Qureshi, A.; Wahid, A.; Qazi, S.; Shahzad, M. K. K.

Longitudinal biomedical datasets, particularly high-resolution microbiome profiles, present unique and complex challenges arising from their extreme sparsity, prevalence of zero inflation, and inherently dynamic, nonstationary behavior. Conventional Recurrent Neural Networks (RNNs) frequently encounter difficulties in these contexts, as they are unable to reliably differentiate between structural zeros and sampling zeros. To address these limitations, we introduce DynaBiomeX, a methodological framework specifically developed to model temporal dependencies within sparse biomedical timeseries data, while maintaining clinical interpretability. The DynaBiomeX framework integrates three parallel architectures BiDirectional Long Short Term Memory (BiLSTM), Attention based Gated Recurrent Unit (GRU), and an adapted Temporal Fusion Transformer (TFT) within a unified dual strategy Clinical Decision Support (CDS) workflow. These models are combined using stacking ensembles, which serve to maximize screening sensitivity by optimizing collective decision boundaries and thereby reducing the likelihood of missed cases. Concurrently, the adapted TFT functions as a specialized Sentinel, enhanced with Gated Residual Networks (GRN), to actively filter out stochastic noise and robustly validate predictions. This design helps the model separate real biological signals from random sampling noise. We tested on a multi modal dataset from 1,871 hematopoietic cell transplantation (HCT) patients to detect gut dysbiosis. Our benchmarks comparisons reveal that the model architectures exhibit distinct functional differences. Stacking ensembles achieved the best discriminative performance ROCAUC = 0.912, making them well suited for risk screening. In contrast, the noise filtering TFT reached perfect precision with no false positives Precision = 1.0 and strong reliability (MCC = 0.646; p<0.001). Ablation studies show that the model can find hidden microbial patterns even without clinical covariates ROCAUC > 0.81. DynaBiomeX couples sensitive ensemble screening with precise transformer validation to robustly analyze sparse longitudinal data. Validated on microbiome dysbiosis, this framework offers a template for zero inflated domains like single cell sequencing and EHR monitoring.