Dyslipidemia Prediction

In this repository, we utilize machine learning and deep learning techniques to predict Apolipoprotein C-III (ApoC-III), which is the measurement affecting indirectly blood lipid disease. Concretely, TG and LDL-C are two main factors causing Dyslipidemia by clinging to the vessel wall obstructs blood circulation. Lipoprotein lipase (LPL) is an enzyme in other to hydrolysis TG while ApoC-III inhibits LPL. They operate and regulate the body continuously. Decreasing ApoC-III makes reducing TG and LDL-C so helps reduce subclinical atherosclerosis in the body. Conversely, it is harmful to the body by increasing TG in the blood. We incorporate Single-nucleotide polymorphism (SNP) of patients with treatments that are drugs that help reduce ApoC-III. Compatible metrics for each person are different. Therefore, we select the treatment which has the minimum ApoC-III score will be proposed as the main treatment method for different patients.

Contents

1. Introduction
2. Prerequisites
3. Install
4. Getting started
5. Repo structure
6. Results

1. Introduction

Desciption of Dataset

• The dataset is collected 384 SNP which are extracted from 978 different patients.
• Each SNP contains one of four genotypes AA, AB, BB, NC (allen B is superior to allen A and NC is represented for unknown).
• There are three kinds of treatments: 𝑓𝑒𝑛𝑜𝑓𝑖𝑏𝑟𝑖𝑐 𝑎𝑐𝑖𝑑, 𝑓𝑒𝑛𝑜𝑓𝑖𝑏𝑟𝑖𝑐 𝑎𝑐𝑖𝑑 𝑠𝑡𝑎𝑡𝑖𝑛, 𝑠𝑡𝑎𝑡𝑖𝑛 𝑎𝑙𝑜𝑛𝑒.
• Prediction value outcome is a real value which is represented for ApoC-III score.

Models

Regression models used in this project as below:

• Machine Learning Models: Linear Regression, Regularization L2 Linear Regression, Regularization L1 Linear Regression, ElasticNet, MultiTask ElasticNet, Support Vector Regression, Decision Tree Regression, XGBoost, AdaBoost, ExtraTreeRegressor, GradientBoostingRegressor, RandomForest, LightGBM.
• Deep Learning Models: Neural Network, Convolutional Neural Neural.

2. Prerequisites

• Python >= 3.6
• PyTorch >= 1.7
• Sklearn >= 0.23
• Other dependencies described in requirements.txt

3. Install

Create Conda virtual environment:

conda create -n dyslipidemia python=3.8 -y
conda activate dyslipidemia
pip install -r requirements.txt

4. Getting started

Run all models on public dataset

python main.py all --mode public

There are several control parameters:

• model: Run mlmodel or dlmodel or all [mlmodel, dlmodel, all]
• mode: Choose dataset [public, private]

(And more parameters are configed in args)

5. Repo structure

• dataset: Dataset used in this repository
• lib
  • models:
    • abstract_model.py: Abstract class for models
    • cnn_regression.py: Convolutional Neural Network model
    • dl_model.py: Run deep learning model
    • ml_model.py: Run machine learning model
    • nn_regression.py: Neural Network model
  • config.py: Configuration
  • runner.py: Executes all models by mode
• utils
  • load_data.py: Loads dataset from dataset folder and transform by mode.
• logging: Contain log file folder
• plot_EDA.ipynb: Explore Data Analysis and Plot char.
• main.py: Config and run the runner.

6. Results

The result of public and private datasets shows in the table below:

Machine Learning Models

Model Name	Split [80-20] (public)	K-fold (public)	Split [80-20] (private)	K-fold (private)
Linear Regression	25.045	24.672	25.503	29.723
Regularization L2 Linear Regression	24.962	24.591	25.403	29.552
Regularization L1 Linear Regression	23.791	23.421	23.581	27.320
ElasticNet	20.585	19.963	16.773	16.748
MultiTask ElasticNet	22.425	22.064	21.207	24.094
Support Vector Regression	20.948	20.141	17.234	17.208
Decision Tree Regression	27.862	29.044	46.827	43.348
XGBoost	23.142	21.558	25.630	24.260
AdaBoost	21.265	20.337	16.922	17.409
ExtraTreeRegressor	21.223	20.223	19.532	19.603
GradientBoostingRegressor	22.202	21.393	25.986	27.510
RandomForest	20.991	20.490	20.534	19.927
LightGBM	22.138	21.184	18.057	18.676

Deep Learning Models

Model Name	Split [80-20] (public)	Split [80-20] (private)
Neural Network	22.181	19.254
Convolutional Neural Neural	22.363	22.207