br Experimental design materials and

Experimental design, materials and methods
Conflict of interest
Acknowledgments The work is supported by the Faculty Research Fund (FRG1/13–14/010; BIOL3840101) to WKFT and General Research Fund (261610), Research Grant Council Hong Kong to CKCW.
Experimental design, materials and methods
Measurement protocol Three concentrations doses 0.1, 0.3 and 1vol.% were used to prepare the dilutions in water for the pure analytes. The same analyte concentrations were used to prepare mixtures. The gas p53 tumor suppressor included samples of pure ethanol (‘lab’ attribute eth-0.1, eth-0.3 and eth-1), samples of pure acetone (ace-0.1, ace-0.3 and ace-1), samples of binary mixtures of ethanol and acetone (ace-0.1-eth-0.1, ace-0.1-eth-0.3, ace-0.3-eth-0.1, ace-0.1-eth-1 and ace-1-eth-0.1) and samples of water dilutions without any analyte (pure air). Hence, the total number of classes is 12. The choice of these analytes and concentrations was not affected by any particular application constraint, except that the sensors of selected models show consistent and diverse responses among the gas classes. The number of samples per class among 58 samples is the following: The measurements were split into 5 batches (‘batch’ attribute), where each batch contained records approximately for all gas classes given in a random order. All the batches were acquired within a time period of 4 days to minimize the effect of the long-term internal and environmental noise present in the system. The number of samples per batch among 58 samples is the following: The measurement protocol was the following: we delivered 10μL of the corresponding dilution to the vessel using a micropipette. The vessel was connected to the ventilator ‘Source’ outlet. After 3min of exposition, the source of the gas vapor was removed from the vessel, and the recovery phase started. During the recovery phase, the system was sampling room air for 2 additional minutes to record the decay in the sensors signals. Note that 2min of recovery phase was not sufficient to recover the sensors baseline and re-establish again the initial conditions in the gas chamber. Hence, although we acquired 2min of recovery phase, the system was pumping air until the sensors recovered the baseline and the whole gas sample was exhausted from the gas chamber.
Signal-processing methods The readout data was the output voltage of the sensors׳ conditioning circuit. The 16 acquired time-dependent voltages were converted to resistances according to the voltage-divider scheme and the corresponding load resistor. Hence, each data point in the array described the resistance of a sensor R(t) at a certain time of measurement t. The resistance values in the data set were normalized by subtracting the baseline value R0=R(t0) at the starting point of the measurement t0 and scaling by factor R0, (R(t)−R0)/R0. Note that such format of the measured raw data allows for comparison of responses among different sensors. The recorded time-series signal for each sensor were acquired at the sampling frequency of 25Hz during 5min, resulting in 7500 data points per time-series of a single sensor. Previously to computing the low-frequency and high-frequency features, the raw data were pre-processed by a set of digital filters. A median filter was used to remove the spikes in the signals. Then we employed two Butterworth filters of 3rd order: a low-pass filter (cut-off frequency 0.01Hz) and a high-pass filter (pass-frequency 0.07Hz) to extract the low/high frequency parts from the original signals, respectively. Note that these low/high frequency signals (output of the two Butterworth filters) are not distributed within the data set. For feature extraction implemented in [1], both low-frequency and high-frequency sensor signals were divided by respiratory cycles, where each cycle was processed independently. Thus, a feature is referred to as a feature by respiratory cycle. Since high-frequency signals showed oscillatory behavior similar to a sine wave curve, we decided to follow a straightforward strategy for feature extraction in this case. We used amplitude of the high-frequency signal (oscillation) at every respiratory cycle as a feature. Low-frequency trajectories had a monotonic behavior, and we used the magnitude of the low-frequency signal as a feature at every respiratory cycle. The value was taken at the same time of oscillation, where the amplitude of the high-frequency signal was measured. Note that the low-frequency and high-frequency features were computed only for the first 13 respiration cycles. Fig. 1 illustrates the feature extraction flow for a single transient of sensor No. 7.