IDENTIFICATION OF BIOMARKERS ASSOCIATED TO LIVER COMPLICATIONS IN TYROSINEMIC TYPE - 1 PATIENTS: A MULTICENTRIC MACHINE LEARNING APPROACH.

This project focuses on developing a model that generalizes the biochemical phenotype of (TYROSINEMIC TYPE - 1) patients with altered alpha-phetoprotein. The code comprises preprocessing, exploratory and predictive analysis of health data for two cohorts from Chile and Italy. The extracted data undergoes a series of operations, such as loading into a Pandas DataFrame, filtering, and saving, for subsequent usage. Furthermore, the project leverages a machine learning pipeline powered by XGBoost, Optuna, SHAP, and Ray libraries. The datasets showed predictive capacity enabling models to generalize the biochemical phenotype of type-1 tyrosinemia. Also, characterizing features in terms of their importance for predictions revealed different rankings depending on which test set was used.

Project Structure

The project comprises two primary parts:

Part 1: Data Preprocessing and Extraction

The src/main.py script hosts the operations for data extraction, processing, and saving.

1. Logging Configuration: Set up logging to record significant script execution events in the workflow.log file.
2. Function Definitions: Define sheet_to_dataframe and compare_dataframes for data extraction and comparison, respectively.
3. Data Extraction: Establish a connection to Google Sheets using the gspread package, extract datasets, and load them into DataFrames.
4. Data Filtering: After loading, select predefined features from each DataFrame and drop rows missing certain features.
5. Data Saving: Save the final cleaned DataFrames as CSV files in the data/ directory.

Part 2: XGBoost, Optuna and Ray-based Machine Learning Pipeline

This section incorporates Python classes and functions that streamline Machine Learning pipelines, with an emphasis on the XGBoost, Optuna, and Ray libraries.

File Structure

• src/main.py: Main Python script.
• credentials/project-30463-38031804e4b0.json: Google Sheets service account credentials.
• data/: Directory for processed data saved as CSV files.

Key Components

Libraries Used

• ray: For distributed and parallel computing.
• optuna: A library for hyperparameter optimization.
• xgboost: A gradient boosting library providing a robust framework for constructing predictive models.
• shap: A library that uses Shapley values to interpret any machine learning model's output.

Classes and Functions

• DataImputer: Imputes missing numerical values in a DataFrame.
• DataSplitter: Performs stratified splitting on a DataFrame.
• ModelInstance: Instantiates, trains, and evaluates a model.
• objective(): Optimizes a model's performance and a given feature's importance.
• make_a_study(): A Ray remote function that fine-tunes hyperparameters.
• make_multiple_studies(): Runs multiple optuna studies for various features and targets.
• launch_to_ray(): Applies the entire pipeline, from data preprocessing to hyperparameter optimization.