Metabolomic Biomarkers for Osteoarthritis

Context

Research project at Memorial University on population-based knee osteoarthritis (OA) datasets. Metabolomics produces a wide field of candidate markers, and simple correlation with disease status is misleading — age, BMI, and metabolite concentrations are tangled together. The question was not “what correlates with OA” but “what plausibly drives it,” which is a causal-structure problem, not a ranking problem.

Constraints

• Correlation is cheap; causal claims are not. Any biomarker finding had to survive scrutiny beyond association.
• Observational, population-based data — no experimental intervention, so causal structure had to be inferred from the joint distribution and defensible modeling assumptions.
• Many candidate variables with real interdependence, which inflates spurious associations if modeled naively.

Approach

I modeled the joint distribution as a Bayesian network — a directed graphical model whose edges encode conditional dependence between covariates, metabolite concentrations, and the OA outcome. The learned structure proposes candidate causal paths rather than flat correlations. I then validated those paths independently: structural equation modeling and path analysis to test the hypothesized directed relationships, and multiple regression to check individual associations while controlling for covariates.

Decisions & Trade-offs

• Bayesian network over a correlation matrix or single regression. The graphical model captures conditional independence — the difference between “X and OA move together” and “X and OA move together only because of BMI” — at the cost of structure-learning complexity and sensitivity to modeling assumptions.
• Cross-validation with SEM and path analysis. Rather than trusting one method’s output, agreement between an inferred network and independently specified structural models raises confidence that a path is real and not an artifact of one algorithm.

Impact

The work surfaced metabolomic markers with defensible directed relationships to knee OA — findings backed by convergent evidence from probabilistic and structural methods rather than a single correlation. Methodologically, it is an end-to-end causal-inference pipeline on messy observational health data: model the dependence structure, then validate the causal claims before believing them.