Predicting mild cognitive impairment from simulated driving performance: A preliminary investigation

Abstract

INTRODUCTION

This study aimed to determine whether simulated driving performance in challenging environmental conditions can predict mild cognitive impairment (MCI).

METHODS

MCI (n = 37) and control (n = 42) participants completed a simulated driving scenario with two challenging sections involving inclement weather and navigating an unfamiliar environment. Speed and steering outcomes were assessed.

RESULTS

The inclement weather logistic regression model distinguished cognitive status with more than 75% accuracy. However, the route planning model was unable to significantly distinguish individuals based on cognitive status. Receiver operating characteristic (ROC) analyses revealed a significant area under the curve (AUC) for the inclement weather model, with an optimal cutoff yielding 78.4% apparent sensitivity.

DISCUSSION

The present findings highlight the potential value of simulated driving in challenging conditions as sensitive functional markers of early cognitive impairment. These preliminary findings provide the foundation for naturalistic and longitudinal studies to establish whether driving assessments can meaningfully capture functional changes associated with the onset of cognitive impairment.

Highlights

• Individuals with and without mild cognitive impairment (MCI) completed challenging simulated driving tasks.

• Driving metrics included speed and steering outcomes and route accuracy.

• Cognitive status was predicted with greater than 75% accuracy.

• Simulators may be a clinical tool to support early detection of cognitive decline.

1. INTRODUCTION

Simulated driving is recognized as powerful tool ¹ , ² , ³ that offers a safe, controlled environment to assess individuals with mild cognitive impairment (MCI) in challenging environmental conditions that cannot be safely tested on the road. Simulators allow systematic manipulation of driving scenarios such as introducing inclement weather, dense traffic, distraction, complex maneuvers, sudden braking, and unfamiliar navigation, which may be particularly sensitive to early cognitive changes. ³ , ⁴ , ⁵ In addition, simulators capture specific driving behaviors that may reflect early cognitive decline and future crash risk. ⁶ , ⁷ Moreover, challenging driving conditions, such as inclement weather and unfamiliar navigation may be particularly relevant for individuals with MCI who often avoid driving under these conditions. ⁸ , ⁹ Therefore, this study will apply the strategic use of challenging conditions in driving simulators for early detection of cognitive decline and identifying individuals in the prodromal stages of dementia.

MCI represents a pre‐dementia stage ¹⁰ , ¹¹ characterized by subtle cognitive decline beyond what would be expected based on age and education. ¹² Although individuals with MCI typically remain independent, ¹² they experience deficits in memory, ¹³ executive function, ¹⁴ , ¹⁵ attention, ¹⁶ and visuospatial processing. ¹⁷ These functions are essential for safe driving, including hazard detection, speed regulation, path finding, and navigation. ¹⁸ , ¹⁹ Driving, therefore, represents a sensitive functional indicator in which early, subtle changes in cognitive decline may become apparent.

Evidence from simulator and on‐road studies shows that individuals with MCI demonstrate impairments in turning, lane and speed control, and braking compared with cognitively intact peers. ³ , ⁴ Previous studies highlight the link between driving performance and specific cognitive domains in MCI. ²⁰ , ²¹ For example, memory, attention, executive function, processing speed, and visual perception are key domains linked with fitness to drive in individuals with MCI and Alzheimer's disease. ²⁰ Furthermore, poor working memory is associated with a higher probability of driving cessation among individuals with MCI. ²¹ Although functional independence is largely preserved in MCI, deficits in driving skills can lead to self‐restriction or driving cessation, ⁵ , ²¹ thereby reducing independence and increasing caregiver burden. Prior research has examined how MCI affects simulated driving performance. However, it remains unknown whether simulator‐derived driving behaviors can be used to identify individuals with cognitive impairment and support earlier detection and timely intervention.

RESEARCH IN CONTEXT

1. Systematic review: A targeted literature review found that driving simulators are well suited for assessing driving performance under challenging conditions that cannot be safely evaluated on‐road. Individuals with mild cognitive impairment (MCI) commonly experience driving difficulties that contribute to self‐regulation and driving cessation. However, the utility of driving simulators for identifying individuals in the prodromal stages of dementia remains undefined.

2. Interpretation: Simulator‐derived driving metrics accurately distinguished individuals with MCI from cognitively intact peers, with speed metrics showing the strongest predictive value. These findings provide preliminary evidence that simulated driving performance may serve as a valid and reliable functional assessment tool for cognitive impairment.

3. Future directions: Future work should validate these simulator‐based metrics against on‐road driving performance and across the cognitive aging continuum. Longitudinal studies are needed to determine if these driving behaviors can predict future Alzheimer's disease or dementia.

This study investigates the central question: Can simulated driving performance in challenging conditions reliably predict MCI? To address this question, the predictive value for cognitive classification of driving outcomes from two challenging conditions were evaluated using logistic regression. Because of established sex differences in MCI prevalence ²² , ²³ , ²⁴ and driving behavior, ²⁵ , ²⁶ , ²⁷ we included sex in the models and considered a sex‐only benchmark model. We hypothesized that simulator‐based driving measures would accurately predict MCI, even when accounting for the effects of sex. Clinically, this work aims to provide preliminary evidence that simulator driving behavior during challenging scenarios can serve as objective behavioral markers of cognitive decline. Ultimately these findings may support the potential of simulator driving for MCI detection and inform evaluations of driving safety, warranting future longitudinal investigations to establish their clinical utility.

2. METHODS

2.1. Participants

Thirty‐seven individuals with MCI and 42 healthy controls participated. Participants with MCI were referred by a board‐certified clinical neuropsychologist from a neurology clinic (co‐author J.D.L.) and diagnosed following comprehensive neuropsychological evaluation in accordance with the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM‐5). Specifically, diagnostic criteria included subjective cognitive concern, objective impairment in one or more cognitive domains, preserved functional independence in daily activities based on self‐report and corroborating reports from a knowledgeable informant, and absence of dementia. The cognitive status of healthy controls was verified using the Montreal Cognitive Assessment (MoCA) ²⁸ and the Dementia Rating Scale‐2 (DRS‐2). ²⁹ , ³⁰ These instruments were selected specifically as they are highly validated, gold‐standard screening tools for cognitive impairment.

Demographic characteristics are presented in Table 1. Inclusion criteria were: (1) age 60–85 years; (2) current or past drivers who recently stopped driving within the past 1 year; (3) ability to operate steering wheel and pedals; and (4) English‐language proficiency. Exclusion criteria included neurological disorders (e.g., stroke, epilepsy, traumatic brain injury, Parkinson's disease), dementia, uncorrected sensory impairment, simulator sickness, substance abuse, major psychiatric illness, untreated sleep, depression, or anxiety disorders, and self‐reported history of simulator sickness. Procedures were approved by the Colorado State University Institutional Review Board (Institutional Review Board [IRB] #2582) and conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent.

TABLE 1.

Descriptive sample characteristics (N = 79).

Variable	MCI	HC	p
N	37	42
Sociodemographic characteristics
Age, years, mean ± SD	75.31 ± 5.55	74.64 ± 4.83	0.571
Years of education, mean ± SD	16.62 ± 2.55	16.95 ± 2.22	0.544
Women, %	29.73	64.29	0.002
Ethnicity, %			0.559
Non‐Hispanic White	97.62	94.60
Hispanic	0	2.70
Other	2.38	2.70
Side dominance, %			0.369
Right	92.86	97.30
Left	7.14	2.70
Cognitive status, mean ± SD
MoCA (score/30)	24.27 ± 3.31	27.17 ± 2.08	< 0.001
DRS‐2 (score/18)	8.14 ± 2.46	10.90 ± 1.51	< 0.001
Baseline driving habits
Driving frequency, %			0.174
<5 days/week	43.24	28.57
5+ days/week	56.76	71.43
Mileage (miles/week), mean ± SD	108.53 ± 116.01	113.01 ± 80.63	0.841

Abbreviations: MCI, Mild Cognitive Impairment; HC, Healthy Controls; DRS‐2, Dementia Rating Scale‐2; MoCA, Montreal Cognitive Assessment; SD, standard deviation.

2.2. Experimental protocol

Participants completed a single baseline session (≈60 min) including sociodemographic questionnaires and a simulated driving assessment.

2.3. Participant characteristics

We collected sociodemographic characteristics (age, sex, ethnicity, education) and assessed cognitive status using the Montreal Cognitive Assessment (or MoCA) ²⁸ and Dementia Rating Scale‐2 (or DRS‐2). ²⁹ , ³⁰ The MoCA is a valid and sensitive screening tool where a score <26/30 indicates impaired cognitive function. The DRS‐2 is an age‐ and education‐adjusted measure in which higher scores indicate better cognition. Finally, we obtained on‐road driving behavior using a questionnaire of weekly driving frequency and mileage.

2.4. Simulated driving course

Driving performance was assessed using a high‐fidelity simulator (NADS MiniSim, The University of Iowa, USA) displayed on three 40‐inch LCD monitors (1920 × 1080, 50 Hz). Participants sat in a professional simulator cab (Simgear, Schaumburg, IL) and used gas and brake pedals with their preferred leg. Prior to testing, participants completed a standardized, 5 min familiarization drive that included key driving tasks including maintaining the speed limit, following a lead vehicle, braking, and making turns. After familiarization, participants completed two challenging scenarios in a fixed order: interstate driving in inclement weather followed by route planning in an unfamiliar environment. The simulation was terminated according to pre‐defined safety criteria if the participant requested to stop or exhibited signs of distress. Simulator sickness was monitored throughout testing using verbal check‐ins at specific intervals. Across both scenarios, lane width was set at 3.66 m (12 ft), traffic density was light with no more than 5 vehicles per mile, and all participants encountered identical traffic events in the same order.

2.4.1. Inclement weather section

The inclement weather section featured a divided highway with sweeping curves and an interchange. The 1.85 mile route included a heavy rain segment that reduced visibility to 16.67 m (50 ft). Posted speed limits (75 mph) were displayed at ≈0.5 mile intervals. Participants entered an interstate highway before the onset of rain (Figure 1A), at which point they activated the windshield wipers.

FIGURE 1.

Simulated driving course. (A) Schematic representation of the inclement weather section. Participants drove on an inclement during rainy weather where the speed limit was 75 mph. (B) Schematic representation of the route planning section. Participants were provided with a map of a city and instructed to plan their route from a stop sign to a pharmacy. Once the participant affirmed, they had planned their route, the map was taken away, and participants attempted to navigate the city to arrive at their destination.

2.4.2. Route planning section

The route planning section featured an urban road geometry with two standard orthogonal intersections and a roundabout. Post speed limits (35 mph) were displayed at every main roadway segment. Participants were shown the map and were asked to plan a route from a predefined origin to a pharmacy destination (Figure 1B). Participants were given as much time as need to review the map and plan the route. When participants indicated they were ready, the map was removed and they started navigating the city. The first intersection required participants to proceed straight and the second intersection required a right turn. At the entrance of the roundabout, participants encountered a predefined traffic event in which they were required to yield to an oncoming vehicle before entering the roundabout. The route distance from origin to destination was ≈1.5 miles.

2.5. Simulated driving outcomes

2.5.1. Speed control

Outcomes included average speed, speed variability, speed error, percent time on‐speed, and percent time not accelerating. Speed variability was calculated as the coefficient of variation (standard deviation / mean *100). Speed error was calculated as the root mean squared error. Percent time on‐speed was defined as the proportion of time driving between 70 and 80 mph. Percent time not accelerating was defined as the proportion of time with no gas pedal input. Higher speed variability, speed error, and time not accelerating, along with lower average speed and time on‐speed, indicate poorer speed control.

2.5.2. Steering control

Steering outcomes included steering wheel variability, number of lane departures, lane departure time, and lane variability. Number of lane departures was defined as the number of instances during which the vehicle crossed the lane boundary. Steering wheel variability was calculated as the standard deviation of steering wheel angle. Lane departure time was defined as the total time the vehicle was outside the lane boundaries. Lane variability was calculated as the standard deviation of lateral vehicle position. Higher values across these steering measures indicate poorer steering control. These metrics are consistent with our prior simulator work in neurological populations ³¹ and have been validated against real‐world driving performance in younger and older adults. ³²

2.5.3. Route Score

Figure 1B shows the route planning section. The pre‐defined correct navigation route began on Holt Dr., and required participants to proceed through intersections including Park Dr. and Queen Dr., and successfully turning onto Martin Dr. to reach the final destination (Hy‐Vee Pharmacy). Route score was determined by the total number of incorrect maneuvers, where higher values indicated poorer navigation.

2.6. Statistical analysis

Group characteristics were compared using independent t‐tests for continuous variables and chi‐square tests for categorical variables. To examine whether driving performance predicted cognitive status, hierarchical binary logistic regressions were conducted separately for each driving section, with group (MCI vs control) as the criterion variables, and sex and driving metrics as predictor variables. Due to established sex differences in MCI prevalence ²² , ²³ , ²⁴ and sex differences in driving behavior in older adults, ²⁵ , ²⁶ , ²⁷ sex was included in Block 1 and driving metrics were included in Block 2. This hierarchical approach allows for a sex‐only benchmark model to understand how much driving metrics improve classification beyond sex alone.

To reduce overfitting and stability of parameter estimates, the number of predictors included in each model was limited relative to the number of MCI cases (maximum of 5 predictors), yielding an approximate observations‐per‐predictor ratio (i.e., n/k ratio) of ≈7, which is considered adequate for small‐sample logistic regression. ³³ To evaluate parameter stability and obtain robust confidence intervals, we used stratified bootstrapping (1000 bootstrap samples stratified for group and sex). Before entering predictors in the model, multicollinearity was assessed among the various speed and steering variables, respectively. For the correlation results, see Table S1.

In the inclement weather section, average speed was highly correlated with speed variability, speed error, and percent time on‐speed (|r| > 0.7). Lane variability correlated with steering wheel variability (r = 0.828) and lane departure time correlated with number of lane departures (r = 0.731). Thus, the inclement weather model included the following predictors: sex, average speed, percent time not accelerating, lane variability, and lane departure time.

In the route planning section, average speed was highly correlated with speed error and percent time on‐speed (|r| > 0.8). Lane departure time was highly correlated with number of lane departures (r = 0.844). No other speed and steering variable showed high collinearity. Given the several speed and steering variables that offer unique variance, we combined the speed and steering variables into composite scores, respectively. These composite scores allowed us to adhere to a maximum of five predictors. The speed control composite score was computed using average speed, speed variability, and percent time not accelerating. The steering control composite score was computed using lane departure time, lane variability, and steering wheel variability. Speed and steering variables were standardized into z‐scores using pooled means and standard deviations across groups. Scores were direction adjusted so that higher values reflected poorer performance and then averaged. In addition, driving metrics, particularly number of lane departures, could be influenced by exposure time or distance driven. In the inclement weather section, participants drove a fixed distance. However, in the route planning section, participants had the freedom to navigate different paths. To account for participant‐level variation exposure, we included total distance driven in the route planning section. Thus, the route planning model predictors included sex, total distance driven, speed control composite, steering control composite, and path score.

Predicted probabilities from each model were entered into receiver operating characteristic (ROC) analyses to compute area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value, and negative predictive value. For interpretability, variables were sign‐reversed where necessary so that higher scores indicated worse performance, ensuring consistency with increased odds of MCI classification.

As a secondary analysis, we conducted multivariate analyses of covariance (MANCOVA) for continuous outcomes and generalized linear model (GLM) with a Poisson distribution for lane departures and route score with sex included as a covariate. Given unequal group sizes and violations of variance‐covariance assumptions, Pillai's Trace was used for MANCOVA inference. Effect sizes were reported as partial eta squared (ηp² ) for MANCOVA and incidence rate ratios (IRRs) for GLMs. According to Cohen, ³⁴ an ɳ_p ² of 0.01 indicates a small effect, ɳ_p ² of 0.06 indicates a moderate effect, and ɳ_p ² of 0.14 indicates a large effect. All analyses were conducted in SPSS v30.0 (IBM, Armonk, NY, USA), with α = 0.05. Power analysis indicated that 70 participants were required to achieve 80% power (α = 0.05) for detecting a moderate‐to‐large odds ratio (OR) with balanced groups. Therefore, we recruited a total of 79 participants (37 with MCI and 42 controls).

2.6.1. Missing values

Two MCI (both male) participants did not complete all the driving tasks. Five controls (1 male, 4 female) did not complete the inclement weather task, and four controls (1 male, 3 female) did not complete the route planning task. These missing data are attributed to simulator intolerance/sickness. Although missing data within each section and group did not exceed 15%, we applied multiple imputation because the data was conservatively considered missing not at random (MNAR).

3. RESULTS

3.1. Sociodemographic characteristics, cognitive status, and baseline driving habits

Age in participants with MCI did not differ significantly different compared with controls (p = 0.571; Table 1). Most participants were White/Caucasian and right‐side dominant, with no significant group differences in years of education. The MCI group demonstrated lower MoCA and DRS‐2 scores as compared with controls (p < 0.001). Consistent with prevalence rates, ²³ , ³⁵ the MCI group contained fewer female participants than the control group (p = 0.002). Relative to the control group, the MCI group did not show significant differences in driving frequency (p = 0.174) or mileage per week (p = 0.841).

3.2. Prediction of cognitive status from simulated driving performance

Table 2 summarizes the logistic regression models by each driving section. Given the differences in sex distribution among the groups, Block 1 in both models serves as the sex‐only benchmark step. Block 1 was significant (χ² (df = 1, n = 79) = 9.62, p = 0.002), and explained a modest proportion of pseudo variance in cognitive status classification (Cox and Snell R ² = 0.115; Nagelkerke R ² = 0.153). At Block 1, the model correctly classified 67.1% of participants overall, with classification of 64.3% for controls and 70.3% for individuals with MCI.

TABLE 2.

Parameter estimates derived from logistic regression.

	B	SE	Wald	df	p	OR	Bootstrap 95% CI for OR
							Lower	Upper
Block 1, Sex‐only benchmark
Sex	−1.448	0.495	8.996	1	0.004	0.235	0.083	0.578
Inclement weather, Block 2
Sex	−1.760	0.699	9.662	1	0.003	0.172	0.033	0.487
Average speed	0.738	0.452	4.634	1	0.043	2.093	0.996	5.878
Percent time not accelerating	0.173	0.409	0.302	1	0.617	1.189	0.533	2.804
Lane variability	0.304	0.814	0.699	1	0.283	1.355	0.153	3.651
Lane departure time	0.726	0.581	2.989	1	0.067	2.068	1.080	9.539
Route planning, Block 2
Sex	−2.005	−0.156	11.494	1	0.002	0.135	0.031	0.378
Speed composite	0.727	0.092	2.657	1	0.104	2.068	0.856	7.516
Steering composite	−0.176	−0.061	0.188	1	0.641	0.838	0.282	1.910
Total distance driven	−0.341	−0.033	0.862	1	0.379	0.711	0.275	1.874
Route score	0.343	0.056	1.157	1	0.273	1.410	0.704	3.280

Abbreviations: B, beta; SE, standard error; df, degrees of freedom; OR, odds ratio; CI, confidence interval.

For inclement weather, Block 2 was significant (χ² (df = 4, n = 79) = 18.32, p = 0.001), and explained a moderate proportion of pseudo‐variance in cognitive status classification (Cox and Snell R ² = 0.298; Nagelkerke R ² = 0.398). The model did not show significant departure from fit (Hosmer and Lemeshow test χ² (8) = 7.219, p = 0.513). The model correctly classified 75.9% of participants overall, with classification accuracy of 78.6% for controls and 73.0% for individuals with MCI. Average speed emerged as a significant predictor of cognitive status (p = 0.031, OR = 2.09), such that slower average speed was associated with an increased odds of MCI classification. Predicted probabilities derived from the logistic regression model were used in an ROC analysis, which demonstrated good discrimination between groups (unvalidated apparent AUC = 0.830, sensitivity = 78.4%, specificity = 73.8%, positive predictive value = 72.5%, and negative predictive value = 79.5%) at an exploratory, sample‐specific optimal probability cutoff of 0.44 (Figure 2). Applying this unvalidated apparent probability cutoff (≥0.44) to classify individuals as having MCI yielded a sensitivity of 78.4%.

FIGURE 2.

Receiver operator characteristic (ROC) curve for the predicted probabilities from the interstate weather logistic regression model. The ROC analysis revealed good discrimination between groups (area under the curve [AUC] = 0.830) and that an unvalidated apparent probability cutoff of 0.44 results in a sensitivity of 78.4% and a specificity of 73.8%.

For route planning, Block 2 was not significant (χ² (df = 4, n = 79) = 7.256, p = 0.123), and explained a modest proportion of pseudo‐variance in cognitive status classification (Cox and Snell R ² = 0.192; Nagelkerke R ² = 0.257). The model did show significant departure from fit (Hosmer and Lemeshow test χ² (8) = 15.608, p = 0.048). The model correctly classified 65.8% of participants overall, with classification of 66.7% for controls and 64.9% for individuals with MCI.

As a secondary sensitivity analysis, we repeated the route planning logistic regression model using individual metrics rather than composite scores (see Table S2). Block 2 was not significant (χ² (df = 8, n = 79) = 10.611, p = 0.225), and explained a moderate proportion of pseudo‐variance in cognitive status classification (Cox and Snell R ² = 0.226; Nagelkerke R ² = 0.302). The model did not show significant departure from fit (Hosmer and Lemeshow test χ² (8) = 4.812, p = 0.777). The model correctly classified 67.1% of participants overall, with classification of 69.0% for controls and 64.9% for individuals with MCI.

3.3. Group comparisons on simulated driving performance

Table 3 shows group effects and sex‐adjusted values across driving sections. The multivariate effect of group was not significant (Pillai's Trace = 0.28, F (17,59) = 1.35, p = 0.197, ɳ_p ²  = 0.28) but was associated with large effect sizes. Although the overall MANOVA was not significant, the univariate comparisons suggest that MCI participants showed slower average speed, higher speed error, lower time on‐speed, and a trend toward higher speed variability relative to controls. The univariate comparisons also suggest that the MCI participants displayed a greater number of lane departures, longer time departed from the lane, and increased lane variability. During the route planning task, the univariate comparisons suggest that MCI participants showed slower average speed, higher speed error, and lower time on‐speed compared with controls.

TABLE 3.

Sex‐adjusted values for all driving simulator outcomes.

Variable	MCI	HC	F (1, 75) / Wald χ 2	p	ɳp2 / IRR
Inclement weather
Average speed, mph	55.7 [53.2, 58.3]	61.7 [59.4, 63.9]	12.257	<0.001	0.140
Speed variability, %	20.6 [17.1, 24.0]	16.4 [13.3, 19.4]	3.286	0.074	0.042
Speed error, mph	22.5 [19.8, 25.1]	16.9 [14.5, 19.3]	9.613	0.003	0.114
Time on‐speed, %	13.8 [7.4, 20.2]	23.7 [18.0, 29.4]	5.331	0.024	0.066
Time no acceleration, %	12.0 [9.6, 14.4]	10.1 [8.0, 12.3]	1.345	0.250	0.018
Steering wheel variability, o	6.2 [2.0, 10.3]	3.8 [0.1, 7.5]	0.707	0.403	0.009
Number of lane departures, count	6.4 [5.6, 7.4]	3.8 [3.3, 4.5]	23.693	<0.001	1.294
Lane departure time, s	25.4 [20.9, 29.9]	14.8 [10.8, 18.9]	12.187	<0.001	0.14
Lane variability, ft	2.0 [1.7, 2.3]	1.6 [1.4, 1.9]	4.157	0.045	0.053
Route planning
Average speed, mph	20.4 [19.4, 21.3]	21.9 [21.0, 22.7]	5.195	0.025	0.065
Speed variability, %	23.5 [20.7, 26.2]	21.5 [19.1, 24.0]	1.068	0.305	0.014
Speed error, mph	15.7 [14.9, 16.6]	14.5 [13.8, 15.2]	5.06	0.027	0.063
Time on‐speed, %	34.2 [26.7, 41.8]	45.1 [38.3, 51.9]	4.558	0.036	0.057
Time no acceleration, %	26.1 [22.2, 30.0]	21.1 [17.6, 24.5]	3.747	0.057	0.048
Steering wheel variability, o	11.9 [10.1, 13.7]	12.8 [11.2, 14.4]	0.495	0.484	0.007
Number of lane departures, count	11.8 [10.6, 13.1]	10.6 [9.6, 11.7]	2.239	0.135	1.131
Lane departure time, s	26.9 [20.7, 33.1]	23.2 [17.6, 28.7]	0.802	0.373	0.011
Lane variability, ft	2.1 [1.9, 2.2]	2.1 [2.0, 2.2]	0.028	0.867	0.000
Total distance driven, miles	1.6 [1.5, 1.7]	1.7 [1.6, 1.8]	2.233	0.139	0.029
Route score, count	1.0 [0.7, 1.5]	0.9 [0.6, 1.2]	0.363	0.547	0.816

Note. To facilitate interpretability, sex‐adjusted group values represent the estimated marginal means derived from the MANCOVA and GLM tests. All values are reported as mean [95% confidence interval]. The F statistic and ɳp² are reported for continuous variables from the MANCOVA and Wald χ ² and IRRs are reported for variables from the GLM tests. Abbreviations: f, female; HC, healthy control; IRRs, incidence rate ratios; m, male; MCI, mild cognitive impairment; mph, miles per hour; ηp², partial eta squared; χ², chi squared; GLM, Generalized Linear Model; MANCOVA, Multivatiate Analysis of Covariance; ft, feet; s, seconds.

4. DISCUSSION

This study examined whether simulated driving performance in challenging conditions can distinguish individuals with MCI from cognitively healthy controls. Driving performance metrics derived from an inclement weather scenario classified cognitive status beyond sex alone with ≈76% accuracy. Of interest, speed control emerged as the most robust predictor of cognitive status across different driving conditions. To our knowledge, this is among the first studies to demonstrate that simulated driving performance under high consequence, yet ecologically valid conditions, can reliably predict and classify MCI status. These findings highlight the potential of simulator‐based assessments as safe and objective platforms to investigate functional vulnerabilities associated with early cognitive decline.

4.1. Simulator‐based driving outcomes in challenging conditions predict early cognitive decline

A central finding is that driving performance in inclement weather classified cognitive status with over 75% accuracy, with slower average speed emerging as the most informative predictor. It is important to note that the simulator driving in inclement weather improved classification beyond sex alone. This association likely reflects the higher‐order cognitive demands of continuously adapting speed in dynamic environments, requiring integration of sensory input, rapid decision‐making, and precise motor execution. These demands are amplified in high‐speed interstate settings and adverse weather, where reduced visibility and traction increase complexity. Research on performance under pressure suggests that such conditions accelerate the breakdown of cognitive control, situational awareness, and precision because the available cognitive resources needed to manage extreme stressors are already diminished. ³⁶ In this context, slower average speed may indicate cautious driving and self‐regulatory strategies in individuals with MCI to compensate during challenging driving conditions. We propose that as cognitive demands increase under challenging driving situations, these compensatory mechanisms may become less effective, allowing subtle cognitive deficits to emerge more prominently.

4.2. Functional assessment of cognitive impairment through driving simulator

Traditional cognitive assessments are often time intensive and may require repeated testing to detect meaningful change. ³⁷ , ³⁸ In contrast, functional assessments may be particularly informative because the distinction between MCI and dementia is based, in part, on impairments evident in everyday function. ¹² Driving is a complex, real‐world instrumental activity of daily living that may capture subtle functional deficits associated with early cognitive decline. Supporting this idea, naturalistic driving studies have shown that continuous monitoring of real‐world driving patterns over time and at scale can predict MCI and dementia onset. ³⁹ , ⁴⁰ Indeed, naturalistic driving represents the gold standard for capturing everyday driving routines and compensatory behaviors with unparalleled ecological validity. However, individuals with cognitive impairment often avoid driving in challenging conditions, ⁵ , ²¹ which may limit the sensitivity of naturalistic approaches to detect deficits that only emerge under acute stress. Simulated driving addresses this gap by enabling the safe, standardized evaluation of driving performance under high‐cognitive‐load and high‐risk conditions (e.g., severe inclement weather) that would be otherwise unsafe to rigorously test on public roads. ¹ , ² Together, these approaches are highly complementary. Naturalistic methods excel at measuring real‐world behavior, whereas simulators are uniquely suited to probe behaviors that may not manifest in self‐regulated, daily driving. Moreover, costly, high‐immersion hardware is not strictly necessary to capture these driving behaviors. More accessible and affordable configurations, such as desktop‐based setups equipped with commercial off‐the‐shelf steering wheels and pedals, may be viable alternatives for clinical and translational applications. Indeed, previous work has demonstrated the feasibility of using portable driving simulators to identify driving impairments in older drivers with MCI. ⁴¹ Adapting high‐demand behavioral challenges to these lower‐cost platforms may provide a practical avenue to evaluate functional vulnerabilities without requiring specialized, high‐cost driving simulators.

With the increasing prevalence of MCI with the aging population, currently available diagnostic approaches can be invasive and often require follow‐up assessments to confirm diagnosis. Driving, a core activity of daily living, is often compromised in individuals with MCI, ³ , ⁴ even when overall functional independence is preserved. Our findings demonstrate that simulated driving metrics can achieve up to 78.4% sensitivity (i.e., classifying those with MCI) using simple, easily captured measures. These metrics can be obtained with low‐cost tools and assessed within controlled simulator environments that systematically probe challenging conditions not consistently encountered on the road. Simulators also enable repeated assessments, supporting longitudinal monitoring of cognitive decline. Together, these preliminary findings support the potential utility of simulated driving as a tool to probe early functional vulnerabilities in individuals with MCI.

4.3. Limitations and further considerations

We acknowledge a few key limitations of the current study. First, despite statistical adjustments, the sex imbalance in our cohort may influence model estimates and limit generalizability. This imbalance may reflect differences in MCI prevalence between male and female participantss ²² , ²³ , ²⁴ as well as sex‐related differences in study enrollment, highlighting the need for more balanced samples in future studies. Second, because of relatively small sample size, we evaluated an MCI cohort without differentiating subtypes (e.g., amnestic/non‐amnestic, single/multi‐domain) limiting our ability to characterize driving patterns across the heterogeneity of MCI. Third, our relatively small sample size constrained model complexity. Because logistic regression analysis in our statistical tool (SPSS) uses the same dataset to build and test the models, performance estimates and specific probability cutoffs are exploratory and likely optimistic. Proper bootstrap optimism correction or cross‐validated performance estimates are needed. External validation and calibration in large, independent cohorts is required before these metrics can be generalized as clinically meaningful thresholds. Fourth, translating these simulator findings to population‐level screening will require scalable naturalistic methodologies (e.g., smartphone telematics, in‐vehicle sensors, and passive monitoring). Crucially, these future large‐scale models must rigorously adjust for a comprehensive suite of covariates, including geography, road type, traffic density, weather, socioeconomic factors, and local driving norms. Fifth, braking measures were excluded due to the infrequency of spontaneous braking in our low‐traffic density scenarios, although they may prove informative in dynamic, high‐traffic environments. Sixth, while map‐based route planning may be sensitive to cognitive deficits by requiring active spatial encoding and retention, it may not fully capture contemporary driving practices that increasingly rely on dynamic GPS‐based navigation. Finally, we did not study the association between specific cognitive domains and simulated driving performance to identify the precise cognitive processes underlying these driving behaviors in MCI.

5. CONCLUSIONS

This proof‐of‐concept study provides preliminary evidence that simulated driving in challenging conditions offers a safe, ecologically valid and functionally meaningful approach for identifying MCI. Speed‐control measures in high‐stake, challenging conditions such as inclement weather, are especially sensitive to early cognitive decline. Although these findings provide initial evidence supporting the potential of simulated driving behaviors to identify MCI, they represent an important first step toward developing objective, functionally relevant markers of cognitive decline. Future validation against naturalistic, on‐road driving, and longitudinal studies is critical to determine their application for screening, cognitive risk assessment, or driving safety decision‐making.

CONFLICT OF INTEREST STATEMENT

The authors declare there are no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All participants provided written informed consent.

Supporting information

Supporting Information

Supporting Information