Gaussian Processes for Monitoring Patients with Mobile Sensors

As wearable physiological sensors become more common, there is a need for algorithms that can use the resulting waveforms to perform robust data analysis. Existing techniques have failed to penetrate into clinical practice due to their perceived lack of robustness. This presentation will argue that the natural framework for inference with noisy, incomplete data is that of Bayesian Gaussian processes. We describe the use of multi-task algorithms for monitoring patients via wearable sensors. Such algorithms must be able to cope with differing sampling rates between sensors; different modalities of data acquisition; and phase offsets between sensors.