DynaBiomeX: An Interpretable Dual-Strategy Deep Learning Framework for Architectural Noise Filtration in Sparse Longitudinal Microbiome Data

Abstract

ObjectiveLongitudinal microbiome datasets present unique challenges due to extreme sparsity, zero-inflation, and non-stationary behavior. Conventional Recurrent Neural Networks (RNNs) struggle to distinguish structural from sampling zeros in these contexts, limiting their utility for Clinical Decision Support (CDS). MethodsWe introduce DynaBiomeX, an interpretable framework specifically developed for sparse biomedical time-series. It integrates Stacking Ensembles (Bi-LSTM, GRU) with an adapted Temporal Fusion Transformer (TFT) in a unified Screener-Sentinel workflow. The Ensembles optimize collective decision boundaries to maximize sensitivity and minimize missed cases. Concurrently, the TFT functions as a Physiological Gatekeeper, utilizing Gated Residual Networks (GRN) to actively filter stochastic noise from real biological signals. We validated this approach on a multi-modal dataset of 1,871 hematopoietic cell transplantation (HCT) patients to detect gut dysbiosis. ResultsStacking ensembles maximized discriminative performance (ROC-AUC = 0.912), effectively serving as high-sensitivity screeners. In contrast, the Adapted TFT functioned as a precision Sentinel, achieving zero false positives (Precision = 1.0) and high stability (MCC = 0.646). Crucially, the TFT demonstrated superior probabilistic reliability with a low Expected Calibration Error (ECE = 0.0085), addressing the "black-box" overconfidence typical of deep learning models. Ablation studies confirmed predictive robustness even without clinical covariates (ROC-AUC > 0.81). ConclusionDynaBiomeX couples sensitive screening with precise, calibrated validation to robustly analyze sparse longitudinal data. Validated on microbiome dysbiosis, this framework offers a scalable template for zero-inflated domains like single-cell sequencing and EHR monitoring. HighlightsO_LIDynaBiomeX integrates Stacking Ensembles and Temporal Fusion Transformers for microbiome risk stratification. C_LIO_LIGated Residual Networks distinguish stochastic sampling zeros from structural zeros in sparse ASV data. C_LIO_LIA novel Screener-Sentinel workflow resolves the trade-off between surveillance sensitivity and diagnostic precision. C_LIO_LIClinical metadata functions as a physiological gatekeeper, filtering noise to achieve zero false positives (MCC 0.646). C_LIO_LIThe framework minimizes alarm fatigue in Auto-FMT decision support by validating latent dysbiosis signals. C_LI