Predictive consequences of alternative representations of metabolic measurements: A controlled synthetic study of categories, longitudinal information and transition preservation
Metabolic categories summarise observed abnormalities, but shared labels may conceal distinctions relevant to future evolution. We examined the predictive consequences of alternative representations under controlled synthetic dynamics. For each of six generating scenarios, twenty independent confirmation cohorts contained 5,115 simulated individuals each, including 1,023 held out. Observations through year 20 supported year-25 prediction under continued loading or reassignment to a specified low-load regime. Conventional categories, current and longitudinal measurements, selected descriptors and an outcome-guided embedding were evaluated with Ridge regression and histogram-based boosting. Paired cohort comparisons retained their original multiplicity families. A descriptive ablation examined phenotype-bearing and recorded etiological inputs. Separate post hoc work assessed monthly category-transition preservation. Under scheduled noisy conventional-panel observation and Ridge regression, the selected current and longitudinal representation (Phi) had lower loss than binary syndrome status and criterion count in all twelve scenario-task comparisons against each comparator. In current-state-sufficient continuation, the Phi-minus-binary-status difference was -0.0156 (pointwise 95% confidence interval -0.0187 to -0.0125). No advantage over current-plus-baseline measurements was established under the joint primary correction, and Phi-derived clusters performed better in nonlinear low-load prediction. Stronger categories and different learners altered some comparisons. Adding recorded etiological indicators produced small, context-dependent changes, including adverse increments. Two constructed states sharing syndrome status and count had next-month positive-category probabilities of 10.16% and 69.54%. An independent analytic argument established failure of strong lumpability for that monthly kernel. These were synthetic model probabilities, not patient risks. Categorisation can conceal future-relevant distinctions without being uniformly inferior for fitted prediction. Performance depends on information access, transformation, task and learner. The findings support explicit evaluation of metabolic representations, but do not establish a predictively sufficient representation of human physiology, clinical replacement of metabolic syndrome or therapeutic benefit.