Can a Novel Web-Based Computer Test Predict Poor Simulated Driving Performance? A Pilot Study With Healthy and Cognitive-Impaired Participants

Introduction

Cognition and Driving

Driving a car is a very challenging instrumental activity of daily living that requires the integration of high-level cognition, vision, and motor function [1]. These three domains are usually evaluated when assessing medical fitness to drive in older drivers, as they are commonly affected by age-related diseases [2]. In this article, we focus on the assessment of driving-relevant cognition in older drivers.

Since driving is a complex activity, driving performance is very sensitive to cognitive impairment [3], which is commonly the result of age-related neurodegenerative disorders (eg, Alzheimer’s disease and other causes of dementia) [4]. The prevalence of neurodegenerative disorders doubles every five years after the age of 65 years [5]. Therefore, health professionals need easy access to screening tests in order to assess fitness to drive. Due to the ageing population in the Western world and increasing numbers of older drivers, identifying drivers at risk without unnecessarily restricting others is a challenging but important task [6]. This task requires reliable and valid cognitive screening tests with good sensitivity and specificity to identify at-risk drivers.

Testing Driving Performance

On-the-road testing (ORT) has been suggested as being a reasonable proxy measure for naturalistic driving in older adults with a range of cognitive impairments [7]. It is the gold standard for measuring driving performance and several authors suggest using it to assess fitness to drive in older drivers [7-10]. Despite its advantages, ORT has limitations: it is time consuming [11] and may have adverse effects that could lead to dangerous driving situations [12]. In addition, researchers cannot control for environmental conditions such as light, weather, traffic, and pedestrians [13].

That is why more recently, driving simulators (DS) have been recommended as a proxy measure for naturalistic driving and they have been introduced to assess fitness to drive of older drivers with and without cognitive impairment [14,15]. Simulators have the advantage of being intrinsically safe, providing excellent controllability, reproducibility, and standardization. Furthermore, they can be installed in specialist centers, their use is less time consuming, and they require fewer organizational demands than ORT. Several studies demonstrate the validity of DS as a proxy for naturalistic driving [13,16-19]. Disadvantages of DS are that they are expensive and large, both of which reduce their accessibility to primary care physicians and older drivers. Furthermore, there is a lack of standard test protocols and cutoff values. Finally, simulator sickness is a rather common side effect, especially for older female drivers. This interferes with DS driving performance [20].

Web-Based Computer Testing of Driving Performance

In a recent study, Rockwood et al [21] used a Web-based dataset to determine the level of cognitive impairment of older Internet users. They concluded that online tracking of people with cognitive impairment can be used to stage dementia. For review, see [22]. Also, when measuring driving performance, some of the limitations of DS can be resolved by the introduction of computer tests. By using standard personal computers with cheap off-the-shelf interface components similar to those that are used in computer gaming, cheap and easily accessible computer tests can be implemented [23]. Compared to DS, they can be easily integrated into the physician’s office. Web-based computer tests use the World Wide Web to distribute the software that can run on local client computers in the physician’s office. Moreover, with Web-based computer tests, the scoring of user performance can be conducted on a central server computer. This allows for generation of a dynamically growing database with normative values and ensures that all users have the same up-to-date normative values. This concept has been successfully introduced by Mills et al [24] for the application of a Web-based computer test to assess driving performance under alcohol and drug influence. The authors mention central scoring algorithms and a central database with normative values as the main advantage of Web-based computer tests. That is why we hypothesized that Web-based computer tests would also be helpful to assess fitness to drive of older drivers and we have developed the Web-based Bern Cognitive Screening Test (wBCST) to assess driving-relevant cognitive performance. The novel wBCST is based on a previously developed computer test [25].

In this pilot study, we investigate whether or not the novel wBCST correlates with DS performance and whether it is able to differentiate between participants with poor and good simulated driving performance. To have a broad and diverse test population, we recruited younger and older healthy participants as well as older participants with cognitive impairment for this study. Hence, this paper first describes the novel wBCST and the DS used, followed by a correlation analysis and a receiver operating characteristics (ROC) [26] analysis to calculate sensitivity and specificity of the novel test to predict driving simulator performance. The discussion and conclusion outline advantages and disadvantages of the wBCST and present possible future applications and research directions.

Methods

Participants

Thirty healthy younger adults (age 22-40 years), 60 healthy older adults (age >50 years), and 15 older (age >50 years) participants with cognitive impairment (Montreal Cognitive Assessment Score [MoCA] <26) [27] were recruited by advertisements in local newspapers and within the local memory clinic. All participants were required to have had a driver’s license for at least two years and to have been driving during the last two years. Exclusion criteria for the study were visual impairment (corrected far visual acuity <0.5 degrees, near visual acuity <0.8 degrees) or significant motor impairment (timed-up-and-go-test >12 seconds) [28]. The study was carried out in accordance with the Declaration of Helsinki and was approved by the local ethics board. Written informed consent was obtained from all participants prior to inclusion. No compensation for participation was provided. Six participants were excluded due to visual impairment. Eighteen participants stopped the DS drive due to simulator sickness (four younger, nine older, and five participants with cognitive impairment). Their data were excluded from further analysis. Due to a malfunctioning of the DS, the data of one healthy older test person were not recorded. The data of the remaining 80 participants were included in the analysis. There were 26 young (range 22-39 years, mean 29.4 years, SD 4.7 years), 44 healthy older (range 54-85 years, mean 68.4 years, SD 5.5 years), and 10 impaired older participants (range 55-87 years, mean 72 years, SD 9.6 years). Trail Making Test A (TMT-A) [29], Trail Making Test B (TMT-B) [29], the MoCA score [27], clock-drawing test (CDT) [30], and timed-up-and-go test [28] were assessed to characterize the participants.

Web-Based Computer Test

A literature review of the most important driving-relevant cognitive functions and how they are affected by cognitive impairment was conducted [6]. This analysis was used to develop a novel computer test [25], which we extended in the context of this study to the wBCST. The wBCST measures eye-hand coordination, selective attention, divided attention, executive function, distance judgment, and speed regulation. It is composed of five subtests, each measuring one of the before-mentioned cognitive competencies. The setup of the wBCST comprises a computer screen showing the test scenario (240B1CS/00 24 inc, Philips Inc), a steering wheel (Driving Force GT, Logitec Inc) with foot pedal, and a personal computer with Windows 7 (Microsoft Inc) operating system. Figure 1 shows a healthy test person taking the wBCST and screenshots of the visual representation of the five subtests. The test persons interact with the wBCST via steering wheel and foot pedal.

Subtest 1 measures selective attention with a visual scene consisting of a simplified street in the center of the screen (Figure 1d) with objects moving from the top down. A red dot moves automatically in the horizontal direction to avoid collisions with oncoming objects and with the roadside. The user is instructed to not touch the steering wheel, but to press the foot pedal whenever a visual target (blue square) appears in the periphery (Figure 1d). In subtest 2, eye-hand coordination is measured and the same street is presented. The test person must use the steering wheel to control the horizontal position of the red dot to avoid collisions with the other objects and with the roadside (Figure 1e). There is no peripheral subtask in this test. Subtest 3 measures divided attention and both tasks of subtests 1 and 2 need to be carried out simultaneously; hence, the user must steer (central task) and react to peripheral stimuli (peripheral task) (Figure 1f). Subtest 4 is designed to measure executive functions and the user must react to more complex peripheral stimuli (green triangle and blue square) while ignoring the movement in the center of the screen (Figure 1g). In subtest 5, distance judgment and speed regulation are assessed. The user can control the velocity of a red dot with the foot pedal. The task as shown in Figure 1h is to cross intersections without colliding with the horizontally moving objects. During the tests, the false positive and false negative errors of the peripheral target detection task and the number of collisions with moving objects as well as with the street border are recorded. The test duration is about three minutes per subtest. With instruction, administration of the entire wBCST takes roughly 20 minutes. For videos of the subtests, see Multimedia Appendices 1 and 2.

‎View this figure

Drive in the Virtual Reality Driving Simulator

A commercially available high-fidelity fixed-frame driving simulator (F12PI-3/A88, Foerst GmbH) with a custom-built virtual driving circuit was used to measure DS performance. A virtual scene was projected by three projectors (Ultra-Short focus LCD projectors, Sanyo) with 1024x786 pixel resolution onto three projection screens (1.8 x 1.4 m) that were installed in front of the driver. DS components utilized by the participants were steering wheel, brake and gas pedals, rear and side mirrors, and turn signals (Figure 2). The virtual driving scene consisted of a street with two lanes in each direction in a suburban environment. The test drive included two left turns at intersections: one with traffic crossing left to right and one with oncoming traffic. Furthermore, it comprised a construction area with road work on both lanes, a roundabout with an unexpected cyclist, a deer hiding behind trees and suddenly jumping onto the road, a car unexpectedly leaving its parking lot, and a child running into the street after a ball. The instruction was to respect traffic rules and to drive as if in a real-world environment.

Participants performed a familiarization run (three minutes) to get used to the handling of the simulator and a test drive (six minutes) during which data were recorded. Once the end of the track was reached, the car stopped automatically. If a participant felt uncomfortable, the DS was stopped. Primary outcome measure of the driving simulator was the number of errors, E_DS (ie, collisions, traffic rule violations, driving in the wrong traffic lane). Secondary outcome measures were mean speed variability, mean lateral acceleration, cumulated time spent on brake, distance to collision, and time to collision. Distance and time to collision are defined by how long (time, respective distance) the virtual car could continue on its current path with constant velocity until a collision would occur.

‎View this figure

Statistical Analysis

The results of the wBCST were ranked and scored. Thus, the score, S_i, of each subtest was calculated as S_i= (N−rank(E_i)) / (N−1) with N being the total number of test samples and E_i being the number of errors in the subtest. With this formula, 1 corresponds to the best possible test results and 0 the worst. Subtest 3 reveals two results, S_3p for the peripheral task and S_3c for the central task. The overall test result, S_wBCST, was calculated as the mean value of S₁ to S₅. The same formula was also used to score the performance in the DS—namely, the score, S_DS, that was calculated out of E_DS, as well as the scores for speed variability, lateral acceleration, time on brake, distance to collision, and time to collision.

Pearson product-moment correlations, step-wise regressions, and associated tests of significance were calculated across S_DS, S_wBCST, and the other secondary measures. The statistical significance of the correlation was computed by transforming the correlation matrix to create a t statistic having n-2 degrees of freedom where n was the number of observations.

In a second step, the number of errors E_DS in the driving simulator was used as a classification criterion to divide the participants into two groups. As suggested by others [31], the mean value plus 1 SD was used as cutoff for one group with good DS performance and another group with poor DS performance. The mean values of the DS and the wBCST for both groups were calculated and the significance of the differences was calculated using a non-parametric Mann-Whitney U-test. Hypotheses were one-sided. Furthermore, a ROC analysis was conducted to determine the sensitivity and specificity of the wBCST when predicting DS performance. A Matlab (The MathWorks Inc) algorithm based on the work of Hanley et al [32] and implemented by Cardillo et al [33] was used to calculate the ROC curve and the area under the ROC curve (AUC). The Wilcoxon test was used to calculate whether the difference from a random classifier was significant. Also, the sensitivity, specificity, and efficiency (fraction of subjects that are correctly classified) were computed for different cutoff values.

Results

User Statistics

The data of 80 participants were included in the data analysis. Ten were assigned, based on the number of errors in the DS using mean value + 1 SD cutoff, to the poor DS performance group (E_DS≥4) and 70 were assigned to the good DS performance group (E_DS<4). With this criterion, all younger participants, 86% (38/44) of the healthy older, and 60% (6/10) of the older participants with cognitive impairment were classified into the good DS group. The mean test performance in MoCA, TMT-A, TMT-B, CDT, and timed-up-and-go-test of the participants with good DS results was better compared to the other group. These differences were statistically significant.

Correlation Analysis of wBCST and DS Performance

Figure 3 shows the results of the correlation analysis with the correlation coefficient r and the associated P values. The wBCST score, S_wBCST, correlates with the DS score, S_DS, with r=0.32 (P=.004). The individual results of the wBCST subtest, S_1…6, correlate with the overall score, S_wBCST, with r values varying between 0.68-0.83 with P<.001. In the DS, the overall score S_DS correlates with the speed variation with r=0.38 (P<.001), the lateral acceleration with r=0.40 (P<.001), the time spent on the brake pedal r=0.13 (not significant, P=.24), the distance to collision with r=0.19 (not significant, P=.19), and the time to collision r=0.31 (not significant, P=.23).

Sensitivity and Specificity of the wBCST in Predicting DS Performance

The ranked normalized wBCST performances of the two groups are represented in Figure 4. The group with poor DS performance performed less well in all tests. The group differences are significant for all subscores and highly significant (P<.001) for the overall S_wBCST score, subtest 2 (S₂), and subtest 5 (S₅).

The ROC curve for using S_wBCST to predict DS performance is shown in Figure 5. The AUC=0.80 is significantly better (P<.001) than a random classifier. The 95% confidence interval of the AUC is 0.68-0.92. A selection of possible cutoff scores and the corresponding sensitivity, specificity, and efficiency values is presented in Table 2.

‎View this figure

‎View this figure

‎View this figure

Discussion

Principal Results

When using the criteria proposed by Cohen [34], the correlation between S_DS and the total score S_wBCST has a medium effect size (r>0.3) and is significant with P=.004, which supports the hypothesis that the wBCST correlates with DS performance. Very good correlations with large effect sizes (r>0.5) were found among the five subtests of the wBCST. This is an important result for the novel test since it indicates good consistency among the wBCST subtests. It could be explained by the uniform visual stimulation material that is used in the subtests (Figure 1). The correlation of the main outcome measure, S_DS, and the secondary outcome measures is medium and statistically significant for the speed variation and the lateral acceleration, but only small and not significant for the time spent on brakes, distance to collision, and time to collision. This was also reported by others [35]. However, it is worthwhile to emphasize here that the novel tests show a good correlation of S_wBCST with existing paper-pencil-based cognitive screening tools, such as the MoCA score, the TMT-A, and the TMT-B, but not with the CDT.

The group differences in the wBCST test and its subscores are all significant, which is a prerequisite for using the wBCST to predict DS performance. The AUC of the ROC curve is slightly larger than 0.80, which is generally considered a good test [33,36]. The ROC curve (Figure 5) and the corresponding table with sensitivity, specificity, and efficiency values allows for selecting a cutoff score that leads to the desired test properties. As mentioned in the introduction, the test should have both a high sensitivity and specificity. When considering Table 2, an appropriate cutoff with a high efficiency could be 0.75, which would lead to 83% sensitivity, 70% specificity, and 81% efficiency. These values are far from perfect, but are within what is to be expected for a test measuring multifaceted characteristics of the cognitive ability to drive safely [37-43].

In the present study, the overall test result, S_wBCST, is calculated simply as the mean of the subscores, S_1…5. It could be that one subtest is more informative than another and, in this case, the subscores should be weighted differently. The sample size of this study is too small to determine the optimal weighting parameters, but this is an interesting question for a future study.

Overall, the novel wBCST was very well accepted by the test population and there were no drop-outs in the wBCST. This is in contrast to the DS with 19 drop-outs (19.1%) due to simulator sickness (18 drop-outs) and technical problems (1 drop-out). Since it is Web-based, the distribution of the test software to different test-sites will be fairly easy and with the central scoring algorithm, data consistency among the different centers can be ensured, which will facilitate validation studies in larger populations.

The online instructions of the test procedure take about 5 minutes. We did not observe difficulties of the test persons to understand the task, except two cognitively impaired test subjects wanted to go through the instructions twice. We concluded that the instructions are clear, but participants should have the option to repeat the instructions.

The steering wheel and the foot pedal are fairly cheap accessories, but to further improve accessibility of the novel test it would be beneficial if the test could also be used with keyboard and mouse. We observed that steering wheel and foot pedal seem to increase the face-validity of the tests, but we would expect that the measured cognitive functions should be independent from the input modality. This could be investigated in a future study.

Limitations

This pilot study has some limitations that need to be mentioned. One is the large number of drop-outs in the DS that might create a selection bias, since it cannot be excluded that people with poor driving performance might be more prone to simulator sickness. Although the current literature on simulator sickness suggests other contributing factors (eg, age, gender), this cannot be excluded [20]. Another limitation is that there are no published and accepted cutoff values for the DS, which makes the selection of the cutoff for the group differentiation difficult. Compared to the DS evaluation, the administration of the wBCST takes 20 minutes, which is quite long. In future studies, one could investigate whether or not test duration could be shortened. Furthermore, the test-retest and inter-rater reliability should be assessed in a future study.

Conclusions

In this pilot study, the novel wBCST looks like a promising test to support clinicians in fitness-to-drive assessments of older drivers. The Web-based distribution and the scoring on a central computer will facilitate further evaluation of the novel test setup. In its current form, the program requires local installation on a client computer in the physician’s office. This is currently not considered a disadvantage, but, of course, one could transfer the test program to run within a Web browser, which would not require local installation. The hardware requirements of the wBCST are very minor and include an office-type personal computer with Windows 7 operating system and a simple steering wheel (eg, Driving Force GT, Logitec Inc). Overall, when considering disadvantages of DS (costs, simulator sickness, space requirements), the authors believe that in many clinical environments the wBCST is better suited to support physicians in fitness-to-drive assessments than a DS. That is why we expect that in the near future, Web-based computer tests will become a valuable and usable tool for fitness-to-drive assessment in older drivers.