On robust risk-based active-learning algorithms for enhanced decision support