Predicting vasovagal reactions to needles from video data using 2D-CNN with GRU and LSTM

When undergoing or about to undergo a needle-related procedure, most people are not aware of the adverse emotional and physical reactions (so-called vasovagal reactions; VVR), that might occur. Thus, rather than relying on self-report measurements, we investigate whether we can predict VVR levels from the video sequence containing facial information measured during the blood donation. We filmed 287 blood donors throughout the blood donation procedure where we obtained 1945 videos for data analysis. We compared 5 different sequences of videos-45, 30, 20, 10 and 5 seconds to test the shortest video duration required to predict VVR levels. We used 2D-CNN with LSTM and GRU to predict continuous VVR scores and to classify discrete (low and high) VVR values obtained during the blood donation. The results showed that during the classification task, the highest achieved F1 score on high VVR class was 0.74 with a precision of 0.93, recall of 0.61, PR-AUC of 0.86 and an MCC score of 0.61 using a pre-trained ResNet152 model with LSTM on 25 frames and during the regression task the lowest root mean square error achieved was 2.56 using GRU on 50 frames. This study demonstrates that it is possible to predict vasovagal responses during a blood donation using facial features, which supports the further development of interventions to prevent VVR.