Akademik Veri Yönetim Sistemi
BIOMEDICAL SIGNAL PROCESSING AND CONTROL, cilt.110, 2025 (SCI-Expanded)
Recent advancements in gut microbiome research have highlighted its significant role in health and disease, underscoring its potential as a biomarker for various chronic conditions. This study presents a novel deep learning-based wellness scoring model designed to assess health status from gut microbiome profiles. The study utilized data from 4,015 samples from a diverse population, encompassing nine disease groups, and an external validation set of 785 samples from publicly available datasets. Our deep learning approach involved two stages: unsupervised representation learning using autoencoders with transformer layers, followed by supervised finetuning for health classification. The model was rigorously validated through 10-fold cross-validation, achieving 85.74% balanced accuracy and 93.38% AUC, as well as through Leave-One-Disease-Out (LODO) validation and external validation, consistently outperforming the Gut Microbiome Wellness Index 2 (GMWI2) in all key metrics. Our findings indicate that the deep learning-based wellness score provides a robust and generalizable measure of health status, effectively capturing the intricate data structures within the gut microbiome. In conclusion, the use of a nationwide cohort that was created under a single protocol mitigated batch effects and biases, enhancing the reliability of the model. While not intended as a standalone diagnostic tool, our wellness scoring model serves as a valuable non-invasive approach for health monitoring and pre-diagnostic check-ups, contributing to personalized and preventive healthcare strategies. The proposed approach paves the way for microbiome foundation models, enabling transfer, zero-shot and few-shot learning.