Multi-Stream Adaptive Feature Integration for Microbiome-Based Multi-Disease Classification

Abstract

Microbiome-based classification of multiple neurological conditions is challenged by sparse features, class imbalance, and heterogeneity between studies. We propose Multi-Stream Adaptive Feature Integration (MASFI), which combines CLR-transformed genus-level features, taxonomy-derived features, and self-supervised taxon-token Transformer representations in a Random Forest classifier with out-of-bag dynamic uncertainty sampling. We evaluated MASFI using 4049 faecal 16S rRNA gene amplicon sequencing samples from participants with Alzheimer’s disease (AD), autism spectrum disorder (ASD), Parkinson’s disease (PD), and healthy controls. Five-fold stratified cross-validation compared MASFI with conventional machine learning and supervised neural baselines. Ablation analyses assessed the contributions of the feature streams, while block- and taxon-level permutation analyses characterised the features used by the fitted classifier. Genera including Romboutsia, Prevotella, and Fusicatenibacter were highlighted as contributors to model predictions. These findings support multi-stream feature integration as an approach to microbiome-based multi-disease classification.