Blood drop patterns and explainable AI reveal livestock anemia signatures

Abstract Anemia significantly limits livestock health and productivity. Existing diagnostic methods remain invasive, infrastructure-dependent, or prone to observer bias. This study evaluated a smartphone-compatible, paper-based blood biosensor combined with machine learning and explainable artificial intelligence for reagent-free image-level anemia classification across goats, sheep, and cattle. For each species, a single drop of whole blood was applied to a glycerol-treated paper substrate, producing bloodspot dispersion patterns associated with packed cell volume (PCV)-defined anemia categories. Standardized smartphone imaging and multi-resolution analysis were used to assess classification performance relative to laboratory hematocrit measurements. Classification performance varied by species, suggesting that species-associated differences in bloodspot morphology and image-pattern consistency may influence the diagnostic information available to image-based models. Sheep showed the highest and most consistent image-level classification performance, while goats and cattle showed more variable species-associated bloodspot patterns. Among the evaluated model configurations, lightweight support vector machine models achieved strong image-level performance with low computational demand. Explainable AI analyses provided complementary evidence that model predictions were associated with spatial bloodspot regions and dominant image-variance components. These findings provide preliminary proof-of-concept evidence for a smartphone-compatible bloodspot imaging framework that may support future development of anemia screening tools for precision livestock management after animal-level and field validation. Because animal-level identifiers were not retained, these results should be interpreted as controlled image-level proof-of-concept performance rather than animal-independent diagnostic accuracy.