Varoščić, Nika (2026) Integrating Host Metadata with Gut Microbiome Abundance Data in Convolutional Neural Networks for Inflammatory Bowel Disease Prediction. Bachelor thesis, Data Science and Society (DSS).
|
PDF
BA5705606NVaroscic.pdf Download (1MB) | Preview |
Abstract
The human gut microbiome is closely tied to host health, and shifts in its composition accompany many diseases, including inflammatory bowel disease (IBD). Advances in sequencing and machine learning have made it feasible to predict diseases from microbiome profiles, yet such models are usually trained on microbial abundance alone, even though host metadata shapes both the microbiome and disease risk. Integrating demographic and lifestyle metadata is one way to improve these models, but it remains underexplored, particularly for deep learning. In this thesis, a convolutional neural network (CNN) pipeline was developed for IBD detection from American Gut Project profiles, comparing an image-only CNN, a metadata-augmented CNN, and metadata-only baselines (multilayer perceptron, random forest, and logistic regression) to assess the contribution of host metadata to predictive performance. The results indicate that inclusion of host metadata can substantially improve classification, raising held-out test AUC from .49–.64 to .76–.78, yet metadata alone performed as well as the combined model, with antibiotic history, age, and body mass index carrying most of the signal. Because these variables also define the healthy-control group, an analysis omitting the three variables showed their contribution to be partly circular. Host metadata is therefore a valuable but confounded component of microbiome-based prediction, underscoring the importance of careful cohort design.
| Item Type: | Thesis (Bachelor) |
|---|---|
| Name supervisor: | Haleem, N. |
| Date Deposited: | 10 Jun 2026 11:47 |
| Last Modified: | 10 Jun 2026 11:47 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/820 |
Actions (login required)
 |
View Item |