Energy-Efficient Missing Data Imputation in Wearable Health Applications: A Classifier-aware Statistical Approach

Dina Hussein, Taha Belkhouja, Ganapati Bhat, Janardhan Rao Doppa

Introduction

Wearable health devices have transitioned from simple step counters to sophisticated clinical tools capable of monitoring chronic conditions like Parkinson’s disease and epilepsy. These devices rely on multi-modal sensor arrays—including accelerometers, electromyography (EMG), and heart rate monitors—to feed machine learning (ML) classifiers that provide real-time health insights. However, the reliability of these systems is frequently compromised by missing data. Sensor malfunctions, battery depletion, or improper device placement can lead to data gaps that significantly degrade the accuracy of health-critical predictions.

The Challenge of Missing Data in Wearables

When sensor data is lost, the performance of standard ML models typically plummets. Current strategies to mitigate this issue generally fall into two categories, both of which are ill-suited for the resource-constrained environment of a wearable device:

  1. Multi-classifier Ensembles: This involves training a unique model for every possible combination of missing sensors. For a device with $N$ sensors, this requires $2^N - 1$ models, creating a prohibitive memory footprint.
  2. Generative Models: Techniques like Generative Adversarial Imputation Nets (GAIN) use complex neural networks to "hallucinate" missing values. While effective, these models are computationally intensive, leading to high energy consumption that rapidly drains wearable batteries.
Impact of missing sensor data on classifier accuracy using a simple zero-filling approach.

As shown in Figure 1, even simple "zero-filling" strategies result in a sharp decline in accuracy as the number of missing sensors increases, highlighting the need for more robust yet efficient recovery methods.

CAM: Classifier-Aware iMputation

To bridge the gap between accuracy and efficiency, researchers have introduced Classifier-Aware iMputation (CAM). The fundamental shift in CAM is its objective: instead of attempting to perfectly reconstruct the missing physical signal, it focuses on providing an imputation that preserves the final decision of the classifier.

Overview of the CAM framework, from offline cluster-based imputation search to real-time classification.

The CAM framework operates in three distinct stages:

  1. Unsupervised Clustering: During the training phase, the system applies k-means clustering to historical sensor data to identify representative patterns or "states."
  2. Offline Imputation Search: For each cluster and each possible combination of missing sensors, CAM performs a one-time optimization. It uses gradient descent to find an "average" imputation pattern that minimizes the difference between the classifier's internal prediction scores (logits) for the full data and the imputed data.
  3. Runtime Look-up Table: At runtime, the device avoids complex calculations. When a sensor failure occurs, the system identifies the current data cluster, retrieves the pre-optimized "average" pattern from a small look-up table, and fills the gap.

To further enhance robustness, the classifier is fine-tuned using a mix of original and CAM-imputed data, ensuring it remains accurate even when operating on substituted values.

Experimental Results

The authors validated CAM across seven diverse health datasets, ranging from physical activity recognition (PAMAP2) to seizure detection (Epilepsy) and stress monitoring (WESAD).

Accuracy and Robustness

CAM demonstrated remarkable resilience, maintaining accuracy within 5% of the "perfect data" baseline with one missing sensor and within 10% with two. Notably, CAM outperformed state-of-the-art generative models like GAIN by over 15% in accuracy on several tasks, particularly as the severity of data loss increased.

Resource Efficiency

The primary advantage of CAM is its minimal overhead:

  • Energy Consumption: CAM requires less than 10 mJ per imputation. In assistive device tasks, it achieved nearly 100% energy savings compared to generative neural network approaches.
  • Memory Footprint: The look-up table approach is extremely lean. For example, on the Epilepsy dataset, CAM required only 0.02 MB of memory, compared to the 1.84 MB required by GAIN—a nearly 100x reduction.

Conclusion

CAM represents a "third way" for missing data recovery in IoT health applications. By prioritizing "classifier-awareness" over signal reconstruction, it achieves clinical-grade accuracy with the efficiency required for long-term wearable use. While extreme data loss remains a challenge, CAM’s ability to turn a complex generative problem into a simple table lookup paves the way for more reliable and long-lasting wearable health monitors. Future work involves real-world clinical deployments to validate these benefits in patient-facing applications.

Reference: Hussein, D., Belkhouja, T., Bhat, G., & Doppa, J. R. (2024). Energy-Efficient Missing Data Imputation in Wearable Health Applications: A Classifier-aware Statistical Approach. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24).

Summary updated

BibTeX
@inproceedings{hussein2024energy,
  author = {Dina Hussein and Taha Belkhouja and Ganapati Bhat and Janardhan Rao Doppa},
  title = {{Energy-Efficient Missing Data Imputation in Wearable Health Applications: A Classifier-aware Statistical Approach}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure