Leveraging Naturalistic Driving Digital Biomarkers for Early Mild Cognitive Impairment Detection: Deep Learning Strategies

Background: Alzheimer disease and related dementias are increasing worldwide, with early detection during the mild cognitive impairment (MCI) stage critical for timely intervention. Driving behavior, which reflects everyday cognitive functioning, has emerged as a promising, noninvasive, and inexpensive digital biomarker when paired with machine learning. However, prior research has often relied on controlled settings, high-level features, or assumptions that fail to capture the sporadic nature of MCI, leaving a gap in modeling naturalistic driving data for robust early detection. Objective: This study aims to address the limitations of prior work by developing deep learning strategies that leverage driving data collected in a naturalistic setting as digital biomarkers for early detection of MCI. Methods: Clinically classified participants (8 with MCI and 14 cognitively normal; N=22) drove their personal vehicles under naturalistic conditions for several consecutive days. A total of 3 participants (2 cognitively normal and 1 MCI) withdrew before completing the experiments. In-vehicle sensors recorded GPS, accelerometer, and gyroscope signals, which were segmented into full trips and turning maneuvers. Three modeling strategies were compared: (1) single-view, (2) feature-level fusion, and (3) model-level late fusion. Classification models were trained and evaluated to assess their accuracy, discriminative ability, and participant-level performance. Results: Models using full-trip data consistently outperformed turn-only inputs, with the best-performing model achieving 78% accuracy and an area under the receiver operating characteristic curve of 77%. Turn-based inputs alone demonstrated limited discriminative power; however, combining them with trip data through late fusion improved performance, though not beyond the full-trip baseline. Participant-level analysis indicated that classification accuracy improved with increased data volume, and trip-wise modeling more effectively captured the episodic nature of MCI than majority-vote aggregation. A frequency-based risk score was proposed as an interpretable and flexible output, enabling practical application in clinical and community settings. Conclusions: Naturalistic driving behavior offers a scalable and noninvasive approach for early cognitive screening. Deep learning models using full-trip naturalistic driving data show promise for detecting MCI, with fusion strategies providing supplementary insights. This framework supports proactive, real-world monitoring of cognitive decline, laying the foundation for digital health interventions in dementia prevention.