Leveraging Neural-Networks, Boosting and Domain-Knowledge to Discover Physiological Indicators with Minimal Sensitivity to Data Resolution

Success of prevention strategies critically depends on capabilities of early detection of emerging abnormalities. Clinical diagnostic technologies cannot be used for continuous 24/7 monitoring to ensure very early preventive alerts. Although long-term continuous recording of physiological data with modern wearable devices is affordable, technological limitations require lower sampling rates leading to significantly lower resolution and deterioration of diagnostics accuracy. Recently, we demonstrated that these problems can be significantly alleviated using hybrid framework where multi-expert ensemble learning enhances individual domain-expert models with further combination of ensemble components via flexible boosting algorithms. Here we extend our framework to handle a wide range of data resolutions without model re-training, which allows to cover wearable devices with different technical specifications and adaptively varied sampling rates. We propose to use neural networks (NNs) for combination of ensemble components for practical fine-tuning on limited data, accuracy enhancement and pre-training on synthetic data covering rare states. Real cardio data from www.physionet.org and synthetic data from physics-based models are used for illustrations.