🏥 Bridging the Generalizability Gap in Diabetes Prediction

A Noise-Resilient Hybrid Imputation-Ensemble (NR-HIE) Framework

"AI models shouldn't just predict; they must adapt to the messiness of real-world clinical data."

---

📌 Project Overview

In medical diagnostics, Missing Not At Random (MNAR) data is a critical challenge. Standard AI models often fail when deployed in real-world settings because they learn from "clean" data but face "messy" patient records.

This project introduces the NR-HIE Framework, a novel architecture designed to bridge the "Generalizability Gap." By combining a Triple-Stream Imputation strategy with Stacked Generalization, we achieved robust performance on completely unseen external populations.

---

🚀 Key Features

Feature	Description
🧠 Hybrid Imputation	A triple-stream approach combining MICE (Global correlations), KNN (Local clusters), and Robust Statistics (Outlier handling).
⚙️ Stacking Ensemble	A Meta-Learner utilizes predictions from XGBoost, Random Forest, and Logistic Regression to minimize variance.
🌍 Generalizability	Validated on an external Bangladeshi Cohort (DiaBD), achieving an accuracy surge instead of a drop.
🔍 Explainable AI	Integrated SHAP (Shapley Additive exPlanations) to provide transparent, biologically valid reasoning for every prediction.

---

📊 Results: The "Generalizability Surge"

Most models degrade on external data. Ours improved.

Metric	Baseline Model (Standard)	NR-HIE (Internal Test)	NR-HIE (External Validation)
Accuracy	75.32%	77.27%	81.62% 🚀
Brier Score	0.1302	0.1273 (High Safety)	-
ROC-AUC	0.74	0.83	0.76

The drop in F1-score on external data is attributed to a prevalence shift, yet the high ROC-AUC confirms the model's robust ranking ability.

---

🛠️ Methodology & Architecture

1. The Triple-Stream Pipeline

Instead of relying on a single imputation method, we generate three "views" of the data:

1. MICE Stream: Captures multivariate relationships.
2. KNN Stream: Preserves local data structure.
3. Robust Stream: Preserves the signal of "missingness."

2. Stacked Generalization

The Level-1 Meta-Learner (Logistic Regression with L2 Regularization) dynamically weights these streams based on their predictive reliability.

---

📉 Explainability (SHAP Analysis)

We verified that the model relies on medically relevant biomarkers rather than noise.

• Top Predictors: Glucose, BMI, Age (Biologically Valid).
• Noise Features: Skin Thickness (Low Impact).

---

📦 Installation & Usage

Prerequisites

• Python 3.8+
• Jupyter Notebook

Setup

1. Clone the repository:

   git clone [https://github.com/hasnainalam596/NR-HIE-Generalizable-Diabetes-Prediction-Framework.git](https://github.com/hasnainalam596/NR-HIE-Generalizable-Diabetes-Prediction-Framework.git)

2. Install dependencies:

pip install -r requirements.txt

3. Run the Notebook:

jupyter notebook notebooks/01_Diabetes_Prediction_NR_HIE.ipynb

📂 Project Structure

├── data/                   # Datasets (Pima & DiaBD)
│   ├── diabetes.csv
│   └── diaBD_bangladesh.csv
│
├── notebooks/              # Jupyter Notebooks (Source Code)
│   ├── Project_Notebook.ipynb
│
├── reports/                # Documentation
│   ├── Project_Report.pdf
│   └── Project_Presentation.pptx
│   └── Project_Proposal.pdf
│
├── images/                 # Charts for README
│   ├── architecture_diagram.png
│   └── shap_summary.png
│
├── requirements.txt        # Libraries List
└── README.md               # Project Details

📜 Future Roadmap

**Demographic Expansion: Retrain on male datasets to mitigate gender bias.

Longitudinal Analysis: Integrate LSTMs for time-series risk forecasting.

Deployment: Develop a Streamlit/Flask web app for doctors.

🤝 Contact & Contribution

Hasnain Alam

Role: Data Scientist & Researcher

GitHub: (https://github.com/hasnainalam596)

Email: alamhasnain457@gmail.com

If you find this research useful, please star ⭐ this repository!