We are developing prototypes that explicate our vision of connecting personal medical data to scienti c literature as well as to emerging grey literature (e.g., community forums) to help people nd and understand information relevant to complex medical journeys. We focus on robust combinations of natural language processing along with linked data and knowledge representation to build knowledge graphs that help people make sense of current conditions and enable new manners of scienti c hypothesis generation. We present our work in the context of a breast cancer use case. We discuss the bene ts of biomedical linked data resources and describe some potential assistive technology for navigating rich, diverse medical content.
As scienti c knowledge continues to grow in size and diversity, it is increasingly di cult to discover and manage information relevant to any particular context. It can be challenging to determine how a statement or report relates to others and to form and evaluate (often competing) hypotheses, e.g. related to diagnosis or treatment paths. Complications grow when content is both structured and unstructured, and when some is from less accredited sources. We aim to expand the boundaries of Linked Science by focusing on evidence modeling from natural language processing techniques (NLP) over broad content and by identifying promising data-driven hypotheses using linked data and nanopublication style encodings. We present this discussion in the context of a breast cancer demonstration use case informed by challenges experienced during a co-author's recent cancer journey. Cancer is a complex disease to manage and treat, often requiring chemotherapy, surgery, radiation, and drugs to reduce recurrence. We show how management of this information by the patient is aided by semantic technologies combined with natural language processing algorithms.
A breast cancer patient wishes to better understand her diagnosis and planned treatment. She is interested in expected chemotherapy side effects, and leveraging experiences of other similar individuals to proactively nd and evaluate promising coping strategies. She reads through oncologist-provided documents about her proposed chemotherapy drugs and uses search engines to nd more about likely adverse e ects that appear detrimental to her quality of life. She nds con icting opinions on the e cacy of di erent coping strategies, and needs to determine an approach to e ectively weigh the possible pros and cons. Managing this information is mentally taxing and can easily overwhelm a patient.
Our patient needs to nd and comprehend potentially con icting evidence about treatment options and side e ects. We propose new software, using a variety of arti cial intelligence tools built on the interoperability principles promulgated by linked data and the Semantic Web, to address these challenges. 2
The patient uses current technologies to obtain information about her treatment strategy and to formulate promising side e ect mitigations. This can be time consuming for anyone, but more so for medically nave patients. Furthermore, technologies such as web forums or social networking sites are becoming increasingly common for discourse between patients as they can often include anecdotal reports, that have not yet been validated through clinical trials, but may be valuable. They are often presented in layperson terms and sometimes attract new patients who may be less medically literate. Due to lack of scienti c rigor, there may be contradictory or unsupported information available, as shown in the following two answers about a mitigation for the very common, taxol-related, nail bed problem:
My onc[ology] nurse told me to rub tea tree oil into my cuticles and
nails every night. It is a natural anti-septic and for whatever reason can
sometimes help prevent nail infections and lifting during taxol. 1
I wouldn't use tea tree oil. A friend did on some cracked skin and it got
worse. 2
The rst suggestion is a common preventive approach for nail problems: tea tree
oil prevents nail infections because \it is a natural anti-septic" and appeals to
authority \my onc nurse told me to...". The second suggestion from a di erent
user in the same thread advises against tea tree oil as \a friend [applied tea
tree oil] on some cracked skin and it got worse." Natural Language techniques
may be used to extract coping strategies for particular conditions but without
deeper knowledge, provenance, and tools, the user may not know how to evaluate
and/or integrate potentially contradictory suggestions. We are extending joint
extraction techiques proposed in [
The Repurposing Drugs using Semantics (ReDrugS) project [
We aim to provide extensive explanation mechanisms since explanation is a key
component of transparent systems and user studies have shown that
explanations are required if agents are to be trusted [
We identify potential evidence on the use of tea tree oil in
chemotherapyinduced nail bed problems. Not only would a patient want to know evidence,
source, and authoritativeness for both views, she might also want the system
further decompose these arguments and present supporting evidence as to the
antimicrobial nature of tea tree oil in more authoritative sources (e.g. [
We claim that we can reuse the ReDrugS content to nd prescription drugs for chemotherapy side e ects. Provenance may be displayed to show that the recommendation is from a validated authoritative source. While that framework was originally designed to nd potential new o -label uses for drugs along with con dence ratings, the explanation component is more critical for our use so that researchers may inspect evidence sources and the methods used to determine the system con dence. Without such explanations, people would have di culty evaluating competing suggestions.
Our systems3 provide explanation drill down so users can obtain as much detail as they desire, thus allowing a patient to nd, for example, if authoritative sources contain prescription drugs for coping with a particular side e ect. Our NL-based extraction work can be used to identify alternative, possibly competing, therapies, e.g. an herbal remedy recommended anecdotally with potentially corroborating authoritative sources. 5
Natural Language Processing can expose some of the unstructured content of medical records as structured content as well as assist in generating linked data from unstructured sources. The ReDrugS framework provides a semanticallyintegrated system combining many di erent structured biomedical resources to generate a broadly reusable knowledge graph. By integrating the natural language and structured knowledge representation approaches, we can obtain a much richer annotated knowledge base that includes source and con dence information. Our prototypes demonstrate some ways that this rich resource may then be used to help patients and their support networks to discover, integrate, and evaluate information relevant to complicated medical situations and to help form transparent and data-driven hypotheses about how to proceed. We believe these e orts demonstrate some opportunities for future AI-enhanced Linked Sciencebased assistants that use the wealth of structured content as well as the growing grey literature collection.
The authors thank Heng Ji and Alex Borgida for their discussions that helped shape this work.
http://tw.rpi.edu/web/project/ReDrugS