BiLSTM-based Predictive Modeling for Diabetes Mellitus

📌 Project Overview This repository contains a modularized deep learning pipeline for predicting diabetes risk using medical features. The core of this project explores the performance of Bidirectional Long Short-Term Memory (BiLSTM) networks on structured medical datasets. While BiLSTMs are traditionally used for sequential data, this research evaluates their ability to capture complex feature relationships in tabular data through high-dimensional mapping and aggressive regularization.

🚀 Engineering Highlights Unlike standard academic scripts, this project is built for robustness and reproducibility: Leakage-Free Pipeline: Data splitting is performed prior to scaling and resampling (SMOTEENN) to ensure zero data contamination. Feature Engineering: Implemented Log Transformations for skewed features (Insulin/Pedigree) and created medical risk-factor interactions to improve signal-to-noise ratio. Advanced Regularization: Utilized L2 Kernel Regularization, Activity Regularization, and Batch Normalization to mitigate the generalization gap inherent in training Deep Learning models on small tabular datasets. Modular Architecture: Separated logic into preprocessing, model_definition, and execution modules for enterprise-grade maintainability.

📊 Performance Benchmarks Note: These results represent the final stabilized model after controlling for overfitting. Metric Training Score Testing Score Accuracy 84.08% 72.08% Recall 90.15% 87.04% Precision 81.20% 58.02% F1-Score 85.44% 69.63%

Research Note: The observation of an ~12% generalization gap is documented as a characteristic of BiLSTM architectures when applied to non-sequential, small-scale tabular data. This project prioritizes Generalization and Honesty over inflated leaked accuracy scores.

🛠 Tech Stack Deep Learning: TensorFlow 2.x, Keras Data Science: Pandas, NumPy, Scikit-learn Resampling: Imbalanced-learn (SMOTEENN) Environment: Python 3.9+

📂 Repository Structure code Text ├── data/ # Contains diabetes.csv ├── src/ │ ├── preprocessing.py # Log transforms, SMOTEENN, and Leakage-free scaling │ └── model.py # BiLSTM architecture with L2/Dropout ├── main.py # Orchestration script with EarlyStopping ├── requirements.txt # Dependency list └── README.md

⚙️ Installation & Usage Clone the repo: code Bash git clone https://github.com/AYESHAASS/Diabetes-Prediction-BiLSTM.git cd Diabetes-Prediction-BiLSTM

Install Dependencies: code Bash pip install -r requirements.txt Execute Pipeline: code Bash python main.py