TL;DR
This paper introduces three novel entropy-based measures derived from in-home IoT sensor data to quantify changes in daily activity patterns. These measures, when combined, show promise in indicating healthcare-related events in individuals with dementia, offering a new avenue for proactive monitoring and early intervention.
What Problem Does This Paper Address?
The paper tackles the challenge of effectively monitoring the well-being of individuals, particularly those with dementia (PLWD), living independently at home. While existing methods often rely on machine learning models to analyze raw sensor data, they may miss subtle yet crucial information about the uncertainty and distribution of activities. The authors propose that by quantifying the complexity and predictability of daily routines using information theory concepts like entropy, they can derive more insightful features that better capture deviations indicative of health issues. This is crucial for developing early intervention strategies and improving the quality of life for PLWD.
Methods
The core of this work lies in applying and adapting three information-theoretic measures to time-series activity data collected from IoT sensors:
- Shannon's Entropy: This standard measure quantifies the uncertainty or randomness in the distribution of locations visited within a home. A higher entropy suggests more varied or unpredictable movement patterns.
- Entropy Rate of a Markov Chain: This measure focuses on the predictability of transitions between locations. By modeling daily movement as a first-order Markov chain, the entropy rate captures how much "new information" is gained from observing the next location, given the current one. This is particularly relevant for identifying deviations from routine.
- Entropy Production Rate (EP): This measure, rooted in non-equilibrium thermodynamics, is adapted to quantify the "dissipation" or irreversibility in the system's transitions. The authors employ a neural network-based approach (NEEP) to estimate this rate from the time-series data, viewing in-home movement as a non-equilibrium system.
These entropy measures are then used as engineered features to be fed into downstream inference and analysis models.
Datasets
The study utilizes a large-scale in-home monitoring dataset collected between December 2020 and March 2022 as part of a dementia care clinical study. This dataset comprises 9,370 person-days of activity data from participants' homes, gathered using a variety of IoT sensors (PIR sensors, smart plugs, thermometers, wearable devices, smart scales, light sensors, and temperature sensors). The data is labeled by a monitoring team who verify alerts and identify healthcare-related events such as accidental falls, abnormal motor function, hospital admissions, UTIs, anxiety, depression, disturbed sleep, agitation, and confusion.
Key Insights and Ideas
- Entropy as a Measure of Routine: The paper highlights that entropy measures can effectively capture the "routineness" of daily activities. Participants with more consistent daily patterns tend to exhibit lower entropy values, while deviations from routine lead to increased entropy.
- Complementary Nature of Entropy Measures: No single entropy measure consistently predicts all healthcare events. The combination of Shannon's entropy, entropy rate, and entropy production rate provides a more comprehensive view, as different events or individuals might manifest changes in different entropy metrics.
- Individualized Patterns: The study observes that even for the same healthcare-related event, different participants may show distinct changes in their entropy measures. This underscores the need for personalized monitoring and analysis.
- Potential for Early Detection: The preliminary results suggest that these entropy measures can act as sensitive indicators of subtle changes in activity patterns, potentially signaling the onset of healthcare-related events before they become critical.
Key Results
The experiments demonstrate that the proposed entropy measures, particularly when used in combination, can indicate the occurrence of healthcare-related events. For instance, in analyzing "anxiety and depression" (C1 in Fig. 7), different participants showed increases in daytime EP, nighttime entropy, or a combination thereof. Similarly, for "disturbed sleep pattern" (C2), changes were observed across various entropy measures, with some participants showing significantly higher nighttime entropy. The paper emphasizes that using a combination of these measures in an inference model is more effective than relying on any single measure for improved prediction accuracy.
Limitations and Future Work
The authors acknowledge that their retrospective analysis is based on historical data. A key limitation is that a single entropy measure might not be sufficient for ideal prediction, necessitating the combination of measures.
Future work includes:
- Deploying these entropy measures within their digital platform (Minder) to evaluate their clinical utility in real-world settings.
- Further investigating the impact of seasonal effects and circadian rhythms on these measures.
- Exploring the use of these entropy measures as engineered features in modern machine learning models to enhance the accuracy of inference and predictive models.
- Conducting more extensive studies to validate the findings across a larger and more diverse population.
Summary updated