Ensemble Learning Approach via Kalman Filtering for a Passive Wearable Respiratory Monitor

Utilizing passive radio frequency identification (RFID) tags embedded in knitted smart-garment devices, we wirelessly detect the respiratory state of a subject using an ensemble-based learning approach over an augmented Kalman-filtered time series of RF properties.

We propose a novel approach for noise modeling using a "reference tag," a second RFID tag worn on the body in a location not subject to perturbations due to respiratory motions that are detected via the primary RFID tag. The reference tag enables modeling of noise artifacts yielding significant improvement in detection accuracy. The noise is modeled using autoregressive moving average (ARMA) processes and filtered using state-augmented Kalman filters. The filtered measurements are passed through multiple classification algorithms (naive Bayes, logistic regression, decision trees) and a new similarity classifier that generates binary decisions based on current measurements and past decisions.

Our findings demonstrate that state-augmented Kalman filters for noise modeling improves classification accuracy drastically by over 7.7% over the standard filter performance. Furthermore, the fusion framework used to combine local classifier decisions was able to predict the presence or absence of respiratory activity with over 86% accuracy.

The work presented here strongly indicates the usefulness of processing passive RFID tag measurements for remote respiration activity monitoring. The proposed fusion framework is a robust and versatile scheme that once deployed can achieve high detection accuracy with minimal human intervention.

The proposed system can be useful in remote noninvasive breathing state monitoring and sleep apnea detection.