Predictive and mechanistic evaluation of ODE-informed machine learning for systemic lupus erythematosus disease-activity stratification.

BACKGROUND: Mechanistic consistency does not necessarily imply improved prediction. We evaluated ordinary differential equation (ODE)-informed machine learning for systemic lupus erythematosus (SLE) disease-activity stratification, assessing predictive utility and mechanistic consistency separately. METHODS: GSE65391 data classified 131 patients as high (n = 56) or low (n = 75) disease activity. A published four-state ODE generated anti-inflammatory, immune-complex, tissue-damage, and pro-inflammatory features. Architecture-matched Random Forest (RF) models were compared using feature ablation and outer-fold evaluation. Directionality, perturbation, and parameter sensitivity assessed mechanistic consistency; two independent SLE datasets supported exploratory proxy comparisons. RESULTS: RF + ODE initially achieved AUROC 0.7133 versus 0.5940 for RF alone (ΔAUROC=+0.1193; 95% CI, 0.0209-0.2140). Removing the outcome-coupled ODE residual reduced AUROC to 0.6708. Archived out-of-fold evaluation and independent outer-fold reimplementation yielded AUROCs of 0.6019 and 0.6529, respectively; neither established robust incremental discrimination. The reimplementation excluded held-out outcomes from predictor generation. Within the calibrated representation, anti-inflammatory, immune-complex, and pro-inflammatory states showed consistent disease-activity directionality, with strong immune-complex-pro-inflammatory coupling (Spearman ρ=0.912). Perturbation and parameter-sensitivity analyses supported stability of selected relationships. External modality-specific proxies showed exploratory directional analogy without validating source-defined states. CONCLUSIONS: The apparent predictive advantage attenuated under stricter evaluation, while selected state relationships remained stable under tested perturbations and parameter variation. The published ODE provided a structured representation for examining dynamical relationships, with incremental predictive value assessed separately. Mechanistic consistency and predictive utility should be evaluated as distinct dimensions.