README

Introduction

The main goal of this project is to analyze the relationship between medullary thyroid carcinoma (MTC) and associated gene/clinical data using machine learning. We employ a RandomForestClassifier to classify and evaluate models on various test datasets, and we use SHAP to interpret feature importance.

Core Workflow

1. Data Loading: Load gene features and clinical data from multiple Excel files.
2. Data Preprocessing: Fill missing values, convert categorical variables (e.g., tumor_grade, sex, Heredity, T, N, M, TNM stage) into numerical codes.
3. Feature Selection: Use RandomForestClassifier together with SelectFromModel to identify the most discriminative features.
4. Model Training & Evaluation:
   • Run K-fold cross-validation to choose the best model on the training set.
   • Evaluate on independent test sets to output performance metrics (precision, recall, confusion matrix, AUC, etc.).
5. SHAP Interpretability: Compute SHAP values to evaluate the contribution and impact of each feature on model predictions.

---

Project Structure

your_project/
│
├─ main.py         # Python script containing the core logic
└─ README.md

Note: The actual names of the Excel files depend on your local files.
The script assumes they reside in the data/ folder.

---

Requirements

• Python 3.7+
• Key Libraries:
  • pandas
  • numpy
  • scikit-learn
  • shap
  • argparse

Install the dependencies via pip:

pip install pandas numpy scikit-learn shap argparse

---

---

Usage

1. Clone or Download
   Clone or download this project to your local environment. Ensure the data/ directory contains the required Excel files.

2. Install Dependencies

   pip install -r requirements.txt

   Or manually:

   pip install pandas numpy scikit-learn shap argparse

3. Run the Script
   From the terminal, navigate to the project folder and execute:

   python main.py --seed 42

   Argument details:

   • --seed: Random seed (default is 10).

4. Output
   The script outputs:

   • Cross-validation scores and the best validation score.
   • Test set performance metrics (precision, recall, confusion matrix, AUC).
   • Prediction results (predicted labels, probabilities, true labels).
   • Selected important features.
   • SHAP values for feature interpretation.

---

Key Functions

• load_data()
  Loads data from multiple Excel files into dictionaries for further processing.

• print_metrics(y_true, y_pred, y_pred_proba)
  Calculates and displays classification metrics including precision, recall, confusion matrix, and AUC.

• print_metrics_conf_matrix(y_test, y_pred, y_pred_proba)
  Similar to print_metrics but emphasizes the confusion matrix.

• train(gene_discovery_data, gene_test1_data, gene_test2_data)
  Trains a RandomForestClassifier with K-fold cross-validation, selects the best model, and evaluates on test datasets.

• Clinical_train_with_slectfeature(gene_discovery_data, gene_test1_data, gene_test2_data, max_features=10, random_state=1)
  Performs feature selection, model training, evaluation, and SHAP-based feature importance analysis.

---

Interpreting Results

• Cross-validation

  • Outputs accuracy scores for each fold in K-fold CV.
  • Selects the model with the best overall cross-validation accuracy.

• Test Set Performance
  Metrics include precision, recall, F1 Score, confusion matrix, and AUC.

• Feature Importance
  Evaluates feature importance using both RandomForest’s built-in feature importances and SHAP values. Higher SHAP values indicate more influential features.

---

Frequently Asked Questions (FAQ)

• Excel File Names or Versions Differ
  Update file paths and sheet names in load_data() accordingly.

• Library Compatibility Issues
  Use virtual environments (e.g., conda, venv) to resolve potential dependency conflicts.

• Missing Data Issues
  Consider alternative strategies (e.g., median, mean, interpolation) if current filling strategy is unsuitable.

• Categorical Variable Encoding
  Update the dictionary r_dict in the script if category mappings need to be modified.

---

Contributing & Extensions

You are encouraged to modify the code by trying different models, feature selection strategies, or visualizations. Contributions, suggestions, and improvements are welcome via issues or pull requests.

---

License

You can release this project under your laboratory’s or company’s preferred license. If not specified, consider using the MIT License.

---

We hope this README helps you and others easily understand and use the project. For further questions or clarifications, please reach out.