Blood biomarker-based classification study for neurodegenerative diseases

Abstract

As the population ages, neurodegenerative diseases are becoming more prevalent, making it crucial to comprehend the underlying disease mechanisms and identify biomarkers to allow for early diagnosis and effective screening for clinical trials. Thanks to advancements in gene expression profiling, it is now possible to search for disease biomarkers on an unprecedented scale.Here we applied a selection of five machine learning (ML) approaches to identify blood-based biomarkers for Alzheimer's (AD) and Parkinson's disease (PD) with the application of multiple feature selection methods. Based on ROC AUC performance, one optimal random forest (RF) model was discovered for AD with 159 gene markers (ROC-AUC = 0.886), while one optimal RF model was discovered for PD (ROC-AUC = 0.743). Additionally, in comparison to traditional ML approaches, deep learning approaches were applied to evaluate their potential applications in future works. We demonstrated that convolutional neural networks perform consistently well across both the Alzheimer's (ROC AUC = 0.810) and Parkinson's (ROC AUC = 0.715) datasets, suggesting its potential in gene expression biomarker detection with increased tuning of their architecture.

Figure 1

Workflow for identification of blood biomarkers. Training and test datasets were standardized separately. Feature selection is applied to training data to generate feature sets of genes using different approaches. Each of these feature sets is used to train five different classification models to distinguish control and disease patients: linear regression (LR), support vector machine (SVM), XGBoost, random forest (RF), and multilayer perceptron (MLP). The feature set and classification model combinations are evaluated in test datasets. Additionally, VAE and convolutional neural network (CNN) models are trained and evaluated separately.

Figure 2

Evaluation scores for different numbers of genes selected using VSSRFE. VSSRFE identified a panel of 159 genes on AD data (A) and a panel of 5 genes on PD data (B) that gave the best prAUC, ROC-AUC, and accuracy scores.

Figure 3

ROC curves for each classification algorithm on PD data.

Figure 4

ROC curves for each classification algorithm on AD data.