Integrated frameworks for knowledge discovery in human-machine complex systems using multiple data stream

Complex human-machine systems where human plays controlling roles are highly dynamic and complicated making the traditional models and methodologies less effective. The operability of such a complex system is affected by the performance and inter-relationships of a wide range of both internal and exogenous variables. The dynamic nature of such systems makes it necessary to apply probabilistic and stochastic models to capture the system variability. In this study, we propose integrated frameworks for two such systems, transportation and healthcare, by applying advanced data analytics, statistical and stochastics models and machine learning methods to extract important knowledge for either prediction or causal analysis. The results can be used for both off-line design of better targeted countermeasures and corrective actions or on-line monitoring for situational awareness which can in turn assist with well-informed control actions.;For the transportation system, we present a novel approach to formulate the real-time traffic safety risk of individual drivers and present data-driven frameworks to predict the drivers' individualized safety risks. In particular, the models take advantage of near-crashes in addition to crashes and is capable of handling different types of variables. We first used the VTTI's 100-car Naturalistic Driving Study (NDS) data to develop an ensemble classifier to classify driving events into the crash and near-crash. We have then extended our methodology and developed a model for the Second Strategic Highway Research Program (SHRP-2) NDS data which is a more comprehensive study with more safety-related variables. Extensive data preparation and feature engineering were necessary to make data ready for model building. For the traffic safety risk prediction, we have used a weighted regularized regression model, to classify the trichotomous driving outcomes in relation to multi-stream safety data. We have further improved the resolution of the classes of driving outcomes by decomposing the class of normal driving. The developed prediction models can be used in advanced driver assistance systems to warn drivers of critical traffic incidents. We have also proposed a hybrid physics/data-driven approach to be used in a personalized kinematic-based Forward Collision Warning (FCW) system. In particular, we have used a hierarchical regularized regression model to estimate the driver's reaction time in relation to his/her individual characteristics, driving behavior and surrounding driving conditions. This personalized reaction time will be then plugged into the Brill's one-dimensional car-following model. We have also developed a simple rule-based algorithm to decide when to use the predicted values in a conservative FCW system.;For the healthcare system, we also develop a quantitative framework to identify the main sources of variation in patient flow. Since 1983, under Health Care Financing Administration (HCFA)'s system each hospital inpatient is classified into predefined Diagnosis-Related Groups (DRGs), and the hospital is paid the amount that HCFA has assigned to each DRG. In other words, irrespective of what the hospital charges for, it will be paid only a fixed price for each DRG through major reimbursement plans. Therefore, it is logical to expect that by reducing the within DRG discrepancies, hospitals can cut cost and improve patient safety and satisfaction. In order to reach this goal, the first step is to identify the main sources of variations. We have used a mixture of first-order n-step Markov models to cluster patients into similar groups and then applied the well-known random forest classifier to identify significant factors affecting the patient sequence among tens or hundreds of potential factors including patient profile and hospital-related variables. We illustrated the applicability of our proposed approach by using a simulated data based on a real-life case study.