An Interpretable Hybrid 3-Tier LSTM Model for Breast Tumor Classification: Balancing Accuracy and Transparency in Clinical Decision Support

This repository contains a reproducible, modular pipeline for tumor classification on the Breast Cancer dataset using a Hybrid 3‑Tier LSTM model. The pipeline includes:

• Clean data loading and preprocessing (drop id and Unnamed: 32 if present)
• Encode target (diagnosis: Malignant M = 1, Benign B = 0)
• Train/test split with optional stratification
• SMOTE on training set only
• Decision Tree feature importance (with figure)
• 3‑Tier LSTM (Dense → LSTM → Dense head) with input reshaped to (samples, timesteps=1, features)
• Full evaluation (Accuracy, Precision, Recall, F1, AUC, Confusion Matrix, ROC curve)
• Explainability: SHAP and LIME (with actual feature names)
• Reproducible configs via config.yaml
• One‑command run via python main.py

Note: You need Python 3.10+ and a working TensorFlow environment. GPU is recommended but not required.

---

Quickstart

# (optional) create venv
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

pip install --upgrade pip
pip install -r requirements.txt

# Run the full pipeline
python main.py

Use the ready conda environment

conda env create -f environment.yml conda activate hybrid_lstm python main.py

All outputs (models, figures, reports) will be saved under outputs/.

---

Project Structure

.
├── config.yaml
├── main.py
├── requirements.txt
├── .gitignore
├── data/
│   └── Cancer_Data.xls
├── src/
│   ├── load_data.py
│   ├── preprocess.py
│   ├── evaluate.py
│   ├── explain.py
│   ├── plots.py
│   └── models/
│       └── lstm_model.py
└── outputs/
    ├── figures/
    ├── models/
    └── reports/

---

Reproducibility Notes

• SMOTE is applied only on the training split to avoid leakage.
• Feature scaling is fitted on train and applied to both train/test.
• Feature names are preserved throughout for correct plotting and explanations.
• SHAP uses a KernelExplainer on a small background sample to keep runtime reasonable on CPUs.
• LIME uses LimeTabularExplainer with consistent feature names and classes.

---

Data set source

-https://www.kaggle.com/datasets/uciml/breast-cancer-wisconsin-data

Cite this article

Nti, I.K., Nyarko-Boateng, O., Zaman, A. et al. Interpretable hybrid three tier LSTM model for accurate and transparent breast tumor classification in clinical decision support. Discov Health Systems 5, 2 (2026). https://doi.org/10.1007/s44250-025-00315-6