**3. Results**

#### *3.1. Tra*ffi*c Conditions*

A combination of 40 km/h of average speed and 20 km/h of standard deviation, which exhibit relatively higher accuracy, sensitivity, specificity, and positive predictive value, was selected. The mean values of four measures at the combination were the highest. It is highlighted in bold in Table 3. From Table 3, we can see that a few cases have an accuracy of over 80%. However, we cannot select them due to low sensitivity and predictive value, which are under 50%. Low sensitivity shows that a large number of cases with high stress are not recognized, and the low predictive value indicates the low effectiveness of the model in detecting high stress. Thus, considering all the effective parameters of the developed model, the most accurate criterion of low vs. high traffic conditions was found to be with average speed of 40 km/h and standard deviation of 20 km/h: Classification model accuracy = 80.3%, sensitivity = 85%, specificity = 78%, and positive predictivity = 70%. Based on these results, we can conclude that traffic conditions (low vs. high traffic) are an important element to classify driving stress levels (low vs. high stress). In addition, we presumed that traffic conditions (average speed < 40 km/h and SD < 20 km/h) could be a clear threshold of driving stress levels (low vs. high stress), compared to the other 19 conditions in Table 3.

Table 5 shows the developed model for the classification of low stress and high stress levels depending on the high traffic. In the traffic condition model, only NOM was used because the variables NOD and NOM are the same.


**Table 5.** Model based on traffic conditions (low vs. high traffic).

It was found that three extracted features (number of occurrences (N), mean amplitude (MeanOM), and maximal amplitude (MaxOM)) of the hand EDA data are significant (all ps<0.05). The mathematical rule of the developed model for the classification of stress levels depends on the probability concept. High stress was recognized by the model if the probability of case i was greater than 0.5. Otherwise the case was classified as low stress.

The goodness of model fit measure evaluates the ability of developed models to explain variation in the dependent variable (high and low stress). As such measures, pseudo-R-squared are usually used in logistic regression analysis. In the presented study pseudo-R-squared of Cox and Snell and Nagelkerke were obtained using IBM SPSS Statistics Software (Version 23). For traffic condition model Cox and Snell and Nagelkerke R-squares are 0.323 and 0.432 respectively. Figure 6 shows the confusion matrix of the analyzed datasets for the classification of high and low stress levels by using the EDA data, depending on the traffic conditions.

**Figure 6.** Confusion matrix of the model using the tra ffic condition datasets.

#### *3.2. Road Type*

In the second method for road type prediction, where driving in cities was recognized as a high stress state and highway as a low stress state, the statistical method accuracy reached 82.9% with sensitivity of 81%, specificity of 84%, and positive predictivity of 65%. Table 6 shows the developed model including all used predictors for road type segments.


**Table 6.** Model based on di fferent types of road segments.

The binary logistic regression model for road type segments shows that Min OD is significant (*p* < 0.05). For road type model goodness of fit was presented by Cox and Snell and Nagelkerke R-square with values of 0.374 and 0.518, respectively. As for the tra ffic condition model, pseudo-R-squared was obtained through SPSS Statistics Software (Version 23). Figure 7 shows the confusion matrix of the model developed by the road types.

**Figure 7.** Confusion matrix of the model using the road type datasets.
