Background: Type 2 diabetes (T2D) is a growing public health problem in Mexico. Lipid profile alterations have been shown to appear years before changes in glycemic biomarkers, and some of the latter are limited in availability, especially in underserved settings. Therefore, anthropometric variables and lipids represent relevant early indicators for the early detection of the disease. This study evaluates the capacity of non-glycemic clinical data—including lipid profile and anthropometric indicators—to detect T2D using machine learning, and compares the performance of different feature engineering approaches. Methods: Using more than a thousand clinical records of Mexican adults, three experiments were developed: (1) a distribution and normality analysis to characterize the variability of lipid variables; (2) an evaluation of the predictive power of multiple atherogenic indices (Castelli I, Castelli II, TG/HDL, and AIP); and (3) the implementation of statistical transformations (logarithmic, quare-root, and Z-standardization) to stabilize variance and improve feature quality. Logistic regression, SVM-RBF, random forest, and XGBoost models were trained on each feature set and evaluated using accuracy, sensitivity, specificity, F1-score, and area under the ROC curve. Results: The AIP index showed the greatest discriminatory power among the atherogenic indices, while normality-based transformations improved the performance of distribution-sensitive models, such as SVM. In the final experiment, the SVM-RBF and XGBoost models achieved AUC values greater than 0.90, demonstrating the feasibility of a diagnostic approach based exclusively on non-glycemic data. Conclusions: The findings indicate that the transformed lipid profile and anthropometric variables can constitute a solid and accessible alternative for the early detection of T2D in clinical and public health contexts, offering a robust methodological framework for future predictive applications in the absence of traditional glycemic biomarkers.