| Summary: | With the increase in volume of wearable sensors, there exists the possibility of personalising patient care, employing automated algorithms. However, automated algorithms are typically less reliable than gold-standard expert labels; the latter are scarce and expensive. In real-life applications, expert labels are not available, and algorithms for processing sensor data must be relied upon, without access to the “ground truth”. It is therefore difficult to choose which algorithms to trust or discard at any point in time, where different algorithms may be optimal for different patients. We propose two fully-Bayesian generative models for fusing labels from (i) independent and (ii) potentially-correlated algorithms. They aggregate outputs of the algorithms in an unsupervised manner, to estimate jointly the assumed bias and precision of each algorithm without access to the ground truth. The latter fused estimate may then be used to infer the underlying ground truth. For the first time in the biomedical context, we show that modelling correlations between annotators, and fusing information concerning task difficulty (i.e., quality of the data), improves these estimates. Also, we adopt a strongly-Bayesian approach to inference using Gibbs sampling to improve estimates over the existing state-of-the-art. Our proposed models were applied to simulated and two publicly-available biomedical datasets, and showed that they outperform all existing approaches recreated from the literature. Our models are robust when dealing with missing values, and are suitably efficient for use in real-time biomedical applications, thereby providing the basis for the reliable use of sensors for personalising the care of the individual.
|
|---|