Research Article
Machine Learning-Based ALS Diagnosis Using a Multi-stage Gene Selection on Gene Expression Data
@INPROCEEDINGS{10.1007/978-3-032-44094-5_13, author={Dinh-Quan Dang and Hai-Long Nguyen and Hai-Chau Le}, title={Machine Learning-Based ALS Diagnosis Using a Multi-stage Gene Selection on Gene Expression Data}, proceedings={Industrial Networks and Intelligent Systems. 12th EAI International Conference, INISCOM 2026, Da Nang, Vietnam, February 26--27, 2026, Proceedings}, proceedings_a={INISCOM}, year={2026}, month={11}, keywords={Machine Learning Gene expression Gene Selection Sequential Forward Feature Selection Amyotrophic Lateral Sclerosis (ALS)}, doi={10.1007/978-3-032-44094-5_13} }- Dinh-Quan Dang
Hai-Long Nguyen
Hai-Chau Le
Year: 2026
Machine Learning-Based ALS Diagnosis Using a Multi-stage Gene Selection on Gene Expression Data
INISCOM
Springer
DOI: 10.1007/978-3-032-44094-5_13
Abstract
Amyotrophic Lateral Sclerosis (ALS) presents a significant diagnostic challenge due to its complex etiology and lack of definitive biomarkers. This study introduces a structured machine learning framework to identify a robust and minimal gene signature from blood-based gene expression data for accurate diagnosis of ALS. We propose a novel multi-stage feature selection pipeline that sequentially integrates Differential Expression analysis (DEGs), Minimum Redundancy Maximum Relevance (mRMR) criterion, and the Max-Min Parents and Children (MMPC) constraint-based algorithm. This process distills the high dimensional transcriptomic data (21,029 genes) into a core set of 27 biologically relevant candidate genes. Subsequently, a Sequential Forward Feature Selection (SFFS) wrapper, evaluated via 3-fold cross-validation, is employed to determine the optimal, classifier-specific gene subset for seven distinct machine learning algorithms (AdaBoost, Decision Tree, k-Nearest Neighbors, Logistic Regression, Random Forest, Support Vector Machine, and XGBoost). Our methodology identified the Support Vector Machine (SVM) paired with a parsimonious gene signature as the optimal diagnostic model. It achieved an Area Under the Curve (AUC) of 87.63% and a sensitivity of 86.55% on an independent test set, substantially outperforming existing benchmarks.