TL;DR
This paper presents a novel smart home system that uses IoT sensors to automatically classify individuals as either normal controls or in the early stages of dementia, based on their daily living activities (ADLs). By personalizing the classification criteria to individual cognitive abilities, the system achieved a high accuracy of 91.20%, offering a more objective and less intrusive alternative to traditional subjective ADL assessments.
What Problem Does This Paper Address?
The global aging population is leading to a significant increase in dementia cases and elderly individuals living alone. Diagnosing dementia relies heavily on assessing Activities of Daily Living (ADLs), but current methods, primarily ADL questionnaires, are subjective and prone to clinician bias, potentially leading to inaccurate assessments. Furthermore, the unique cultural lifestyles and living environments in South Korea, along with the increasing number of elderly living alone, present specific challenges for existing ADL assessment tools. This paper aims to bridge this gap by developing an objective, automated system for early dementia detection using smart home sensor data, tailored to individual cognitive profiles.
Methods
The core of this research lies in leveraging a smart home environment equipped with various IoT sensors (motion, door, temperature-humidity, vibration, lidar, smart plugs) to collect data on participants' ADLs. The study focuses on "inferred ADLs" derived from sensor data, rather than direct observation, to maintain unobtrusiveness. Key ADL categories like cooking, household chores, grooming, taking medications, using household appliances, and indoor wandering were monitored.
The methodology involves several key steps:
- Data Collection: ADL data was collected over an extended period (approximately 13 months) from participants living in smart home environments.
- Feature Engineering: Raw sensor data was pre-processed to extract meaningful features related to ADL activities, including counts, durations, and movement patterns. This involved combining data from multiple sensors to infer specific ADLs.
- Personalization: A crucial innovation is the personalization of ADL anomaly detection criteria. Instead of using uniform standards, the system adapts anomaly detection thresholds based on an individual's cognitive ability, measured by their MMSE score. This accounts for variations in normal daily living patterns among individuals.
- Machine Learning Classification: A Random Forest (RF) algorithm was employed for classifying participants into normal control and early-stage dementia groups. The study compared the performance of a personalized RF model against a non-personalized model.
Datasets
The study utilized data collected from 13 elderly individuals living alone in South Korea: 7 normal controls and 6 with early-stage dementia. The data was gathered over approximately 13 months (May 2020 to June 2021). This dataset, collected from actual residential smart home environments, is unique in its focus on Korean cultural lifestyles and the specific challenges faced by elderly individuals living alone. The paper does not explicitly introduce a new benchmark dataset but rather uses this collected data for training and evaluation.
Key Insights and Ideas
- Personalization is Key: The most significant insight is that tailoring ADL anomaly detection criteria to an individual's cognitive ability (MMSE score) dramatically improves classification accuracy. This acknowledges that "normal" behavior varies significantly between individuals.
- Late-Night Activity Matters: Contrary to expectations, late-night activities (00:00-05:00) and evening hours (17:00-24:00) related to using appliances like electric mats and TVs, as well as general TV watching, emerged as highly impactful features for distinguishing between normal controls and early-stage dementia patients. This suggests that subtle changes in nighttime routines can be strong indicators.
- IoT Big Data Outperforms Specific ADLs: The study found that using a broad range of IoT sensor data, reflecting various aspects of daily living, yielded better classification performance than relying solely on features derived from specific ADL activities. This highlights the power of comprehensive, unobtrusive monitoring.
- Objective ADL Assessment is Achievable: The research successfully demonstrates that smart home sensor data can provide an objective and reliable alternative to subjective ADL questionnaires for dementia diagnosis.
Key Results
- Personalized Model Superiority: The personalized Random Forest model achieved an accuracy of 91.20%, significantly outperforming the non-personalized model, which achieved 84.54% accuracy.
- IoT vs. ADL Features: The IoT-only model (86.80% accuracy) performed better than the ADL-only model (83.47% accuracy), reinforcing the value of comprehensive sensor data.
- Statistical Differences: Significant statistical differences (p < 0.05) were found between normal controls and early-stage dementia groups in the usage counts and durations of most household appliances and ADL activities, with notable exceptions like 'microwave oven' and 'pill organizer' usage counts.
- Gait Speed Difference: A statistically significant difference was observed in gait speeds between the two groups (p < 2.2 × 10–16), with early-stage dementia patients exhibiting a slightly faster average gait speed.
Limitations and Future Work
- Small Sample Size: The study acknowledges that the number of participants was small (13 individuals) due to the challenges of home visits and recruitment during the COVID-19 pandemic. The study period was also relatively short (13 months).
- Future Personalization: The authors plan to conduct further research on automating the personalization process, potentially integrating it with ongoing studies that use hospital records for MCI prediction.
- Domain Knowledge: The study's success in achieving high classification accuracy without relying heavily on pre-defined domain knowledge about ADLs is highlighted as a significant contribution, suggesting that data-driven approaches can be highly effective.
Summary updated