Predictive consequences of alternative representations of metabolic measurements: A controlled synthetic study of categories, longitudinal information and transition preservation

Under scheduled noisy conventional-panel observation in a synthetic longitudinal setting, the Phi current-and-longitudinal representation yields lower Ridge-loss than binary syndrome status and criterion count across all twelve scenario-task comparisons.

• Why it matters: Shared metabolic category labels can hide distinctions that matter for future evolution, so it is unclear which representation choices preserve predictive information under realistic observation schedules.
• What they did: Across six generating scenarios, twenty confirmation cohorts (5,115 simulated individuals each, including 1,023 held out) were used to compare conventional categories, current and longitudinal measurements, selected descriptors, and an outcome-guided embedding with Ridge regression and histogram-based boosting.
• The result: In continued loading, the Phi-minus-binary-status loss difference was -0.0156 (95% CI -0.0187 to -0.0125), and Phi-derived clusters improved nonlinear low-load prediction, while no advantage over current-plus-baseline was established under the joint primary correction.