Aims: Here, we sought to leverage the symmetric projection attractor reconstruction (SPAR) technique to uncover novel insights into the expression of motion-induced nausea, and explore whether machine models can distinguish between malaise and normal from electrocardiogram (ECG)-generated attractor images.
Methods: We obtained the SPAR transforms of ECG signals recorded from healthy participants at rest and during nausea, for a binary image classification task using a total of 6 pretrained deep neural networks (i.e., ResNet-50, ResNet-101, DenseNet-201, Xception, Inception-v3, and Inception-ResNet- v2) with transfer learning. Bayesian optimization with 100 iterations was utilized to tune the hyperparameters mini-batch size, learning rate, momentum, and L2 regularization. To evaluate model performance, we performed a leave-one-participant-out cross-validation (LOPOCV).
Results: We observed that the DenseNet-201 network performed better than other networks as evaluated by accuracy for attractors generated using N = 3, 5, and 6 points (70.14%, 70.14%, and 69.91%, respectively). When examining model performance using the receiver operating characteristic (ROC) curve and the area under a ROC curve (AUC), we found that ResNet-50 and ResNet-101 achieved AUC scores considered acceptable – for 4-point attractors (AUC 0.7091 and AUC 0.7275), respectively. The DenseNet-201 model demonstrated acceptable AUCs for attractors computed using embedding dimensions (N = 3; AUC 0.7325) and (N = 6; AUC 0.7098). Attractor visual examination revealed that the arms of the (3, 1) attractor at baseline are wider than those of the nausea (3, 1) attractors. Further, we observed that baseline (5, 1) attractors are less dense near the center than their nausea counterparts.
Conclusion: Our observations provide new insights into how physiologic characteristics captured via ECG-derived attractor images may be important for distinguishing expression of motion-induced malaise. The SPAR method may help assess the efficacy of new therapeutics for motion sickness management.