Together about Dementia

Nicklas Sindlev Andersen, Marco ChiarandiniIJCAI 2022 · DOI

TL;DR

This paper introduces "Together about Dementia" (SOD), a mobile health app that uses AI to detect when individuals with dementia are wandering and lost, triggering a community-based rescue mission involving relatives and volunteers. The system leverages spatio-temporal anomaly detection on location data to proactively identify at-risk individuals, aiming to improve their safety and reduce caregiver burden.

What Problem Does This Paper Address?

The number of people diagnosed with dementia is rising globally, largely due to an aging population. A significant challenge with dementia is the increased risk of wandering and getting lost, which poses a serious safety concern for individuals and a substantial burden on their caregivers and relatives. While existing mobile health (mHealth) solutions offer location tracking, they often lack proactive detection of wandering behavior and coordinated community response. This paper addresses the need for an intelligent system that can not only track but also predict and respond to a person with dementia becoming lost, thereby enhancing their safety and providing a more comprehensive support network.

Methods

The core of the "Together about Dementia" (SOD) system is its AI engine for automatic wandering detection. This engine operates on two components:

  • Periodic Component: This part processes historical location data to learn typical movement patterns of a person with dementia. It decomposes continuous location streams into trajectories, segments them into geohash sequences, and extracts frequent sequential patterns using the PrefixSpan algorithm. This creates a personalized model of normal behavior.
  • Real-time Component: This component compares the person's current movement pattern against their learned historical patterns. When a new location data point is added, it calculates an anomaly score. If this score exceeds a predefined threshold, it indicates anomalous wandering behavior, triggering an alarm.

The system's architecture is built on a microservices and serverless approach using Kubernetes (k3s). This design choice emphasizes efficiency, reliability, and maintainability. Redis is used as a message queue and broker for real-time communication between components and users. The backend system manages the rescue mission, coordinating alerts to relatives and nearby volunteers.

Datasets

The paper mentions that the automatic wandering detection engine has been tested and compared with an alternative approach using synthetically generated data. User tests have also been conducted in protected environments. The authors make their code-base and test data available.

Key Insights and Ideas

  • Proactive AI-driven detection: Moving beyond simple location tracking, the system's AI actively predicts and identifies wandering behavior before a person is confirmed lost, enabling earlier intervention.
  • Community-powered rescue: The app mobilizes a network of relatives and volunteers, creating a distributed support system that leverages local community resources for safety.
  • Privacy-preserving coordination: Volunteers are not given the exact address of the person with dementia, enhancing privacy while still enabling them to provide assistance.
  • Participatory design: The app's development involved a broad range of stakeholders (PwD, relatives, nurses, administrators, AI experts, developers), ensuring the system is user-centered and practical.
  • Scalable and maintainable architecture: The use of microservices and serverless functions allows the system to scale efficiently and be easily updated.

Key Results

The paper focuses on the demonstration of the system and its architecture. While specific quantitative results on detection accuracy or mission success rates are not detailed in this demonstration paper, the authors state that the AI engine has been tested on synthetic data and the architecture has undergone scalability and reliability load tests. User tests in protected environments have been completed, and the app is being launched on a restricted set of users for further monitoring.

Limitations and Future Work

  • Synthetic Testing: The automatic wandering detection engine has primarily been tested in synthetic settings. The authors acknowledge that the system has only been tested in synthetic settings at the time of writing.
  • Real-world Assessment: The next crucial step is the assessment of real-life rescue mission episodes to evaluate the system's performance in actual scenarios.
  • User Adoption and Engagement: While user tests have been conducted, understanding long-term user adoption and engagement, especially among volunteers, will be important.
  • AI Model Refinement: Continuous refinement of the AI engine and its parameters (like the anomaly threshold $\theta$) will be necessary as more real-world data becomes available.

Summary updated

BibTeX
@inproceedings{andersen2022together,
  author = {Nicklas Sindlev Andersen and Marco Chiarandini},
  title = {{Together about Dementia}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/846},
}
Figure