LCID: LLM-powered Cognitive Impairment Diagnosis over Long-term Multimodal Sensor Data with Hyperdimensional Encoding and Attentional Fusion

With the accelerating global population aging, health monitoring has gradually become a critical issue in public healthcare. Cognitive impairment (CI), as a prevalent health concern among the elderly population, exhibits age-related progression in symptom severity and may ultimately develop into serious cognitive disorders such as Alzheimer's disease and dementia. The rapid advancement and widespread application of sensing technologies have driven significant progress in behavior perception and early disease diagnosis based on sensor-monitored data. To address these challenges, this paper proposes LCID, a CI diagnosis framework leveraging large language models (LLMs) based on long-term multimodal sensor data. While LLMs possess domain-specific expertise and exceptional reasoning capabilities, directly inputting raw sensor data into LLMs is suboptimal due to constraints in inference efficiency and text prompt length limitations. Therefore, LCID adopts a hierarchical framework. First, raw data are transformed into abstract representations through semantic perception, for which we design a multimodal feature fusion strategy integrating hyperdimensional encoding mechanisms. Subsequently, long-term activity sequences are constructed to enable CI diagnosis leveraging LLMs. We also design and further optimize the structure of prompts. Experimental results on a real-world dataset demonstrate the promise of our framework in utilizing LLMs to assist accurate CI diagnosis.