A comprehensive two-phase study combining network science and deep learning to analyze patient care pathways in chronic disease management
This repository presents a complete analysis of Electronic Health Record (EHR) data for chronic disease management, structured in two complementary research phases:
Phase 1 applies network analysis to identify critical healthcare infrastructure and care coordination patterns.
Phase 2 leverages Graph Neural Networks to predict multimorbidity risk and recommend preventive care interventions.
- Diabetes Mellitus
- Cardiovascular Disease (Heart Disease)
- Cancer
- Chronic Kidney Disease (CKD)
Identifying universal hubs, bridges, and community structures in patient care networks
- Analyzed 262,833 patient transitions across 120 care facilities
- Identified 8 universal hub sites handling 83.7% of all transitions
- Detected 23 functional communities (modularity Q=0.67)
- Quantified 34 bridge facilities critical for care coordination
NetworkX PageRank Betweenness Centrality Louvain Algorithm Plotly Interactive Visualizations
Predicting multimorbidity risk using patient similarity networks and deep learning
| Model | AUC-ROC | Accuracy | F1-Score |
|---|---|---|---|
| GCN | 0.9745 | 96.50% | 0.9810 |
| GIN | 0.9734 | 95.00% | 0.9730 |
| GraphSAGE | 0.8965 | 96.25% | 0.9795 |
- Built patient-patient similarity graphs from 1,162 patients
- Trained 5 GNN architectures (GCN, GraphSAGE, GAT, GIN, APPNP)
- Developed care recommendation system identifying preventive care gaps
- Achieved 94% performance with only 30% labeled data
PyTorch Geometric Graph Neural Networks Semi-Supervised Learning Collaborative Filtering Ablation Studies
| Category | Technologies |
|---|---|
| Languages | Python 3.8+ |
| Data Processing | Pandas, NumPy, SQL |
| Network Analysis | NetworkX 2.8 |
| Deep Learning | PyTorch 1.12, PyTorch Geometric 2.0 |
| Machine Learning | scikit-learn |
| Visualization | Plotly, Matplotlib, Pyvis |
| Algorithms | PageRank, Betweenness Centrality, Louvain, GCN, GraphSAGE, GAT, GIN, APPNP |
healthcare-network-analysis/
│
├── README.md ← You are here
│
├── project-1-network-analysis/ ← Phase 1: Network Analysis
│ ├── README.md
│ ├── Network_Analysis_CODE.ipynb
│ ├── Cancer_Graph.html
│ ├── Diabetes_Graph.html
│ ├── HeartDisease_Graph.html
│ ├── Chronic_Kidney_Disease_Graph.html
│ ├── Multi_Disease_Interactive_Graph.html
│ └── Network_Communities.html
│
└── project-2-gnn-prediction/ ← Phase 2: GNN Prediction
├── README.md
├── GNN_Multimorbidity_Prediction.ipynb
└── visualizations/
├── ROC_Curves_Model_Comparison.png
├── Feature_Ablation_Study.png
├── TSNE_Embedding_Visualization.png
└── ...
git clone https://github.com/Mahendra-logics/healthcare-network-analysis.git
cd healthcare-network-analysiscd project-1-network-analysis
pip install pandas numpy networkx plotly pyvis
jupyter notebook Network_Analysis_CODE.ipynbcd project-2-gnn-prediction
pip install torch torch-geometric networkx scikit-learn pandas numpy
jupyter notebook GNN_Multimorbidity_Prediction.ipynb- Capacity Planning: Strategic resource allocation at universal hub facilities
- Care Coordination: Targeted interventions at bridge sites connecting communities
- Comorbidity Management: Integrated care clinics for overlapping disease networks
- Risk Stratification: Proactive identification of high-risk multimorbidity patients
- Care Gap Analysis: Preventive service recommendations based on peer utilization
- Population Health: Scalable risk assessment for large patient cohorts
- Novel Methodology: Patient-patient similarity network construction from EHR data
- Comparative Analysis: Systematic evaluation of 5 GNN architectures for healthcare
- Semi-Supervised Learning: High accuracy with limited labeled data
- Interpretability: Comprehensive ablation studies revealing feature importance
- Hub Infrastructure: 8 universal sites form the backbone of multi-disease care delivery
- Disease Patterns: CKD patients navigate most complex pathways (16.1 avg sites/patient)
- Community Structure: Strong modularity (Q=0.67) indicates functional specialization
- Bridge Sites: 34 critical connectors vulnerable to bottlenecks
- Model Performance: GCN achieves 97.45% AUC for multimorbidity prediction
- Feature Importance: Behavioral features (network position, care utilization) outperform demographics
- Data Efficiency: 30% labeled data sufficient for 94% of maximum performance
- Network Robustness: Models maintain 85%+ AUC even with 75% missing edges
- Clinical Utility: Care recommendation system identifies specific preventive gaps
Bala Mahendra Pothabathula
M.S. Computer Science | University of South Florida
LinkedIn | Email
Collaborators:
- Karthikeya Moturi
- Gowtham Sai Chimmana
This research was conducted as part of CIS 4930/CAI 5155 (Network Analysis & Machine Learning with Graphs) at the University of South Florida, Department of Computer Science.
This project is for educational and research purposes. Please cite appropriately if referencing this work.
- Phase 1: Network Analysis - Critical infrastructure identification
- Phase 2: GNN Prediction - Deep learning risk prediction
For questions, collaborations, or further information about this research:
Email: pothabathula@usf.edu
LinkedIn: [linkedin.com/in/bala-mp]
GitHub: @Mahendra-logics
⭐ If you find this work useful, please consider starring this repository!
---
## 📝 **How to Add This Main README**
### **Step 1: Go to your main repository**
https://github.com/Mahendra-logics/healthcare-network-analysis
### **Step 2: Create new README file**
Click **"Add file"** → **"Create new file"**
### **Step 3: Name the file**
In the "Name your file..." box, type:
README.md