Estimating cognitive workload while driving using functional near infrared spectroscopy (fNIRS)

Kurzfassung

Introduction and Aim: We envision that driver assistive systems which adapt their functionality to the driver’s cognitive state could be a promising approach to reduce road accidents due to human errors [1]. Workload is an important cognitive state because a cognitive overload or underload results in a decrease in human performance which may result in fatal incidents while driving. Here, we investigate if it’s possible to predict variations of cognitive workload levels (WL) from fNIRS brain activation measurements while driving to inform adaptive assistive systems in future cars. Methods: In our study, we implemented the n-back working memory task with several continuously changing load levels as a speed regulation task into a realistic driving simulation in order to control and manipulate cognitive workload. We introduced five different workload levels (i.e. 0-back to 4-back) and speed signs every 20 seconds while the participant was driving on a highway with concurrent traffic. Depending on the current n-back task, the participant was supposed to remember the previous ‘n’ speed sequences and adjust his speed accordingly. A detailed explanation for the n-back experimental paradigm can be found in [2]. FNIRS data were recorded from the frontal and parietal cortices using a 32-channel neuroNIRX-system from five participants during the course of the whole experiment which lasted for around 30 minutes. Results: We used the multivariate linear regression approach by combining fNIRS data from all channels to predict WL using the elastic net regularization model which combines the L1 and L2 penalties of the lasso and ridge regression techniques to get a continuous estimate of workload over time. Fig.1 depicts a plot of WL induced by the n-back task (red curve) and the WL predicted by the model (blue curve) for an example participant. The correlation between the two curves is almost 0.8 for a 10-fold cross-validation. For the remaining four participants, we achieved a correlation between 0.6 and 0.8 which were all statistically significant (p < 0.05). Fig.1: Ten-fold cross-validated prediction of workload from deoxyhemoglobin fNIRS measurements using multivariate regression analysis in an example participant Discussion & Conclusion: It can be seen from Fig.1 that the predicted WL more or less follows the induced WL and could be used to predict if the WL is either increasing or decreasing. There are certain regions in the blue curve where the model seems to over- or underestimate WL which may be due to the incomplete model which currently neglects WL imposed by the concurrent driving task in the changing traffic situations. As a next step, we plan to introduce some non-linearity into the workload model to get much better prediction rates. Acknowledgements: This work was funded by a grant of the Volkswagen Foundation and the Ministry of Science and Culture of Lower Saxony to the Research Centre on Critical Systems Engineering for Socio-Technical Systems. Anirudh Unni and Klas Ihme contributed equally. References [1] Parasuraman, R. (1987). Human-computer monitoring. Human Factors, 29, 695-706 [2] Unni et al. (2015). Brain activity measured with fNIRS for the prediction of cognitive workload. 6th IEEE Conference on Cognitive Infocommunications, 349-354