TL;DR
This paper introduces a novel "Glass Box" Markov Chain model to detect subtle changes in the daily activity patterns of people living with dementia (PLWD) using in-home IoT sensor data. By analyzing kitchen activity as a proxy for eating and drinking habits, the model aims to proactively identify potential health issues like malnutrition and dehydration, paving the way for more timely and personalized care.
What Problem Does This Paper Address?
Malnutrition and dehydration are significant concerns for individuals with dementia (PLWD), often leading to accelerated cognitive decline and increased hospitalizations. These issues are frequently linked to extreme changes in eating and drinking behaviors. However, there's a critical lack of objective methods to detect these behavioral shifts in real-time, especially within the familiar environment of a person's home. This leaves a gap in care, where interventions are often reactive rather than proactive. The COVID-19 pandemic, with its disruptive impact on daily routines, provided a unique opportunity to study how such external events influence the behavior of PLWD and to test methods for detecting these changes.
Methods
The researchers employed a two-pronged approach:
- Quantitative Analysis of COVID-19 Impact: They used linear mixed-effects (LME) modeling to analyze in-home kitchen activity data from 21 PLWD households over a 499-day period, encompassing the COVID-19 pandemic and its associated lockdowns. This allowed them to quantify the impact of the pandemic on daily routines.
- Novel "Glass Box" Markov Chain Model: To address the need for proactive anomaly detection, they developed a "Glass Box" analytical approach. This involves:
- Proxy for Behavior: Using kitchen activity (detected by IoT sensors like motion, fridge door, and smart plugs for appliances) as a proxy for eating and drinking habits.
- Transition Matrices: Deriving "transition matrices" from sequences of sensor recordings. These matrices represent behavioral patterns by quantifying the probabilities of moving between different kitchen activities (e.g., from fridge to kettle).
- Sliding Window Algorithm: Applying a sliding window technique to extract these transition matrices over time.
- Dissimilarity Measure: Using the Frobenius distance to quantify the dissimilarity between transition matrices from different time periods (e.g., a baseline period versus a current period). A higher dissimilarity indicates a significant change in behavior.
- Explainability: The "Glass Box" nature of the model means that the detected changes can be traced back to specific shifts in the transition probabilities, making the model's decisions interpretable.
Datasets
The study utilized in-home monitoring data collected from:
- 73 households of PLWD: This larger cohort was used for the proof-of-concept study of the Markov Chain model.
- 21 households of PLWD: This subset was used for the quantitative analysis of the COVID-19 pandemic's impact on kitchen activity.
The data was collected using Develco IoT devices, including passive infrared motion sensors, fridge door sensors, and smart plugs for kitchen appliances. All data was anonymized and made publicly available.
Key Insights and Ideas
- COVID-19 as a Natural Experiment: The pandemic provided a unique, real-world scenario to observe how external disruptions significantly alter the daily routines of PLWD, particularly their kitchen-related activities.
- Kitchen Activity as a Proxy: The study highlights the utility of using readily available IoT sensor data from the kitchen as a reliable proxy for complex behaviors like eating and drinking, which are difficult to measure directly.
- Proactive vs. Reactive Care: The core idea is to shift from reactive interventions (responding after a health crisis) to proactive monitoring and early detection of behavioral changes that may signal impending health issues.
- "Glass Box" Explainability: The proposed Markov Chain model is designed to be interpretable, allowing caregivers and researchers to understand why a change is flagged, rather than just receiving an alert. This is crucial for building trust and facilitating appropriate responses.
- Individualized Behavioral Profiles: The model's ability to be applied to each household individually allows for the creation of unique behavioral profiles, enabling the detection of deviations from an individual's "normal" patterns.
Key Results
- COVID-19 Impact: The study found a significant increase in daytime kitchen activity and a significant decrease in nighttime kitchen activity during the COVID-19 pandemic. This change preceded official lockdown measures, suggesting a proactive behavioral adaptation by PLWD.
- Markov Model Efficacy: Case studies demonstrated that the Markov Chain model could effectively detect:
- Sustained Changes: Gradual shifts in behavior over time, as seen during the COVID-19 pandemic periods.
- Cumulative Changes: The accumulation of behavioral shifts, potentially linked to successive health events.
- Episodic Changes: Short-term, distinct behavioral alterations, such as those related to sleep disturbances or medication effects.
- Nighttime Activity Decline: A statistically significant decrease in nighttime kitchen activity was observed ($t$(147) = -2.90, $p$ < 0.001), which is a key finding suggesting potential disruptions in sleep-wake cycles or evening routines.
Limitations and Future Work
- Proxy Measures: The study acknowledges that kitchen activity is a proxy and doesn't directly measure food or fluid intake. Future work could integrate data from other sensors (e.g., smart cups, scales) for a more comprehensive picture.
- Generalizability: While the model is presented as "Glass Box" and generalizable, further validation across diverse populations and a wider range of adverse health events is needed.
- Algorithm Tuning: The algorithm's parameters (window size, step frequency) were curated through trial-and-error. More systematic methods for parameter optimization could be explored.
- Future Directions: The authors propose deploying this algorithm within the "Minder" platform to generate real-time alerts for carers, which can then be validated. They also aim to expand the model to incorporate data from sensors throughout the entire household and to conduct qualitative studies to gather feedback on alert usefulness.
Summary updated