Conceptualization, I.A.K. and S.S. (Sajid Shah); Data curation, S.S. (Sara Shafique) and I.A.K.; Funding acquisition, M.E.; Investigation, S.S. (Sara Shafique), S.S. (Sajid Shah), and W.J.; Methodology, I.A.K., R.N.J. and M.E.; Resources, R.N.J.; Software, S.S. (Sara Shafique), S.S. (Sajid Shah) and W.J.; Validation, W.J.; Writing—original draft, S.S. (Sara Shafique) and I.A.K.; Writing—review & editing, R.N.J. and M.E. All authors have read and agreed to the published version of the manuscript.

Figure 1. Process of gaze detection with the developed application.

Figure 1. Process of gaze detection with the developed application.

Figure 2. Gaze interaction and bounding box formation.

Figure 2. Gaze interaction and bounding box formation.

Figure 3. Gaze avoidance. (a) Looking left, (b) looking above, (c) looking right, (d) looking below.

Figure 3. Gaze avoidance. (a) Looking left, (b) looking above, (c) looking right, (d) looking below.

Figure 4. Experimental setup.

Figure 4. Experimental setup.

Figure 5. Overall view of gaze interaction and SAD relationship.

Figure 5. Overall view of gaze interaction and SAD relationship.

Figure 6. Relationship between SAD and the gaze interaction of female participants.

Figure 6. Relationship between SAD and the gaze interaction of female participants.

Figure 7. Relationship between SAD and the gaze interaction of male participants.

Figure 7. Relationship between SAD and the gaze interaction of male participants.

Figure 8. Comparison between gaze avoidance behavior of females.

Figure 8. Comparison between gaze avoidance behavior of females.

Figure 9. Comparison between gaze avoidance behavior of males.

Figure 9. Comparison between gaze avoidance behavior of males.

Figure 10. Confusion matrix of dataset. Class 0, 1, 2, 3, and 4 stand for ‘No SAD’, Marked, Moderate, Severe, and Very Severe groups, respectively.

Figure 10. Confusion matrix of dataset. Class 0, 1, 2, 3, and 4 stand for ‘No SAD’, Marked, Moderate, Severe, and Very Severe groups, respectively.

Table 1. Gaze interaction detection method to measure the SAD in a live communication studies comparison.

Table 1. Gaze interaction detection method to measure the SAD in a live communication studies comparison.

RefLive
EnvironmentAutomatic CodingRelationship with SAD SubclassesDataset of SADSAD—Gaze Avoidance Relationship Proved[33]✓××××[8]✓×××✓[18]×✓××✓[19]✓××××

Table 2. Data collected at various locations per participants’ convenience.

Table 2. Data collected at various locations per participants’ convenience.

S. NoLocationPercentage of ParticipantsDescription1.Home10%The data was collected at the locations where it was most convenient for the participants. This step was especially important to collect data from women as due to social and cultural barriers most women could not travel across to various locations for the experiment2.University12.8%3.College77.2%

Table 3. The conversation contents and procedure.

Table 3. The conversation contents and procedure.

Conversation DomainTask ExplanationTime FrameIntroduction ConversationThe observer and participant were introduced to each other.1 minSocial Support ConversationThe participants were asked to choose something that they want to change about themselves.2 minConflict ConversationThe participants were asked to tell something about their best friends and what they want to change in them.Minutes

Table 4. Pair sample independent t-test to show the difference between looking below sub-behavior and all other sub-behaviors.

Table 4. Pair sample independent t-test to show the difference between looking below sub-behavior and all other sub-behaviors.

Test PairtDfSig. (2-tailed)Cohen’s DLooking Below—Gaze Interaction−7.427490.000−1.05Looking Below—Looking Above4.054490.0000.57Looking Below—Looking Right2.630490.0110.37Looking Below—Looking Left4.096490.0000.58

Table 5. Recall, precision score of SAD classes.

Table 5. Recall, precision score of SAD classes.

ClassRecallPrecision0—No SAD0.800.731—Marked0.700.882—Moderate0.900.753—Severe0.800.804—Very Severe0.800.89

Table 6. Multi-class confusion matrix output for TP, FP, FN, and TN.

Table 6. Multi-class confusion matrix output for TP, FP, FN, and TN.

ClassTPFNFPTN0—No SAD823371—Marked731392—Moderate913373—Severe822384—Very Severe82139

Table 7. Multi-class confusion matrix output.

Table 7. Multi-class confusion matrix output.

ClassSpecificityPositive
Likelihood RatioNegative
Likelihood RatioPositive
Predictive ValueNegative
Predictive ValueAccuracy0—No SAD92.50%10.670.2272.73%94.87%90.00%1—Marked97.50%28.000.3187.50%92.86%92.00%2—Moderate92.50%12.000.1175.00%97.37%92.00%3—Severe95.00%16.000.2180.00%95.00%92.00%4—Very Severe97.50%32.000.2188.89%95.12%94.00%