A NEURAL NETWORK-BASED SYSTEM FOR PREDICTION OF COMPUTER USER COMFORT

Abstract

This article proposes a neural network-based system for prediction of computer user comfort with respect to the existing settings of the workstation. In this context, anthropometric measures and the existing measures of a computer workstation were related to back-support comfort, distance comfort, keyboard comfort, monitor comfort, and seat comfort using two distinct modeling approaches—multiple linear regression and artificial neural network. The purpose of this article was to compare and contrast the resulting models. The data from 144 computer workstations were used and a total of 12 different data types such as shoulder to elbow, eye to buttock, pan height, monitor height, or distance from the chair were recorded. While multiple linear regression could not be used to adequately predict the computer workstation comfort, the neural network was deemed superior. This approach allows ergonomists to aid in the decision-making process of computer environment design and the prediction of the health risk in an occupational environment.

Computer-related repetitive disorders have become one of the largest problems facing ergonomists and the medical community because it is developing in epidemic proportions within the occupational environment. In previous years, the prevalence of neck and upper extremity complaints among computer users is high (Tittiranonda, Burastero, and Rempel, 1999; Flodgren, Heiden, Lyskov, and Crenshaw, 2007). Studies have shown that factors such as repetitive and stereotype movements, fixed postures, static load, insufficient recovery time, muscular fatigue, and psychosocial stress increase the risk of developing musculoskeletal symptoms (Flodgren et al., 2007; Tittiranonda et al., 1999; Jensen, 2003; Juul-Kristensen, S⊘gaard, Str⊘yer, and Jensen, 2004).

In the wake of the expanding use of computers, concerns have been expressed about their potential health effects. Researches have shown that these symptoms can result from problems with the equipment, workstations, office environment, or job design, or from a combination of these (Murray, Mass, Parr, and Cox, 1981). Identification and modeling of the likelihood of computer-related repetitive strain injuries (CRRSIs) by using traditional modeling tools is frequently difficult. A lack of understanding of what differentiates the disability potential of different work designs can lead to inefficient allocation of resources for prevention purposes. The etiology of musculoskeletal disorders associated with work is qualitative and more attention to modeling techniques may allow the development of models that explicitly describe the risk associated with various work designs so that specific, quantitative workplace assessments can be made (Dempsey and Westfall, 1997).

A number of debilitating medical conditions such as carpal tunnel syndrome, tendonitis, thoracic outlet syndrome, tenosynovitis, and low-back disorders can be caused and aggravated by using computers in an unhealthy manner. Much of the CRRSIs were solved by ergonomic perspective in literature (e.g., Kamienska, 1993; Arnt, 1983; Carter and Banister, 1994; Clift, 1989; Grandjean, 1987; Kroemer, 1992; Murray et al., 1981; Pascarelli and Quilter, 1994; Peterson, 1989; NIOSH, 1991; Houtman and Kwantes, 1999; Punnett and Bergqvist, 1997). Computer-related repetitive strain injuries have been researched by Cakir, Hart, and Stewart (1979), OSHA (1999), and Ulusam, Dülgeroğlu, and Kurt (2001).

OBJECTIVE

The risk potential for CRRSIs due to computer tasks can be conceptualized in view of the mathematical foundation. Computer-related repetitive strain injuries may occur quite suddenly and in a nonlinear fashion.

Practitioners are interested in identifying quick methods for evaluating the risk of computer-related ergonomic injury. There are many factors impacting an individual's likelihood of developing injury. There are many potential factors (as well as many unknowns) regarding how people are affected from workstation design. It is not easy to handle the relationships between factors. The task of the industrial ergonomist is fairly difficult because the potential risk factors that may contribute to the onset of CRRSIs (e.g., carpal tunnel syndrome) interact in a complex way, and require him/her to apply elaborate data measurement and collection techniques for a realistic work condition analysis.

The main objective of this study was to develop an artificial neural network-based predictive system, which could predict comfort of the computer user according to existing settings of computer workstations (physical workplace measurements). The proposed system provides a faster response and a reduced cost compared to human experts. Such a system could be very useful in hazard analysis and injury prevention from computer tasks. In addition to this, it can be useful data to check the existing working environment.

LITERATURE ON THE USE OF ARTIFICIAL NEURAL NETWORKS FOR ERGONOMIC PROBLEMS

There are some neural network-based models to evaluate for various ergonomic problems such as automobile seat comfort prediction with statistical model and neural network process and estimating anthropometric measurements by adaptive neuro-fuzzy inference system. (e.g., Karwowski, Jarvinen, and Zurada, 1992; Arcand, 1994; Jung and Park, 1994; Lim, Fok, and Tan, 1996; Harden, Crumpton, and Killough, 1996; Zurada, Karwowski, and Marras, 1997; Carnahan and Redfern, 1998; Lin and Hwang, 1999; Chen, Kaber, and Dempsey, 2000; Chung, Lee, Inseok, Dohyung, and Sang, 2002; Kiryu, Shibai, Hayashi, and Tanaka, 2002; Kaya et al., 2003; Lee, Karwowski, Marras, and Rodrick, 2003; Chen, Kaber, and Dempsey, 2004; Kolich, Seal, and Taboun, 2004).

Chen et al. (2004) used neural networks for data classification problems. They developed a method to predict the risk of inquiries in industrial jobs. Jung and Park (1994) predicted and analyzed the human reach posture with backpropagation method. Arcand (1994) examined neural networks as a tool for optimizing the placement of icons in an interface. Lin and Hwang (1999) used neural networks to achieve dynamic task allocation in a flexible manufacturing system. Chen et al. (2000) modeled a neural network for prediction of musculoskeletal disorder risk such as low-back disorder risk associated with industrial jobs. Kiryu et al. (2002) estimated the ratings of perceived exertion from objective indices such as muscular fatigue-related indices and the heart rate, using a feed-forward type artificial neural network. Lim et al. (1996) coped with investigation of the posture and motion by using neural network. Harden et al. (1996) investigated the use of artificial neural networks as a tool for predicting the need for accommodations in the workplace. Carnahan and Redfern (1998) developed a low back injury risk classifier using evolutionary computation. Chung et al. (2002) used a neural network approach to predict the whole-body postural stresses from the body joint motions.

Techniques such as artificial neural networks (ANNs) have been proposed as methodologies to develop risk models of the often-complex relationships between task, workplace, and worker characteristics and the risk of work-related musculoskeletal disorders. Recently, ANNs have been investigated as to their effectiveness for modeling Low Back Disorders (LBD) risk (Karwowski et al., 1992; Zurada et al., 1997; Chen et al., 2000).

The researchers have demonstrated the use of advanced statistical methods including logistic regression and generalized additive models (e.g., Dempsey and Westfall, 1997; Marras et al., 1993) and ANNs (e.g., Zurada et al., 1997) to predict musculoskeletal disorder risk associated with occupational exposures. These methods have produced varying prediction accuracy depending upon the task factors examined and the accuracy of the historical data used for model development. However, neural networks have been shown to yield superior predictive capability in comparison to, for example, multiple linear regression models.

METHOD

Experimental Data for Model Development

The experimental data for model development were collected by students at University of Gaziantep, Industrial Engineering Department in regard of a project of Human Factors Engineering lecture. The students collected the existing settings of computer users. The data from 144 computer workstations such as bank and laboratory were used (some data were eliminated). A total of 12 different data type were recorded by students. The collected data were shoulder to elbow (measure from user elbow tip to the top of the user's shoulder), elbow to fingertip (measure from the user's elbow tip to the tip of the user's fingers), shoulder to buttock (while seated upright, measure from the top of the seat pan to the top of the user's shoulder), Eye to buttock (while seated upright, measure from the top of the seat pan to the user's eye level), Back to buttock (while seated upright, measure the top of the seat pan to the user's belt-line), Knee to foot (measure from the back of the user's knee to the underside of the user's heel), Back support height (measure from the center of the back support to the floor), arm support height (measure from the top of the arm support to the floor), pan height (measure from the top of the seat pan to the floor), monitor height (measure from the top of the monitor screen to the floor), keyboard height (measure from the top of the keys to the floor), distance from chair (measure from the back support to the front of the keyboard).

In order to evaluate the comfort of the computer user in his/her existing workstation, collected data were entered “Kairos” expert system (http://www.peaktechnology.co.nz). The Kairos computer safety is a toolkit to investigate the problems of musculoskeletal disorders among computer users. To protect against computer-related health problems, computer users need to set their workstations up correctly and to work safely. It monitors the users' mouse and keyboard use and warns them when they're working in an unhealthy manner. The software is effective within even the most deadline-driven workplaces, including call centers. It also provides a tool that enables each computer user to ergonomically customize his or her workstation. To create the real data of the existing setting of workplaces in the real world condition, we used an expert system called a Kairos 7.2 (Peak Technology Ltd., Wellington, New Zealand). Naturally, the analyses of the Kairos expert system was compared to the developed neural network prediction model.

The collected data were evaluated with a trial version of Kairos 7.2. Kairos 7.2 gave a workstation analysis of each computer user. A total of 144 evaluations were obtained. Kairos's workstation reports involve back support height, arm support height, pan height, monitor height, keyboard height, and distance from chair. In this study, these workstation records (analyses) were regarded as comfort indicators. According to the Kairos 7.2 results, arm support and distance from chair comfort indicators can be too close, okay, or too far. Back support comfort, pan comfort, monitor comfort, and keyboard comfort indicators can be too low, okay, or too high.

The second was a neural network developed with the help of Alyuda NeuroIntelligence 2.2 (Neo Digital, Inc., 2001–2008, Los Altos, CA). As part of this process, input neurons were created for every input variable (shoulder to elbow, elbow to fingertip, shoulder to buttock, eye to buttock, back to buttock, knee to foot, back support height, pan height, monitor height, keyboard height, and distance from chair) and an output neuron was created for the output variable (back comfort, distance comfort, keyboard comfort, monitor comfort, and pan comfort).

Training, Test, and Validation Datasets

Training, test data, and validation sets were selected for back support comfort, pan comfort, monitor comfort, keyboard comfort, and distance from chair separately. For back support comfort, the order of 100 (69.44%) existing setting of computer users in the training set were randomized. The 22 (15.28%) and 22 (15.28%) of existing settings of computer users in the test set and validation set were randomized respectively. For the remaining comforts, some randomization was applied with small differences. The purpose of breaking the data into a training set, test, and validation set was to provide a check on a real-world situation. The training set was used to train the neural network, a procedure that reduces the least square error between the correct response and the actual response until an acceptable overall error is reached. New data that the system had not been exposed to were then presented to the network, and its performance on these new data patterns was evaluated.

Network Input Parameters

Each observation in the training data contained the 11 variables that described the existing settings of a user workstation. These variables were as follows: (1) shoulder to elbow, (2) elbow to fingertip, (3) shoulder to buttock, (4) eye to buttock, (5) back to buttock, (6) knee to foot, (7) back support height, (8) pan height, (9) monitor height, (10) keyboard height, and (11) distance from chair. Measurements of arm support height were removed due to abnormality of data.

The prediction variables were five comfort indicators. These are back support comfort, distance comfort, pan comfort, monitor comfort, and keyboard comfort. These variables were used only as a teacher's response during the network's training using a batch back-propagation algorithm. A batch back propagation for networks of any size can be used. A batch back propagation algorithm is the most popular algorithm for training of multi-layer perceptions. The performance of a batch back-propagation algorithm depends on convergence, tune-up of the learning rate and momentum parameters, and high probability of getting caught in local minima.

Network Architecture

The best feed-forward neural networks architecture with batch back-propagation have been searched. All tested network architectures contained from 4 to 28 neurons in a single hidden layer, and 3 neurons in the output layer. The searching of the best network architecture for comfort indicators was made in order. In all calculations, the neurons were fully connected. For example, for the network architecture of back support comfort, the least number of misclassifications was obtained for the network with 11 neurons and 2 neurons in a single hidden layer and three output layers (Figure 1), respectively. Therefore, this network structure was chosen for further consideration. The network accepted 11 inputs (previously mentioned) and 3 outputs trained for values are too low, okay, and too high, respectively. The network architecture of the back support comfort and the other network architectures of comfort indicators have been given in Table 1. In the next section, the remaining comfort prediction results were presented in terms of network performance and quality of comfort predictions.

FIGURE 1 The architecture of the neural network used for back support comfort prediction system.

TABLE 1 The Architecture of the Neural Networks Used for System Training Purposes

	Input	Hidden	Output
Output variables	neuron	layer neuron	neuron
Back support comfort	11	2	3
Distance comfort	11	4	3
Keyboard comfort	11	28	3
Monitor comfort	11	28	3
Pan comfort	11	11	3

The learning rate and the momentum used in training were 0.1 for all comfort indicators. The learning rate is a control parameter, which affects the changing of weights. The bigger learning rate causes larger weight changes by each iteration. The momentum is a control parameter that effects the changing of weights. The greater the momentum, the more the current weight change is affected by the weight change that took place during the previous iteration.

RESULTS

Network Performance and Quality of Comfort Predictions for Back Support Comfort

After training of the network for back support comfort, the developed prediction system could predict 98 out of 100 cases in training sets correctly. In other words, correct classification rate (CCR) of the proposed system for a training set could attain 98%. The CCR of the validation set is 90.90%. At the end of the training, the mean of the training set, validation set, and test set of CCR are 91%, 95.45%, 86.36%, respectively. Sample confusion matrix for back support comfort is given in Table 2. Training speed (iteration/second) was 409.858618. There are four different stopping conditions including the maximum number of iterations, minimum absolute error, minimum relative error, and minimum error improvement thresholds. The stopping condition is determined by minimum error improvement thresholds and if the minimum error improvement threshold is less than or equal to 3%, training is stopped. At the end of the iteration number (63855), the desired stop condition has been attained for back support comfort training data. The same stopping conditions were applied to the rest of the training sets.

TABLE 2 Confusion Matrix for Back Support Comfort with Training Set, Validation Set, and Test Set

	Training set			Validation set			Test set
Target/output	Too low	Too high	OK	Too low	Too high	OK	Too low	Too high	OK
Too low	78	3	0	18	0	0	18	0	0
Too high	1	13	0	0	3	0	0	1	0
OK	3	2	0	1	0	0	1	2	0

According to network statistics, the weights of the importance (input importance) are 0.009495% for shoulder to elbow, 0.30389% for elbow to fingertip, 0.650787% for shoulder to buttock, 0.482542% for eye to buttock, 24.051404% for back to buttock, 20.256105% for knee to foot, 41.442627% for back support height, 3.525766% for pan height, 6.256698% for monitor height, 5.581848% for keyboard height, and 1.493451% for distance from chair.

Network Performance and Quality of Comfort Predictions for Distance Comfort

After training of the network for distance comfort, the developed prediction system could predict 100 out of 100 cases in a training set correctly. The CCR of the validation set is 86.956522%. At the end of the training, the mean of the training set, validation set, and test set of the CCR are 95.918367%, 91.304348%, 95.652174%, respectively. Sample confusion matrix for distance comfort is given in Table 3. The training speed (iteration/second) was 359.049276. At the end of the iteration number (20035), the desired stop condition has been attained for distance comfort data.

TABLE 3 Confusion Matrix for Distance Comfort with Training Set, Validation Set, and Test Set

	Training set			Validation set			Test set
Target/output	Too far	Too close	OK	Too far	Too close	OK	Too far	Too close	OK
Too far	87	0	0	21	0	0	22	0	0
Too close	0	2	3	1	0	0	1	0	0
OK	1	0	5	1	0	0	1	0	0

According to network statistics, the weights of the importance (input importance) are 10.630778% for shoulder to elbow, 2.46053% for elbow to fingertip, 5.157042% for shoulder to buttock, 4.268731% for eye to buttock, 8.473599% for back to buttock, 0.24711% for knee to foot, 1.314193% for back support height, 0.584615% for pan height, 3.358437% for monitor height, 0.375626% for keyboard height, and 63.129338% for distance from chair.

Network Performance and Quality of Comfort Predictions for Keyboard Comfort

After training of the network for keyboard comfort, the developed prediction system could predict 97.959184 out of 100 cases in a training set correctly. The CCR of the validation set is 73.913043%. At the end of the training, the mean of the training set, validation set, and test set of the CCR are 97.959184%, 73.913043%, and 73.913043%, respectively. Sample confusion matrix for keyboard comfort is given in Table 4. Training speed (iteration/second) was 143.676235. At the end of the iteration number (20172), the desired stop condition has been attained for keyboard comfort data.

TABLE 4 4Confusion Matrix for Keyboard Comfort with Training Set, Validation Set, and Test Set

	Training set			Validation set			Test set
Target/output	Too low	Too high	OK	Too low	Too high	OK	Too low	Too high	OK
Too low	66	0	0	12	0	0	14	0	0
Too high	1	22	0	4	3	0	4	3	0
OK	1	0	8	2	0	2	2	0	0

According to network statistics, the weights of the importance (input importance) are 11.990836% for shoulder to elbow, 5.329607% for elbow to fingertip, 13.155575% for shoulder to buttock, 1.611735% for eye to buttock, 4.730807% for back to buttock, 12.664113% for knee to foot, 6.605948% for back support height, 6.980112% for pan height, 5.760722% for monitor height, 18.175792% for keyboard height, and 12.994754% for distance from chair.

Network Performance and Quality of Comfort Predictions for Monitor Comfort

After training of the network for monitor comfort, the developed prediction system could predict 100 out of 100 cases in training sets correctly. The CCR of the validation set is 95.652174%. At the end of the training, the mean of the training set, validation set, and test set of CCR are 100%, 95.65174%, and 91.304348%, respectively. Sample confusion matrix for monitor comfort is given in Table 5. Training speed (iteration/second) was 140.347315. At the end of the iteration number (8477), the desired stop condition has been attained for monitor comfort data.

TABLE 5 Confusion Matrix for Monitor Comfort with Training Set, Validation Set, and Test Set

	Training set			Validation set			Test set
Target/output	Too low	Too high	OK	Too low	Too high	OK	Too low	Too high	OK
Too low	85	0	0	22	0	0	20	1	0
Too high	0	11	0	0	0	0	1	1	0
OK	0	0	2	1	0	0	0	0	0

According to network statistics, the weights of the importance (input importance) are 1.451825% for shoulder to elbow, 1.526486% for elbow to fingertip, 12.147447% for shoulder to buttock, 12.973352% for eye to buttock, 1.644906% for back to buttock, 7.210242% for knee to foot, 2.312249% for back support height, 0.46942% for pan height, 57.68025% for monitor height, 0.58375% for keyboard height, and 2.000071% for distance from chair.

Network Performance and Quality of Comfort Predictions for Pan Comfort

After training of the network for pan comfort, the developed prediction system could predict 100 out of 100 cases in the training set correctly. The CCR of the validation set is 73.913043%. At the end of the training, the mean of the training set, validation set, and test set of the CCR are 96.938776%, 86.956522%, 78.26087%, respectively. Sample confusion matrix for seat comfort is given in Table 6. Training speed (iteration/second) was 250.734071. At the end of the iteration number (10581), a desired stop condition has been attained.

TABLE 6 Confusion Matrix for Pan Comfort with Training Set, Validation Set, and Test Set

	Training set			Validation set			Test set
Target/output	Too low	Too high	OK	Too low	Too high	OK	Too low	Too high	OK
Too low	48	1	0	11	2	0	14	4	1
Too high	1	19	0	0	4	0	0	1	0
OK	0	1	28	1	0	5	0	0	3

According to network statistics, the weights of the importance (input importance) are 7.85719% for shoulder to elbow, 3.672326% for elbow to fingertip, 1.311055% for shoulder to buttock, 4.192335% for eye to buttock, 5.379389% for back to buttock, 39.527461% for knee to foot, 8.950568% for back support height, 8.216675% for pan height, 9.659929% for monitor height, 5.055615% for keyboard height, and 6.177458% for distance from chair.

Descriptive statistics, together with a one-way ANOVA (degree of significance was set to 0.05) were used to determine which of the inputs could be used to evaluate the comfort of the computer workstations. For the one-way ANOVA, the independent variables were (1) back support comfort, (2) distance comfort, (3) monitor comfort, (4) keyboard comfort, and (5) pan comfort. Dependent variables were (1) shoulder to elbow, (2) elbow to fingertip, (3) shoulder to buttock, (4) eye to buttock, (5) back to buttock, (6) knee to foot, (7) back support height, (8) pan height, (9) monitor height, (10) keyboard height, and (11) distance from chair.

Data from 144 workstations were used to develop a multiple linear regression model with a Minitab Inc. (Minitab Inc., 2006, PA, USA) statistical software package. The relationship between each of the 11 input variables and back support comfort, distance comfort, monitor comfort, keyboard comfort, and pan comfort (output variables representing comfort) was examined using Pearson product moment correlation coefficients (Table 7). In terms of back support comfort, only six of the input variables (shoulder to buttock, eye to buttock, back to buttock, knee to foot, back support height, keyboard height) were statistically related to the back support comfort. Only three of the input variables (shoulder to elbow, elbow to fingertip, and distance from chair) were statistically related to the distance comfort. For keyboard comfort, only five of the input variables (shoulder to buttock, eye to buttock, back to buttock, pan height, keyboard height) were statistically related. Only six of the input variables (shoulder to elbow, shoulder to buttock, eye to buttock, back to buttock, knee to foot, monitor height) were statistically related to monitor comfort and finally only four of the input variables (knee to foot, pan height, keyboard height, distance from chair) were related to pan comfort. The significance level is 0.05 for p value. The result obtained was that input variables were linearly related to back support comfort, distance comfort, keyboard comfort, monitor comfort, and seat comfort.

TABLE 7 Relationship Between Predictor Variables and Comfort Variables

	Back support		Distance		Keyboard		Monitor		Pan
Predictor variable	r	p	r	p	r	p	r	p	r	p
Shoulder to elbow	−0.079	0.348	0.257	0.002	−0.082	0.330	−0.205	0.014	−0.047	0.574
Elbow to fingertip	−0.125	0.135	0.201	0.016	−0.028	0.741	0.003	0.975	−0.012	0.887
Shoulder to buttock	−0.312	0.000	0.075	0.374	0.324	0.000	−0.375	0.000	−0.071	0.397
Eye to buttock	−0.163	0.051	0.144	0.085	0.180	0.031	−0.404	0.000	−0.078	0.355
Back to buttock	−0.209	0.012	−0.046	0.581	0.354	0.000	−0.292	0.000	−0.111	0.184
Knee to foot	−0.388	0.000	0.049	0.561	0.141	0.091	−0.269	0.001	−0.540	0.000
Back support height	0.259	0.002	−0.051	0.546	0.119	0.155	−0.140	0.095	−0.028	0.742
Pan height	0.022	0.795	0.085	0.312	0.179	0.032	0.076	0.366	0.184	0.027
Monitor height	0.002	0.977	−0.108	0.196	0.095	0.258	0.381	0.000	0.082	0.327
Keyboard height	−0.214	0.010	−0.097	0.248	−0.162	0.052	−0.108	0.200	−0.228	0.006
Distance from chair	0.062	0.463	−0.534	0.000	0.041	0.627	0.127	0.128	0.339	0.000

It is possible to develop a multiple linear regression model with significant input variables (p ≤ 0.05) for all comfort variables. The linear model of back support comfort and the rest of the comfort variables can be expressed, respectively, as follows:

The outputs from Minitab for the our prediction problem are presented in Table 8. Examination of these outputs lead to the following observations. In terms of back support comfort, the standard error of the estimate is 0.540027 for back support comfort, 0.438777 for distance comfort, 0.775927 for keyboard comfort, 0.419843 for monitor comfort, and 0.616084 for pan comfort. These values are a measure of the amount that the actual values differ from the fitted values. The regression equation explains 40.3% of the variation in back support comfort, 15.6% of the variation in distance comfort, 23.1% of the variation in keyboard comfort, 50.9% of the variation in monitor comfort, and 37.4% of the variation of pan comfort (Table 8). The regression slope coefficients of all the comforts were tested to determine whether they were different from zero. The computed F values for all comforts (Table 8) are used to test for the significance of the regression. The large F ratio and the small p value show that the regression is significant. In this case, the t statistic of −2.01 for shoulder to buttock variable and its p (0.047) value indicate the coefficient of shoulder to buttock is not significantly different from zero (accept H _o : β1 = 0). The small t statistic of −0.95 for eye to buttock variable and its large p (0.345) value indicate the coefficient of eye to buttock is not significantly different from zero. The t statistics of −0.75 for back to buttock and its large p (0.456) value indicate that the coefficient of eye to buttock is not significantly different from zero. The large t statistics of −5.30 of knee to foot and its small p (0.000) value indicate the coefficient of the knee to foot is significantly different from zero (reject H _o : β₁ = 0). The large t statistics of 6.56 of back support height and its small p (0.000)

TABLE 8 One-Way ANOVA for Comfort Inputs

	Source	Sum of squares	df	Mean square	F	Sig.
Back support	Regression	26.9843	6	4.4974	15.42	0.000
	Residual error	39.9532	137	0.2916
	Total	66.9375	143
S = 0.540027 R-Sq = 40.3% R-Sqadj = 37.7%
Distance	Regression	4.9840	3	1.6613	8.63	0.000
	Residual error	26.9535	140	0.1925
	Total	31.9375	143
S = 0.438777 R-Sq = 15.6% R-Sqadj = 13.8%
Keyboard	Regression	24.9084	5	4.9817	8.27	0.000
	Residual error	83.0846	138	0.6021
	Total	107.9931	143
S = 0.775927 R-Sq = 23.1% R-Sqadj = 20.3%
Monitor	Regression	25.0110	6	4.1685	23.65	0.000
	Residual error	24.1487	137	0.1763
	Total	49.1597	143
S = 0.419843 R-Sq = 50.9% R-Sqadj = 48.7%
Pan	Regression	31.4634	4	7.8659	20.72	0.000
	Residual error	52.7588	139	0.3796
	Total	84.2222	143
S = 0.616084 R-Sq = 37.4% R-Sqadj = 35.6%

value indicate the coefficient of the back support height is significantly different from zero. The t statistics of −1.87 for keyboard height and its p (0.063) value indicate that the coefficient of keyboard height is not significantly different from zero. In terms of distance comfort, it has been seen that the coefficients of elbow to fingertip and the distance from that chair are significantly different from zero. The coefficients of back to buttock and keyboard height in the regression equation for keyboard comfort are significantly different from zero. The coefficients of back to buttock, knee to foot, and monitor height are significantly different from zero in terms of monitor comfort and finally the coefficients of knee to foot, pan height, and distance from chair are significantly different from zero. In summary, the coefficients of shoulder to buttock, eye to buttock, back to buttock, and keyboard height can be dropped from the regression function of back support comfort. The coefficient of shoulder to elbow can be dropped from the regression function of distance comfort. The coefficient of shoulder to buttock, eye to buttock, and pan height can be dropped from the regression function of keyboard comfort. The coefficient of shoulder to elbow, shoulder to buttock, and eye to buttock can be dropped from the regression function of monitor comfort. The coefficient of keyboard height can be dropped from the regression function of pan comfort.

It has been seen that some of the relationships were nonlinear. This became apparent when the input variables were plotted against the comfort variables in a separate exercise. This was not a surprising result. Since one-to-one relationships were not as straightforward as one would hope, the data were assumed to be beset with several important interactions and levels of nonlinearity. This triggered the application of a neural network approach.

Whereas there are five comfort indicators, only pan comfort prediction results were shown in Tables 9-11. There were many results (data) of the neural system-based prediction of the remaining comfort indicators for training set, validation set, and test set.

TABLE 9 Results of the Neural System-Based Prediction of Pan Comfort for Training Set

							Back
	Shoulder to	Elbow to	Shoulder to	Eye to	Back to	Knee to	support	Pan	Monitor	Keyboard	Distance	Class of	Model
User	elbow	fingertip	buttock	buttock	buttock	foot	height	height	height	height	from chair	comfort	prediction
0	34	40	57	75	27	52	71	42	117	78	61	Too low	Too low
1	33	42	60	76	30	54	74	42	116	77	59	Too low	Too low
2	32	41	49	68	26	54	75	47	116	76	60	Too low	Too low
5	31	37	60	77	25	53	70	43	117	79	63	Too low	Too low
6	32	35	57	73	26	65	69	43	117	79	45	Too low	Too low
7	29	37	53	74	29	55	72	43	115	77	49	Too low	Too low
8	35	38	50	66	25	64	67	43	124	77	47	Too low	Too low
10	29	44	54	67	23	57	63	42	111	77	64	Too low	Too low
11	27	36	49	66	27	46	75	48	112	77	58	Too low	Too low
13	31	39	56	79	22	55	63	41	113	76	48	Too low	Too low
14	31	45	55	82	28	45	65	39	108	71	69	Too low	Too low
16	33	41	57	73	29	47	65	41	111	73	67	Too low	Too low
17	35	43	61	77	29	49	63	41	107	71	67	Too low	Too low
18	37	43	64	75	33	47	67	41	108	73	71	Too low	Too low
19	33	41	63	76	27	49	71	44	114	77	73	Too low	Too low
21	37	43	63	75	31	45	63	43	109	75	69	OK	OK
22	35	45	67	81	29	43	69	45	111	79	71	OK	OK
25	34	29	54	77	20	40	72	46	100	74	66	Too high	Too high
26	38	25	52	77	17	48	65	52	110	74	75	Too low	Too low
27	29	26	53	74	24	35	67	41	100	77	67	Too high	Too high
28	25	24	65	76	26	39	70	65	108	60	77	Too high	Too high
29	35	30	55	78	22	41	73	45	100	74	66	Too low	Too low
30	31	36	65	77	29	47	63	45	109	73	71	OK	OK
33	37	30	61	79	27	45	67	43	109	75	73	OK	OK
34	35	48	54	68	28	43	74	42	107	72	62	Too low	Too high
35	36	47	52	66	21	46	70	42	119	74	60	Too high	Too high
36	38	48	56	78	32	47	76	42	120	75	72	Too high	Too high
39	39	50	58	78	34	52	78	42	113	70	73	Too high	Too low
40	34	45	53	66	22	43	68	42	114	70	64	Too high	Too high
41	35	47	56	70	28	46	69	42	110	70	68	Too high	Too high
43	34	46	53	68	24	44	68	42	116	70	66	Too high	Too high
45	35	45	67	81	29	44	71	49	112	81	60	Too low	Too low
46	31	41	65	72	29	47	76	64	109	83	55	OK	OK
47	34	44	56	80	25	44	75	57	105	79	61	OK	OK
48	37	40	64	75	35	47	73	48	106	78	58	Too low	Too low
50	37	47	61	77	27	45	67	44	120	75	44	OK	OK
52	31	37	60	75	28	53	69	41	118	71	38	Too low	Too low
53	35	44	61	72	29	50	77	41	117	74	47	Too low	Too low
54	38	48	65	83	30	46	71	48	108	76	78	Too high	Too high
55	34	46	53	70	21	46	63	45	106	76	78	OK	OK
56	31	44	53	70	24	43	64	41	112	74	69	OK	OK
58	33	44	51	78	25	40	62	41	85	67	65	OK	OK
60	33	46	60	67	37	54	70	50	125	73	67	Too low	Too low
61	29	42	52	78	24	38	67	47	127	70	65	Too high	Too high
64	30	43	56	82	28	44	64	40	107	71	69	Too low	Too low
65	31	40	59	76	27	53	69	41	108	71	65	Too low	Too low
68	36	43	64	75	33	47	67	41	108	73	71	Too low	Too low
69	34	41	63	76	27	49	71	44	114	77	73	Too low	Too low
70	32	41	65	77	29	47	63	45	109	73	71	OK	OK
71	35	43	63	75	31	45	63	43	109	75	69	OK	OK
72	36	45	67	81	29	43	69	45	111	79	71	OK	OK
73	36	47	61	79	27	45	67	43	109	75	73	OK	OK
74	35	47	57	77	25	48	72	44	104	78	57	Too low	Too low
77	37	49	60	78	23	47	73	45	95	60	43	Too low	Too low
80	39	52	62	80	25	49	66	45	103	77	41	Too low	Too low
81	37	49	60	79	23	47	66	45	104	77	49	Too low	Too low
82	38	54	58	75	23	46	66	45	104	77	56	Too low	Too low
83	35	45	60	76	24	46	66	45	103	76	61	Too low	Too low
85	40	42	63	73	37	59	70	50	125	73	62	Too low	Too low
87	46	49	61	83	37	54	72	49	123	86	65	Too low	Too low
88	27	42	55	66	24	38	67	47	127	70	61	Too high	Too high
89	36	41	50	66	25	48	68	49	118	69	75	OK	OK
90	34	46	53	70	21	46	63	45	106	76	78	OK	OK
92	45	49	54	76	25	44	76	42	99	73	51	OK	OK
93	40	47	56	78	27	44	76	42	99	73	73	OK	OK
95	30	39	56	69	23	48	68	45	100	73	69	Too low	Too low
96	29	31	48	63	26	46	68	45	100	73	57	OK	OK
97	33	37	53	70	26	48	72	45	100	73	67	Too low	Too low
98	25	34	41	57	19	43	68	45	100	73	56	OK	OK
101	30	40	55	69	25	48	67	41	104	76	60	Too low	Too low
103	30	41	50	69	21	36	59	41	104	76	65	Too high	Too high
104	28	43	53	74	33	41	77	46	111	75	70	Too high	Too high
105	32	50	57	77	42	47	70	40	116	76	84	Too low	Too low
106	29	34	59	88	33	50	87	46	118	75	55	Too low	Too low
109	35	48	66	81	32	44	72	49	109	75	79	Too high	Too high
110	28	36	57	84	35	51	85	45	119	72	60	Too low	Too low
112	34	43	56	73	28	48	75	51	117	73	68	Too high	Too high
113	30	42	57	71	26	43	62	41	107	70	61	OK	OK
114	38	48	65	83	30	46	71	48	108	76	78	Too high	Too high
116	31	44	53	74	24	43	64	41	112	74	69	OK	OK
118	33	44	51	71	25	40	62	41	85	67	65	OK	OK
119	30	40	50	67	25	48	68	49	121	75	83	OK	OK
120	33	46	60	78	37	54	70	50	125	73	67	Too low	Too low
121	29	42	52	69	24	38	67	47	127	70	65	Too high	Too high
122	33	46	60	80	37	54	72	49	123	86	75	Too low	Too low
123	40	52	66	84	33	47	73	45	122	76	60	Too high	Too high
126	36	48	59	79	24	52	73	44	103	76	57	Too low	Too low
127	33	45	55	71	22	44	66	45	104	77	57	Too low	Too low
129	40	52	63	82	23	56	78	45	104	79	57	Too low	Too low
130	34	47	54	76	25	49	68	45	103	77	57	Too low	Too low
132	32	40	51	68	20	42	64	45	104	75	56	Too low	Too low
134	37	46	65	81	28	45	71	43	102	74	75	Too low	Too low
135	35	47	55	72	23	47	63	45	103	74	70	OK	OK
136	38	47	57	77	26	44	77	43	101	72	73	OK	Too high
138	32	45	54	74	25	43	64	40	112	74	69	OK	OK
140	30	40	50	67	25	46	68	49	118	85	83	OK	OK
142	33	44	51	71	25	40	62	41	100	73	70	OK	OK
143	29	38	49	70	17	36	61	40	99	73	65	Too high	Too high

TABLE 10 Results of the Neural System-Based Prediction of Pan Comfort for Validation Set

							Back
	Shoulder to	Elbow to	Shoulder to	Eye to	Back to	Knee to	support	Pan	Monitor	Keyboard	Distance	Class of	Model
User	elbow	fingertip	buttock	buttock	buttock	foot	height	height	height	height	from chair	comfort	prediction
4	30	33	52	68	26	49	71	44	117	78	60	Too low	Too low
9	33	40	59	79	31	40	74	43	113	77	66	Too high	Too high
20	31	41	65	77	29	47	63	45	109	73	71	OK	Too low
24	31	26	54	73	22	39	69	66	108	60	77	Too high	Too high
32	35	30	67	81	29	43	69	45	111	79	71	OK	OK
38	38	46	56	74	24	45	72	42	115	74	69	Too high	Too high
44	37	44	61	75	31	45	61	58	114	72	43	OK	OK
51	35	41	61	76	27	50	71	44	114	72	41	Too low	Too low
57	40	47	56	74	27	44	76	42	99	73	73	OK	OK
63	40	52	66	80	33	47	73	45	122	76	60	Too high	Too high
67	36	43	61	77	29	49	63	41	107	71	67	Too low	Too low
76	39	51	62	81	23	54	75	44	103	78	57	Too low	Too low
79	34	48	56	75	23	45	66	45	102	76	58	Too low	Too low
86	33	44	53	73	22	40	58	35	114	63	52	Too low	Too high
94	33	36	60	82	29	52	70	46	101	74	60	Too low	Too low
100	29	36	51	70	24	51	68	46	104	76	40	Too low	Too low
107	29	45	57	78	31	58	72	41	106	78	71	Too low	Too low
111	38	41	62	73	35	58	71	50	121	74	66	Too low	Too low
117	40	47	56	78	27	44	76	42	99	73	73	OK	OK
125	39	51	62	81	23	54	75	44	103	78	57	Too low	Too low
128	32	41	54	74	22	47	66	45	104	77	59	Too low	Too low
137	29	40	58	79	27	43	70	45	105	79	80	OK	OK
141	30	41	51	70	17	38	69	49	118	85	83	Too low	Too high

TABLE 11 Results of the Neural System-Based Prediction of Pan Comfort for Test Set

							Back
	Shoulder to	Elbow to	Shoulder to	Eye to	Back to	Knee to	support	Pan	Monitor	Keyboard	Distance	Class of	Model
User	elbow	fingertip	buttock	buttock	buttock	foot	height	height	height	height	from chair	comfort	prediction
3	42	43	50	71	20	51	60	43	115	77	57	Too low	Too low
12	31	36	56	69	31	54	76	48	112	77	57	Too low	Too low
15	33	39	60	75	28	53	69	41	108	71	65	Too low	Too low
23	37	47	61	79	27	45	67	43	109	75	73	OK	OK
31	36	29	54	73	22	41	73	44	105	75	70	Too low	Too high
37	37	49	55	74	28	45	74	42	115	74	68	Too low	Too high
42	40	48	56	72	30	45	71	42	115	70	68	Too low	Too high
49	34	41	56	72	29	47	65	41	109	72	42	Too low	Too low
59	30	40	50	71	25	48	68	49	121	75	83	OK	OK
62	33	46	60	69	37	54	72	49	123	86	75	Too low	Too low
66	32	41	57	73	29	47	65	41	111	73	67	Too low	Too low
75	36	46	56	78	24	48	70	46	95	60	42	Too low	Too low
78	32	41	54	74	22	47	66	45	104	77	59	Too low	Too low
84	46	49	61	83	37	54	72	49	123	86	49	Too low	Too low
91	36	47	53	66	21	46	70	53	119	74	61	Too high	Too high
99	33	43	64	77	33	50	74	46	104	80	67	Too low	Too low
102	32	40	54	72	26	49	71	41	104	76	64	Too low	Too low
108	33	55	56	74	27	44	60	38	105	64	61	Too low	Too high
115	34	46	53	70	21	46	63	45	106	76	78	OK	OK
124	35	47	57	77	25	48	72	44	104	78	57	Too low	Too low
131	33	42	53	70	21	44	64	45	104	77	49	Too low	Too low
133	33	42	50	65	21	42	64	45	103	76	57	Too low	OK
139	35	47	62	80	37	54	72	50	118	85	83	Too low	Too low

CONCLUSION AND DISCUSSION

This research found that an artificial neural network could be used to adequately predict computer user comfort with respect to the existing settings of a workstation. In addition to an artificial neural network, a multiple linear regression model was developed to predict computer user comfort. The neural network was deemed superior to the regression model because it explained more of the variance in comfort with a lower average error. The neural network's ability to deal with interaction effects is offered as the principle reason for its superior performance.

The neural network, which considers a larger number of inputs, is also more useful to settings of a computer workstation because constraints that are an inevitable part of the design process are less limiting. That is, the more options design teams (ergonomists) have for how to improve comfort, the better. Note also that the regression model is inadequate in explaining the comfort of a computer user. The neural network considers all the existing settings of the workstation. For this reason, the neural network is capable of determining how a particular set of inputs will affect a more focused subset of a target population.

Based on the neural networks, the most important predictors were back support height for back support comfort, distance from chair for distance comfort, keyboard height for keyboard comfort, monitor height for monitor comfort, and knee to foot for pan comfort. This raises an interesting discussion topic in that how and why these dominant measurements effect the comfort of computer users.

Based on the multiple linear regression model, it has been seen that some physical measurements are more active and significant than others. For example, shoulder to elbow, elbow to fingertip, and distance from chair are significant in predicting distance comfort.

Having established that a neural network can be used to predict comfort of a computer user, future research should focus on the derivation of an optimized set of inputs (i.e., a set of inputs that will result in maximum comfort). These can be done with optimization algorithms. Once completed, the findings could be published in the form of computer workstation comfort design guidelines/standards. Perceptions of computer workstation comfort are constantly changing. A comfortable computer workstation from 1980, for example, would probably not be considered comfortable today. This suggests that prediction models, in order to remain useful, will need to be periodically updated. Ergonomists, computer users, and computer suppliers will play a vital role in specifying new computer workstation designs.

In order to protect against computer-related health problems and injuries, employers need to be sure that their computer users are aware of the hazards of computer use and have the skills and tools to set their workstations up correctly and to work safely. The neural network-based system can be used specifically to prevent computer fatigue and computer-related health problems. It is also beneficial to assist ergonomic positioning of your computer and prediction of the health risk in an occupational environment. To the best of our knowledge, this article describes the first neural network-based approach applied to analyzing the existing settings of a computer workstation and user.