Quick Medical Reference (QMR) (Introduction to Medical Informatics) (http://www.cpmc.columbia.edu/edu/textbook) functions 1. diagnosis: enter findings (present, absent, don't know) get ranked diagnosis list suggest what findings to assess next can assert or defer specific diagnoses will then look for second diagnosis (putting aside findings explained by first) 2. teaching: given finding, what diseases cause it given disease, what findings are seen given disease, what other diseases are associated critique - comments on student's suggested Dx 3. spreadsheet: given one finding, how should it be worked up compare two diseases for similarities, differences suggest how to differentiate two diseases what diseases are similar to a disease (homology) what if patient had pulmunary disease and diarrhea what would be the common thread? KNOWLEDGE BASE began as INTERNIST-I by J Myers turned into QMR by R Miller now a commercial product 25 person years consists of findings, diseases, and disease profiles vocabulary findings (4600): history, physical exam, laboratory diseases (640): from internal medicine stored in a classification hierarchy completer - uses lexical lookup to find terms disease profiles formal description of a disease in terms of findings an average of 85 findings per disease also maps diseases to diseases (causes, associations) expressed as frequencies, evoking strengths, imports each disease takes 2 weeks full time using 100 papers includes references in the profiles evoking strength (ES) how many patients with this finding have the disease similar to P(D|T) 0 = non-specific finding (could be any disease) 1 = minimally suggests disease (rare or unusual) 2 = mildly suggests disease 3 = half the time suggests disease 4 = most often suggest disease 5 = always suggests disease (pathognomonic) as probability 1 = 0 - 5 2 = 5 - 35/40 3 = 35/40 - 65 4 = 60 - 85/90 5 = 85/90 - 100 eg, 50% of patients with white cells in the urine have urine infection, so ES=3 frequency (F) how often do we see the finding in cases of the disease similar to sensitivity, P(T|D) 1 = finding is rarely seen in cases of the disease 2 = significant minority of cases 3 = about half of cases 4 = majority of cases 5 = essentially all cases as probability 1 = 0 - 5 2 = 5 - 35/40 3 = 35/40 - 65 4 = 60 - 85/90 5 = 85/90 - 100 eg, 80% of patients with kidney infection have side pain, so F=4 import how important the finding is (must it be explained) 1 = finding occurs in normal persons and can be ignored ... 5 = finding must be explained by a diagnosis disease links diseases suggest each other with ESs and Fs also specify the type of link caus - one disease causes the other pced - one disease precedes the other by months pdis - one disease predisposes patient to develop other coin - two diseases occur together without known cause sys - one disease includes the other (eg, systemic dis) eqiv - two distinct forms of same disease eg, aortic stenosis causes congestive heart failure properties tell how costly it is to elicit findings avoid contradicted findings (stress test during MI) avoid redundant findings (ask for HCT after RBC) common sense (pregnant male) PROBLEM SOLVING METHOD (inference engine) generate a score for each diagnosis (higher=better) for pos findings in disease, add weighted Ess like P(D|T) 0 => 1 1 => 4 2 => 10 3 => 20 4 => 40 5 => 80 for pos findings not in disease, add weighted imports (unexplained by disease) 1 => -2 2 => -6 3 => -10 4 => -20 5 => -40 for neg findings that should be in disease, add Fs P(-T|D) = 1 - (T|D) 1 => -1 2 => -4 3 => -7 4 => -15 5 => -30 take diseases with score above a relative threshold among selected diseases, classify as: competitors - patient only has one alternatives - patient may have both must choose among top score and its competitors if top > second+90 then done else ask for other findings mode = pursue, rule out, or discriminate once one diagnosis is selected drop findings explained by diagnosis try to explain leftover findings with second Dx stop when leftover findings have import <= 2 SIMPLIFIED EXAMPLE pneumonia bronchitis URI (ES F) (ES F) (ES F) cough 2 4 3 4 3 4 sputum 3 4 2 3 1 1 fever 0 4 0 1 0 2 chest pain 2 3 1 2 1 1 patient with cough, sputum, fever, and chest pain pneumonia = 40 bronchitis = 34 URI = 28 patient with cough alone pneumonia = 10-37 = -27 bronchitis = 20-12 = 8 URI = 20-6 = 14 comparison to Bayes Bayes uses P(D), P(T|D), and P(~T|~D) QMR uses only ES approx P(D|T) and F approx P(T|D) P(D|T) = prevalence, specificity, sensitivity P(T|D) = sensitivity thus prevalence is folded ESs (cannot alter per locale) study on ESs was done define ES as P(D|T) but scoring algorithm adds ESs => more like P(T|D)/P(T|~D) = likelihood compared several diseases with different P(D) found empirically that it was in between QMR-DT is an attempt to turn QMR KB into probabilities EVALUATIONS looked at 19 very difficult published cases (CPC) 19 patients, 43 diagnoses QMR got 25 / 43 MDs taking care of patient got 28 /43 expert got 35 / 43 consultation service diagnosing 20 difficult real cases ward team: on list 60%, first choice 30% consult MD: on list 80%, first choice 50% QMR: on list 85%, first choice 60% concluded that QMR helpful when run by experienced MD using QMR compared how users enter findings found positive findings well correlated negative findings poorly correlated means use of program may not be reliable