Abstract
Wearable health monitors are essential for continuous physiological tracking, yet they frequently suffer from missing sensor data due to battery depletion, hardware malfunction, or loss of skin contact. Standard imputation methods are either too computationally expensive for low-power devices or fail to maintain diagnostic accuracy. This paper presents Classifier-Aware iMputation (CAM), a novel framework that shifts the computational burden of data reconstruction to an offline training phase. By utilizing unsupervised clustering and a pre-calculated lookup table of "optimal average" patterns, CAM preserves classification performance with minimal energy overhead. Evaluated across seven diverse health datasets, CAM outperforms generative AI baselines in both accuracy and efficiency, making it a viable solution for resource-constrained medical wearables.
What Problem Does This Paper Address?
Modern wearable health devices—ranging from smartwatches to specialized medical patches—rely on multi-modal sensor arrays to detect seizures, monitor stress, or track physical activity. However, these sensors are inherently unreliable. Factors such as battery failure or sensor displacement often lead to missing data streams.
When data from one or more sensors is lost, standard machine learning models typically experience a drastic decline in accuracy. Simple strategies like zero-filling fail to maintain the integrity of the diagnostic output as the number of missing sensors increases.
Existing approaches to mitigate this are often impractical for wearable deployment:
- Multi-classifier systems: These require storing a unique model for every possible combination of missing sensors, leading to exponential memory growth ($2^M$ for $M$ sensors).
- Generative Models (e.g., GANs): While they can reconstruct missing signals, they are too power-hungry for the limited battery life of tiny wearable devices.
Methods: Classifier-Aware iMputation (CAM)
The authors propose CAM, which optimizes for classification accuracy rather than perfect signal reconstruction. The process involves several key stages:
- Unsupervised Clustering: During the offline training phase, sensor data is grouped into clusters representing distinct physiological states (e.g., "walking" vs. "resting").
- Classifier-Aware Search: Instead of reconstructing the exact missing signal, CAM identifies an "average" imputation pattern for each cluster that ensures the final classifier still makes the correct prediction by minimizing the difference in "logits" (raw scores).
- Lookup Table Generation: These optimal patterns are stored in a compact lookup table on the device.
- Runtime Imputation: When a sensor fails at runtime, the device identifies the current data cluster and retrieves the corresponding "best-fit" replacement from the table.
Key Results and Insights
The CAM framework was validated across seven diverse health-related datasets, including physical activity (PAMAP2), seizure detection (Epilepsy), and stress detection (WESAD). Key findings include:
- Performance Stability: CAM maintained accuracy within 5-10% of the "gold standard" (full data availability), even with multiple sensor failures.
- Superior Efficiency: CAM achieved 15% higher accuracy than the GAIN generative model while using significantly less power. In hardware tests on an Odroid-XU3 board, CAM consumed less than 10 mJ per imputation.
- Memory Savings: Most datasets required less than 1 MB of memory, whereas generative approaches often required 10x to 100x more space.
- Preservation over Reconstruction: The study demonstrates that for diagnostic tasks, a static "average" pattern for a specific state is often sufficient to keep a classifier on track, eliminating the need for complex generative models.
Conclusion
CAM provides a robust, energy-efficient solution for handling missing data in wearable health applications. By focusing on preserving the classifier's decision-making process rather than perfect signal reconstruction, it enables high-fidelity monitoring on low-power hardware. Future work will focus on real-world clinical deployments to validate robustness against "in-the-wild" sensor failures.
Summary updated