ML-Therapy-Prediction

Machine Learning Pipeline for Clinical Therapy Scheme Prediction

---

Project Overview

ML-Therapy-Prediction is a modular machine learning project for predicting neoadjuvant and adjuvant therapy schemes based on clinical data.
The pipeline includes robust data cleaning, feature engineering, correlation-based reduction, single-label and multi-label model training, and evaluation.

Key capabilities:

• Automated data preprocessing, including missing value imputation and categorical encoding.
• Feature selection by removing highly correlated categorical variables using Cramér's V and numerical correlation using Pearson's coefficient.
• Support for both single therapy prediction and multi-label classification tasks.
• Baseline model benchmarking using LazyPredict.
• Custom pipelines with feature importance visualization and optional hyperparameter tuning.

---

Directory Structure

ML-Therapy-Prediction/
├── Single_Therapy_ML/
│   ├── main.py
│   ├── pipeline_utils.py
│   ├── preprocess.py        
│   └── correlation_visualizer.py
│
├── Multi_Therapy_ML/
│   ├── main.py
│   ├── correlation_visualizer.py
│   ├── pipeline_utils.py     
│   └── preprocess.py
|
└── README.md               # Project documentation (this file)

Requirements

• Python ≥ 3.8
• pandas
• numpy
• scikit-learn
• seaborn
• matplotlib
• xgboost
• scikit-multilearn
• lazypredict

Install all dependencies with: pip install -r requirements.txt

Usage

1. Data Preprocessing

Preprocess raw data and generate cleaned and encoded datasets.

This script will:

• Clean and standardize string fields
• Handle missing values with median or mode imputation
• Remove redundant categorical features based on Cramér’s V
• Remove redundant numerical features based on Pearson correlation

Outputs:

• Cleaned Excel file
• Encoded numerical dataset
• Cramér’s V and correlation matrix printed to console and visualized as a heatmap.

2. Single-Label Classification

Train and evaluate single-label models. cd Single_Therapy_ML python main.py This module:

• Runs LazyPredict for quick benchmarking
• Trains a chosen classifier (e.g., RandomForest, Bagging, ExtraTrees)
• Plots and exports feature importance
• Compares actual vs. predicted labels in an Excel report.

3. Multi-Label Classification

Train and evaluate multi-label models. cd ../Multi_Therapy_ML python main.py This module:

• Encodes multiple target columns using OneHotEncoder
• Supports MultiOutputClassifier, ClassifierChain, and LabelPowerset
• Benchmarks each approach for Subset Accuracy and Hamming Loss
• Plots and exports top features for interpretability
• Performs hyperparameter tuning with GridSearchCV.

Results

All results including:

• Cleaned datasets
• Feature importance ranking
• Prediction vs. actual comparison