Analysis of Gender Differences in Online Handwriting Signals for Enhancing e-Health and e-Security Applications

If we get a look at signal processing, in speech signal analysis, it is well-known that gender differences exist. The clearest example is pitch. In general, women speak at a higher pitch, about an octave higher than men. This is mainly due to morphological aspects such as the vocal tract length and oral cavities dimension. This implies significant differences. For instance, in [4], the authors studied the effect of age and gender on emotion recognition applications. They compared the performance of four different models and presented the relationship between age/gender and emotion recognition accuracy. Experimental results showed that using a separated emotion model for each gender and age category gives a higher accuracy compared with using one classifier for all the data.

In [5], the authors identified that the associations between hearing loss (HL) and cognitive impairment varied according to gender in older community dwellers, suggesting that different mechanisms are involved in the etiology of HL.

Another different topic is behavioral sex differences. In [6], the authors found that many popular conceptions of behavioral sex differences are founded more on personal beliefs than on facts. Some of these beliefs include that women speak more, speak faster, leave more sentences unfinished, and operate at a simpler conceptual level than men.

In online handwriting signal analysis, the gender influence has attracted less interest from the scientific community, when compared to other signals such as speech. Most of the existing literature dealing with gender and handwriting signals is aimed to gender classification [7‐14]. For example, in [14], Likforman-Sulem et al. employed the support vector machine (SVM) classifier and modeled a set of dynamic handwriting features extracted from a sentence copy task in order to stratify a cohort into males and females. They reached 65% accuracy, suggesting that the gender effect is present but with a relatively small impact. This finding is in line with the study of Liwicki et al. [11], who utilized Gaussian mixture models trained on kinematic and spatial features and reached a classification accuracy of 64%. The team also reported that even though this accuracy is low, it outperforms the classification based on the offline handwriting, i.e., an accuracy of 55%. On the other hand, in [13], the authors quantified an offline handwritten text by spatial (geometric) features and achieved 74% gender classification accuracy using the Random Forest algorithm. Although this accuracy slightly differed across languages (English and Arabic), it again highlighted that the gender effect in handwriting processing should be taken into account and further explored. For example, the conclusions of the Kaggle competition (which hosted more than 190 teams) suggest that the effect of gender is also associated with the age of the writer [12].

Nevertheless, besides the above-mentioned studies, some important questions concerning the gender classification must be addressed, the most important one being are there gender/age differences in the handwriting of healthy subjects that must be taken into account when analyzing pathologies such as Alzheimer’s or Parkinson’s disease? The first step would be to identify these differences. If they exist, it is important to obtain deep knowledge about them, in a similar way that we know that the pitch of males/females is significantly different. In the case of age differences, they can be alleviated by the fact that control groups in pathology analysis are usually paired by age, or, in the case of imbalanced datasets, age is considered as a confounding factor and could be regressed out. However, it has not been deeply analyzed whether significant differences exist in handwritten tasks performed by males and females. If differences exist in speech signals, then why not in handwritten tasks?

The second important question is about the correlation between different handwritten tasks performed by the same user: if a user exerts a high pressure and exhibits higher writing speed than the standard population, is this result independent of the performed task? And, are there gender differences?

In this paper, we want to address the question about differences in online handwritten tasks due to gender. We are mainly interested in temporal, kinematic, and dynamic features such as time in-air/on-surface, pressure, acceleration, and complexity (measured by Shannon entropy). In order to respond to the questions, we performed a set of experiments on the BIOSECUR-ID database [15].

For a recent description of the state of the art, applications and future trends in online handwriting analysis, see [16, 17]. Usually, the scientific literature has ignored the separate analysis by gender, but we can forecast increasing attention to this issue, as improved classification accuracies in e-health and e-security could be obtained by considering the gender issue. From this point of view, we consider that a good starting point is the analysis of different online handwritten tasks performed by healthy subjects (when performing the same task). Obviously, this limits the analysis of other levels of information such as differences in word usage and expressions, which could be evaluated by means of a handwritten free text analysis. Nevertheless, this requires an acquisition of a new database, which is beyond the scope of this paper.

We also skip some interesting questions, which should be addressed in future research work, such as the following analysis of gender dependence skills in e-security and e-health applications:

We have analyzed the BIOSECUR-ID [15] database, which consists of 330 subjects (46% females, 54% males), who provided handwritten samples in four acquisition sessions. The participants performed the following tasks:

The database was acquired with the Wacom Intuos3 A4 tablet in combination with the Wacom Inking pen. It provides 5080 dpi and 1024 pressure levels and a spatial accuracy ± 0.25 mm.

Figure 1 shows the population distribution for both genders. Experimental mean and standard deviation reveal that both populations present a similar distribution.

The parameters used to carry out the analysis were:

Table 1 shows handwriting features for different tasks using recordings from session 1.

Worth to mention that males and females have been educated during the schooling period to write texts in the same way. However, the signature is something personal, chosen by the individuals, and is therefore more prone to gender differences.

The next figures depict scatter plots for a specific feature and a couple of tasks. Males are depicted in red and females in green. Worth to mention that when comparing different tasks performed by the same user, there is a large correlation as the dots appear aligned. Figure 2 shows the five analyzed features (t_up, t_down,\(\overline{p }\), \(\overline{s }\), \(\overline{a }\)) for the text in cursive letters (in x coordinate) vs numbers (in y coordinate). Figure 3 shows the same previous features for the cursive text (in x coordinate) vs capital letters (in y coordinate). Figure 4 is analogous with the text in capital letters versus numbers, and Fig. 5 compares genuine signatures versus skilled forgeries performed by the same user. Figure 6 compares cursive letters vs. signature, and Figs. 7, 8, 9, 10 capital letters vs. signature.

Table 2 shows Pearson’s correlation coefficients for a specific feature when performing two different tasks (for instance, the first cell in the upper left corner of the table represents the correlation between the t_up required to perform a text in cursive letters versus numbers).

Considering the following range of values:

We observe:

Another interesting question is whether there are significant differences between a specific task and a feature when performed by males and females. In order to respond to the question, and since the features do not have normal distribution, we performed the Mann–Whitney U test with the null hypothesis that the two samples come from distributions with equal medians. Table 3 summarizes the p values for different tasks and features.

Considering as significant those results with p < 0.05, we observe: