Automating Systematic Reviews of Environmental Health Literature with Machine Learning

点击翻译按钮获取中文摘要

Project Summary IDOC Software proposes the development of Artificial Intelligence (AI) algorithms, together with user-friendly software, for facilitating the efficient production of Systematic Reviews (SRs) in the field of Environmental Health (EH). SR is the “Gold Standard” for assessing evidence to be used for decision making in a variety of health contexts, including health care, public health and environmental health. SRs synthesize evidence from studies that meet eligibility criteria based on the decision being made (such as hazard identification or risk assessment). All relevant studies need to be considered in an SR, meaning that all of the potentially related articles must be evaluated one by one. For example, if the SR question relates only to 40-65 year old women, then studies containing men or containing women outside this age range must be excluded from the final set of articles used to draw a conclusion. The time (and expense) involved in screening potentially thousands of citations is substantial, often taking a team of screeners months to complete. This severely limits the numbers of SRs that can be conducted and threatens timely decisions by policy makers. AI has tremendous potential to accelerate the conduct of SRs by automatically recognizing words that relate to eligibility criteria, however there are significant challenges. In the field of EH the same study populations, exposures, and health outcomes can be described with many different combinations of words and phrases. It is difficult for AI algorithms to generalize language in the way needed to overcome the complexity inherent in these scientific communications. IDOC Software has developed algorithms capable of deducing connections between words and phrases. These learned connections are formed around a EH framework, or ontology, known as PECO: Population, Exposure, Comparator, and Outcome. The software maps key words and phrases in an article onto these categories and then highlights these terms in the article text via color-coding. A screener then need not read an entire article to determine if it meets the eligibility criteria. Instead, the screener scans the “P” colored words to determine if the population studied meets the “P” inclusion criteria. Then the “E” colored words can be evaluated, and so on. This accelerates the rate at which a screener can evaluate articles manyfold. The challenge for the AI algorithms is to then find all the PECO words and phrases and accurately categorize them. High accuracy requires taking into account causal and other relationships between the words and phrases. Advances in machine learning and natural language processing achieved in Phase I on article titles and abstracts, and then on the full text of articles in Phase II, will result in more efficient conduct of SRs, reducing costs and time, and thereby furthering the goal of making timely evidence-informed decisions and policy to protect public health from unsafe environmental exposures.