A Hybrid AI Framework for Sensor-Based Personal Health Monitoring towards Precision Health

Mbithe NzomoAAAI 2024 · DOI

Abstract

As wearable technology becomes ubiquitous, the challenge shifts from data collection to the generation of actionable, clinically relevant insights. This paper presents a hybrid neuro-symbolic AI framework designed to bridge the gap between raw sensor data and precision health. By integrating machine learning for pattern recognition with symbolic reasoning and Bayesian networks for expert knowledge and uncertainty management, the framework provides an explainable, agent-based approach to monitoring conditions such as atrial fibrillation and chronic stress.

Introduction

The proliferation of wearable sensors has enabled continuous monitoring of physiological signals, yet the transition from simple data tracking to "precision health" remains elusive. Current health monitoring systems often operate in silos, providing data points without the necessary clinical context. To move toward a proactive healthcare model, AI systems must not only detect abnormalities but also predict future adverse outcomes and provide explanations that align with medical expertise.

This research addresses three primary hurdles in the current landscape: the lack of process interoperability within clinical workflows, the limitation of systems to detection rather than prediction, and the "black box" nature of deep learning models which hinders clinical trust. We propose a hybrid architecture that treats sensor data as part of a broader ecosystem including lifestyle, medical history, and expert guidelines.

The Hybrid AI Architecture

The core of the proposed framework is a neuro-symbolic approach that combines the strengths of data-driven and knowledge-driven AI. This integration is managed through an agent-based paradigm using the Beliefs-Desires-Intentions (BDI) framework.

Neuro-symbolic Integration

While deep learning (the "neuro" component) excels at extracting complex patterns from high-frequency signals like ECG or EEG, it lacks the ability to reason about medical logic. Conversely, symbolic AI (ontologies and knowledge graphs) excels at representing expert rules and clinical guidelines but cannot process raw signal data. By combining these, the framework can interpret a detected heart rate spike through the lens of a patient's known medical history and current clinical protocols.

Reasoning Under Uncertainty

Medical diagnosis is inherently probabilistic. To address this, the framework utilizes Bayesian Networks to model causal relationships and handle the uncertainty inherent in sensor noise and biological variability. This allows the system to provide a confidence level for its findings and explain the causal chain that led to a specific recommendation.

The BDI Agent Model

The system is structured as a "Personal Health Agent." Using the BDI model, the agent maintains:

  • Beliefs: The current state of the user based on sensor data and history.
  • Desires: Long-term health goals (e.g., maintaining sinus rhythm).
  • Intentions: Specific actions or interventions (e.g., alerting a physician or suggesting a breathing exercise).

Clinical Interoperability and Ontologies

A significant contribution of this work is the focus on "process interoperability." Most systems focus on data formats (e.g., HL7 FHIR), but this framework utilizes specialized ontologies to represent clinical workflows. This ensures that the AI’s suggestions are grounded in established medical practice, making the output more useful for healthcare providers who must integrate these insights into standard care paths.

Use Cases: Atrial Fibrillation and Stress

The framework has been initially validated through a use case focused on Atrial Fibrillation (AF). By applying the hybrid model to electrocardiogram (ECG) data, the system successfully distinguished AF from other arrhythmias while providing explainable justifications for its classification.

Ongoing work is expanding this model to stress monitoring. This application requires multi-modal data—combining ECG with electroencephalogram (EEG) signals—and utilizes Dynamic Bayesian Networks to track the evolution of stress states over time. This shift represents the framework's move from static detection to dynamic, sequential decision-making.

Discussion and Ethical Considerations

Preliminary results indicate that the hybrid approach significantly enhances the explainability of AI-driven diagnostics. However, the move toward autonomous health monitoring introduces ethical risks, most notably the "cascade of care." False positives generated by AI can lead to a cycle of unnecessary tests, increased healthcare costs, and patient anxiety. The framework aims to mitigate this by incorporating human-in-the-loop participation and rigorous clinical grounding.

Conclusion

The proposed hybrid AI framework represents a step toward a more transparent and clinically integrated form of personal health monitoring. By combining the pattern recognition of machine learning with the logical rigor of symbolic AI, we can create systems that do not just monitor data, but actively support precision health through explainable, predictive, and context-aware interventions.

Summary updated

BibTeX
@inproceedings{nzomo2024hybrid,
  author = {Mbithe Nzomo},
  title = {{A Hybrid AI Framework for Sensor-Based Personal Health Monitoring towards Precision Health}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30403},
}
Figure