Discovery of early-alert indicators using hybrid ensemble learning and generative physics-based models

Abstract

Early detection of developing abnormalities or treatment effects could critically enhance success of prevention and treatment strategies. While many advanced technologies are available for accurate clinical diagnostics, their wide 24/7 usage required for early preventive alerts including detection of emerging intermittent patterns is not feasible. Although modern wearable devices offer affordable continuous recording of physiological data, data collected over long-term necessarily have significantly lower resolution due to technological limitations leading to sharp accuracy deterioration of mainstream diagnostic techniques. Recently, we demonstrated that some of these challenges can be resolved by hybrid framework where boosting algorithms are used for enhancement of existing domain-expert models with further non-linear combination of boosted ensemble components via deep learning or other machine learning algorithms. While normal-abnormal differentiation performance of such hybrid indicators was confirmed using real cardio data from www.physionet.org, their applicability to more challenging problem of early-stage detection of emerging abnormalities or treatment effects remain unknown since long-term transition data from normal to abnormal states is not available. Here we propose a framework for verification and enhancement of indicator abilities for such early detection using simulated transition paths obtained by sampling real normal/abnormal data and employing realistic synthetic data generated by physics-based models. Robust performance of our hybrid indicators was confirmed in cases where other existing approaches fail.