A Multi-Hypothesis Monitoring Architecture: Application to Ambulatory Physiology

Kristian Kersting, Benoît Vettier, Laure Amate, Catherine GarbayStairs 2012 · DOI

TL;DR

This paper introduces a novel multi-agent system architecture designed to monitor human activity by managing multiple concurrent hypotheses about the user's state. This approach tackles the inherent uncertainty in sensor data and context, aiming to detect deviations from expected behavior rather than identifying a single "correct" situation, making it particularly relevant for applications like ambulatory health monitoring.

What Problem Does This Paper Address?

Monitoring human activity, especially in real-world, ambulatory settings, is challenging due to two primary issues: the influence of context and the inherent limitations of sensor data. Sensor data can be noisy, ambiguous, or incomplete, making it difficult to definitively determine a person's current state. Furthermore, understanding human activity requires considering not just the raw data but also the broader context, including the environment, temporal sequences of events, and user expectations. Traditional monitoring systems often struggle with this ambiguity, attempting to pinpoint a single most likely situation. This paper addresses the need for a more robust approach that can handle uncertainty by considering multiple plausible interpretations (hypotheses) simultaneously and dynamically adapting to changing conditions.

Methods

The core of the proposed solution is a Multi-Agent System (MAS) that employs a multi-hypothesis monitoring architecture. This architecture is built around a dynamic population of "hypothesis-agents," each representing a potential situation. These agents operate in a cycle of verification (evaluating their likelihood against incoming sensor data) and prediction (generating new hypotheses when current ones become weak).

Key methodological innovations include:

  • Normative MAS: The system incorporates organizational rules and filters to govern agent behavior and information flow, ensuring compliance with high-level requirements and dynamic adaptation.
  • Situation Network: A knowledge base of interconnected situation models, organized by abstraction levels, allows for the decomposition of complex scenarios into simpler, data-driven states.
  • Focus, Anticipation, and Exploration: These mechanisms guide the hypothesis generation process. "Focus" evaluates data-level hypotheses within a given context. "Anticipation" generates successor hypotheses when data changes. "Exploration" allows the system to move from data-level hypotheses to higher-level, more meaningful ones.
  • Confidence Ranges and Filters: Hypotheses are assigned confidence values, categorized into "low," "medium," and "high" ranges. Filters dynamically adjust system parameters (e.g., thresholds for confidence ranges) to manage the number of hypotheses, balance sensitivity with computational cost, and ensure system efficiency.
  • Adaptive Control: The system is designed for adaptive control, allowing parameters and rules to be modified on-the-fly, either autonomously or through human intervention, to adapt to evolving situations and requirements.

Datasets

The paper mentions the use of data collected from a monitoring belt providing synchronized Heart Rate, Breath Rate, Skin Temperature, and 3-axis accelerometry. This data was tailored to a specific subject and collected in cooperation with physiologists. Future data collection is planned to include environmental data like altitude, outside temperature, and sound volume to enrich context interpretation. The paper does not introduce new benchmarks but rather uses this collected data for simulating real-time interpretation.

Key Insights and Ideas

  • Embracing Uncertainty: The most significant insight is the shift from trying to find a single "true" situation to managing and comparing multiple concurrent hypotheses. This acknowledges the inherent ambiguity in real-world data.
  • Dynamic Hypothesis Management: The system's ability to dynamically generate, evaluate, and replace hypotheses based on incoming data and evolving contexts is crucial for robust monitoring.
  • The Power of Abstraction Levels: Organizing situation models across different abstraction levels (States and Micro-scenarios) allows for both detailed data interpretation and the understanding of broader, meaningful situations.
  • Normative Control for Adaptability: The use of normative MAS with filters provides a flexible and powerful mechanism for regulating system behavior, adapting to changing requirements, and ensuring operational efficiency.
  • Focus on Discrepancies: The goal is not just to recognize events but to detect discrepancies between hypothesized situations and expectations, which is a more nuanced approach to monitoring.

Key Results

The paper presents simulation results demonstrating the system's dynamics. Figure 5 illustrates the number of likely and unlikely hypotheses over time, showing how the "Efficiency filter" dynamically adjusts parameters to manage the number of hypotheses. A drop in unlikely hypotheses is observed when the system becomes less tolerant, followed by the generation of new hypotheses to replace failing ones. Figure 6 demonstrates the influence of verification frequency on hypothesis confidence, showing how tuning this parameter can affect the system's tolerance to low-likelihood hypotheses and the number of anticipated hypotheses. These results highlight the system's expressiveness and adaptivity potential in interpreting ambulatory physiological data.

Limitations and Future Work

The authors acknowledge that the knowledge base (situation network) must be learned and computed off-line, which is outside the scope of this paper. They also note that the complexity of the interpretation system increases with the richness of the hypotheses, leading to more intricate parameter interdependencies.

Future work will focus on:

  • Training Adaptive Control Rules: Discovering heuristics for the system's adaptive control, as parameter interdependencies can be costly to manage manually.
  • Specification of Alert Rules: Further work is needed on specifying alert rules, particularly for comparing expected scenarios with hypothesized timelines.
  • Enriching Context: Incorporating additional data sources like environmental data to enhance context interpretation.
  • More Sophisticated Models: Developing more sophisticated models for physiological data and scenarios.

Summary updated

BibTeX
@inproceedings{kersting2012multi,
  author = {Kristian Kersting and Benoît Vettier and Laure Amate and Catherine Garbay},
  title = {{A Multi-Hypothesis Monitoring Architecture: Application to Ambulatory Physiology}},
  booktitle = {Stairs},
  year = {2012},
  doi = {10.3233/978-1-61499-096-3-348},
}
Figure