Sustainable Wearables for Health Applications and Beyond via Uncertainty-Aware Energy Management

Dina Hussein, Chibuike E. Ugwu, Ganapati Bhat, Janardhan Rao DoppaIJCAI 2025 · DOI

TL;DR

This paper tackles the challenge of powering wearable health devices sustainably by developing an "uncertainty-aware" energy management system. It introduces a novel method for predicting future energy harvesting with guaranteed accuracy and uses this to make smarter energy allocation decisions, significantly improving device reliability and performance.

What Problem Does This Paper Address?

The United Nations' Sustainable Development Goals (SDGs) emphasize good health and well-being, which increasingly rely on wearable Internet of Things (IoT) devices for continuous health monitoring. However, a major bottleneck for these devices is their limited battery capacity, leading to frequent recharging and hindering widespread adoption. While energy harvesting (EH) from ambient sources like light and motion offers a solution, these sources are highly unpredictable. Existing energy management (EM) strategies often struggle with this uncertainty, leading to suboptimal decisions and potential device failure. This paper aims to bridge this gap by providing a robust EM system that can reliably operate wearable devices by intelligently managing energy in the face of unpredictable EH.

Methods

The core of the proposed solution lies in two key innovations:

  • Uncertainty-Aware Energy Harvesting Prediction: The authors introduce an "upper-bound calibrated multi-target conformal prediction" (UC-MTCP) method. This technique goes beyond simple point predictions of future energy harvest. Instead, it generates prediction intervals with provable coverage guarantees, meaning the actual energy harvest is highly likely to fall within the predicted range. This is crucial for understanding the bounds of available energy.
  • Overhead-Aware Energy Management: The predicted uncertainty regions from UC-MTCP are then fed into an energy management (EM) algorithm. This EM algorithm uses an "overhead-aware sampling" strategy. It intelligently samples potential energy allocation decisions within the uncertainty bounds, balancing the need to explore various scenarios with the limited computational resources of wearable devices. To make the final decision, a lightweight machine learning model is trained to mimic an "Oracle" (an ideal decision-maker with perfect knowledge of future energy).

Datasets

The authors evaluated their approach using two real-world activity datasets:

  • ARAS dataset: This dataset includes activity data from four users in multiple homes over 30 days.
  • Mannheim dataset: This dataset includes activity data from six users over two weeks.

To create a comprehensive dataset for evaluation, both datasets were extended to cover a six-year period (2015-2020) by shuffling and augmenting the original data. Typical EH levels from motion and ambient light irradiance were combined with the activity data to generate ground truth EH values. The data was split into training (2015), validation (2016), and testing (2017-2020) sets.

Key Insights and Ideas

  • Uncertainty is not just noise, it's information: The paper strongly advocates for explicitly modeling and leveraging uncertainty in energy harvesting predictions, rather than trying to eliminate it. This uncertainty provides valuable bounds for decision-making.
  • Conformal prediction is a powerful tool for wearables: The application of conformal prediction for multi-target regression in the context of EH forecasting is novel and demonstrates its effectiveness in providing guaranteed coverage for resource-constrained devices.
  • Smart sampling is key for limited resources: The overhead-aware sampling strategy is critical for making the EM algorithm practical on wearable devices, ensuring that exploration of energy scenarios doesn't consume too much power itself.
  • ML can bridge the gap between sampling and optimal decisions: Training a lightweight ML model to learn from sampled decisions allows the system to approximate optimal energy allocation without needing perfect future knowledge.

Key Results

  • Guaranteed Coverage: The UC-MTCP approach achieved over 90% coverage with tight prediction intervals, consistently meeting or exceeding the target coverage levels.
  • Near-Optimal Decisions: The proposed EM algorithm produced decisions that were, on average, within 2 Joules of an optimal Oracle.
  • Improved Utility: Compared to baseline methods, the proposed approach achieved at least a 25% improvement in application utility.
  • Low Overhead: The algorithmic overheads (UC-MTCP, IGP, ML model) were found to be minor and feasible for deployment on resource-constrained IoT devices.

Limitations and Future Work

  • Deployment in Real-World Health Settings: The authors plan to deploy wearable IoT devices in real-world health settings, such as for movement disorder monitoring, to validate the approach with actual user data.
  • Scalability to More Complex Scenarios: While the current work focuses on specific EH sources (light and motion), future work could explore more complex EH scenarios and a wider range of wearable applications.
  • User Behavior Modeling: The current model implicitly accounts for user behavior through activity data. More explicit modeling of user behavior could further enhance prediction accuracy and energy management.

Summary updated

BibTeX
@inproceedings{hussein2025sustainable,
  author = {Dina Hussein and Chibuike E. Ugwu and Ganapati Bhat and Janardhan Rao Doppa},
  title = {{Sustainable Wearables for Health Applications and Beyond via Uncertainty-Aware Energy Management}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1079},
}
Figure