<![CDATA[Diabetic foot problems continue to be a significant cause of morbidity among people with diabetes, requiring accurate and scalable methods for early identification. Infrared thermography is a non-invasive technique for detecting asymmetry in plantar temperature, which is associated with the risk of inflammation. However, the use of infrared thermography is limited by the small size of labeled datasets in the biomedical domain, posing issues for the generalizability of models and the reliability of validation. In this work we aim at studying the effectiveness of a carefully verified handcrafted-feature based model for diabetic foot thermogram categorization with limited data. The suggested pipeline combines multi-descriptor feature representation, Recursive Feature Elimination (RFE) and margin-based classification by means of Support Vector Machines in a nested cross-validation setting, together with bootstrap stability analysis. Experiments on a publicly accessible plantar thermogram dataset (N = 167) show strong performance with holdout AUC of 0.942 and nested cross-validation AUC of 0.959 ± 0.045. Calibration findings show a low Brier score of 0.065 with balanced sensitivity (0.88) and specificity (0.889). Further study shows that discriminative performance stems from distributed gradient–texture interactions, not sparse feature subsets. The results show that a careful validation and a systematic design of the representation can overcome the limitations of data and allow the development of reliable thermographic screening systems.]]>
