Experimental Validation

The startle respone varies both between individuals, as shown in Figure 3, and for each individual at different times, as shown in Figure 7. The robustness of the startle detection algorithm was tested on 27 responses representing both differences between individuals and differences for an individual at different times. In a startle eliciting experiment, subjects were seated, wearing a skin conductivity sensor with electrodes on the index and middle finger and asked to listen to three stimuli (50 millisecond audio bursts of 95 dB white noise). These stimuli were designed to elicit three startle responses. Of the eleven subjects who participated in the study, nine subjects responsed to the stimulus, but two habituated rapidly as shown in Figure 3 and were eliminated.

As the threshold of the startle detection algorithm was varied the number of missed responses and the number of falsely detected responses varied. If the threshold is set too high, some of the startle responses are not detected, and if the threshold is set too low, then both false positives and false negatives result as the threshold passes below the skin conductance baseline, as shown in Figure  4. For this pool of responses, a zero error rate was found with a threshold of 0.0005 microseimens. However, using this same method, the threshold could be customized for an individual user from multiple responses over time.

  

Figure: Varying the threshold changes the accuracy of the algorithm. Figure (a) shows missed responses from a threshold that is too high. Figure (b) shows false positive responses from a threshold that is too close to baseline. These three responses are from a single user, showing that threshold variation is important both across individual's and within the variation of an individuals response,

  

Figure: Adjusting the threshold creates two kinds of errors, false positives when the threshold is too low (but intersecting the signal) and missed responses when the signal is too high or low. Figure (a) is a record of the number of false positives for the startle responses as the threshold is varied, Figure (b) is a record of the responses missed by the algorithm.