Driving into diagnosis: Leveraging high-resolution telematics and sensorimotor profiling to identify preclinical Alzheimer's disease

Abstract

Background

Early detection of preclinical Alzheimer's disease (AD) could expand preventative care. Current biomarkers are costly, invasive, or lack generalizability. Driving and sensorimotor performance may reveal prodromal changes.

Objective

We tested whether features from high-frequency driving trips detect preclinical AD and whether demographic, genetic, or sensorimotor data improve accuracy.

Methods

Drivers aged ≥ 65 (n = 254) from Driving Real-World In-Vehicle Evaluation System (DRIVES) completed cerebrospinal fluid Aβ₄₂/Aβ₄₀ and amyloid Positron emission tomography (PET) to label amyloid positive (preclinical AD) or negative. A GPS datalogger recorded location (1 Hz) and accelerometer/gyroscope (20 Hz) data between June 2022 and January 2024. Eleven driving features (e.g., average speed, jerk, idle time, turns) were extracted per trip. Vision, hearing, olfaction, gait, and grip strength were assessed. TabNet models classified amyloid status using (1) driving only, (2) driving plus age and APOE ε4, and (3) driving plus age, APOE ε4, sex, and education. LightGBM models evaluated sensorimotor features. Performance was measured on a 20% held-out test set (AUC, accuracy, precision, recall, F1).

Results

The top-performing model (driving, age, APOE ε4, sex, education) achieved an AUC of 0.84, accuracy of 0.85, and F1 score of 0.85. Key predictors were idle time, turns, and average jerk. Sensorimotor models performed modestly (AUCs of 0.66 [sensory alone] and 0.67 [sensory and sociodemographic]), with grip strength and word-in-noise scores as the top contributors.

Conclusions

A high-frequency trip's driving telemetry, combined with age and APOE ε4 status, discriminates preclinical AD, outperforming multisensory measures. Driving offers a scalable, digital biomarker to complement conventional testing. Monitoring may enable population-level screening for older adults at risk.

Keywords

Alzheimer's disease Apolipoproteins E biomarkers machine learning naturalistic driving noninvasive detection older adults

Introduction

Alzheimer's disease (AD) is a progressive neurodegenerative disorder prevalent among older adults (age ≥ 65) that affects multiple regions of the brain, including the cerebral cortex and hippocampus.¹ AD is characterized by the pathological accumulation of amyloid-β (Aβ) plaques and neurofibrillary tangles, which contribute to the cognitive decline and dementia observed in patients.² Globally, over 55 million individuals are currently living with dementia, with approximately 10 million new cases diagnosed annually.³

The cognitive domains impacted by AD include executive function, memory, language, and visuospatial skills. AD progresses from a preclinical phase, which is largely asymptomatic, to mild cognitive impairment, which is characterized by minor deficits in memory, language, or judgment, and finally to the clinical stage, which impacts instrumental and basic activities of daily living.⁴ While cognition is often the primary sequela, sensorimotor function is equally important, with evidence from several studies indicating that declines in olfaction, hearing, vision, and gait may precede the onset of cognitive symptoms by as much as 15 years.^5–7

Due to the slow onset and debilitating nature of AD as it progresses from mild to advanced stages, it imposes a significant socioeconomic burden on families and society, including direct and indirect costs of caregiving, medication, and hospitalization ($781 billion in 2025 in the United States),⁸ among other personal and social burdens.^1,9 Given these high costs, research has shifted toward earlier screening and identification of AD for timely intervention and care planning. Pharmacological efforts focused on monoclonal antibody therapies aim to slow the cognitive decline of prodromal AD.¹⁰ The fundamental premise behind this approach is that the preclinical phase, during which patients remain cognitively normal (∼15–20 years), underscores the importance of early detection so that interventions can be taken to delay symptom onset. Positron emission tomography (PET) imaging, utilizing tracers, cerebrospinal fluid (CSF), and blood-based biomarkers, supports early detection. However, these tools are not widely accessible or accepted by patients due to their invasive procedures and high costs, or their requirement for continued validation in the case of plasma biomarkers.¹¹

Noninvasive methods for early detection involve wearable technologies and sensor-based solutions, such as gait, physical activity, or eye movement tracking.¹² These approaches are based on evidence indicating that subtle behavioral, cognitive, and motor changes associated with AD may begin before clinical symptoms emerge.¹³ A critical aspect of these complex behavioral and cognitive changes is reflected in spatial navigation and driving abilities.¹⁴ If driving is impaired during the asymptomatic preclinical AD stage, it may serve as a novel, noninvasive tool for early detection.

Preliminary studies suggest that a decline in driving performance is associated with AD biomarkers.^15,16 Evidence from both on-road and simulator assessments indicates that driving performance is influenced by the severity and progression of AD. Overall, AD patients are more likely to receive a failure rating on a standardized road test than healthy control drivers.¹⁷ However, on-road and simulator assessments of driving performance do not fully represent real-life driving experiences. A recent innovation involves using dataloggers plugged into a vehicle to monitor and evaluate naturalistic behavior during a trip. This approach yields real-time data that can be used to examine risk, safety, and decline over time. Drivers with prodromal AD commit more driving errors and have a smaller driving space relative to healthy individuals.¹⁸ Two studies demonstrated that driving indicators alone can be used to differentiate between older drivers with and without preclinical AD.^19,20 Moreover, when combined with demographics, these indicators exhibit greater predictive power, further supporting their potential as a valuable tool for early detection.¹⁹ Integrating machine learning techniques can facilitate the prediction of AD prior to the appearance of symptomatic stages, providing a noninvasive and easily accessible means for early detection.²¹ Sensorimotor measurements, such as vision, olfaction, hearing, and walking performance, represent a valuable complement to GPS-based driving data. Emerging evidence suggests that these sensorimotor functions may be impaired years before the clinical manifestation of AD.^13,22 Consequently, they are potential physiological markers for early AD detection in its preclinical stage.

The study investigated whether daily single naturalistic driving trips, rather than monthly aggregated behavior,²³ could accurately detect preclinical AD and whether incorporating established sociodemographic predictors could improve prediction accuracy. Additionally, we investigated the potential of leveraging machine learning to classify preclinical AD using sensorimotor data collected across multiple sessions. This approach offers promising insights into early, noninvasive detection methods for AD.

Methods

Participants

A subset of participants was recruited from The Driving Real-World In-Vehicle Evaluation System (DRIVES) Project (n = 254), a longitudinal study on aging, driving, and preclinical AD at Washington University School of Medicine.^24,25 The inclusion criteria were as follows: (1) age 65 or older, (2) cognitively normal at baseline (as indicated by a score of 0 on the Clinical Dementia Rating [CDR^®] scale), (3) willingness to complete biomarker testing using CSF collection and/or PET scan, (4) self-report driving at least once per week, (5) drive a vehicle with a working Onboard Diagnostics-II port (OBD-II), and (6) have a valid driver's license. Written informed consent was obtained from all participants, following a study protocol approved by the Washington University School of Medicine Human Research Protection Office (202010214, 202003209).

Participant characteristics

Sociodemographic information, including age, sex, and education level, was collected. APOE ε4 carrier status was determined from blood-based genotyping. APOE ε4 is associated with an increased risk of developing AD and is considered a major genetic risk factor. Having one copy of the APOE ε4 allele increases the risk, and having two copies (one from each parent) further increases the risk.

Clinical assessment

Each participant completed a battery of annual clinical, neurological, motor, and sensory functioning and neuropsychological testing. Annual cognitive functioning was assessed via a PACC score.^26–28 The brief neuropsychological battery computes the mean z-score of animal naming (semantic fluency²⁹), the Free and Cued Selective Reminding test (episodic memory^30,31), and the Trail Making Test Parts A and B (processing speed, executive function³²). This is a robust cognitive measure we have used for over seven years.³³ These measurements included visual acuity of the monocular and binocular vision, and contrast sensitivity of binocular vision using the King-Devick Test application on an iPad.³⁴ Grip strength of the right and left hands was recorded by taking the average of three trials per hand with a calibrated dynamometer.³⁵ To assess mobility and walking speed, an average of three ten-meter walk test trials was recorded.³⁶ Auditory function was assessed using the Word-In-Noise test, and olfactory function using the Odor Identification Test from the National Institute of Health (NIH) Toolbox.³⁷ Additionally, biomarkers were obtained every two to three years based on the participant's follow-up. Participants were categorized into two groups based on their amyloid status, as determined by the CSF Aβ₄₂/Aβ₄₀ ratio or a PET scan, in accordance with previously published studies.^38,39

Driving data and measures

Data were collected from June 30, 2022, to January 3, 2024; additional data collection is ongoing for future use. Driving data were collected using a commercial GPS data logger (G2 Tracking Device™, Azuga Inc., San Jose, CA) plugged into a vehicle's OBD-II port. This device worked in conjunction with a custom data pipeline in The DRIVES Project.^24,25 This integrated system captured essential parameters of each driving trip, including vehicle number, trip date, trip time, and accelerometer and gyroscope data along the x, y, and z axes at a frequency of 20 Hz. GPS coordinates (latitude and longitude), engine RPM, and vehicle speed were recorded at 1 Hz. This increased sampling rate of 1 s, compared to the prior 30-s rate for GPS data, is newly introduced in our efforts to record the highest fidelity driving data, here termed “high frequency driving data”.⁴⁰ The high-frequency approach allowed data analysis and predictions from a single driving session. Any trip data (defined as ignition on to off) containing fewer than 100 s (equivalent of 100 samples recorded with a frequency of 1 Hz) was excluded from the analysis. Additionally, we excluded any trip that was shorter than one-third of a mile (535.91 meters), as such sessions may lack sufficient data reflective of typical driving behavior and could introduce variability into the dataset and driving features.

Driving features

A set of features was extracted from the daily driving data to train the machine learning models, capturing various aspects of driving behavior during a single trip. These features include average speed, maximum speed, speed variance, the number of sudden accelerations, the number of sudden decelerations, route length, driving duration, time of day, erratic lateral movements, average jerk, idle time, maximum distance from home, and the number of turns. Table 1 provides a summary of these features along with their descriptions. These features were selected based on relevant literature about their ability to represent critical aspects of driving behavior.^21,41 Effect sizes and significance tests were used to explore group differences, while feature importance was used for model interpretation.

Table 1.

A summary of driving features.

Characteristics	Indicator	Abbreviation	Description
Driving Space	Route length	RouteL	Distance in meters from home position to the destination
	Driving duration	DriveDu	How long the drive lasted
	Time of day	ToD	Time of day that the trip happened, morning, afternoon, evening, and night
	Max distance from home	MaxD	Maximum distance in meters from the starting position of the trip
	Number of turns	nTurns	Total number of turns in a trip
Driving performance	Average speed	AvgS	Average speed (in Km/h) calculated over all the speeds recorded during a trip at 1 Hz
	Maximum speed	MaxS	Maximum speed (in Km/h) calculated over all the speeds recorded during a trip at 1 Hz
	Speed variance	VarS	Variance of speed calculated over all the speeds recorded during a trip at 1 Hz
	Number of sudden accelerations	nSuddAcc	Total number of sudden acceleration events during a trip when acceleration is above 8 mph/s
	Number of sudden decelerations	NSuddDe	Total number of sudden acceleration events during a trip when decelerations is below -8 mph/s
	Erratic lateral movements	ErrLat	Total number of instances when x-axis accelerometer absolute value is above 2 standard deviations of the average x-axis accelerometer during a trip
	Average jerk	AvgJ	Average of absolute change in the acceleration over all time points of a trip

Machine learning models and evaluation metrics

An 80-20 split was implemented for training and testing purposes for the driving and sensory models, ensuring a robust evaluation of model performance across diverse feature sets. This is a common approach in machine learning where 80% of the data is used for training the model, and the remaining 20% is used for testing the model. The training data is used to teach the model to make predictions, while the testing data is used to evaluate how well the model performs.

A binary classification model was implemented using features extracted from the driving data and labels (positive vs. negative) derived from the CSF Aβ₄₂/Aβ₄₀ ratio or PET. TabNet is an innovative deep-learning model specifically designed for tabular data, leveraging a sequential attention mechanism to selectively focus on important features at each decision step, making it both highly interpretable and efficient.⁴² This model is particularly effective for large datasets as it can process vast amounts of tabular data swiftly with a high degree of accuracy. Unlike traditional models that might require extensive feature engineering or complex ensembles, TabNet simplifies the pipeline and enhances model performance through its built-in feature selection capabilities. Its ability to handle high-dimensional data with ease and its robustness in feature utilization make it a superior choice for complex, data-driven tasks. We systematically trained the models using multiple sets of hyperparameters (grid and randomized search) to identify and select the configuration that yielded the highest performance. We trained three separate models: (1) driving features, (2) driving features combined with participant age and APOE ε4 status, and (3) driving features with age, APOE ε4, sex, and education.

Additionally, we trained two distinct LightGBM machine learning models for the binary classification of participants with and without preclinical AD using the sensorimotor data. LightGBM is a gradient-boosting framework that uses tree-based learning algorithms, recognized for its efficiency in handling high-dimensional data involving thousands of data points.³² The first model incorporated only sensorimotor features, while the second model included sensorimotor features alongside the additional variables. The decision to use TabNet for driving data and LightGBM for sensorimotor data was based on the specific strengths of each model with respect to the type of data being handled. TabNet is well-suited for complex, high-dimensional tabular data like driving patterns, while LightGBM is effective at handling structured data with both categorical and numerical features, making it ideal for sensorimotor data. TabNet and LightGBM models are comparable using the same performance metrics.

We used five performance metrics to evaluate the machine learning models: AUC, accuracy, precision, recall, and F1. We used the AUC of the receiver operating characteristic (ROC) as our primary evaluation metric during model training. The ROC curve is a graphical representation that illustrates the diagnostic ability of a binary classifier system as its discrimination threshold is varied. It plots the true positive rate (sensitivity) against the false positive rate (1-specificity) at various threshold settings. The AUC provides a single scalar value that summarizes the overall ability of the model to discriminate between positive and negative cases across all thresholds. During the evaluation phase, accuracy was the primary metric used to assess model performance, which represents the proportion of true results (both true positives and true negatives) among the total number of cases examined. We also employed precision and recall measures to assess our models’ performance further. Precision quantifies the accuracy of positive predictions made by the model, essentially measuring the proportion of true positives among all positive predictions (including false positives). On the other hand, recall measures the ratio of true positive cases to the sum of true positive and false negative cases, effectively assessing the model's ability to identify all instances of the target class. We also measured the F1 score as an additional metric to evaluate model performance. The F1 score is a harmonic mean of precision and recall, providing a single metric that balances the model's precision and recall. This measure is especially relevant because it offers a more comprehensive view of the model's effectiveness, especially when the distribution of classification (with or without preclinical AD) is imbalanced.

Results

Overview of participant demographics and clinical data

A summary of demographic (n = 254) information is presented in Table 2. Among the demographic variables analyzed, a significant group difference was observed only for age (p = 0.006). Box plots of driving and sensorimotor indicators for Class 0 (participants without preclinical AD) and Class 1 (participants with preclinical AD) are presented in Figure 1. The boxes represent the interquartile range (IQR), with the central line showing the median and the whiskers extending to 1.5 times the IQR from the quartiles. Group differences for the features presented in Figure 1 were analyzed, revealing statistically significant differences (p < 0.05) between participants with and without preclinical AD for all features. However, Cohen's d values for these features were consistently below 0.2, indicating small effect sizes.

Table 2.

Participant demographic information.

	Without preclinical AD (n = 180)	With preclinical AD (n = 74)	p
Age (y)	77.82 ± 5.33	79.86 ± 5.32	p = 0.006
Education (y)	16.77 ± 2.22	16.78 ± 2.24	p = 0.956
Sex (% female)	48.88	52.70	p = 0.679
APOE ε4 allele	22.5	67.4	p = <0.001

Driving and sensory features ranking

Feature importance (Figure 2) was derived from two distinct sets of models: those based solely on driving data and those based on sensory data. Within the models that rely solely on driving data, the three most significant features identified are average speed, idle time, and number of turns. These factors suggest that aspects of driving behavior and driving space are critical in influencing the model's outcomes. However, age emerged as the most influential feature when age and APOE were incorporated into the model.

In the models based on sensory data, grip strength, words in noise, and the ten-meter walk test initially stood out as key indicators. These features reflected the sensorimotor capabilities of individuals and were closely tied to their preclinical AD classification. Nonetheless, when APOE ε4 was included in the sensory data models, it emerged among the top three most important features; age became the sixth most important indicator. This underscores the overriding influence of age and APOE ε4 status in both driving and sensorimotor models.

Machine learning models

Metrics evaluating the performance of the driving data and sensorimotor models—AUC, accuracy, precision, recall, and F1 score are presented in Table 3. Integrating age and APOE into the models enhanced the prediction accuracy and F1 scores. Specifically, the top-performing driving model, which incorporates driving features, demographic factors (age, sex, and education), and genetic factors (APOE ε4), achieved an accuracy of 85% and an F1 score of 85%. Incorporating only age and APOE ε4 resulted in an accuracy of 78% and an F1 score of 76%. In terms of the two sensorimotor models, the inclusion of age and APOE also impacted performance. With these factors incorporated, the models reached an accuracy of 72% and an F1 score of 37%. Conversely, without age and APOE, the accuracy and F1 scores were lower, at 71% and 28%, respectively. In both sensorimotor models, the low F1 scores were driven by recall.

Table 3.

Presents the performance metrics—accuracy, precision, recall, F1 score, AUC—for the driving data models and the sensorimotor models.

Driving	Models	Accuracy	Precision	Recall	F1 score	AUC
	Driving features only	0.62	0.55	0.62	0.57	0.5011
	Driving features + Age + APOE	0.78	0.78	0.78	0.76	0.8208
	Driving features + Age + APOE + Sex + Education	0.85	0.86	0.85	0.85	0.8400
Sensorimotor
	Sensory features only	0.712	0.5	0.19	0.28	0.66
	Sensory features + Age + APOE + Sex + Education	0.726	0.55	0.29	0.37	0.67

Discussion

This study builds upon previous work that utilized GPS-based telematics data as a cost-effective and noninvasive method to identify individuals with preclinical AD by examining data from high-frequency driving data.^19,21 It highlights the promise of naturalistic driving data as a digital biomarker for early detection in older adults. Additionally, this study introduced a novel machine learning approach that can identify signs of preclinical AD with approximately 82% accuracy by analyzing data from a single driving trip. This represents a substantial improvement over previous studies, which required aggregated data from one month of driving as a single datapoint for training the classification models.^21,23 This change highlights the approach's efficacy and enhances its practicality for widespread screening, potentially facilitating early intervention strategies for individuals at risk.

We employed a machine learning approach to manage large sets of driving data for classification tasks using CSF Aβ₄₂/Aβ₄₀ or PET scan positivity as labels. Each model was equipped with a feature set specifically tailored to capture the complexities present in everyday driving behaviors. To further evaluate the performance of our models, we incorporated precision and recall as additional metrics. Together, these metrics reveal nuanced details about the models’ effectiveness and balance in predicting AD.

Our findings reveal a noteworthy yet expected observation: the performance of our models was enhanced when age and APOE ε4 status were incorporated as features. Both age and APOE ε4 genotype are putative predictors of AD⁴³; the risk of AD increases with age, and the presence of one or two APOE ε4 alleles increases the risk of AD by multiple factors. This enhancement in model performance aligns with existing research on other biomarkers,^44,45 demonstrating that the ability to detect preclinical AD improves when combined with age and APOE ε4 status.

An essential finding of this study is identifying the most important driving features that influence classification accuracy. Our results indicate that the three models’ top three driving features differ. For instance, in the first model, average speed, idle time, and number of turns emerged as the most essential features. In the second model, average jerk, idle time, and number of turns were prioritized, while the third model highlighted idle time, number of sudden moves, and number of turns. It is important to note that our TabNet model selectively utilized a subset of features to train the model with the best performance metric. This observation is consistent with TabNet's design of sparsity regularization, which employs attentive feature selection to promote sparsity and concentrate on the most informative features.⁴² The automated attention to sparsity optimization allowed for resulting models that were parsimonious, interpretable, robust, and avoided overfitting. Out of the ten driving features, only five appeared among the top three features across the three models. We have categorized these driving indicators into two distinct groups—driving performance and driving space. Jerk, speed, and sudden moves are classified under driving performance, while idle time and number of turns are considered elements of driving space. This distinction aligns with previous research, underscoring that driving space and performance are crucial for the machine learning classification of preclinical AD. Consistent with findings from a 2012 study, our results show that individuals in the early stages of dementia exhibit a restricted driving space, as shown in driving features such as route length, driving duration, and number of turns.⁴⁶ Similarly, spatial navigation studies have reported reduced navigable space for individuals with early-stage AD.^14,47 In our analysis, individuals with preclinical AD experienced a reduced maximum distance duration from home, highlighting a tendency towards a restricted driving space.

A prior study of 161 healthy aging adults demonstrated a relationship between driving space and attentional control; individuals with decreased attentional control exhibited a reduced driving space.⁴⁸ This finding aligns with previous research correlating attention tests, such as the Useful Field of View (UFoV), and driving self-regulation.^49,50 These individuals might have recognized a decline in their cognitive capacities and subsequently adjusted their driving patterns accordingly. In our study, driving duration—a prominent feature in our machine learning model —is an indicator of driving space. This suggests that the group with preclinical AD similarly regulated their driving space, either consciously or in response to a perceived decline in cognitive ability. Speed and jerk emerged as significant features, with jerk shown to be an identifier of aggressive driving.⁵¹ One plausible explanation for jerk being a prominent feature is the altered processing speed observed in individuals with preclinical AD. A decline in processing speed, typical in the preclinical AD group, might impair the ability to respond effectively to complex and rapidly changing events, such as a car changing lanes or a traffic light turning red. This impairment could lead to increased jerk, reflecting the abruptness and potentially unsafe adjustments in driving. This finding aligns well with previous research that has established a clear relationship between cognitive ability and driving performance.⁵² Notably, they did not find a direct relationship between processing speed and driving performance.⁵³ This might be attributed to laboratory-based speed measures not translating well into real-world driving scenarios.⁵⁴ Real-world driving demands continuous adjustment and quick responses, which traditional laboratory tests may not adequately capture.

Machine learning models were also built to detect preclinical AD, including sensorimotor data, but no driving data. The sensorimotor machine learning models did not perform as well as the driving data models. While each sensorimotor feature exhibited statistical significance, effect sizes for each were small (under 0.2). Feature importance was more useful in determining the predictive power of each sensorimotor indicator. Grip strength of each hand, ten-meter walk, words-in-noise, and odor score ranked among the most important sensorimotor features for predicting preclinical AD. Sensorimotor deficits (in grip strength, hearing, gait, etc.) among cognitively normal older adults may be influenced more by general aging and health factors than by early AD pathology.

A key innovation in our approach is the incorporation of high-frequency data. Specifically, GPS data were recorded at 1 Hz, while accelerometer and gyroscope data were captured at 20 Hz. This enhanced data collection frequency allows for a more detailed and accurate representation of driving behaviors, which is crucial for our study's robust analysis and modeling. Another unique aspect of our research is the model's predictive capability from a single driving trip. The advancement in our data collection capacity has enabled us to effectively extract the driving features of a single trip for model training, which is suitable for single-trip classification. This dramatically enhances our model's immediacy and applicability in real-world scenarios. We can achieve classification immediately after a single driving trip with high-frequency recording. This approach significantly improves the timeliness and efficiency of our detection capabilities.

As we collect more data, the likelihood increases of identifying driving trips that may indicate disruptions in complex behaviors, such as those associated with driving. Notably, not every trip made by individuals with preclinical AD, indicated by PET scan and CSF Aβ_42/40 levels, exhibits abnormal driving patterns. However, with increased recorded trips, the probability of detecting such abnormalities in driving indicators also rises. A further step in our research involves exploring automatic feature extraction by processing driving data with models adept at handling time-series data. Recurrent Neural Networks (RNNs) are posited to support past input to support more accurate predictions with sequential GPS data and have been successfully implemented to detect aggressive driving behavior.⁵⁵ Given this success, employing deep RNN networks to identify specific driving behaviors associated with preclinical AD is plausible.

Driving is a dynamic, context-dependent activity that integrates perception, motor control, attention, and route selection in ways that often elude capture by single summary cognitive scores. In a prior study using low-frequency data, only attentional control showed a small, domain-specific association with change in driving space (e.g., fewer trips, fewer unique destinations, smaller radius of gyration) over ∼24 months.⁴⁸ Episodic memory, working memory, and processing speed did not predict changes in naturalistic driving, and no domain predicted adverse event rates (e.g., hard braking, sudden acceleration). Effect sizes were modest (≈1% variance in change; ∼6 fewer trips/month after one year for an individual 1 SD below the attentional control mean), and self-report measures related to driving were only weakly aligned with naturalistic metrics. These findings indicate that global composites aggregating across domains, such as the PACC, are unlikely to display strong, omnibus associations with real-world driving features in cognitively unimpaired older adults. Consistent with the study's focus on scalable, sensor-derived features and to avoid target leakage from clinic-based cognition into telematics-only models, we omitted new PACC–driving analyses here; instead, we reference this prior evidence and interpret our results within that context.

This study has limitations that should be considered. First, while most participants were the primary drivers of their vehicles, they may have loaned their vehicles out to a friend or family member. Consequently, a small number of trips may have been conducted by friends, spouses, or family members; however, given the volume of data for each participant, any potential noise is largely mitigated. Second, because all participants lived in the Greater St Louis Area, the results may not be generalizable to populations in other regions. A third limitation, which presents a significant opportunity for enhancing future research, concerns the criteria for including or excluding specific driving trips based on their duration. Short trips may not adequately capture the driving behaviors typical of individuals with preclinical AD that longer trips might reveal. These shorter trips could paradoxically yield misleading data, detracting from the model's effectiveness despite the intention of enriching it with more data. Identifying optimal trip durations for inclusion in the analysis will be crucial for improving the accuracy and reliability of future studies, with the goal being single-session identification of individuals with preclinical AD.

Conclusion

Our study indicates driving is a complex behavior that may exhibit subtle changes during the preclinical phase of AD. While the machine learning model incorporating driving behavior, age, and APOE ε4 status demonstrated the highest performance in predicting preclinical AD, the model based solely on driving indicators still offers considerable efficacy. This is noteworthy given its noninvasive nature, ease of installation, and accessibility, making it a practical choice for widespread use. In future research, key sensorimotor measures should be incorporated into driving data and demographics models to determine whether a sufficiently effective model can be built without including APOE ε4 status.

Footnotes

Acknowledgements

We thank The DRIVES Project participants for their dedication and contributions to science, without whom this work would not be possible. We also thank members of The DRIVES Project study team for their help with data collection.

ORCID iDs

Ramkrishna K Singh

Semere Bekena

Ganesh M Babulal

Ethical considerations

The study protocol was approved by the Washington University in St Louis Institutional Review Board (IRB #202010214, 202003209). All research procedures were conducted in accordance with the World Medical Association Declaration of Helsinki.

Consent to participate

Written informed consent to participate was obtained from all participants before enrollment.

Consent for publication

Not applicable.

Author contribution(s)

Behdad Parhizi: Data curation; Formal analysis; Methodology; Writing – original draft.

Ramana Kolady: Writing – review & editing.

Mia T Vogel: Writing – review & editing.

Ramkrishna K Singh: Writing – review & editing.

Semere Bekena: Writing – review & editing.

Raja Chokkalingam: Resources; Writing – review & editing.

Yiqi Zhu: Writing – review & editing.

Ganesh M Babulal: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Project administration; Resources; Supervision; Validation; Visualization; Writing – original draft; Writing – review & editing.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was funded by the NIH and National Institute on Aging (NIH/NIA) grants R01AG068183 (GMB), R01AG056466 (GMB), R01AG067428 (GMB), T32HL130357 (MTV). The content is solely the responsibility of the authors and does not necessarily represent the official position of the funder.

Declaration of conflicting interests

The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: Dr. Babulal is an Editorial Board Member of this journal but was not involved in the peer-review process of this article nor had access to any information regarding its peer review.

Data availability statement

The data supporting this study's findings are available from the corresponding author upon reasonable request.

References

Masters

Bateman

Blennow

, et al. Alzheimer's disease. Nat Rev Dis Primers 2015; 1: 15056.

Knopman

Helene

Petersen

, et al. Alzheimer disease (primer). Nat Rev Dis Primers 2021; 7: 33.

Nichols

Steinmetz

Vollset

, et al. Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: an analysis for the Global Burden of Disease Study 2019. Lancet Public Health 2022; 7: e105–e125.

Lane

Hardy

Schott

. Alzheimer's disease. Eur J Neurol 2017; 25: 59–70.

Albers

Gilmore

Kaye

, et al. At the interface of sensory and motor dysfunctions and Alzheimer's disease. Alzheimers Dement 2015; 11: 70–98.

Merten

Paulsen

Chappell

, et al. Sensory and motor functions improve long-term predictions of neurodegeneration and cognitive decline. Innov Aging 2023; 7: 193.

Merten

Pinto

Paulsen

, et al. Midlife hearing, vision, olfactory, and motor function improve the long-term prediction of cognitive decline and onset of cognitive impairment. Alzheimers Dement 2023; 19: e071813.

Zissimopoulos

Dana . The Cost of Dementia in 2025, https://schaeffer.usc.edu/wp-content/uploads/2025/04/The-Cost-of-Dementia-in-2025.pdf (2025).

Deb

Sambamoorthi

Thornton

, et al. Direct medical expenditures associated with Alzheimer's and related dementias (ADRD) in a nationally representative sample of older adults-an excess cost approach. Aging Ment Health 2018; 22: 619–624.

10.

Doody

Thomas

Farlow

, et al. Phase 3 trials of solanezumab for mild-to-moderate Alzheimer’s disease. N Engl J Med 2014; 370: 311–321.

11.

Hampel

Cummings

, et al. Blood-based biomarkers for Alzheimer’s disease: current state and future use in a transformed global healthcare landscape. Neuron 2023; 111: 2781–2799.

12.

Vrahatis

Skolariki

Krokidis

, et al. Revolutionizing the early detection of Alzheimer’s disease through non-invasive biomarkers: the role of artificial intelligence and deep learning. Sensors 2023; 23: 4184.

13.

Kourtis

Regele

Wright

, et al. Digital biomarkers for Alzheimer’s disease: the mobile/wearable devices opportunity. NPJ Digit Med 2019; 2: 9.

14.

Coughlan

Laczó

Hort

, et al.

Spatial navigation deficits—overlooked cognitive marker for preclinical Alzheimer disease?

Nat Rev Neurol 2018; 14: 496–506.

15.

Kostyniuk

Molnar

. Self-regulatory driving practices among older adults: health, age and sex effects. Accid Anal Prev 2008; 40: 1576–1580.

16.

Roe

Babulal

Head

, et al. Preclinical Alzheimer's disease and longitudinal driving decline. Alzheimers Dement (N Y) 2017; 3: 74–82.

17.

Hird

Egeto

Fischer

, et al. A systematic review and meta-analysis of on-road simulator and cognitive driving assessment in Alzheimer’s disease and mild cognitive impairment. J Alzheimers Dis 2016; 53: 713–729.

18.

Eby

Silverstein

Molnar

, et al. Driving behaviors in early stage dementia: a study using in-vehicle technology. Accid Anal Prev 2012; 49: 330–337.

19.

Babulal

Johnson

Fagan

, et al. Identifying preclinical Alzheimer’s disease using everyday driving behavior: proof of concept. J Alzheimers Dis 2021; 79: 1009–1014.

20.

Roe

Stout

Rajasekar

, et al. A 2.5-year longitudinal assessment of naturalistic driving in preclinical Alzheimer's disease. J Alzheimers Dis 2019; 68: 1625–1633.

21.

Bayat

Babulal

Schindler

, et al. GPS Driving: a digital biomarker for preclinical Alzheimer disease. Alzheimers Res Ther 2021; 13: 115.

22.

Murphy

. Olfactory and other sensory impairments in Alzheimer disease. Nat Rev Neurol 2019; 15: 11–24.

23.

Bayat

Roe

Schindler

, et al. Everyday driving and plasma biomarkers in Alzheimer's disease: leveraging artificial intelligence to expand our diagnostic toolkit. J Alzheimers Dis 2023; 92: 1487–1497.

24.

Babulal

Addison

Ghoshal

, et al. Development and interval testing of a naturalistic driving methodology to evaluate driving behavior in clinical research. F1000Res 2016; 5: 1716.

25.

Babulal

Traub

Webb

, et al. Creating a driving profile for older adults using GPS devices and naturalistic driving methodology. F1000Res 2016; 5: 2376.

26.

Donohue

Sperling

Salmon

, et al. The preclinical Alzheimer cognitive composite: measuring amyloid-related decline. JAMA Neurol 2014; 71: 961–970.

27.

Mormino

Papp

Rentz

, et al. Early and late change on the preclinical Alzheimer's cognitive composite in clinically normal older individuals with elevated amyloid β. Alzheimers Dement 2017; 13: 1004–1012.

28.

Weintraub

Carrillo

Farias

, et al. Measuring cognition and function in the preclinical stage of Alzheimer's disease. Alzheimers Dement (N Y) 2018; 4: 64–75.

29.

Radanovic

Carthery-Goulart

Charchat-Fichman

, et al. Analysis of brief language tests in the detection of cognitive decline and dementia. Dement Neuropsychol 2007; 1: 37–45.

30.

Grober

Buschke

Crystal

, et al. Screening for dementia by memory testing. Neurology 1988; 38: 900–900.

31.

Rosenberg

Ryan

Prifitera

. Rey auditory-verbal learning test performance of patients with and without memory impairment. J Clin Psychol 1984; 40: 785–787.

32.

Armitage

. An analysis of certain psychological tests used for the evaluation of brain injury. Psychol Monogr 1946; 60: i–48.

33.

Aschenbrenner

Gordon

Benzinger

TLS

, et al. Influence of tau PET, amyloid PET, and hippocampal volume on cognition in Alzheimer disease. Neurology 2018; 91: e859–e866.

34.

Galetta

Liu

Leong

, et al. The King-Devick test of rapid number naming for concussion detection: meta-analysis and systematic review of the literature. Concussion 2015; 1: CNC8.

35.

Hamilton

Balnave

Adams

. Grip strength testing reliability. J Hand Ther 1994; 7: 163–170.

36.

Peters

Fritz

Krotish

. Assessing the reliability and validity of a shorter walk test compared with the 10-meter walk test for measurements of gait speed in healthy, older adults. J Geriatr Phys Ther 2013; 36: 24–30.

37.

Gershon

Wagster

Hendrie

, et al. NIH Toolbox for assessment of neurological and behavioral function. Neurology 2013; 80: S2–S6.

38.

Volluz

Schindler

Henson

, et al. Correspondence of CSF biomarkers measured by Lumipulse assays with amyloid PET. Alzheimers Dement 2021; 17: e051085.

39.

Vos

SJB

Xiong

Visser

, et al. Preclinical Alzheimer's disease and its outcome: a longitudinal cohort study. Lancet Neurol 2013; 12: 957–965.

40.

Babulal

Stout

Benzinger

TLS

, et al. A naturalistic study of driving behavior in older adults and preclinical Alzheimer disease: a pilot study. J Appl Gerontol 2019; 1: 14.

41.

Roe

Bayat

Babulal

. Associations between plasma, imaging, and cerebrospinal fluid biomarkers with driving behavior and cognitive tests: implications for biomarker usefulness. J Alzheimers Dis Rep 2023; 7: 1095–1102.

42.

Arik

SÖ

Pfister

. Tabnet: attentive interpretable tabular learning. Proc AAAI Conf Artif Intell 2021; 35: 6679–6687.

43.

Saddiki

Fayosse

Cognat

, et al. Age and the association between apolipoprotein E genotype and Alzheimer disease: a cerebrospinal fluid biomarker-based case-control study. PLoS Med 2020; 17: e1003289.

44.

Schindler

Bollinger

Ovod

, et al. High-precision plasma β-amyloid 42/40 predicts current and future brain amyloidosis. Neurology 2019; 93: e1647–e1659.

45.

Dayon

Wojcik

Núñez Galindo

, et al. Plasma proteomic profiles of cerebrospinal fluid-defined Alzheimer’s disease pathology in older adults. J Alzheimers Dis 2017; 60: 1641–1652.

46.

Davis

Papandonatos

Miller

, et al.

Road test and naturalistic driving performance in healthy and cognitively impaired older adults: does environment matter?

J Am Geriatr Soc 2012; 60: 2056–2062.

47.

Allison

Babulal

Stout

, et al. Alzheimer disease biomarkers and driving in clinically normal older adults: role of spatial navigation abilities. Alzheimer Dis Assoc Disord 2018; 32: 101–106.

48.

Aschenbrenner

Murphy

Doherty

, et al. Neuropsychological correlates of changes in driving behavior among clinically healthy older adults. J Gerontol B Psychol Sci Soc Sci 2022; 77: 1769–1778.

49.

O’Connor

Edwards

Small

, et al.

Patterns of level and change in self-reported driving behaviors among older adults: who self-regulates?

J Gerontol B Psychol Sci Soc Sci 2012; 67: 437–446.

50.

Okonkwo

Crowe

Wadley

, et al. Visual attention and self-regulation of driving among older adults. Int Psychogeriatr 2008; 20: 162–173.

51.

Feng

Bao

Sayer

, et al. Can vehicle longitudinal jerk be used to identify aggressive drivers? An examination using naturalistic driving data. Accid Anal Prev 2017; 104: 125–136.

52.

Merickel

High

Dawson

, et al. Real-world risk exposure in older drivers with cognitive and visual dysfunction. Traffic Inj Prev 2019; 20: S110–S115.

53.

Antin

Guo

Fang

, et al. The influence of functional health on seniors’ driving risk. J Transp Health 2017; 6: 237–244.

54.

Lin

Chen

D-G

Vance

, et al. Trajectories of combined laboratory-and real world-based speed of processing in community-dwelling older adults. J Gerontol B Psychol Sci Soc Sci 2013; 68: 364–373.

55.

Talebloo

Mohammed

Far

. Deep learning approach for aggressive driving behaviour detection. arXiv preprint arXiv:211104794, 2021.