Calibration and Classifier Design for IoT Healthcare Applications using Federated Learning

Recent trends in health care informatics rely on federated learning to preserve privacy and improve communication efficiency. Applications, for example, remote patient monitoring and fitness tracking, use IoT devices for communication and on-device computations. Often these devices use low-cost sensors for signal acquisition, leading to calibration errors and loss in classification accuracy. In this work, we propose an algorithm to jointly calibrate and design classifiers for a finite number of IoT devices using the federated learning architecture. We correct the gain and offset deviations using a linear calibration model, thus preparing the data for device-aware classifier design. By employing on-device training, we learn individual classifiers and communicate the compressed classifier parameters to learn a global classifier model. Further, the global classifier is recalibrated to minimize the joint mean square error of the designed classifiers. We validate the calibration and classifier design algorithms on synthetic data and an available fitness tracker dataset with several users.