Identifying major depressive disorder in older adults through naturalistic driving behaviors and machine learning

Chen Chen, David C. Brown, Noor Al-Hammadi, Sayeh Bayat, Anne Dickerson, Brenda Vrkljan, Matthew Blake, Yiqi Zhu, Jean-François Trani, Eric J. Lenze, David B. Carr, Ganesh M. BabulalNpj Digit. Medicine 2025 · DOI

TL;DR

This study leverages naturalistic driving data from older adults and machine learning to identify Major Depressive Disorder (MDD). The findings suggest that driving patterns can serve as a digital biomarker for depression, offering a novel, objective, and potentially scalable method for early detection and intervention.

What Problem Does This Paper Address?

Diagnosing Major Depressive Disorder (MDD) in older adults is a significant challenge. Symptoms of depression can mimic or be masked by other age-related conditions like mild cognitive impairment (MCI) and Alzheimer's disease (AD), leading to delayed or incorrect diagnoses. Furthermore, older adults may be less likely to report their distress due to stigma or a belief that symptoms are a normal part of aging. This diagnostic ambiguity can result in poorer health outcomes and increased risks, including motor vehicle crashes. Traditional screening methods often lack specificity and sensitivity. The paper aims to address this by exploring whether objective, real-world driving behavior data, analyzed with machine learning, can provide a more accurate and accessible way to identify depression in this population.

Methods

The study employed machine learning techniques, specifically XGBoost and logistic regression, to analyze naturalistic driving data.

  • XGBoost: This ensemble learning algorithm was chosen for its ability to handle complex datasets, minimize loss using second-order gradients, and its scalability and accuracy. The researchers used GridSearchCV with 10-fold GroupKFold cross-validation to tune hyperparameters (learning rate, maximum depth, gamma, subsample) and optimize for recall, aiming to capture as many true depression cases as possible.
  • Logistic Regression: This served as a baseline model to compare against XGBoost, particularly for demographic and antidepressant-related features.
  • Feature Engineering: Driving data was aggregated monthly and included features such as the rate of hard braking and cornering events, number of short and medium-length trips, and distance traveled. Demographic data (age, sex, education) and medication information (total classes of medications, antidepressant use) were also incorporated as features.

Datasets

The study utilized a dataset of naturalistic driving data collected over two years from 157 older adults.

  • Participants: 81 participants were diagnosed with MDD, and 76 were non-depressed.
  • Data Collection: In-vehicle GPS dataloggers captured driving behaviors.
  • Data Aggregation: Data was aggregated monthly for each participant.
  • Comparison: The study did not introduce a new benchmark dataset but rather used existing data to evaluate the proposed machine learning models.

Key Insights and Ideas

  • Driving as a Digital Biomarker: Naturalistic driving behavior shows significant potential as an objective, functional digital biomarker for identifying depression in older adults. This moves beyond subjective self-reports and clinician assessments.
  • Subtle Behavioral Changes: Even subtle changes in driving patterns, like increased hard braking or less predictable routes, can be indicative of underlying depression.
  • Medication Matters: The inclusion of medication data, particularly the total number of medication classes, significantly improved the model's performance in identifying depression. This highlights the complex interplay between medical conditions, treatments, and functional abilities.
  • Demographics' Limited Role: While age and sex are known risk factors for depression, their direct inclusion in the models, especially with logistic regression, led to overfitting and poorer generalization. This suggests that driving behavior and medication are more potent predictors in this context.
  • Scalability and Accessibility: The use of readily available GPS technology and machine learning offers a pathway to scalable and cost-effective screening for depression, potentially reaching a wider population.

Key Results

  • Driving Features Alone: A model using only driving features achieved a good F1 score of 0.82, 80% specificity, and a high AUC of 0.84.
  • Best Performing Model: The top-performing XGBoost model combined total classes of medications and driving features, achieving a high F1 score of 0.81, a comparable AUC of 0.86, precision of 0.74, and recall of 0.90. The high recall is particularly valuable for minimizing false negatives.
  • Feature Importance: For the best model, the most important features included the rate of hard cornering, total medication classes, number of trips between 1-5 miles, rate of hardcore braking events, and number of trips less than 1 mile.
  • Logistic Regression Performance: Logistic regression models, especially those incorporating demographics, performed less effectively, with an F1 score of 0.63 for demographics and antidepressants combined.

Limitations and Future Work

  • Confounding Factors: The study acknowledges limitations such as the potential for shared vehicle use to introduce noise, reliance on self-reported medication data (which may not be entirely accurate or up-to-date), and the lack of detailed electronic health record (EHR) data on psychiatric history or the specific purpose of antidepressant use. Over-the-counter medications were also not fully accounted for.
  • Comorbidities: The presence of cognitive decline or preclinical AD in participants was not explicitly controlled for, which could have influenced both driving behavior and depressive symptoms.
  • Generalizability: The depressed group was younger and more likely to be female and on antidepressants, which may limit the generalizability of the findings to the broader older adult population.
  • Future Directions: The authors suggest future research should:
  • Incorporate EHR data and a more comprehensive medication list to better understand their impact.
  • Utilize more sophisticated modeling techniques, such as deep learning, to capture more complex patterns.
  • Recruit larger, more diverse samples and extend observation periods.
  • Develop methods to distinguish between drivers and ensure data quality.
  • Integrate additional contextual data (e.g., cognitive assessments) to disentangle the effects of cognitive impairment and depression.

Summary updated

BibTeX
@article{chen2025identifying,
  author = {Chen Chen and David C. Brown and Noor Al-Hammadi and Sayeh Bayat and Anne Dickerson and Brenda Vrkljan and Matthew Blake and Yiqi Zhu and Jean-François Trani and Eric J. Lenze and David B. Carr and Ganesh M. Babulal},
  title = {{Identifying major depressive disorder in older adults through naturalistic driving behaviors and machine learning}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-01500-w},
}
Figure