Hypothesis

Integrating Disentangled Interpretable Multimodal Attention Fusion (DIMAF) with Deep Conditional Transformation Models (DCTMs) and median‑SHAP will reveal non‑proportional, time‑varying hazard pathways that link specific multi‑omic signatures (e.g., mitochondrial DNA methylation, circulating proteomic inflammation clusters) to exceptional longevity, outperforming standard Cox, joint models, and gradient‑boosting baselines on centenarian cohorts.

Mechanistic Rationale

DIMAF learns modality‑specific and shared latent representations via intra‑ and inter‑modal attention, isolating omics dimensions that drive survival signal while suppressing noise [2]. DCTMs extend the Cox framework by modeling the conditional distribution of survival times through flexible Bernstein polynomials, capturing non‑linear and non‑proportional hazard shapes that arise when biomarker trajectories interact with age‑dependent processes [1]. When DIMAF‑derived latent factors are fed as time‑varying covariates into a DCTM, the model can estimate how each disentangled omics module modulates the baseline hazard across the lifespan, a capability absent in current siloed approaches. Median‑SHAP anchors explanations to the median predicted survival, mitigating bias from right‑skewed censoring distributions common in longevity data [3][4].

A further mechanistic layer posits that mitochondrial epigenetic drift alters the dynamic coupling between oxidative‑stress proteomics and inflammasome activation, producing a time‑dependent hazard inflection point around the ninth decade. This inflection would appear as a change‑point in the Bernstein polynomial coefficients estimated by the DCTM, detectable only when the mitochondrial methylation latent DIMAF axis is included.

Testable Predictions

1. Predictive Performance: A DIMAF‑DCTM median‑SHAP pipeline will achieve a higher concordance index (C‑index ≥0.88) and integrated AUC (≥0.86) on a held‑out longevity cohort (e.g., New England Centenarian Study with multi‑omics and quarterly clinical measures) compared to:
   • Standard Cox with baseline omics
   • Extended Cox with time‑varying covariates
   • Joint models of longitudinal biomarkers and survival
   • Gradient‑boosting survival models (XGBoost, GBM) and Random Survival Forests
2. Non‑Proportional Hazard Detection: The estimated Bernstein polynomial coefficients will show a statistically significant change in slope at ~85‑90 years only when the mitochondrial methylation DIMAF latent variable is included (p < 0.01, likelihood‑ratio test).
3. Median‑SHAP Interpretation: Median‑SHAP values will identify the mitochondrial methylation axis and a specific inflammasome proteomic cluster as the top two shared latent features contributing to hazard reduction in long‑lived survivors, with SHAP values consistently directional across bootstrap resamples.
4. Simulation Validation: Survival times generated via the Lambert W function with continuous time‑varying covariates mimicking the identified omics‑hazard dynamics will be recovered by the DIMAF‑DCTM model with parameter bias <5%, confirming model identifiability.

Experimental Design

• Data: Multi‑omics (whole‑genome methylation, plasma proteomics, transcriptomics) and longitudinal clinical covariates (IL‑6, CRP, gait speed) from ≥1,200 individuals aged ≥80, with ≥5‑year follow‑up and known vital status.
• Preprocessing: Normalize each omics layer, apply DIMAF to obtain shared and modality‑specific latent factors (dimensionality = 20 per modality).
• Modeling: Fit a DCTM where the conditional transformation function depends on the DIMAF latent factors as time‑varying covariates; estimate via maximum likelihood with Bernstein polynomial order selected by cross‑validation.
• Interpretation: Compute median‑SHAP for each latent factor at decadal intervals; assess stability via 1,000 bootstrap samples.
• Comparison: Train baseline models on identical data; evaluate using time‑dependent C‑index, integrated Brier score, and calibration plots.
• Falsifiability: If the DIMAF‑DCTM does not surpass the best baseline by ≥0.03 in C‑index, or if no change‑point in hazard coefficients is detected around the ninth decade, the hypothesis is refuted.

Expected Impact

Successful validation would provide a mechanistically interpretable framework that bridges high‑dimensional omics with flexible hazard modeling, uncovering timing‑specific biological drivers of exceptional longevity and offering a template for other complex aging phenotypes.