Recent advances in neural network extensions of Cox proportional hazards models—such as ResDeepSurv’s residual self-attention for non-linear risk functions[1], CBNN’s explicit handling of time-varying covariates[2], and SMAHP’s AFT-based mediation for high-dimensional exposures[5]—provide a powerful framework for longitudinal aging research. Yet no study has integrated these tools to test whether dynamic interactions between mitochondrial DNA (mtDNA) copy number and microglial activation states mediate mortality risk beyond the sum of their individual effects. We hypothesize that, in human aging cohorts, time-varying mtDNA copy number influences hazard of death through a non-linear, microglia-dependent pathway that is detectable as a significant attention-weighted interaction term in a middle-fusion deep survival model.

To test this, we will assemble a longitudinal multi-omic dataset (e.g., UK Biobank aging subset or Framingham Heart Study offspring cohort) with repeated measures of whole-genome sequencing (mtDNA copy number), plasma proteomics (inflammatory markers such as sTREM2, YKL-40), epigenomics (DNAm PhenoAge), and metabolomics over a 10‑year follow‑up. Using a middle‑fusion strategy[3], we will train an autoencoder to learn a shared latent representation from the omics layers, then feed this representation into a SAVAE‑Cox network[4] that incorporates attention mechanisms to capture time‑varying covariate interactions. The attention layers will output weighted scores for each omic feature at each time point, allowing us to isolate the contribution of the mtDNA × microglial activation interaction.

Our primary testable prediction is that the attention‑weighted product of mtDNA copy number (log‑scaled) and a microglial activation signature (e.g., weighted sum of sTREM2, CCL4, and CST7) will significantly improve model fit (likelihood ratio test, p < 0.01) and increase concordance index (ΔC‑index ≥ 0.02) compared to a baseline model containing only the main effects of these features and established risk factors (age, sex, smoking, cardiovascular disease burden). We will further validate the mediation pathway using SMAHP‑style AFT decomposition to estimate the indirect effect of mtDNA on mortality through the microglial signature, testing whether the indirect hazard ratio exceeds 1.0 with bootstrap confidence intervals that do not cross null.

Falsifiability is built in: if the interaction term fails to reach statistical significance after correcting for multiple omic features, or if the indirect mediation effect is null, the hypothesis is rejected. Additionally, we will perform a permutation test shuffling the temporal ordering of mtDNA and microglial measures; a true biological interaction should lose predictive power when temporally misaligned, whereas a spurious correlation would persist.

Mechanistically, we posit that declining mtDNA copy number elevates reactive oxygen species, triggering NLRP3 inflammasome activation in microglia, which in turn amplifies neuroinflammation and systemic cytokine release, creating a feed‑forward loop that accelerates frailty and mortality. The deep model’s attention weights are expected to highlight periods where mtDNA decline precedes spikes in microglial markers, providing temporal causality clues that pure statistical models miss. By linking a concrete bioenergetic‑immune axis to survival predictions through interpretable attention, this work bridges the methodological gap highlighted in the literature and offers a testable, biologically grounded model of aging-related mortality risk.