Predicting Seemingly Unpredictable Events Using Pattern Classification of Physiological Data
Julia A. Mossbridge, Northwestern University Department of Psychology
Distinct physiological changes preceding randomly selected and seemingly unpredictable arousing (vs. calm) stimuli arriving 3–10 seconds in the future have been described in peer- reviewed journals using five different physiological measures (skin conductance, heart rate, blood volume, EEG, fMRI) in at least twelve different laboratories. A meta-analysis of experiments examining such anomalous anticipatory activity reports a small but highly significant overall effect (Mossbridge et al., under review). Here we examined the utility of pattern classification in predicting seemingly unpredictable upcoming events based on physiological data preceding those events.
In the first experiment, 20 undergraduates were fitted with 64 EEG electrodes. Participants used their right hand to press the left mouse button as soon as they saw a “1” or heard a low tone, and the right button if they saw a “2” or heard a high tone. All stimuli were randomized. Raw EEG was current source–density transformed and artifacts were removed before analysis. Based on previous work by others revealing predictive alpha patterns preceding visual stimuli, we used as dependent variables the phase of the mean (across-trial) peak alpha (7.5–12 Hz) frequency for each person at each electrode relative to 550 ms preceding the onset of the upcoming stimulus presentation (left vs. right). A random forest pattern classification algorithm was used to classify the dependent variables according to the nature of the upcoming stimulus presentation. To be conservative, the classifier was executed using the actual data versus a randomized version of the same data. Using data from the entire group of participants averaged across trials, the classifier was able to predict the upcoming stimulus class with 76% accuracy (p < 0.012; generalization performance). Importantly, two left parietal electrodes (opposite the button-press hand) were most critical for classification.
In the second experiment, in each of 625 trials, author JM tried to guess a future “target” image from four photos. Following this selection, a random number generator randomly selected one of the photos as the target and displayed it. The dependent variables were the mean baselined heart rate (HR) and skin conductance (SC) during each of the ten seconds preceding the presentation of feedback (correct vs. incorrect). The same classification algorithm was used in an attempt to predict, for this one participant, the physiological “anticipatory signature” of the upcoming event, but it was not able to predict the upcoming feedback on a trial-by-trial basis. One possible explanation is that over time the experimenter’s response to the stimuli became less dramatic, consistent with the results of the meta-analysis revealing a decrease in effect size accompanying an increase in the number of trials performed.
Pattern classification of EEG signals suggests that physiological prediction of seemingly unpredictable future events is both possible and informative. Future work will attempt to reduce classification error rate for group data and examine the best paradigms for producing classifiable data on a trial-by-trial basis, and address response shifts over time.