⚠️ Important Notes

• 🎓 Educational Data: Synthetic dataset for learning purposes only
• ❌ Not for Clinical Use: Do not apply insights to real patients
• 📚 Research Only: Suitable for methodology development and education
• ✅ Reproducible: All code, data, and results version controlled