Association of Daily Driving Behaviors with Mild Cognitive Impairment in Older Adults Followed Over 10 Years

Abstract

Background and Objectives:

Driving integrates multiple cognitive, sensory, and motor systems and may serve as a real-world indicator of functional decline in aging. Older adults with mild cognitive impairment (MCI) often experience subtle driving changes before formal dementia diagnosis, but longitudinal, real-world evidence is limited. This study examined whether naturalistic driving data can differentiate older adults with MCI from those with normal cognition (NC) over time and evaluated the discriminative ability of driving features compared with conventional risk factors.

Methods:

We conducted a prospective, observational cohort study of community-dwelling older drivers enrolled in the DRIVES Project at Washington University. Participants underwent annual Clinical Dementia Rating^® (CDR), neuropsychological testing, and APOE ε4 genotyping. Driving behaviors were captured daily for up to 40 months using GPS-enabled in-vehicle dataloggers, recording trip frequency, duration, distance, time of day, speeding, hard braking, and spatial mobility (entropy, maximum distance, radius of gyration). Longitudinal changes were analyzed with linear mixed-effects models adjusting for age, sex, race, education, and APOE ε4. Logistic regression with ROC analysis evaluated discrimination between older adults with MCI and NC compared to conventional sociodemographic and genetic markers.

Results:

Among 298 participants (MCI n=56; NC n=242; mean age 75.1 years; 45.6% female), groups were similar in age, sex, race, and APOE ε4 status at baseline, and most driving behaviors. Over time, MCI drivers showed greater reductions in monthly trip count (MCI:−0.501, SE:0.21, 95% CI[−0.923,−0.083] vs NC: −0.523, SE:0.09, 95% CI[−0.709,−0.337]; p<.001), nightly trips (MCI:−0.334, SE:0.17, 95% CI[−0.675,0.001] vs NC:−0.339, SE:0.07, 95% CI[−0.480,−0.197]; p<.001), and random entropy (MCI: −0.008, SE:0.004, 95% CI[−0.016,−0.001]; NC:−0.014, SE:0.002, 95% CI[−0.017,−0.011]; p<.001). Key features like medium trip distance, speeding events, entropy, and maximum distance classified MCI vs NC drivers (AUC 0.82, 95% CI 0.75–0.89). Adding demographics, APOE ε4, and cognitive composite improved AUC to 0.87 (95% CI 0.81–0.93).

Discussion:

MCI was associated with progressive declines in driving frequency, complexity, and spatial range, supporting naturalistic driving data as a potential unobtrusive digital biomarker for early cognitive decline. Limitations include a predominantly white, highly educated sample and a lack of external validation, warranting cautious interpretation. Continuous monitoring could augment clinical assessments, inform driving safety decisions, and guide interventions to preserve mobility in aging.

Introduction

Older adults (age ≥ 65) represent an expanding proportion of the driving population. In the United States in 2022, older adults comprised 22% (52 million) of drivers and are projected to reach a quarter of all drivers by 2050.¹ Driving as an instrumental activity of daily living is critical for independence and community mobility. However, advanced age is accompanied by increased crash risk per miles driven and more significant injury and mortality from crashes. In 2022, 9,100 older drivers died in fatal traffic crashes, and 270,000 were treated in emergency departments for crash-related injuries.² Age-related declines in sensory, motor, and cognitive functions (e.g., vision loss, slowed reflexes, impaired memory, and executive function) may adversely affect driving ability.³

Among age-related health factors, cognitive functioning is critical for driving ability and safety in older adults. Mild cognitive impairment (MCI), indicating measurable objective cognitive decline, affects approximately 22% of older adults⁴, while 10% of older adults have dementia. Thus, nearly one-third of older adults live with some degree of cognitive impairment.⁵ MCI also confers a high risk of progression to prodromal Alzheimer’s disease and related dementias ([ADRD), with about 10–20% of older adults with MCI developing dementia annually.⁶ Consequently, a substantial number of drivers on the road have mild neurocognitive deficits or are in the prodromal stages of ADRD.⁷

Driving studies have documented how older adults with MCI or early dementia consistently perform worse on road tests and driving simulators than adults with normal cognition (NC), and are at a higher rate of at-fault crashes.^8–10 One longitudinal analysis found a nearly five-fold increase in motor vehicle crashes during the three years preceding a dementia diagnosis.¹¹ Moreover, neuropathological studies have revealed AD changes in the brains of older drivers who suffered fatal crashes without ever being diagnosed, suggesting that incident cognitive impairment contributes to collision risk.^{12, 13} These findings underscore that deterioration in driving safety often begins during the prodromal stages of cognitive impairment, well before overt dementia is recognized clinically.^{14, 15}

Early identification of at-risk older drivers is a public health priority. However, there is no consensus on screening, let alone formal testing of driving ability.¹⁶ The American Academy of Neurology’s 2010 Evaluation and Management of Driving Risk in Dementia identified the Clinical Dementia Rating ^® (CDR) as the strongest tool for identifying patients with dementia at an increased risk for unsafe driving and a crash.¹⁷ Older drivers with increasing severity (CDR 0.5–1) have a higher relative risk for failing an on-road driving test compared to those with a CDR 0. Additional reviews have highlighted the importance of rating dementia severity.¹⁸ However, early identification of unsafe drivers remains challenging, time-consuming, and requires expert neurological interviews and examinations. Older adults with a diagnosis of early dementia or MCI are often unaware of their declining abilities and continue driving. At the same time, current screening services evaluate only a small fraction of those with cognitive concerns.¹⁹ In practice, formal fitness-to-drive assessments or cognitive screenings are typically initiated only after concerning crashes or near-misses, memory deficits, or family reports.³ A considerable number of older drivers with MCI go undetected until a crash or near-miss forces discussion. This gap highlights the need for proactive and scalable approaches to monitor driving safety in an at-risk population.

Advances in digital technology, specifically in-vehicle sensor systems like dataloggers, offer a promising solution to this problem.²⁰ Modern dataloggers are increasingly equipped with GPS receivers, accelerometers, gyroscopes, and other telemetry features that continuously record driving behavior.²¹ Enabling these devices for disease detection affords unobtrusive, continuous surveillance of driving performance, yielding objective real-world data in real-time. Notably, subtle changes in driving behavior that often accompany early cognitive decline may be detectable with such technology. For instance, an older driver in the early stages of impairment might begin driving more slowly, making fewer trips, avoiding conditions like nighttime or highway driving, or exhibiting lapses in traffic sign recognition.²² Observing these behaviors can be challenging for family or clinicians; in-vehicle dataloggers can capture and quantify them as potential early warning signs of decreasing driving fitness.²³ Driving metrics such as variability in speed control, lane-position deviations, delayed reaction to external stimuli, and navigation patterns differ between cognitively intact drivers and those with underlying neurodegeneration.²⁴

Traditional clinical indicators such as older age, female sex, lower educational attainment, poorer cognitive test performance, and APOE ε4 carrier status are well-established correlates of mild cognitive impairment (MCI) and AD risk.^25–27 Other studies have used MRI, PET, and CSF biomarkers to discriminate these groups, but these biomarkers are burdensome.^28–31 Additionally, these markers are largely cross-sectional and often reflect disease burden only after substantial neuropathological change has occurred. Moreover, these indicators require in-person assessments, specialized testing, or genotyping, which may limit scalability in primary care or community settings. Moreover, these indicators provide limited insight into how cognitive decline manifests in day-to-day life.

In contrast, naturalistic driving behavior represents a continuous, real-world functional phenotype that is highly sensitive to subtle neurocognitive changes. Driving is a complex instrumental activity of daily living that integrates multiple cognitive, sensory, and motor systems in real time. By quantifying driving patterns longitudinally, we can capture early functional decline that may emerge before or concurrently with measurable cognitive deficits on conventional tests. Additionally, objective, sensor-derived driving data can be collected passively, continuously, and unobtrusively in the community, offering a low-burden means to monitor functional status at scale. If driving behaviors provide discrimination comparable to, or better than, conventional clinical markers, they could augment existing screening approaches, help identify at-risk drivers earlier, and enable proactive interventions before safety is compromised. This is particularly valuable for older adults who may not seek cognitive evaluation until significant functional decline has occurred.

Longitudinal monitoring of driving performance provides a window into an individual’s functional abilities that complements annual clinic-based cognitive assessments. It is unclear whether the daily driving behavior of older adults with MCI differs from that of older adults with normal cognition, whether this varies over time, and whether driving behavior can discriminate between the two groups. This study has two aims: 1) to examine whether there are group differences in driving behavior trajectories over time, comparing older adults with normal cognition with older adults with MCI, and 2) to evaluate whether driving behaviors alone can better discriminate MCI and NC older drivers compared to conventional clinical risk factors. We hypothesize that older drivers with MCI would show overall reduced driving frequency (number of trips taken/month, miles driven/trip), limited trips during night, and a shrinking driving footprint area over time. We also hypothesize that driving features would be able to discriminate between drivers with MCI from drivers with normal cognition, with excellent discrimination (Area Under the Curve [AUC] ≥ 0.80) comparable to demographics, cognitive functioning, and APOE ε4.

Methods

Subject Sample

Participants are enrolled in ongoing studies in the DRIVES Project at Washington University School of Medicine. The DRIVES Project is a 12-year NIH-funded cohort that examines how aging, preclinical AD biomarkers, and MCI impact driving performance and driving behavior among older adults.³² There are two cohorts enrolled: healthy control participants were required to be at least 65 years old with normal cognition at baseline, rated a 0 on the CDR^®,³³ while older adults with MCI were 65 or older with a CDR of 0.5 or greater. All participants, regardless of cognitive status, had to drive a non-adapted vehicle at least weekly and agree to complete annual clinical, neurological, and neuropsychological assessments. Self-reported data on race/ethnicity and sex were collected at the baseline visit.

Standard Protocol Approvals, Registrations, and Patient Consents

The Washington University Institutional Review Board approved all informed consent and protocols (IRB# 201706043, 202010214, 202003209) for this research study. Written consent was obtained for all participants. The DRIVES Project studies were conducted in accordance with the International Conference on Harmonization Guidelines for Good Clinical Practice and the principles of the Declaration of Helsinki.

Clinical Assessment

The CDR combined information from separate interviews with participants and their collateral sources. The global CDR (0, 0.5, 1, 2, 3) and CDR sum of boxes (0–18) were derived from scores in memory, orientation, judgment, problem-solving, community affairs, home and hobbies, and personal care. Participants with a CDR 0 and who did not progress to 0.5 were classified as having normal cognition, while participants who enrolled in the study with baseline CDR 0.5 and stayed 0.5 or progressed to 1.0 were grouped as MCI. Participants who progressed to CDR 0.5 at any point were excluded entirely from the analysis, including all driving data collected before progression, to ensure that the CN group remained free of cognitive impairment across the observational period. Additionally, since CDR 1.0 participants are not prevalent in the cohort, they were classified as MCI for these analyses. Participants also completed neuropsychological assessments, including the Trail Making Test A and B, Category Fluency, Verbal Fluency, Free and Cued Selective Reminding Task: Free Recall Score. A Preclinical Alzheimer’s cognitive composite (PACC) score is a validated and standardized set of cognitive tests developed to detect subtle cognitive decline in individuals who are clinically normal but may have underlying Alzheimer’s disease pathology.³⁴ A baseline PACC score³⁵ was computed by standardizing the scores of the four sub-tasks using their means and standard deviations and then calculating the mean of each subject’s standardized scores. Both CDR and neuropsychological assessment of cognitive function were completed at baseline and annual follow-up visits for each participant—the datalogger was plugged in at baseline. Additionally, since Apolipoprotein ε4 (APOE ε4) is an established genetic risk factor for AD, genotyping was completed at baseline to establish heterozygosity or homozygosity for the ε4 allele.

Naturalistic Driving Data Collection

A commercial vehicle datalogger, the Azuga G2 Tracking device (Azuga, Inc.), was plugged into participants’ vehicles’ onboard diagnostic port (OBDII) during the participants’ baseline visit. Daily driving behavior was continuously captured from ignition on until ignition off. Tabular data captured date, time, speed, latitude, and longitude, alongside event-based threshold alerts for instances of speeding or hard braking. The Driving Real-World In-Vehicle Evaluation System (DRIVES) methodology^{20, 21} analyzed various metrics longitudinally, including the total number of trips, average distance traveled, trips undertaken at night, occurrences of speeding, instances of hard braking, entropy (unpredictability of the unique locations visited), and radius of gyration (speed of turning and steering relating to stability), among others.^{23, 36} Data (1/1/2015–1/31/2025) were aggregated daily for each participant, processed, and reduced to summarize specific behaviors over monthly epochs.

Statistical Analysis

Participants were required to have at least 12 consecutive months of naturalistic driving data to account for seasonality and to have completed at least two annual in-clinic assessments (including the CDR and neuropsychological testing). Longitudinal data analyses were truncated at 40 months from baseline for all participants and assumed a linear relationship between aggregated monthly naturalistic driving variables across groups. Older adults with normal cognition were required to be CDR 0 at each follow-up assessment and were excluded (n=23) from the analyses if they progressed to 0.5. Since MCI status was defined at baseline, analyses of discrimination are time-aggregated (12-month) discrimination analysis with respect to cognitive status and are not predictive of future onset. For each participant, baseline was defined as the date of their first in-person clinical assessment and installation of the in-vehicle datalogger, which occurred during the same visit. Baseline dates varied across individuals due to rolling enrollment. In longitudinal analyses, month 1 corresponded to each participant’s baseline date, and follow-up months were counted sequentially thereafter. To standardize follow-up length across participants, analyses included a maximum of 40 months of post-baseline driving data per individual, regardless of their calendar date of enrollment.

A linear time term was chosen based on prior DRIVES Project findings and visual inspection of smoothed trajectories, which indicated approximately monotonic changes in monthly aggregated driving behaviors over the 40-month follow-up period, and to provide a parsimonious, interpretable model within this timeframe. A random coefficients model (linear mixed model) examined the average rate of change in the driving outcomes based on CDR groups (NC vs. MCI), adjusting for baseline age, race, education, sex, and APOE ε4 status. This model allowed y-intercepts and slopes (monthly rate of change) to vary randomly between participants and fitted a separate regression line for each subject. In this analysis, the fixed-effect intercept represents the estimated baseline mean of the driving/mobility variable for the reference group (normal cognition), adjusted for covariates, while the random intercept captures participant-specific deviations from that mean. The fixed-effect coefficient for CDR group membership (MCI vs. NC) tests whether there is a statistically significant difference in baseline means between groups. The interaction between each group and time was examined to test for differences in slope over time. Estimated means of each participant’s driving mobility variables were obtained from the random coefficients model. Estimated means of each participant’s driving mobility variables were obtained by averaging individual-level predicted values from the adjusted random coefficients (linear mixed-effects) models, thereby reflecting the covariate distribution in the study sample rather than setting covariates to fixed values. To aid interpretation, we generated descriptive LOESS-smoothed curves of participant-level estimates to visualize group trajectories over time. These figures are intended for descriptive purposes only; all statistical inferences are based on the mixed-effects models presented below.

In developing the logistic regression models, we considered the number of events per predictor variable (EPV) given our 56 MCI cases. Data from the first 12 months of driving were averaged into a single value for each variable. To reduce the risk of overfitting, we employed a conservative variable selection strategy, retaining only driving variables with an individual receiver operating characteristic curve and AUC greater than 0.60, and exhibiting low multicollinearity (Pearson’s r < 0.80). This resulted in four driving predictors in the base model, aligning with published recommendations for maintaining a minimum EPV of 10. The models were intended as exploratory and hypothesis-generating rather than definitive predictive tools. Later modeling steps examined whether discriminating ability can be enhanced by adding demographic variables (baseline age, race, education, sex), APOE ε4 status, and PACC score. Separate analyses were conducted comparing progressors to MCI and older adults with normal conditions. All statistical analyses were two-tailed at a significance level of 0.05 and performed with SAS 9.4 (SAS Institute, Cary, NC).

Data Availability

Anonymized data not published within this article will be made available by reasonable request from any qualified investigator. Given the sensitive nature of the naturalistic driving data collected (e.g., latitude, longitude), aggregates of the monthly level data will be shared.

Results

The sample included 56 older adults with MCI (CDR 0.5 [n=53]; CDR 3 [n=1]) and 242 older adults with normal cognition (CDR=0). On average, drivers with normal cognition had more years of education (MCI vs. NC: 15.96±2.66 vs. 16.79±2.34, p = 0.021). There was a significant mean difference in the PACC score (−0.37±0.76 vs. 0.10±0.66, p = <0.001) and CDR sum of boxes (1.88±1.16 vs. 0.02±0.11, p < 0.001). Older adults with normal cognition had slightly longer follow-up time in months (32.34±8.82 vs. 39.04±4.16, p < 0.001). There were no statistically significant differences between groups in age, sex, race, or APOE ε4 (Table 1).

Table 1.

Participant characteristics at baseline

	MCI (n=56)	NC (n=242)	P value
Age at baseline (years)	74.46 (5.71)	73.05 (4.79)	0.057
Education (years)	15.96 (2.66)	16.79 (2.34)	0.021
Female, N (%)	21 (37.5%)	127 (47.52%)	0.184
Non-Hispanic white, N (%)	48 (85.71%)	210 (86.78%)	0.830
Cognitive composite	−0.37 (0.76)	0.10 (0.66)	<.001
APOE ε4, N(%)	16 (28.57%)	83 (34.30)	0.436
CDR at Baseline			<.001
0	-	242 (100%)
0.5	53 (98.20%)	-
1	3 (1.78%)	-
CDR Sum of boxes	1.88 (1.16)	0.02 (0.11)	<.001
Follow-up time (months)	32.34 (8.82)	39.04 (4.16)	<.001

Abbreviations: Normal cognition; CDR, clinical dementia rating

In the linear mixed-effects models (Table 2), at baseline, there were no significant group differences in total monthly trip count or average monthly driving distance in miles. However, the older adults with MCI showed a slower reduction in monthly trip count (MCI:−0.501, SE:0.21, 95% CI[−0.923,−0.083] vs NC:−0.523, SE:0.09, 95% CI[−0.709,−0.337]; p<.001), though both groups declined significantly over time. Similar patterns were observed for driving distance, with a modest but statistically significant difference between groups (slope: MCI:−0.143[0.065], 95%CI:[−0.217,0.014] vs. NC:−0.118[0.028], 95%CI:[−0.172,−0.062]), p < 0.001). Driving behavior stratified by trip distance revealed that older adults with normal cognition made more medium-distance trips (5–9.9 miles) at baseline (MCI:48.14[10.56], 95%CI:[27.35,68.92] vs. NC:52.51[10.41, 95%CI:[32.01,73.02]], p = 0.015). Over time, however, older adults with MCI experienced a more rapid decline in long trips over 10 miles (slope: MCI:−0.100[0.038], 95%CI:[−0.176,0.025] vs. NC:−0.090[0.017], 95%CI:[−0.123,−0.054], p < 0.001). While short trips (less than 1 mile and 1–4.9 miles) did not differ significantly at baseline, both groups demonstrated a declining trajectory, with steeper change among older adults with normal cognition.

Table 2.

Driving behavior outcomes across older adults with or without mild cognitive impairment, where each variable represents the mean number per trip or monthly summary.

	Y-intercept				Slope
Naturalistic driving	MCI		NC			MCI		NC
	Mean	SE	Mean	SE	P	Mean	SE	Mean	SE	P
Total trip count	226.35	44.65	236.74	44.03	0.177	−0.501	0.21	−0.523	0.09	<0.001
Average dist. miles	67.15	13.61	70.23	13.41	0.221	−0.143	0.07	−0.118	0.03	<0.001
Trips less than 1 mile	54.58	20.83	51.96	20.54	0.401	−0.128	0.09	−0.125	0.04	0.002
Trips 1 to 4.9 miles	75.13	23.12	81.74	22.82	0.085	−0.150	0.10	−0.181	0.04	<0.001
Trips 5 to 9.9 miles	48.14	10.56	52.51	10.42	0.015	−0.045	0.04	−0.082	0.02	<0.001
Trips 10 to 19.9 miles	36.28	8.60	38.22	8.37	0.223	−0.100	0.04	−0.089	0.02	<0.001
Trips over 20 miles	18.85	5.75	19.33	5.64	0.696	−0.060	0.03	−0.049	0.01	<0.001
Mean trip time seconds	7599.20	1239.39	8020.57	1218.49	0.092	−8.760	7.30	−15.42	3.13	<0.001
Total trips start at day	58.80	17.71	62.92	17.43	0.223	−0.040	0.12	−0.16	0.05	0.003
Total trips start at night	171.40	29.75	178.31	29.30	0.210	−0.334	0.17	−0.339	0.07	<0.001
Mean Sudden Accel./trip	0.04	0.09	0.06	0.09	0.477	−2.8E-4	8.0E-4	−2.3E-4	2.0E-4	0.812
Mean Hard braking/trip	0.18	0.09	0.18	0.09	0.921	−1.1E-4	1.7E-4	−1.5E-4	1.0E-4	0.497
Mean Hard cornering/trip	0.40	0.13	0.44	0.13	0.108	4.6E-3	2.1E-3	5.4E-3	0.9E-3	<0.001
Mean Speeding/trip	0.73	0.41	1.09	0.34	0.002	−2.3E-3	3.0E-3	−12.0E-3	0.001	<0.001
Radius of gyration	143.84	237.48	−32.21	232.20	0.001	1.86	1.70	1.41	0.73	0.089
Max distance driven	5472.88	1985.07	3523.47	1950.79	<0.001	38.02	14.01	16.27	6.05	0.001
Random entropy	8.70	0.70	9.44	0.69	<0.001	−8.2E-3	4.1E-3	−13.6E-3	1.7E-3	<0.001

Abbreviations: MCI, mild cognitive impairment; NC, normal cognition; SE, standard error; Accel., acceleration

Temporal driving patterns showed no baseline differences in the number of trips initiated during the day or night; however, older adults with MCI experienced more pronounced declines in daytime trip starts (slope: MCI:−0.040[0.117], 95%CI:[−0.270,0.190] vs. NC:−0.159[0.046], 95%CI:[−0.251,−0.067], p = 0.003), consistent with early self-restriction. Mean trip duration, measured in seconds, was not significantly different at baseline, but the rate of decline was more modest in the MCI group (slope: MCI:−8.760[7.301], 95%CI:[−23.121,5.600] vs. NC:−12.422[3.130], 95%CI:[−21.584,−9.259], p < 0.001). Driving safety-related behaviors, including average sudden accelerations, hard braking, and hard cornering per trip, were comparable between groups at baseline. While no significant group-by-time interaction was observed for sudden acceleration or braking events, the MCI group exhibited a modest but significant increase in the rate of hard cornering over time (slope: MCI:0.004[0.002], 95%CI:[ 0.000,0.009] vs. NC:0.005[0.001], 95%CI:[0.003,0.007], p < 0.001). Speeding events per trip were significantly lower in MCI drivers at baseline and declined less steeply over time compared to CN drivers, suggesting differential patterns of risk-related behavior and self-regulation (Table 2).

Spatial mobility and navigational variability also diverged between groups. At baseline, MCI drivers had significantly higher maximum distances driven (MCI: 5472.88[1985.07], 95%CI:[1568.10,9377.65] vs. NC: 3523.47[1950.79], 95%CI:[−314.13,7361.07]; p < 0.001) and greater radius of gyration (MCI:143.84[237.48], 95%CI:[−323.23,610.91] vs. NC: −32.21[233.20], 95%CI:[−490.90,426.49]; p < 0.001), suggesting initially broader spatial range. Over time, both groups showed reductions in these measures, though the decline in maximum distance driven was significantly steeper in the MCI group (MCI: 38.19[14.01], 95%CI:[−10.44,65.59] vs. NC: 16.27[6.05], 95%CI:[4.35,28.20]; p < 0.001) (Figure 1). Notably, random entropy, a measure of destination variability and trip unpredictability, was statistically significantly lower in the MCI group at baseline and declined over time (MCI:−0.008[0.004], 95%CI:[−0.016,−0.006] vs. NC:−0.014[0.002], 95%CI:[−0.018,−0.011]; p < 0.001), indicating progressive constriction in routine and environmental engagement.

Figure 1.

Driving behavior across frequency (e.g., Average distance), spatial mobility (Radius of gyration), and destination variability (Random entropy); x-axis-months (0–30) and y-axis-estimated mean (A-F).

Finally, receiver operating curves and AUC discriminated older drivers with and without MCI, where driving variables (trips between 5–9.9 miles, overspeeding, maximum distance traveled, random entropy) alone yielded an AUC=0.82 (95% CI, 0.75–0.89). With the addition of demographic variables (age at baseline, sex, race, education), APOE ε4, and the PACC score, the AUCs improved significantly, ranging from 0.84 (95% CI, 0.78–0.91) to 0.87 (95% CI, 0.81–0.92) (Figure 2). The driving model alone outperformed the sociodemographic model 0.65 (95% CI, 0.57, 0.73) and sociodemographic with APOE and PACC score 0.73 (95% CI, 0.66–0.80) with significantly lower AUCs than those without driving variables. We assessed the same models between progressors (Supplemental Table 1) and the MCI and NC groups. Driving along was able to discriminate progressives from NC with an AUC 0.64 (0.52, 0.76), but a higher AUC [0.79 (0.68, 0.91)] for progressors and MCI. The addition of sociodemographic model APOE ε4, and the PACC score increased to 0.87 (0.78, 0.95) between progressors and MCI, and 0.82 (0.74,0.89) between progressives and NC older drivers (Supplemental Tables 2, 3).

Figure 2.

Area under the receiver operating curves (AUC) for discriminating older adults with mild cognitive impairment and adults with normal cognition. Legends show the AUC as each variable type is added to the model: Blue = driving variables, Red = demographic variables, Cyan = apolipoprotein ε4 status, Brown = cognitive composite score.

Discussion

This decade-long study of naturalistic driving behavior among older adults found that drivers with MCI exhibited significant and progressive changes in their daily driving patterns for 40 months of logging data compared to older adults with normal cognition. While baseline differences between groups were modest across most metrics, unique trajectories emerged over time. Older adults with MCI demonstrated earlier and steeper declines in driving frequency, trip complexity, spatial mobility, and destination variability, even after adjusting for age, education, sex, race, and APOE ε4 status. These findings underscore the utility of in-vehicle sensor technology to detect meaningful functional changes that may serve as early indicators of cognitive decline and inform assessments of driving fitness in aging populations.

While most baseline differences between MCI and NC drivers were small, this may reflect a combination of preserved driving ability and early compensatory strategies.³⁷ Some driving measures in our study, such as reductions in trip frequency, daytime driving, and speeding, are likely indices of self-regulatory behaviors intended to reduce risk. In contrast, other measures, such as increased hard cornering or variability in route entropy, may reflect underlying decrements in driving quality and performance capacity. This distinction between adaptive self-regulation and performance-based decline is important, as it may inform how clinicians interpret naturalistic driving data in the context of cognitive status. Understanding the interplay between these domains could help identify drivers who are compensating successfully versus those whose changes signal emerging safety concerns (e.g., MCI).

Our results extend prior cross-sectional simulator-based studies¹⁰ and road tests¹⁹ documenting impaired driving performance in older adults with MCI or prodromal AD. The current study builds on this foundation by providing longitudinal, real-world evidence of subtle behavioral shifts that precede overt functional impairment. Most notably, older adults with MCI drove fewer miles over time, made fewer medium- and long-distance trips, and demonstrated more significant declines in trip initiation during the daytime. This is consistent with past reported self-regulatory behavior³⁷ or environmental risk reduction.^{38, 39} Additionally, measures of navigational diversity, such as maximum distance driven, radius of gyration, and entropy, were initially higher among older adults with MCI but declined more steeply over time. These metrics may indicate a decline in spatial exploration and community engagement, which have been associated with cognitive decline and reduced quality of life in older adults.^40–42

The excellent discriminative ability of datalogger-derived driving features enabled the identification of older adults with MCI. Key variables, such as trip count in the 5–9.9-mile range, instances of speeding, maximum distance traveled, and entropy, in combination, yielded an AUC of 0.82, suggesting a high discriminatory ability. Adding demographic, genetic, and PACC score covariates further enhanced classification accuracy (AUC up to 0.86), highlighting the added discriminatory value of behavioral biomarkers when integrated with clinical and cognitive data. The AUC was slightly higher for discriminating progressors from MCI and slightly lower for identifying progressors from older adults with normal cognition. These results suggest that passive driving behavior monitoring could complement standard diagnostic approaches and help identify individuals who are currently impaired and are at future risk for cognitive and functional driving decline without formal testing or overt complaints. These AUC values are internal estimates without external validation and may be optimistic given the limited number of events. The results should therefore be interpreted as preliminary evidence of potential discriminative value rather than conclusive predictive performance.

These findings also raise important questions about how older adults with MCI adapt their driving behavior over time. A more significant reduction in adverse events like speeding and trip frequency among older adults with MCI may reflect compensatory strategies or early insight into declining fitness-to-drive.⁴³ However, this adaptation may also be driven by reduced confidence, caregiver influence, or subtle executive function impairment, which limits the ability to plan or execute complex trips.⁴⁴ These behavioral signals may precede traditional formal testing of cognitive functioning, offering opportunities for early intervention through occupational therapy and rehabilitation centered on mobility planning or enhanced clinical surveillance.^{45, 46}

Continuous, unobtrusive in-vehicle data collection enables a granular understanding of daily driving behavior that is not possible through clinic-based assessments alone. As the population of older drivers grows, digital phenotyping using datalogger data can fill a critical gap between cognitive screening and driving cessation, enabling more personalized and proactive decision-making. Early detection of declining driving skills would enable clinicians and families to intervene proactively by thoroughly evaluating modifiable factors, implementing safety adaptations or restrictions, or facilitating a planned transition to non-driving mobility before a serious incident occurs. Ultimately, integrating digital sensors, such as dataloggers, into geriatric and neurology practice may help identify early disease, opening an early window for intervention to maintain older adults’ mobility and autonomy, while reducing preventable crashes and injuries on the road. This approach represents a novel and clinically actionable strategy for identifying at-risk drivers early and developing timely interventions that preserve safety and quality of life for aging individuals. Additionally, sensor-derived driving metrics could serve as objective endpoints to gauge the effectiveness of interventions (for example, cognitive rehabilitation programs or advanced driver-assistance technologies) to prolong safe driving.

This study has several strengths, including its 10-year prospective design, well-characterized sample of older adults, rigorous diagnostic ascertainment using the CDR, objective assessment of driving, and robust multivariate modeling approach. However, several limitations are noted. This sample is highly educated, predominantly white, and does not represent the national population. Second, the group of older adults with MCI was relatively small (excluding those without 12 full months of driving data), which may limit the generalizability of the findings, particularly when considering the variability of driving behaviors in more diverse or larger populations. Similarly, the small number of progressors was insufficient to form a third group to characterize the transition stage from normal cognition to MCI. Third, driving behavior may be influenced by unmeasured factors such as caregiver recommendations, geographic location, or vehicle age, make, model, and social support, which were not controlled for in this study. Fourth, data on participants’ medical histories or other potentially relevant covariates, such as visual impairment, motor symptoms, sleep disorders, or medication use, which are known to influence driving behavior, were not included. These factors could have contributed to the observed differences in driving patterns between groups, independent of cognitive status. Fifth, although driving patterns were monitored longitudinally, clinical diagnoses were made annually, and cognitive transitions within that interval could not be captured in real-time. Sixth, PET amyloid and tau biomarker data were not available for the MCI group, which could have enabled stratification of participants by pathological status and allowed for direct correlation between digital driving metrics and biological disease markers. Future work will examine this relationship to determine whether observed behavioral changes are primarily driven by neurodegenerative pathology, as indicated by biomarkers, or by other age-related or comorbid factors. Seventh, descriptive analysis of contemporaneously collected driving and clinical data, the AUC values reflect within-sample discrimination rather than externally validated predictive performance and should be interpreted accordingly. Discrimination metrics were based on the development dataset alone, which can overestimate performance, and that bootstrap validation⁴⁷ reduces but does not eliminate the need for external validation. Finally, while dataloggers offer valuable objective data, they cannot capture qualitative aspects of driving performance (e.g., lane maintenance, hazard recognition) that are observable in road tests or simulators.

Conclusions

Digital driving biomarkers are becoming more ubiquitous, offering an opportunity for older drivers and their healthcare teams to utilize them. Our findings suggest that digital driving biomarkers hold promise for early identification of cognitive impairment and may enhance existing approaches for assessing fitness-to-drive in older adults. Future studies should explore how these metrics evolve from normal cognition to MCI to dementia, evaluate their predictive validity in diverse populations, and test interventions to prolong safe driving through tailored education or support. Clinicians, researchers, and policymakers must collaborate to integrate digital monitoring into routine care while upholding ethical standards for autonomy, privacy, and informed decision-making.