Fig shows the features selected from all

From Fig. 3 we can see a couple of interesting aspects: Firstly, most of the highest MI between features and class occurred at earlier time indices (T = 1 and 2) instead of around maximal response of the sensors (T = 15). Further, we observe significant information about the class in the sensors’ responses at later time indices (T = 19 and 20), which occurs during baseline recovery phase of analysis. These MI results are contrary to both our intuition and the common practice in E-nose literature of selecting features near maximal ha tag response [35]. Secondly, while sensors of the same type (that is, sensors 3, 4, 5 , 9, 10, 11 , and 6, 7 ) have similar results in Fig. 3, there are still some differences in the amount of information they hold about the class. This difference is especially obvious for the 2 WO3 sensors (S6 and S7). This could be due to the differences in these sensors when they were manufactured, though a more likely reason is that the physical location of these sensors in the instrument (they are not placed next to each other) plays a role here. Lastly, Sensor 8 (SnO2 doped with Cu) shows no information about the classes for the second half of the temperature cycle (when sensor surface temperature is approximately 260–300 °C), which indicates this sensor do not respond differently to the individuals’ breath samples at low temperature.