Problem

The global demographic shift towards an aging population presents significant challenges for driving safety, especially in car-dependent regions. Older adults often experience gradual declines in vision, attention, and reaction time, which can subtly reduce driving safety. Existing assessment methods, such as infrequent clinic visits or basic screening tools, offer only limited snapshots and fail to capture real-world driving capabilities. Data-driven approaches face hurdles in accurately identifying senior driving patterns due to diverse environments and highly personalized driving styles. Moreover, the absence of reliable contextual information (like traffic or weather) makes it difficult to differentiate age-related performance changes from situational factors. Many current black-box models also lack explanatory value, making it hard to link observed behaviors to underlying cognitive processes. This creates a critical gap, leaving early signs of driving capability decline undetected in real-world settings and posing significant risks to older drivers.

Method

The paper introduces AURA (Aging Understanding & Risk Assessment), an Artificial Intelligence of Things (AIoT) framework designed for continuous, real-world assessment of driving safety in older adults. AURA leverages a multi-faceted approach, starting with richer in-vehicle sensing to capture both vehicle control patterns and subtle in-cabin behavioral features like head movements and gaze shifts. It employs multi-scale behavioral modeling, fusing fast control actions with fine-grained motion features and long-term habits using temporal convolution, self-attention, and hierarchical temporal models to discern consistent behavioral drift from benign fluctuations. Critically, AURA integrates context-aware analysis, embedding environmental data such as traffic density, weather, and road geometry into its reasoning pipeline through cross-modal alignment and attention-based fusion, allowing for interpretation of behaviors relative to situational demands. For explainability and actionability, AURA uses a "Semantic Abstraction" approach, developed with geriatricians, to translate raw driving traces into transparent, cognitively grounded explanations. This involves decomposing sensor data into micro-behaviors, mapping them to specific cognitive domains, and generating structured, auditable diagnostic reports. The system is designed for edge execution with optimized models (distillation, quantization) and ensures privacy through Federated Learning, transmitting only differentially private gradient updates to the cloud.

Results

Through controlled CARLA simulator observations, the study found that older adults exhibited a risk-averse driving profile, characterized by slower increases in brake usage, reduced throttle, and more frequent, wider head-scanning motions to compensate for visual and attentional declines. Route-level analysis highlighted significant differences in stability, reaction time (especially at intersections), and route deviation (around curves) compared to younger drivers. Analyzing real-world data from the LongROAD dataset revealed that older adults steadily reduce monthly driving distance, trip frequency, and night driving, demonstrating self-regulation. However, these adaptive behaviors complicate the inference of underlying cognitive decline. The DRIVES dataset showed that cognitively impaired seniors brake more often and drive at lower speeds, but these patterns were unstable week-to-week, influenced by context and compensatory strategies. Furthermore, individual driving habits often masked subtle cognitive changes, making it hard to distinguish impairment from personalized styles. Context-aware analysis demonstrated that senior driving behavior is highly influenced by the environment; for example, the presence of a lead vehicle or rainy conditions significantly altered throttle usage, braking, and head-scanning patterns, underscoring the necessity of contextual understanding for accurate assessment.

Implications

AURA provides a crucial framework for fostering safer driving among older adults by offering continuous, individualized, and transparent assessments, bridging the gap between current episodic evaluations and the dynamic reality of age-related changes. This system empowers clinicians with objective diagnostics and equips families with actionable insights, promoting a balance between individual independence and public safety. Beyond the immediate scope of AURA, the paper outlines forward-looking research directions for next-generation aging and driving safety systems. These include developing Elderly Cognitive World Models to understand how cognition evolves over time and influences driving behavior, fostering Federated Embodied Learning where vehicles collaboratively improve shared foundational models while preserving individual privacy, and exploring Digital Therapeutics and Active Health by transforming in-vehicle systems into "active cognitive clinics" that deliver adaptive, just-in-time cognitive stimulation. These advancements aim to leverage AIoT to create a more inclusive society, supporting older adults in maintaining their dignity and independence through intelligent and compassionate systems.