<p>Feline sporotrichosis is a zoonotic subcutaneous mycosis of major relevance in endemic regions of Latin America, particularly in Brazil. Current diagnostic methods are limited by turnaround time, specificity, and cost. In this context, spectroscopy combined with machine learning has emerged as a promising alternative. This study aimed to develop and validate predictive models for the rapid and minimally invasive diagnosis of feline sporotrichosis using ATR-FTIR spectroscopy applied to plasma samples. Seventy-five cats were included and classified by fungal culture as the reference standard. Plasma samples were analyzed by ATR-FTIR (4000–400&#xa0;cm⁻¹), with 20 spectra acquired per animal. Data were evaluated using six machine learning algorithms under three validation frameworks: spectral-level internal testing, Leave-One-Patient-Out (LOPO) cross-validation, and independent external validation (<i>n</i> = 15 animals). On the internal test set, Random Forest, KNN, and LightGBM showed similarly strong discriminative ability (AUC ≥ 0.998), with no significant differences among them. The LOPO cross-validation confirmed biological generalization, with SVM achieving the highest AUC (0.896). External validation confirmed consistent performance, with Random Forest and LightGBM achieving an AUC of 1.000 and excellent probabilistic calibration (low Brier scores). Spectral variable importance and SHAP analyses revealed that discrimination was primarily driven by regions associated with carbohydrates (1040–1050&#xa0;cm⁻¹) and lipids (1746–1751&#xa0;cm⁻¹), reflecting the host’s systemic metabolic and immunological response to the infection. These findings suggest that ATR-FTIR combined with machine learning provides consistent and generalizable diagnostic performance, representing a rapid and minimally invasive approach within the One Health framework.</p> Graphical abstract <p></p>