Model Validation
A model must be validated against data independent of those used for parameter estimation. For ME/CFS models, validation takes three complementary forms.
1 Predictive Accuracy
The strongest form of validation is accurate prediction of previously unobserved phenomena. A model parameterized from metabolomic data that correctly predicts CPET outcomes, or a model fitted to cytokine kinetics that correctly predicts treatment response timescales, demonstrates explanatory power beyond curve-fitting. Cross-validation (holding out subsets of data for testing) provides a weaker but practically useful form of predictive validation.
2 Biological Plausibility
Model predictions must be consistent with established biology. A model predicting negative concentrations, physiologically impossible reaction rates, or cell counts exceeding total body lymphocyte numbers is biologically implausible regardless of its fit to data. Plausibility constraints can be enforced formally through parameter bounds or informative priors in Bayesian estimation.
3 Structural Identifiability
Before fitting a model to data, it is essential to verify that the model structure permits unique parameter estimation in principle (structural identifiability) and in practice given measurement noise (practical identifiability). A structurally unidentifiable model contains parameters or parameter combinations that cannot be determined from any amount of data, indicating model overparameterization. Techniques for identifiability analysis include the differential algebra approach, the generating series method, and profile likelihood analysis. All models in Part V are accompanied by identifiability assessments in Appendix Mathematical Model Details.
What constitutes sufficient validation for a computational model of ME/CFS pathophysiology? Unlike pharmacokinetic models, which can be validated against drug concentration time courses, ME/CFS models predict emergent phenomena (fatigue, post-exertional malaise) that lack precise molecular definitions. Establishing consensus validation criteria—analogous to FDA guidance for physiologically-based pharmacokinetic models—is a prerequisite for clinical translation of ME/CFS models.
The BRANDO meta-epidemiological evidence (Section Blinding Failures in ME/CFS Treatment Research of Chapter Controversies and Debates in ME/CFS Research) provides quantitative priors on blinding bias applicable to any unblinded trial with subjective outcomes. A formal Bayesian framework could incorporate: (1) an informative prior on bias magnitude from Hróbjartsson 2014 (\(\mathcal{N}(0.56, 0.15)\)), (2) a diagnostic contamination prior from Wormgoor 2021 criteria-stratification data, and (3) the trial-specific likelihood. The posterior would give the true treatment effect after accounting for both biases. Applied to Zhao et al. (2026, SMD = 0.85), the posterior 95% credible interval for exercise efficacy in PEM-positive ME/CFS would be predicted to span zero.