An Efficient Machine Learning Framework for Disease Gene Prediction in Parkinson’s Disease and Bladder Cancer

Abstract

Machine learning (ML) has been increasingly used in disease prediction, leveraging both phenotype and genotype data. However, genotype data have received comparatively less attention due to limited availability, whereas phenotype data have been more extensively studied. While breast cancer research is abundant, studies on other cancers, such as bladder cancer, and neurological diseases like Parkinson’s disease, remain limited. High-dimensional datasets pose challenges, including lengthy processing times, overfitting, an excess of features, and difficulties in classification. This study introduces a framework that integrates phenotype and genotype data for cancer prediction, aiming for high accuracy with a minimal number of relevant features. The framework consists of three main procedures: feature selection (FS), cancer prediction (CP), and identification of cancer-associated genes/features (CAG/F). FS employs a hybrid LEDF approach, combining the empirical distribution function (EDF) with three embedded methods: lasso regression selection (LRS), ridge regression selection (RRS), and random forest selection (RFS). EDF acts as a resampling tool with external (EEDF) and internal (IEDF) components that merge as E/IEDF. Features are selected based on classification accuracy using both union and intersection methods. CP applies multiple ML models with cross-validation to enhance prediction accuracy. Lastly, CAG/F identifies cancer-associated genes/features following the FS and CP steps. The algorithms E/IEDF-RFS, E/IEDF-LRS, and E/IEDF-RRS demonstrated excellent performance for RNA gene and dermatology datasets, achieving 100\% accuracy. E/IEDF-RFS reached 94.58\% accuracy for Parkinson’s Disease2, while EEDF-LRS performed best for DNA data with 94.85\% accuracy. E/IEDF-RRS showed 96.43\% accuracy for Parkinson’s Disease1 using RF classifiers, and IED-RFS and E/IEDF-LRS achieved 98.42\% accuracy for the BreastEw dataset.