Risk prediction model using eye movements during simulated driving with logistic regressions and neural networks

Background: Many studies have found that eye movement behavior provides a real-time index of mental activity. Risk management architectures embedded in autonomous vehicles fail to include human cognitive aspects. We set out to evaluate whether eye movements during a risk driving detection task are able to predict risk situations. Methods: Thirty-two normally sighted subjects (15 female) saw 20 clips of recorded driving scenes while their gaze was tracked. They reported when they considered the car should brake, anticipating any hazard. We applied both a mixed-effect logistic regression model and feedforward neural networks between hazard reports and eye movement descriptors. Results: All subjects reported at least one major collision hazard in each video (average 3.5 reports). We found that hazard situations were predicted by larger saccades, more and longer fixations, fewer blinks, and a smaller gaze dispersion in both horizontal and vertical dimensions. Performance between models incorporating a different combination of descriptors was compared running a test equality of receiver operating characteristic areas. Feedforward neural networks outperformed logistic regressions in accuracies. The model including saccadic magnitude, fixation duration, dispersion in x, and pupil returned the highest ROC area (0.73). Conclusion: We evaluated each eye movement descriptor successfully and created separate models that predicted hazard events with an average efficacy of 70% using both logistic regressions and feedforward neural networks. The use of driving simulators and hazard detection videos can be considered a reliable methodology to study risk prediction. (C) 2020 Elsevier Ltd. All rights reserved.