This project is a submission for the RWEsearch Hackathon by Healthark. It leverages Centers for Medicare & Medicaid Services (CMS) data to build an AI-driven solution that predicts 30-day hospital readmission risk for beneficiaries. [cite_start]The final output is an interactive Streamlit dashboard designed to provide actionable insights for healthcare providers. [cite: 27]
[cite_start]The core of this project is a machine learning model that analyzes patient demographics, treatment histories, and prescription data to identify individuals at high risk of readmission. [cite: 26] The final model is a fine-tuned XGBoost Classifier that achieved an outstanding 97% recall and a 63% F1-score, demonstrating a powerful ability to find at-risk patients.
- Data Integration: Combines beneficiary summaries, inpatient claims, outpatient claims, and prescription drug events into a single, analysis-ready dataset.
- Advanced Feature Engineering: Creates over 40 features, including patient-level aggregations, one-hot encoded categoricals, and sophisticated provider-level performance metrics.
- Robust Model Training: Employs a rigorous pipeline including hyperparameter tuning (
GridSearchCV) and advanced techniques to handle class imbalance (scale_pos_weight). - Interactive Dashboard: A user-friendly Streamlit application that visualizes the model's key insights and performance metrics.
1. Clone the repository:
git clone https://github.com/krushanu27/healthark.git
cd healthark2. Install the required libraries:
pip install -r requirements.txt3. Place the data files: This project uses the CMS Linkable 2008–2010 Medicare Data Entrepreneurs' Synthetic Public Use File (DE-SynPUF). Due to their size, the data files are not included in this repository. Please download them and place them in the project's root directory.
4. Run the Streamlit app:
streamlit run dashboard.pyAfter extensive experimentation with multiple architectures (Random Forest, Neural Networks) and data balancing techniques (SMOTE, class weighting), the champion model was a tuned XGBoost Classifier with the following performance on the test set:
| Metric | Score |
|---|---|
| F1-Score (Readmitted) | 63% |
| Recall (Readmitted) | 97% |
| Precision (Readmitted) | 47% |
| Overall Accuracy | 94% |
The model's ability to identify 97 out of 100 at-risk patients makes it a highly valuable tool for clinical intervention.