Comparing Machine Learning and Deep Learning Approaches for ALS Diagnosis Using Blood-Derived Biomarkers and Synthetic Data

Amyotrophic lateral sclerosis (ALS) is a relentlessly progressive neurodegenerative disease characterized by the loss of upper and lower motor neurons. Recent advances suggest that blood-derived biomarkers (BDBs) could enable earlier diagnosis and objective monitoring of disease progression. In this work, we assess the ability of both classical machine learning and deep neural networks to distinguish ALS patients from healthy controls (HC) using BDB measurements. To overcome the twin obstacles of small sample size and class imbalance in clinical datasets, we synthetically expanded our data by generating two balanced cohorts: one 10-fold and one 20-fold larger than half of the original dataset. Models were trained exclusively on these synthetic cohorts and then tested on the remaining 50% of the real data (60 records). Among all approaches, AdaBoost achieved the highest performance, with up to 96.7% accuracy, 0.9792 F1-score, and 0.9944 AUROC on the 20-fold set. The LGBMClassifier also performed strongly, particularly in precision and specificity (both 100%), while the deep learning model showed more variability between folds. Importantly, synthetic data preserves patient privacy while still enabling effective model development and evaluation. These findings demonstrate that carefully generated synthetic data can powerfully augment limited clinical datasets and pave the way for robust, blood-based computational diagnostics in ALS.