Data Review (Clinical and Research)
                              
            (Introduction to Medical Informatics)
         (http://www.cpmc.columbia.edu/edu/textbook)

LAST REVIEWED:  1 October 1997

CLINICAL RESEARCH NEEDS

clinical care (treat patients) would usually be: get me
     result of patient's blood tests
     "intra-patient query"
          pt1: K, ...
     ==>  pt2: K, K, Na, ...
          pt3: Na, ...
          ...

clinical research
     "cross-patient queries"
          K: pt1, pt2, ...
     ==>  Na: pt2, pt3, ...
          ...

uses of cross-patient queries (actually can be used for care
        and research)
     find patient whose name you forgot
     look at patients similar to your own patient (prognosis)
     calculate likelihood of disease given findings
     clinical research screening
     find cases for teaching rounds
     hospital administration (how is length of stay)

for example:
     want to assess the association in male patients between
     hypokalemia (low potassium) and death from myocardial
     infarction (heart attack)
          need to put several simpler queries together

assume we have a database of coded clinical data; how would
     we approach answering the above query?

find patients with
     A) sex=male                              6840
     B) myocardial infarction                  717
     B) (may need a superclass)
     C) hypokalemia (absent)
     C) potassium<3.5 mg/dl                   1976
     D) died after 1/1/85                      579
     E) A, B, C, and D == male, MI, lowK, died  16
     F) A, B, and C == male, MI, lowK           82
     G) A and B == male, MI                    382
     H) G not C == male, MI, never lowK        308
     I) H and D == male, MI, never lowK, died   28

therefore 16/82 (20%) died with lowK vs. 28/308 (9%) died
     without lowK

conclusion: hypokalemia is associated with death from MI
     not causal

QUERYING CODED DATA

query = retrieve data from the database, subject to constraints
     (storage of data has analogous problems)

there is a problem for novice users
     want to make up queries on the fly (ad hoc)
     do not know what things are called (vocabulary)
     do not know where or how things are stored (schema)
     do not know how to phrase a question
          do not know how to write programs
          do not know how to ask the computer for data

MENU-DRIVEN SYSTEMS

problem to solve: user does not know what is available or
        how to say it
     vocabulary is too big
     does not provide procedural (or syntactic) support

set of menus and selections to guide you through the
        available choices
     provide structure; limit choices at any one point
     (eg, CIS, library menu)

for intra-patient queries
     can use menu and then browse full list (CIS)

for cross-patient queries
     too many types of data to browse a large list (paper)

therefore, follow menu with vocabulary searches
     more constrained set of terms
     can use alphabetical, browsing, ...

(context dictates the structure of the interface)

examples from paper

disadvantage:
     limited to what the menus provide
     to repeat a query, need to go back through menus

QUERY LANGUAGE / TOOLS

allow user to type in a question directly

1. Structured Query Language (SQL)

SQL is a prime example for relational databases

SELECT <columns> FROM <tables> WHERE <constraints on rows>

SELECT POTASSIUM FROM ELECTROLYTES
     WHERE PATIENT=123
     AND DATE > "1993-01-01"

good
     relatively declarative (say what rather than how or where)
     common standard among databases (learn it once)
     relatively understandable
     can create ad hoc queries
     can save queries and repeat later on

bad
     
     still need to know what things are called and how
          things are stored (vocabulary, table names, column
          names, necessary joins)
     
     specific to relational databases

2. Medical Query Languages (MQL)

example of a language written especially for medicine
     based upon COSTAR outpatient clinical database

eg,
     WHEN CODE1 IS BLOOD PRESSURE AND
          DATE1 IS AFTER TODAY - 1 YEAR
     WHEN CODE2 IS BLOOD PRESSURE AND
          DATE2 FOLLOWS DATE1
     LIST NAME, UNIT NUMBER
     
     (what is missing from this query)
     
     infers reasonable relations among attributes (ie, adds
             schema knowledge)
          CODE1 and DATE1, NAME and CODE1
     
     handles time
          IS AFTER, 1 YEAR

good
     less schema knowledge needed (only vocabulary)
     better handling of time

bad
     specific to COSTAR databases
     still requires knowledge of Boolean logic


3. Query By Example (QBE)

create a picture of a relational table
     graphical query language
     fill constraints into the table
     indicate which columns you want back

good
     need not know specific syntax

bad (like SQL)
     still need to know vocabulary, tables, columns
     specific to relational databases

4. Natural Language Question Answering System

allow natural language queries
     not need to know rigid syntax
     produce coded queries (SQL) for the database

possible because domain is limited, and because know
     database and schema

even with a theoretically perfect system, it is still
     difficult to get data
     eg, need for filtering
     
     retrieval for potassium
          first one is too old to be relevant
          second must be an error, since out of range
          third one is an invalid test
          
          time         value
          1981-01-03   2.7
          1993-09-09   50.1
          1993-09-10   hemolyzed

RESEARCH SYSTEMS

clinical research
     advance medical science by adding to medical KB
     
     study a disease and its manifestations
     uncover relationships among diseases, findings, ...
     estimate prognosis
     assess effectiveness of therapy
     
     experiments limited by ethical and practical concerns
     often extend over many years (eg, 30 years)
     
     randomized controlled clinical trial
          compare intervention to control
          prospective
          avoid bias (randomize, blinding)
     
     retrospective studies
          look at data already collected for another purpose
          subject to bias

functions of a research system
     
     efficient storage of data, with DB utilities
     
     make it easy for researcher to query data
     
     maintain data quality on input
          correctness = entry errors
          completeness = missing data
          consistency = same finding -> same code
     
     statistical support
     
     screening for candidates to enter study (must be linked
          to clinical database)

general database packages
     cheap and readily available on all computers
     easy to use
     lack sophisticated statistical processing
     lack medicine-specific DB support (eg, time, studies)
     
     eg, Revelation, dBase, ...

general statistical packages
     cheap and readily available on all computers
     easy to use
     use only simple files to store data
     inflexible for new statistical techniques
     
     descriptive statistics
          eg, mean, std deviation, range
          summarize quantity of data in a number
          overall properties, look for associations
     
     graphical display
          eg, scatter plot
          summarize quantity of data in a graph
          look for associations and biases
     
     analytic statistics
          eg, regression
          test specific hypotheses, complex descriptors
     
     eg, BMDP, SPSS, SAS

medical research systems
     more expensive to build or buy
     
     specific to medical research
          longitudinal studies (data over long time)
          time as a special part of syntax
     
     Medlog - interactive, time-oriented database used to
          store and analyze longitudinal clinical data
     
     ClinQuery - real-time cross-patient access to Beth-
          Israel's clinical database (well used)

current research
     1. large quantity of data (no human can process it)
     2. sharing data among institutions
     
     automated abstraction
          turn raw data into relevant high level concepts
          data -> syndromes -> diseases
     
     automated discovery of medical relationships
          time-lagged associations in DB may imply causality
          generate hypotheses
          test with retrospective experiments on DB
     
     common schema, vocabulary, query language

related reading:
     Safran C, Porter D, Rury CD, Herrimann FR, Lightfoot J,
     Underhill LH, et. al. ClinQuery: searching a large
     clinical database. M.D. Computing 1990;7(3):144-53.