1 Introduction

Medical records mean all the documents containing data, assessments and information of any kind on a patient’s status and clinical development throughout the care process. They include identifying of all the professionals who have intervened and aim to obtain the fullest possible integration in the clinical documentation of every patient, at least in the area of each center [1]. The current legislation allows the clinical information to be processed in electronic format, an informed consent being the only exception [1]. Since the introduction of Electronic Health Records (EHR) in the National Health Service, health personnel have become users of video display terminals (VDT).

VDT is alphanumeric or graphical screens, regardless of the method of visual representation used, and represent one of the most typical elements of computerized work equipment. Most workplaces with VDT are designed for sedentary work. Nevertheless, with health personnel the workplace design frequently includes VDT and working standing up; this happens, for example, with workstations for the administration of medication or with a portable terminal, which makes it possible to work beside the bed.

In Spain, the regulation on health protection of the workers exposed to the risks derived from working with VDT is fundamentally set out in the Royal Decree 488/1997 [2]. The National Institute of Safety and Hygiene at Work (Instituto Nacional de Seguridad e Higiene en el Trabajo, INSHT) provides criteria and recommendations for the interpretation and application of the said Royal Decree [3]. According to this regulation “workers using VDT” are, among others, those who effectively work more than 4 h a day or 20 h a week with this equipment. The “portable” systems are excluded when they are not continuously used at work.

Working with VDT has brought multiple benefits, but it also implies a series of risks for the health of the workers, since it demands certain requirements of physical and mental load. It can cause ocular and visual disorders, musculoskeletal disorders or other alterations, such as mental fatigue. The use of VDT during long periods of time has been associated with intense eye strain [4, 5] as well as with changes in the ocular surface and in the condition of the tear film [6, 7]. In these working places, Kroemer [8] considers the screen, the source document and the keyboard as visual targets. If these targets are separate, in direction or distance, the eyes must refocus constantly while they sweep from one target to another, with continuous changes in posture and convergence. This also requires good coordination of eye movements in order to merge the images of both eyes and to obtain a suitable binocular vision [4]. In addition, during computer work blinking frequency decreases, increasing the evaporation rate of the tear film, which compromises the good condition of the ocular surface. The American Optometric Association [9] defines computer vision syndrome (CVS) as a group of eye and vision-related problems that result from prolonged computer use. The most common symptoms associated with CVS are eyestrain, blurred vision, dry eyes and headaches, among others. In Spain, there are two instruments that have recently been developed and validated to measure this syndrome [10, 11].

The prevalence of CVS is high, presenting variations in different studies, according to the characteristics of the sample, the method and the instruments used for data collection. It is related to individual and work-related factors. The higher the exposure to VDT, the greater the prevalence of CVS. Ranashinghe et al. [12], estimate the prevalence of CVS’s at 67.4% among a group of 2210 computer office workers of Sri Lanka. A lower prevalence (53%) has been observed in a study carried out with a sample of 426 office workers in Spain [13] and in a group of 476 call center operators in Brazil (54.6%) [14]. According to the bibliographical review, to date no studies have been identified that evaluate the effects on visual health caused by exposure to VDT in health personnel. The most common occupations in the samples studied in the revised bibliography include office workers, bank employees, high-tech workers, graphical editors and call center operators. Only two studies carried out in Turkey included a sample of workers who used computers from two hospitals. The first one [15] included secretaries, computer operators and hospital data management system users. The second one [7] did not specify whether they were health personnel or not.

Nowadays, machine learning techniques are used extensively in medicine. Classification is a popular method that plays an important role in different medical diagnoses. Classification methods can help to integrate computer techniques in healthcare environments to improve medical services quality and efficiency. In recent years, some studies have employed classification machine learning techniques in different medical applications. For example, support vector machines (SVM) [16] have been employed in order to classify patients who will survive in the long term and those that not in oral squamous cell carcinoma cases, using as input variables clinicopathological parameters and molecular markers. Also, another study employed artificial neural networks for breast cancer diagnosis [17]; the model developed is able to determine which women are more likely to suffer from a particular kind of tumor before they undergo a mammography, which is helpful in order to increase the time from one screening control to the next. Among the classification techniques, genetic algorithms are adaptive search methods for finding optimal or near-optimal solutions, premised on the evolutionary ideas of natural selection. This methodology has already been employed in medical research [18], a remarkable application is the one that employed a genetic algorithm-based model to detect longitudinal changes in white matter fiber bundles of patient with multiple sclerosis.

The aim of the present work was to select the characteristics of the subject most relevant for the occurrence of CVS and then to develop a classification model for its prediction in health personnel. For this purpose, the proposed algorithm combines SVM and genetic algorithms.

2 Materials and methods

2.1 Case of study

This was an observational cross-sectional epidemiological study, based on the completion of two self-administered questionnaires among health personnel of the Monte Naranco Hospital (Oviedo, Spain). This hospital is specialized in geriatrics and palliative care, and it started using EHR in 2007.

The study included health personnel that were using VDT at work in the following occupational categories: physicians and surgeons, residents, nurses, advanced practice nurses (APNs) in training and auxiliary nurses. Exclusion criteria applicable to all participants were: not being a user of VDT at work, being a student, working in another occupational category, having seniority of less than a year, and currently suffering from, and/or being under treatment for, a diagnosed ocular disease.

Of a total of 172 workers, 151 (87.79%) took part in the study. In the end the sample that fulfilled the inclusion and exclusion criteria was finally of 139 workers from 20 different services or hospitalization units. The reasons of exclusion were that they were suffering the following diseases when the questionnaires were collected: dry eye (3), amblyopia (1), non-surgically controlled cataracts (2), vitreous diseases (3) and retinal disease (1). Also, one worker was excluded because he had been in his job for less than 1 year and another one for the lack of information about their seniority.

All the participants signed an informed consent form accepting their participation, in which data confidentiality was guaranteed during the entire process. The study was authorized by the local health authority and approved both by the Research Ethics Committee of the Principality of Asturias and by the Ethics Committee of the University of Alicante, Spain (the coordinating institution of the study), in accordance with the tenets of the Declaration of Helsinki.

2.2 Data collection

The health workers who took part in the study answered the following self-administered questionnaires:

  1. 1.

    Anamnesis and History of Exposure Questionnaire, which was specifically developed for this study, to gather information about sex, age, use of ophthalmics and/or contact lenses use, history of eye disease and treatment, thereof previous eye surgeries, occupational categories (physicians and surgeons including residents, nurses including APNs in training and auxiliary nurses) and seniority, work schedule (morning shifts, evening shifts, rotating shifts without nights, rotating shifts including nights and morning shifts with on-call shifts), current departments and nursing units and its seniority, information about the ease of use of the software application and daily VDT usage at and outside work.

  2. 2.

    The Computer Vision Syndrome Questionnaire (CVS-Q), designed and validated by Seguí et al. [10], was used to measure perceived ocular and visual symptoms during or immediately following computer work. This questionnaire evaluates the frequency (never, occasionally or often/always) and the intensity (moderate or intense) of 16 ocular and visual symptoms: burning, itching, foreign body sensation, tearing, excessive blinking, eye redness, eye pain, heavy eyelids, dryness, blurred vision, double vision, difficulty in focusing for near vision, increased sensitivity to light, colored halos around objects, feeling that eyesight is worsening, and headaches. Subjects with a score of 6 or more on the questionnaire are classified as symptomatic (suffering CVS).

2.3 Support vector machines

The SVM are machine learning techniques. Among other mathematical models for similar problems [19,20,21], SVM are used to model physical systems through the adaptation of their parameters [22,23,24,25]. These methods are broadly known for their usage in classification and regression problems [26, 27]. In the case of the present research, the SVM are used as a classificator. This technique has been selected due to its well-known performance. The performance of SVM relies on the adjustment of the model to data previously taken from the system to be modeled, as its training data set. To train the SVM to model a classification problem, the vectors from the training data are used to map hyperplanes that define the separation of classes. The output estimation provided by the trained SVM can be formulated as:

$$\hat{y}_{i} = a^{T}\Phi \left( {x_{i} } \right) + b$$
(1)

where \(x_{i}\) corresponds to the input vectors from the training set. The function \(\varPhi \left( {x_{i} } \right)\) linearizes the influences between inputs and outputs. In this scenario, the parameters are \(a\) and \(b\), which are a vector of the same dimension as the image of \(\varPhi\), and a coefficient, respectively. These parameters are determined by finding an optimized solution to the following problem and with restrictions:

$$\mathop {\hbox{min} }\limits_{{a,\varepsilon ,\eta_{i} ,\eta_{i}^{{\prime }} }} \frac{1}{2}a^{T} a + c\left( {\frac{1}{N}\mathop \sum \limits_{i = 1}^{N} \left( {\eta_{i} + \eta_{i}^{{\prime }} } \right) + v\varepsilon } \right)$$
(2)
$$a^{T} \varPhi \left( {x_{i} } \right) + b - y_{i} \le \varepsilon + \eta_{i}$$
$$y_{i} - a^{T} \varPhi \left( {x_{i} } \right) - b \le \varepsilon + \eta_{i}^{{\prime }}$$

where \(c\) is a regularization parameter, \(\varepsilon\) is the tolerance error for each input \(x_{i}\). Both \(\eta\) and \(\eta^{{\prime }}\) are the slack variables, which take positive values. Finally, \(v\) is a parameter for the adjustment of the tolerance. Therefore, the output of SMV [28] can be expressed as:

$$\hat{y}_{i} = F\left( x \right) = \mathop \sum \limits_{i = 1}^{N} \left( {\beta_{i}^{{\prime }} - \beta } \right)\varPhi \left( {x_{i} } \right)^{T} \varPhi \left( x \right) + b$$
(3)

In this expression, \(\beta\) and \(\beta^{{\prime }}\) are the Lagrange multiplayers corresponding to the restrictions above. In this context, the kernel function \(K\), can be defined as \(K\left( {x_{i} ,x_{j} } \right) = \varPhi \left( {x_{i} } \right)^{T} \varPhi \left( {x_{j} } \right).\) Consequently, the estimate of SVM turns into the following expression

$$\hat{y}_{i} = \mathop \sum \limits_{i = 1}^{N} \left( {\beta_{i}^{{\prime }} - \beta } \right)K\left( {x_{i} ,x} \right) + b$$
(4)

In general, the SVM for classification can be determined with the parameters \(c\) and \(v\), since, as said before, \(a\) and \(b\) can be obtained as the optimal solution to the quadratic problem [26]. Depending on the sort of function chosen as kernel, other parameters should be determined as well.

In the case of the present research, four different kernel functions are employed. One of them is the linear kernel, which is the sum of cross products. The limitations of linear learning machines are well known and in general, to solve real-world applications, more complex hypothesis spaces are required. Kernel representations are the solution as they allow data to be projected into a high-dimensional features space [29].

According to the Mercer’s theorem [30], let \(X\) be a finite input space with \(K\left( {X,z} \right)\) a symmetric function on \(X\). Then \(K\left( {X,z} \right)\) is a kernel function if and only if the matrix \(K\) is a positive semidefinite or, in other words, has nonnegative eigenvalues.

Please note that the use of kernels can overcome the curse of dimensionality in both computation and generalization.

For the reasons set out in the paragraphs above, three other kinds of nonlinear kernels are employed in the present research:

They are as follows:

  • Polynomial: \(\left( {\phi \left( {x^{1} \mu } \right) + 1} \right)^{p}\), where \(p\) represents the degree.

  • Radial basis function: \(\exp \left( { - \sigma \left\| {x - \mu } \right\|^{2} } \right)\)

  • Hyperbolic tangent: \(\tanh (\phi (x^{1} \mu ) + 1)\)

In the equations listed above, \(\phi\) and \(\sigma\) represent the scaling parameters. As is well known in existing literature, the performance of the kernel depends on the problem to be solved. This is the reason why four different ones are tested in our algorithm.

Finally, we would like to remind the reader of the kernel trick which says: “Given an algorithm which is formulated in terms of a positive definite kernel \(K\), one can construct an alternative algorithm by replacing \(K\) by another positive definite kernel \(K.\)

2.4 Genetic algorithms

In general, in artificial intelligence, an evolutionary algorithm is a kind of algorithm that is based on the evolution of certain possible solutions set by means of a metaheuristic optimization algorithm. The genetic algorithms are evolutionary procedures developed to simulate the evolution of a population in terms of optimizing the survival of the next generation. These algorithms were first developed for chromosomic studies [31], but now genetic algorithms work with the premise of improving fitness over the iterations until a solution for an optimization problem is reached. For each generation, the adjustments to the elements are based on four basic genetic operators used as criteria: crossover, mutation, reproduction and elitism [32]. The optimization problem and the iterations of the genetic algorithm can be formulated as follows [33]:

For a function \(f:D \to R\), and a set of restrictions, the minimization problem consists of looking for the best value \(x\) in the domain \(D\) such that \(f\left( x \right) \le f\left( y \right)\) for all \(y\) in the domain \(D\). The value \(x^{{\prime }}\) in the domain \(D\) is a local minimum of \(f\), when a neighborhood \(N\left( {x^{{\prime }} } \right)\) of \(x^{{\prime }}\) exists where for all \(z\) in \(N\left( {x^{{\prime }} } \right)\), \(f\left( {x^{{\prime }} } \right) \le f\left( z \right)\).

In this scenario, the genetic algorithm begins with an initial population \(P_{0} = \left\{ {I_{0}^{1} , \ldots ,I_{0}^{n} } \right\}\). Each step of the algorithm, the objective function is calculated along with its correspondent performance measures and then generates a new population of which the elements are selected with a determined rule from the four genetic operators defined above. After \(m\) steps, the population is denoted as \(P_{m}\). The algorithm stops when the performance measures are not significantly improved in a chosen number of iterations.

The crossover operation is defined as follows. Let two individuals \(I_{1}\) and \(I_{2}\) of the population \(P_{m}\). A new individual \(I_{12}^{{\prime }}\) (string) is creating by randomly choosing a crossover point in the string of \(I_{1}\) and \(I_{2}\) and splitting them and recombining the two parts of the individuals in such a way that the length continues being the same.

The mutation operation consists on given certain individual \(I_{1}\), the change in certain values of its individual bit components (either from ‘0’ to ‘1’ or from ‘1’ to ‘0’). In order to avoid convergence problems, the mutation rate must be low.

Finally, the elitism mechanism in a GA simply involves passing to the next generation of a certain amount of those individuals of a certain generation that have performed the best performance when evaluated according to the fitness function.

2.5 The proposed algorithm

The aim of the proposed algorithm is to find the adequate variables and parameters with which the SVM will model the proposed classification problem properly. The iterations of the genetic algorithm are focused on maximizing the value of the AUC (area under the ROC curve) of each of the SVM models calculated. The steps of the algorithm are shown in Fig. 1.

Fig. 1
figure 1

Algorithm diagram

Full size image

The algorithm begins with the setting of the parameters for the genetic algorithm, such as crossover, mutation, elitism and population size. After this, an initial population should be created. As was stated by Galán et al. [33], although the setting depends heavily on in the data considered, a range of optimal parameters can be determined; for the present research, the probability values applied for crossover were those from 0.5 to 1 in steps of 0.1, while the mutation probability employed values of 0.1, 0.2 and 0.3. The elitism probabilities were 0.01, 0.05, 0.1 and 0.2. Different sizes for population of 10, 25 and 50 individuals were considered.

The set of the initial population must be a vector with size of the possible variation of parameters for the SVM. We will consider different types of initial populations, depending on the type of kernel. To avoid this problem, we consider branching the algorithm. Specifically, we will consider four ways in parallel; for each branch, the genetic algorithm is performed with each possible type of kernel. Now the sets of initial populations can be created. Each item of the initial population has the following form, depending on the kernel used: Linear: \(\left( {c,v,x_{1} , \ldots ,x_{k} } \right)\). Polynomial: \(\left( {\gamma ,\alpha_{0} ,\alpha ,c,v,x_{1} , \ldots ,x_{k} } \right)\). Radial basis: \(\left( {\gamma ,c,v,x_{1} , \ldots ,x_{k} } \right)\). Sigmoid: \(\left( {\gamma ,\alpha_{0} ,x_{1} , \ldots ,x_{k} } \right)\). The parameters common to all the populations are: the cost \(c\) of constraints violation which corresponds to the constant of the regularization term in the Lagrange formulation of the SVM, ranging from \(10^{ - 2}\) to \(10^{2}\); the tolerance error \(v\) for the determination of the SVM which takes values from \(5 \times 10^{ - 4}\) to \(5 \times 10^{ - 3}\); variables \(x_{1} , \ldots ,x_{k}\), with values 0 or 1, depending if the variable is taken into account for the SVM model or not. For the other parameters, \(\gamma\) is needed in all kernels except linear and it takes values that range from \(1/\left( {2 \cdot datadimension} \right)\) to \(1/\left( {datadimension} \right)\); \(\alpha_{0}\) works as coefficient for the sigmoid and polynomial kernel (from − 1 to 1), and finally, \(\alpha\) for the degree determination of the polynomial kernel (from 3 up to 5). The SVM model is then estimated with the previous values for each of the branches, and its AUC with a validation subset of the data is calculated. The stop criteria will be satisfied if the AUC does not change more than 0.01% in the last 100 iterations of the algorithm.

When the stop criteria are not satisfied, a new population has to be created. This is performed through implementing the crossover, mutation and elitism. With the elitism process, the populations were sorted based on their value of the AUC, and only the ones with higher AUC were considered. Crossover will combine sections of the chosen populations, and finally, with the mutation process an aleatory modification is introduced to the population. Then the whole process is repeated until the stop criteria are satisfied.

3 Results

3.1 Study population and prevalence of CVS

The characteristics of the participants are summarized in Table 1. The mean age of the 139 workers included in the study was 46.75 ± 10.40 years (range 22–64 years), and 82% were women. A majority (46.8%) of the study population worked in a geriatric, palliative or internal medicine service. Most of the workers were employed as a nurse (63.3%). In the occupational category and in the current department or nursing unit of work, the mean seniority was 17.85 ± 10.72 and 12.19 ± 10.61 years, respectively. They worked with VDT an average of 4.93 ± 1.92 h a day and used a computer outside work an average of 1.43 ± 1.15 h a day.

Table 1 Characteristics of the study population and prevalence (P) of computer vision syndrome
Full size table

The prevalence of CVS was 60.4%. A higher prevalence was detected among women (63.2%), those who were between 41 and 50 years of age (72.3%) and individuals who worked as a nurse (64.8%), more than 4 h a day with PVD (65.3%) or with rotating shifts including nights (71.8%). The same happened with wearers of ophthalmic or contact lenses (62.6 and 80.0%, respectively) and those who had undergone ocular surgery (66.7%), as well as with individuals who used VDT outside work (62.9%) or those who did not consider the software application as “easy” (83.3%). The most frequently occurring symptoms included “feeling that eyesight is worsening” (64.7%), “headaches” (61.9%), “difficulty in focusing for near vision” (61.9%), “increased sensitivity to light” (58.3%) and “itching” (56.1%).

3.2 Classification model

The chosen model was selected due to its AUC. The value of this performance metric was 0.9433029, giving a high performance over the validation data. The average AUC of the models for each iteration is shown in Fig. 2. For each iteration the AUC is computed and the average value presented in this figure. As can be observed, the performance of the models increases quickly in the first iterations, and afterward oscillates between the iterations 400 and the 900. Then it raises again until the performance level is high, and finally the average value remains stable for, at least 100 iterations and consequently the algorithm stops, determining the model chosen. This model is set with a sigmoid kernel, and parameters \(\gamma = 2.1 \times 10^{ - 4}\), \(\alpha_{0} = 0\), \(c = 1\) and \(v = 5 \times 10^{ - 4}\). Figure 3 shows a color map in which the AUC variations are compared for different values of \(\gamma\) and \(v\) when the sigmoid kernel is employed for \(\alpha_{0} = 0\) and \(c = 1\).

Fig. 2
figure 2

Average AUC of the estimated models over the genetic algorithm iterations

Full size image
Fig. 3
figure 3

Color map of AUC for different values of γ and v when the sigmoid kernel is employed for \(\alpha = 0\) and \(c = 1\)

Full size image

The variables identified by the model as the most relevant in the development of CVS, together with their description, are summarized in Table 2.

Table 2 Variables and categories selected by the model as the most relevant for the development of CVS
Full size table

The AUC was compared with another performance metric, the Youden index, to corroborate the robustness of the model. The comparison is performed via the correlation between both metrics, giving a result of 0.8594404, which implies a high degree of correspondence between the calculated models.

4 Discussion and conclusions

According to the model developed, workers who used a computer outside work and those exposed to VDT more hours a day at the workplace were more likely to suffer CVS. In other studies the total number of hours a day using VDT (both at work and outside the workplace) was found to be related to visual discomfort [34, 35], and prolonged VDT use at work was also connected to an increased risk of developing CVS [12, 13]. Our results are similar to those of Tauste et al. [13] who observed that workers who use VDT more than 6 h a day at work had a prevalence of CVS of 63% and therefore were more likely to develop the syndrome (OR 2.28; 95% CI 0.96–5.37). Likewise Uchino et al. [36] identified that more than 8 h-a-day VDT users had a significantly increased risk of dry eye disease (OR 1.94; 95% CI 1.22–3.09); this coincides with two of the categories identified in this study: people with dryness and those who use lubricant eye drops.

The model also associates the female gender with a higher risk of developing CVS as well. Many studies have reported an association between the female gender and the prevalence of CVS [12,13,14]. However, when considering individual symptoms, Ranasinghe et al. [12] reported higher prevalence of red eyes, changes in visualizing colors and excessive tearing (p < 0.05) in men. According to our results, working as a nurse is also associated with CVS; this may be explained because 95.5% in this occupational category are women.

In our study, age is also identified as a factor related to CVS. The prevalence of CVS was higher among those aged from 41 to 50 years. The average age of those with CVS (45.92 ± 10.06 years) was lower than the average age of those without CVS (47.12 ± 11.31 years). Rosenfield [37], in a review of the scientific literature on CVS, stated that it is unclear whether asthenopia during computer use is associated with age. As a matter of fact, there are studies where no significant change with age was observed in computer-related symptoms [12, 35].

Another relevant result of the model is that contact lens wearers are more likely to suffer CVS. This is consistent with previous studies. Tauste et al. [13] found, in the group of workers exposed to the computer for more than 6 h a day, that contact lens wearers were almost five times more likely to suffer CVS (OR 4.85; 95% CI 1.25–18.80). Kojima et al. [38] also observed lower tear reservoir volumes in contact lens wearers, as well as higher visual symptom scores associated with blurred vision and visual difficulties.

Finally, a work schedule with shifts including nights was also associated by the model with the development of CVS. Previous studies [39,40,41] suggest an association between rotating night shift work and several diseases, including cardiovascular disease, cancer risk, diabetes, hypertension, chronic fatigue, developing excess weight and obesity, sleeping problems and early spontaneous pregnancy loss.

Taking into account that the AUC means the ability of the model to detect computer vision syndrome and also the results obtained by means of the Youden index (a single statistic that captures the performance of dichotomous diagnostic test), the results of the proposed model are promising.

Nevertheless, these results should be interpreted with caution given the limitations of our study. The first of these is that it was a cross-sectional design, and we cannot be sure that the cause precedes the effect. A second one is that we did not include ophthalmic examinations that inform us of the workers’ refractive state.

Despite these limitations, the use of a validated questionnaire to measure CVS is a particular strength of this research, and this is the first study that describes the relationship between CVS and its associated factors among health personnel by means of genetic algorithms and support vector machines. Both machine learning techniques are combined. SVM are used in order to perform classifications due to its well-known performance, while genetic algorithms are employed in order to optimize the SVM parameters.

Finally, we would like to remark that the main contribution of the present research is a novel hybrid methodology which is able to determine the most important variables in a classification problem.