Intelligent Histories: Detecting Personalized Risk with Longitudinal Surveillance

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Vast amounts of longitudinal data accumulating in electronic health information systems present an untapped opportunity to improve medical screening and diagnosis. Yet doctors typically do not have the time to thoroughly review historical records during a brief clinical encounter, and even when they do, they may find it difficult to rapidly identify long-term patterns across multiple types of data. As a result, the full potential of the electronic health record is not utilized, and conditions that are not easy to diagnose from a single clinical encounter are often missed. For example, abuse and depression may go unrecognized for years as they are masked by other acute conditions that form the basis of clinical encounters, when in retrospect, a review of the longitudinal record may show a discernable pattern. The NLM's Strategic Vision calls for a systems approach to health care that uses next generation electronic health records to facilitate patient-centric care, automated decision support, longitudinal records for patient monitoring, and generation of alerts and reminders. The goal of this project is to answer this call by realizing the full potential of longitudinal medical information to improve medical decision-making. This will be accomplished by developing Intelligent Histories - Dynamic Bayesian Network models of an individual's longitudinal medical information. Building on methods developed for population health surveillance systems, Intelligent History models will be incorporated into a personalized risk surveillance system that will proactively monitor patients' longitudinal histories for long-term risk-associated patterns. The system will present the information in a targeted, contextualized fashion to clinicians, enabling rapid identification of long-term patterns of risk. The work will be carried out in four stages: (1) Developing Intelligent Histories, Bayesian Network risk models that incorporate an individual's multi-year longitudinal coded medical information, including diagnoses, procedures, prescriptions, and laboratory results. The performance of these models will be evaluated and compared with other existing approaches; (2) Extending these models to include explicit representation of temporal trends and relationships including the development of Markov-model based Dynamic Bayesian Network models; (3) Integrating these models into a prototype personalized risk surveillance system that generates alerts and presents the clinician with a tailored view of a patient's longitudinal history. (4) Conducting a formative evaluation to determine whether the prototype system can improve clinicians' abilities to detect and estimate clinical risk. We seek to improve medical decision-making, allowing for earlier detection of clinical conditions, and facilitating a more personalized and systematic approach to medicine.