FEATURE SELECTION APPLIED TO THE TIME-FREQUENCY REPRESENTATION OF MUSCLE NEAR-INFRARED SPECTROSCOPY (NIRS) SIGNALS: CHARACTERIZATION OF DIABETIC OXYGENATION PATTERNS

Diabetic patients might present peripheral microcirculation impairment and might benefit from physical training. Thirty-nine diabetic patients underwent the monitoring of the tibialis anterior muscle oxygenation during a series of voluntary ankle flexo-extensions by near-infrared spectroscopy (NIRS). NIRS signals were acquired before and after training protocols. Sixteen control subjects were tested with the same protocol. Time-frequency distributions of the Cohen's class were used to process the NIRS signals relative to the concentration changes of oxygenated and reduced hemoglobin. A total of 24 variables were measured for each subject and the most discriminative were selected by using four feature selection algorithms: QuickReduct, Genetic Rough-Set Attribute Reduction, Ant Rough-Set Attribute Reduction, and traditional ANOVA. Artificial neural networks were used to validate the discriminative power of the selected features. Results showed that different algorithms extracted different sets of variables, but all the combinations were discriminative. The best classification accuracy was about 70%. The oxygenation variables were selected when comparing controls to diabetic patients or diabetic patients before and after training. This preliminary study showed the importance of feature selection techniques in NIRS assessment of diabetic peripheral vascular impairment.