Fusion of regional and pixel-level plantar pressure features for Diabetic peripheral neuropathy detection
Résumé
Diabetic peripheral neuropathy (DPN) alters foot biomechanics in subtle and progressive ways, often undetected by standard clinical tests in early stages. In this study, we propose a machine learning framework that integrates regional, global, and pixel-level features extracted from normalized plantar pressure recordings to detect DPN in different stages. The relative importance of different regions of the feet (forefoot, midfoot, rearfoot) was emphasized in the classification task, which provides clinically interpretable insights into neuropathic progression. The dataset includes 54 patients divided into four groups: patients with diabetes without PN; with signs of mild PN; with signs of severe PN; and a control group without diabetes. In total, 321 features were extracted: 21 clinically meaningful regional and global features, and the 300 most informative pixellevel features. These were reduced using Principal Component Analysis (PCA) before classifier training. Among all the models tested, the CatBoost achieved the highest accuracy of 92.92%. Notably, the model maintained strong sensitivity across all classes, including the different stages of DPN. Our findings demonstrate that a fusion of pressure-based features, analyzed with machine learning and dimensionality reduction, offers a reliable and interpretable tool for early screening and staging of diabetic neuropathy.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |
