DynaBiome: Interpretable Unsupervised Learning of Gut Microbiome Dysbiosis using Temporal Deep Models

Purpose: Gut microbiome dysbiosis is a contributing factor to various diseases and a critical determinant for autologous fecal microbiota transplantation (Auto-FMT) eligibility assessment. Current dysbiosis classification approaches rely predominantly on supervised learning with manually annotated labels, single-time-point analysis, and black-box models lacking clinical interpretability. This study proposes an unsupervised, explainable framework, DynaBiome, to predict gut dysbiosis states for Auto-FMT eligibility determination. Methods: The framework employs an LSTM autoencoder architecture with integrated sequential layers that capture temporal microbiome dynamics in 14-day windows. The model reconstructs normal microbiome patterns, with high reconstruction errors that indicate possible dysbiotic sequences. SHAP-based interpretability identifies contributing genera at specific time points. Ensemble learning methods are applied to traditional classifiers trained on reconstruction error features. Results: The initial LSTM autoencoder achieved high dysbiotic sensitivity (99% recall) but exhibited over-detection with low non-dysbiotic recall (38%). Threshold optimization and ensemble learning significantly improved classification balance. Logistic Regression demonstrated optimal performance (ROC AUC 0.7976). The Averaged Probabilities Ensemble achieved best generalization (ROC AUC 0.7759), demonstrating 6.9% improvement over Isolation Forest while achieving 95.7% of supervised baseline performance. Conclusion: Integrating unsupervised temporal feature extraction with supervised ensemble methods provides clinically robust and interpretable dysbiosis prediction, overcoming limitations of single time-point approaches while eliminating the need for labelled data.