Unravelling Disease Presentation Patterns in ALS Using Biclustering for Discriminative Meta-Features Discovery

Amyotrophic Lateral Sclerosis (ALS) is a heterogeneous neurodegenerative disease with a high variability of presentation patterns, impacting patient care and survival. Given the heterogeneous nature of ALS patients and targeting a better prognosis, clinicians usually estimate disease progression at diagnosis using the rate of decay computed from the Revised ALS Functional Rating Scale (ALSFRS-R). In this context, we aim at unravelling disease presentation patterns by proposing a new Biclustering-based approach, termed Discriminative Meta-features Discovery (DMD). These patterns (Meta-features) are composed of discriminative subsets of features together with their values, allowing them to distinguish and characterize subgroups of patients with similar disease presentation patterns. The proposed methodology was used to characterize groups of ALS patients with different progression rates (Slow, Neutral and Fast) using Biclustering-based Classification and Class Association Rule Mining. The patterns found for each of the three progression groups (described either as important features used by a Random Forest or as interpretable Association Rules) were validated by ALS expert clinicians, who were able to recognize relevant characteristics of slow, neutral and fast progressing patients. These results suggest that our general Biclustering approach is a promising way to unravel disease presentation patterns and can be applied to similar problems and other diseases.