BEYOND GLUCOSE: IDENTIFYING PRE-DIAGNOSTIC DIABETIC SIGNAL AND BIOMARKERS IN FELINE DIABETES

Abstract

Feline diabetes is a multifactorial metabolic disease that develops gradually and is often di-agnosed only after persistent hyperglycemia appears. However, emerging evidence suggests that pre-diagnostic physiological changes may be detectable through multivariate clinical biomarkers beyond glucose alone. In this work, we investigate whether a diabetic signal exists in routine feline bloodwork prior to diagnosis and how it evolves over time. Using longitudinal veterinary records, we build predictive models across multiple temporal windows leading up to diagnosis and evaluate several classifiers, including sparse support vector machines and ensemble feature-selection methods. Across all approaches, predictive information is distributed across a small, stable set of secondary biomarkers, including SC_ALKP, SC_BUN_CREA, and SC_CHLORIDE, which retain discriminative power even without glucose. We also examine disease structure using a geometric framework based on angles between class- conditional subspaces. Embedding these subspaces into a low-dimensional space via multidimen- sional scaling reveals a gradual decline in class separability as time from diagnosis increases, con- sistent with classifier performance trends. This provides a geometric view of disease progression as a smooth shift in physiological structure rather than an abrupt change. Across experiments, predictive performance remains above chance up to approximately 150–180 days prior to diagnosis, suggesting that clinically meaningful signal exists well before overt hyper- glycemia. Together, these results support feline diabetes as a distributed and temporally evolving condition, and show that sparse and geometric methods can identify pre-diagnostic signal.