[Paper Review] Age Predictors Through the Lens of Generalization, Bias Mitigation, and Interpretability: Reflections on Causal Implications

Debdas Paul, Elisa Ferrari+2 more|arXiv (Cornell University)|Mar 17, 2026

Genetic Associations and Epidemiology0 citations

TL;DR

The paper analyzes how invariant representations and adversarial bias-mitigation can improve out-of-distribution age prediction from transcriptomic data, clarifying limits on causal interpretation and demonstrating alignment with intervention effects.

ABSTRACT

Chronological age predictors often fail to achieve out-of-distribution (OOD) gen- eralization due to exogenous attributes such as race, gender, or tissue. Learning an invariant representation with respect to those attributes is therefore essential to improve OOD generalization and prevent overly optimistic results. In predic- tive settings, these attributes motivate bias mitigation; in causal analyses, they appear as confounders; and when protected, their suppression leads to fairness. We coherently explore these concepts with theoretical rigor and discuss the scope of an interpretable neural network model based on adversarial representation learning. Using publicly available mouse transcriptomic datasets, we illustrate the behavior of this model relative to conventional machine learning models. We observe that the outcome of this model is consistent with the predictive results of a published study demonstrating the effects of Elamipretide on mouse skeletal and cardiac muscle. We conclude by discussing the limitations of deriving causal interpretation from such purely predictive models.

Motivation & Objective

• Motivate robust chronological age prediction across heterogeneous environments ( tissues, cohorts, protocols) beyond standard ERM/SLR.
• Clarify the role of invariance, bias mitigation, and fairness in age prediction under distributional shifts.
• Propose and evaluate an adversarial representation learning framework that promotes domain-invariant features.
• Investigate the causal interpretation of predictive age models and the limitations therein.
• Illustrate framework behavior using publicly available mouse transcriptomic datasets and relate findings to an intervention study.

Proposed method

• Formulate age prediction under multiple environments and discuss invariance and stability of conditional mechanisms across environments.
• Introduce a domain-adversarial learning framework with a latent representation that minimizes domain-discriminability while predicting age.
• Incorporate an l1-filtering layer for interpretable feature attribution within the adversarial setup.
• Ground the approach in domain adaptation theory (HΔH-divergence) and discuss implications for generalization and potential causal interpretations.
• Apply the framework to mouse transcriptomic data and compare against conventional models.
• Discuss limitations in deriving causal conclusions from purely predictive models.

Figure 1: Different associations between $X$ and $Y$ as adopted from Figure 12 of [ 7 ] . A marginal correlation is the weakest form association that ignores dependencies among covariates. A stronger form is the regression relevant coefficients which captures partial correlation (non-zero correlatio

Experimental results

Research questions

• RQ1 Can invariant representations across heterogeneous environments yield robust age predictions in the face of distribution shifts?
• RQ2 To what extent does an adversarial domain-adaptive framework mitigate confounding and improve out-of-distribution performance for aging clocks?
• RQ3 Do predictive age models imply causal interpretations, or are they better understood as capturing stable statistical regularities of age-related biology?
• RQ4 How do intervention studies (e.g., Elamipretide effects) align with predictions from adversarial age predictors?

Key findings

• Adversarial domain-adaptive representations can differentiate treatment vs. control groups in mouse studies where conventional models fail.
• The adversarial framework’s predictions align with reported rejuvenation effects in external age-predictor studies, suggesting potential utility of ensemble approaches.
• Invariance-based representations improve out-of-distribution generalization by reducing reliance on dataset-specific correlations tied to sample attributes.
• Purely predictive models remain limited in causal inference without explicit interventional validation or structural causal assumptions.
• The framework provides a theoretically grounded path toward causal-oriented interpretations while acknowledging inherent limitations.
• A practical interpretation layer (l1-filtering) aids in attributing predictive signals to interpretable features.

Figure 2: For detailed caption, see Section 12