Multi-Armed Bandit for Personalized Drug-Dosage Optimization

(Individual Academic Project – 2025)

A complete, reproducible proof-of-concept that uses classical multi-armed bandit (MAB) algorithms to recommend optimal drug-dosage combinations from historical patient data.

Project Overview

• Goal: Identify the best drug + dosage pair for three major patient groups (mental health, chronic pain, metabolic/chronic disorders) while minimizing cumulative regret.
• Action space: 5 clinically relevant drugs × 3 standardized dosages (0.5, 1.0, 1.5) → 15 arms per group.
• Algorithms implemented from scratch:
  • ε-greedy
  • UCB1 (tunable confidence parameter)
  • Thompson Sampling (Beta-Bernoulli / Gaussian version)
• Evaluation: 100 independent Monte Carlo runs × 1000 pulls per algorithm and patient group.
• Key result: Thompson Sampling consistently outperforms the others, reducing regret by 12–27 % depending on the clinical group.

Features

• Single, fully executable Google Colab notebook (runs in < 4 minutes on the free tier)
• No external dependencies beyond standard Python data-science libraries
• Fixed random seed (42) → 100 % reproducibility
• Publication-ready plots (cumulative regret curves, boxplots, optimal-arm heatmaps)
• Clean, heavily commented code suitable for teaching or collaboration

Repository Contents

├── main.ipynb                  ← Complete Colab notebook (open the badge above)
├── historical_patient_data.csv ← Sample synthetic dataset (real data redacted)
├── requirements.txt            ← pip install -r requirements.txt
├── README.md                   ← This file
└── figures/                    ← Exported plots (optional)

How to Run

1. Click the “Open in Colab” badge above, or
2. Clone the repo and open main.ipynb in Jupyter/Colab locally.

git clone https://github.com/yourusername/multi-armed-bandit-drug-optimization.git
cd multi-armed-bandit-drug-optimization
jupyter notebook main.ipynb

Requirements

pandas>=2.0
numpy>=1.24
matplotlib>=3.7
seaborn>=0.12
scipy>=1.10
scikit-learn>=1.3

Install with:

pip install -r requirements.txt

Future Collaboration (Open Invitation)

This prototype is deliberately lightweight and interpretable so it can be handed directly to clinical or pharmacological research groups for:

• Validation on real, richly phenotyped patient cohorts
• Extension to narrow disease clusters (e.g., treatment-resistant depression, fibromyalgia, diabetic nephropathy)
• Integration of safety constraints and prospective testing

I am actively seeking medical collaborators for the next phase.

License

MIT License – feel free to use, modify, and cite in academic work (please reference this repository).

Author

Vahe Hambardzumyan

⭐ Star this repo if you find it useful for teaching reinforcement learning in healthcare!