SYSTEM RESPONSE TO THE QUESTION “WHAT IS THE BEST DRUG TREATMENT FOR CHRONIC PROSTATITIS

Table 1: System response to the question “What is the best drug treatment for chronic prostatitis?”

to make decisions about patient care. As shown

systems for decision support represent a poten-

by previous work (Cogdill and Moore, 1997; De

tially high-impact application. From a research

perspective, the clinical domain is attractive be-

Groote and Dorsch, 2003), citations from the

cause substantial knowledge has already been cod-

MEDLINE database (maintained by the U.S. Na-

ified in the Unified Medical Language System

tional Library of Medicine) serve as a good source

(UMLS) (Lindberg et al., 1993). The 2004 version

of clinical evidence. As a result of these findings,

of the UMLS Metathesaurus contains information

our work focuses on MEDLINE abstracts as the

about over 1 million biomedical concepts and 5

source for answers.

million concept names. This and related resources

allow us to explore knowledge-based techniques

3 Question Answering Approach

with substantially less upfront investment.

Conflicting desiderata shape the characteristics of

Naturally, physicians have a wide spectrum of

“answers” to clinical questions. On the one hand,

information needs, ranging from questions about

conciseness is paramount. Physicians are always

the selection of treatment options to questions

under time pressure when making decisions, and

about legal issues. To make the retrieval problem

information overload is a serious concern. Fur-

more tractable, we focus on a subset of therapy

thermore, we ultimately envision deploying ad-

questions taking the form “What is the best drug

vanced retrieval systems in portable packages such

treatment for X?”, where X can be any number of

as PDAs to serve as tools in bedside interac-

diseases. We have chosen to tackle this class of

tions (Hauser et al., 2004). The small form factor

questions because studies of physicians’ behavior

of such devices limits the amount of text that can

in natural settings have revealed that such ques-

be displayed. However, conciseness exists in ten-

tions occur quite frequently (Ely et al., 1999). By

sion with completeness. For physicians, the im-

leveraging the natural distribution of clinical in-

plications of making potentially life-altering deci-

formation needs, we can make the greatest impact

sions mean that all evidence must be carefully ex-

with the least effort.

amined in context. For example, the efficacy of a

Our research follows the principles of evidence-

drug is always framed in the context of a specific

based medicine (EBM) (Sackett et al., 2000),

sample population, over a set duration, at some

which provides a well-defined model to guide the

fixed dosage, etc. A physician simply cannot rec-

process of clinical question answering. EBM is

ommend a particular course of action without con-

a widely-accepted paradigm for medical practice

sidering all these factors.

that involves the explicit use of current best ev-

idence, i.e., high-quality patient-centered clinical

Our approach seeks to balance conciseness and

research reported in the primary medical literature,

completeness by providing hierarchical and inter-

active “answers” that support multiple levels of

UMLS concepts. UMLS has an extensive cov-

erage of drugs, falling under the semantic type

drill-down. A partial example is shown in Fig-

P

S

and a few oth-

ure 1. Top-level answers to “What is the best drug

treatment for X?” consist of categories of drugs

ers. All such entities are identified as candidates

and each is scored based on a number of features:

that may be of interest to the physician. Each cat-

egory is associated with a cluster of abstracts from

its position in the abstract, its frequency of occur-

rence, etc. A separate evaluation on a blind test

MEDLINE about that particular treatment option.

set demonstrates that our extractor is able to accu-

Drilling down into a cluster, the physician is pre-

rately recognize the interventions in a MEDLINE

sented with extractive summaries of abstracts that

abstract; see details in (Demner-Fushman and Lin,

outline the clinical findings. To obtain more detail,

2005; Demner-Fushman and Lin, 2006 in press).

the physician can pull up the complete abstract

text, and finally the electronic version of the en-