Secondary use of EMRs for surgical complication surveillance

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Recent statistics indicate that worldwide almost 234 million major surgical procedures are performed each year with the rates of major postsurgical complications (PSCs) range from 3% to 16% and rates of permanent disability or death range from 0.4% to 0.8%. Early detection of PSCs is crucial since early intervention could be lifesaving. Meanwhile, with the rapid adoption of electronic medical records (EMRs) and the accelerated advance of health information technology (HIT), detection of PSCs by applying advanced analytics on EMRs makes it possible for near real-time PSC surveillance. We have developed a rule-based PSC surveillance system to detect most frequent colorectal PSCs near real-time from EMRs where a pattern-based natural language processing (NLP) engine is used to extract PSC related information from text and a set of expert rules is used to detect PSCs. Two challenges are identified. First, it is very challenging to integrate a diverse set of relevant data using expert rules. In the past, probabilistic approaches such as Bayesian Network which can integrate a diverse set of relevant data have become popular in clinical decision support and disease outbreak surveillance. Can we implement probabilistic approaches for PSC surveillance? Secondly, a large portion of the clinical information is embedded in text and it has been quite expensive to manually obtain the patterns used in the NLP system since it requires team effort of subject matter experts and NLP specialists. In the research field, statistical NLP has been quite popular. However, decision making in clinical practice demands tractable evidences while models for statistical NLP are not human interpretable. Can we incorporate statistical NLP to accelerate the NLP knowledge engineering process? We hypothesize that a probabilistic approach for PSC surveillance can be developed for improved case detection which can integrate multiple evidences from structured as well as unstructured EMR data. We also hypothesize that empirical NLP can accelerate the knowledge engineering process needed for building pattern- based NLP systems used in practice. Specific aims include: i) developing and evaluating an innovative Bayesian PSC surveillance system that incorporates evidences from both structured and unstructured EMR data; and ii) incorporating and evaluating statistical NLP in accelerating the NLP knowledge engineering process of pattern-based NLP for PSC surveillance. Given the significance of HIT, our study results will advance the science in developing practical NLP systems that can be translated to meet NLP needs in health care practice. Additionally, given the significance of PSCs, our study results will address significant patient safety and quality issues in surgical practice. Utilizing automated methods to detect postsurgical complications will enable early detection of complications compared to other methods and therefore have great potential of improving patient safety and health care quality while reducing cost. The results could lead to large scale PSC surveillance and quality improvement towards safer and better health care.