A Binary Cascade Framework for Gut Microbiome Classification in Inflammatory Bowel Disease: Few-Shot LLMs, Co-Abundance Graphs, and Clinical Narrative Encoding

Subtyping inflammatory bowel disease (IBD) from gut microbiome sequencing data is a clinically demanding problem. Cross-cohort variability is large, and models trained on one dataset often fail on another. In this work, we propose a binary cascade pipeline that splits the three-class problem —healthy control, Crohn’s disease (CD), and ulcerative colitis (UC) — into two independent binary classifiers optimized at each stage separately. A key component of the pipeline is a clinical narrative encoding strategy (P4) that converts centered log-ratio (CLR) microbial abundances into structured biomedical text, allowing a large language model (LLM) to apply its pre-existing biomedical knowledge without any fine-tuning. We evaluate Gemini-2.5-Flash under two encoding schemes, six few-shot sizes (k ∈ {2, 4, 6, 8, 10, 14}), and two example-selection strategies across three random seeds. The pipeline is benchmarked against a co-abundance graph neural network (GNN) and five supervised classifiers on an internal holdout set and two external cohorts: Dataset 30 (n = 30) and IBD423 (n = 423). P4 encoding achieves Macro-F1 of 0.539 ± 0.024 on Control vs. IBD and 0.585 ± 0.011 on CD vs. UC, without any gradient updates. On external CD vs. UC validation, Gemini reaches F1 = 0.873, matching the best supervised baseline while outperforming all tree-based and graph-based methods. Centroid-based few-shot selection outperforms random selection by 5.2 percentage points. The results indicate that grounding microbial measurements in clinical language is a practical route to interpretable, training-free IBD classification.

@article{Najih2026,
  title     = {A Binary Cascade Framework for Gut Microbiome Classification in Inflammatory Bowel Disease: Few-Shot LLMs, Co-Abundance Graphs, and Clinical Narrative Encoding},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Nouhaila En Najih and Ahmed Moussa},
  doi       = {10.14569/IJACSA.2026.0170785},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170785}
}