Results

We did experiments on the certain theme. This table shows the results.

	Crafts Hobbies	Day Time TV Shows	Movie Genres	School Subjects
Fold Num	11	14	11	13
Accuracy	29/32	36/41	30/35	35/39
	90.63%	85.37%	85.71%	89.74%
F1	93.03%	85.00%	88.37%	90.47%
AUC	0.6042	0.4952	0.6122	0.5294
Label Ratio 0/1 (Unit: set)	28960/14282	6363/7660	29530/18636	25506/24367
	202.77%	83.07%	158.46%	104.7%
Male/Female	11/21	11/30	10/25	11/28
	90.91%/90.48%	90.91%/83.33%	70%/92%	81.82%/92.86%
Sensitivity	0.9091	0.9091	0.7	0.8182
Specificity	0.9048	0.8333	0.92	0.9286
MCI/NC	20/12	20/21	21/14	22/17
	100%/75%	85%/85.71%	90.48%/78.57%	86.36%/94.12%
Sensitivity	1	0.85	0.9048	0.8636
Specificity	0.75	0.8571	0.7857	0.9412

MC-ViViT achieves over 85% accuracy on all four themes. The highest one is 90.63% from Crafts Hobbies. The accuracy of School Subjects reaches 89.74%, which is close to 90%. Three of the F1 Scores are more than 88%, only the F1 score of Day Time TV Shows is 85%. The accuracy and F1 Score solidly support that MC-ViViT has robust performance. The other evidence comes from AUC Score. The average AUC of all four themes is 0.56. Given that not all the clips has MCI features in the video, even for subjects with MCI, it is reasonable to predict parts of their video sequences as NC. Due to this reason, some subjects will be correctly predicted as MCI with low scores.