MetaWearS: A Shortcut in Wearable Systems Lifecycle with Only a Few Shots

Alireza Amirshahi, Maedeh H. Toosi, Siamak Mohammadi, Stefano Albini, Pasquale Davide Schiavone, Giovanni Ansaloni, Amir Aminifar, David AtienzaarXiv 2024 · DOI · arXiv

Abstract

MetaWearS is a meta-learning framework designed to overcome the "data hunger" and high energy costs of deploying deep learning on wearable health devices. By utilizing prototypical networks, it achieves high diagnostic accuracy with minimal labeled data and reduces the energy required for model updates by over 400x compared to traditional retraining methods. This approach bridges the gap between clinical-grade diagnostics and energy-constrained consumer wearables.

Introduction: The Wearable Lifecycle Bottleneck

The deployment of deep learning on wearable health monitoring systems—such as smartwatches for cardiac arrhythmia detection—is currently hindered by three critical bottlenecks. First, data scarcity: training robust models requires massive amounts of labeled biomedical data, which is expensive to collect for specific wearable form factors. Second, the update dilemma: adapting a model to a new user typically requires extensive retraining to avoid overfitting. Third, energy constraints: transmitting a fully retrained model over Bluetooth Low Energy (BLE) is a significant battery drain.

Comparison of traditional wearable model lifecycles (a) versus the MetaWearS few-shot approach (b).

MetaWearS addresses these issues by providing a unified framework for data-efficient training and energy-efficient updating, effectively creating a "shortcut" in the system lifecycle.

Methodology: Prototypical Networks for Wearables

The core of MetaWearS is based on Prototypical Networks, a form of few-shot meta-learning. Instead of a standard classifier, the model learns an embedding space where biomedical signals (EEG or ECG) are projected.

The Prototypical Mechanism

In this embedding space, each diagnostic class is represented by a "prototype"—the mean vector of the embedded support samples. Classification occurs by calculating the squared Euclidean distance between a new query sample and these prototypes.

The meta-learning framework showing episodes of meta-training and meta-testing with support and query sets.

Prototypical Updating

The most significant innovation is the update phase. Rather than retraining millions of neural network weights, MetaWearS simply updates the prototype vector. Because a prototype is a single vector, it is orders of magnitude smaller than a full model, drastically reducing communication overhead. The system was validated on the X-HEEP RISC-V edge platform to measure real-world energy consumption.

Experimental Results

The framework was evaluated on two critical medical use cases: Epileptic Seizure Detection (EEG) and Atrial Fibrillation (AF) Detection (ECG), using a mix of large-scale hospital datasets (TUSZ, PhysioNet) and wearable-specific datasets.

  • Accuracy with Minimal Data: MetaWearS achieved 70% AUC for seizure detection and 82% AUC for AF detection with only a few samples. This represents a performance improvement of up to 45% AUC over standard transfer learning in low-data scenarios.
  • Incremental Personalization: Adding just 16 minutes of new labeled data during the update phase boosted AUC by up to 5.3%.
  • Energy Efficiency: The prototypical update mechanism reduced energy consumption by 456x for seizure detection and 418x for AF detection compared to traditional full-model updates over BLE.

Conclusion

MetaWearS demonstrates that meta-learning can successfully bridge the gap between clinical equipment and consumer electronics. By focusing on moving "targets" (prototypes) in an embedding space rather than retraining feature extractors, the framework bypasses the primary battery and data barriers of modern wearables. Future work will focus on testing against noisier "in-the-wild" data and expanding to other signals like PPG and EMG.

Summary updated

Abstract

Wearable systems provide continuous health monitoring and can lead to early detection of potential health issues. However, the lifecycle of wearable systems faces several challenges. First, effective model training for new wearable devices requires substantial labeled data from various subjects collected directly by the wearable. Second, subsequent model updates require further extensive labeled data for retraining. Finally, frequent model updating on the wearable device can decrease the battery life in long-term data monitoring. Addressing these challenges, in this paper, we propose MetaWearS, a meta-learning method to reduce the amount of initial data collection required. Moreover, our approach incorporates a prototypical updating mechanism, simplifying the update process by modifying the class prototype rather than retraining the entire model. We explore the performance of MetaWearS in two case studies, namely, the detection of epileptic seizures and the detection of atrial fibrillation. We show that by fine-tuning with just a few samples, we achieve 70% and 82% AUC for the detection of epileptic seizures and the detection of atrial fibrillation, respectively. Compared to a conventional approach, our proposed method performs better with up to 45% AUC. Furthermore, updating the model with only 16 minutes of additional labeled data increases the AUC by up to 5.3%. Finally, MetaWearS reduces the energy consumption for model updates by 456x and 418x for epileptic seizure and AF detection, respectively.
BibTeX
@misc{amirshahi2024metawears,
  author = {Alireza Amirshahi and Maedeh H. Toosi and Siamak Mohammadi and Stefano Albini and Pasquale Davide Schiavone and Giovanni Ansaloni and Amir Aminifar and David Atienza},
  title = {{MetaWearS: A Shortcut in Wearable Systems Lifecycle with Only a Few Shots}},
  year = {2024},
  eprint = {2408.01988},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.01988},
}
Figure