A Mathematical Approach to Medical Diagnosis Application Homer R. to Congenital Heart Disease Warner, M.D., Ph.D., Alan F. Toronto, M.D., L. George Veasey, M.D., and Robert Stephenson, Ph.D., Salt Lake City DIAGNOSIS of disease from clinical data is considered by the medical profession as a art which may be mastered only after years of careful study and extensive personal experience. THE subtle Although rapid advances are being made in the development of new and improved methods for acquiring objective information from a patient concerning his illness, similar progress has not been made in analyzing and improving the logical proc¬ ess by which a diagnosis is deduced from this in¬ formation. The present study was undertaken to find an explicit mathematical expression for this logical process, with the hope that such an expres¬ sion might (1) improve the accuracy of diagnosis in certain fields of medicine, (2) lead to a more scientific approach to the teaching of medical diag¬ nosis, and (3) provide a means, with the help of an electronic computer, for relieving the physi¬ cian of the task of storing and recalling for practical in diagnosis an ever-increasing mass of statis¬ tical data. The derivation of such an equation is herein presented and its usefulness illustrated in its application to the diagnosis of congenital heart disease from clinical data. An equation of conditional probability is derived to express the logical process used by a clinician in making a diagnosis from clinical data. Solutions of this equation take the form of a differential diagnosis. The probability that each disease represents the correct diagnosis in any particular patient may be calculated. Sufficient statistical data regarding the incidence of clinical signs, symptoms, and electrocardiographic findings in patients with congenital heart disease have been assembled to allow application of this approach to differential diagnosis in this field. This approach provides a means by which electronic computing equipment may be used to advantage in clinical medicine. use Theory That the logical process involved in medical as a problem in con¬ diagnosis could be expressed ditional probability ' was suggested by Ledley and Lusted.2 The problem consists of estimating the likelihood or probability of event yi occurring in the presence of another event, x. In this paper the event y, is one disease among a series of diseases Read before the Scientific Sessions of the American Heart Association, St. Louis, Oct. 24, 1960. From the Cardiovascular Laboratory of the Latter-day Saints Hospital and the departments of physiology and electrical engineering of the University of Utah. Dr. Warner is an Established Investigator of the American Heart Association. yk, assumed to be mutually exclusive, yi, y2, and the event x is a set of clinical findings Xi, x2, Xj, which will here be called symptoms even though physical signs and electrocardiographic findings are included. The probabihty of yi is de¬ . . . . . . fined by N Equation where N occur in yi a 1: P ^j (yi' y2' yi • • • yk) is the number of times disease yi would large random sample of patients with diseases yi, y2, Downloaded From: http://jama.jamanetwork.com/ by a New York University User on 05/29/2015 = . . . yk- N, (y%, y2, yk) s . •. (P ) is simply yi the incidence of disease yi in this subpopulation consisting only of people having one of these dis¬ eases. The probability of symptoms x occurring in a patient with disease yi is given by of individual symptoms let us consider the case of two symptoms, xa and xb. It might be argued that for xa to be truly independent of xb, the probability of xa must not be influenced by the presence of xtJ; that is N Equation 2: xy =_1 N P x|y 1 where N yi also patients with disease is the number of xy having symptoms Equation 8: y. * Dividing the numer¬ right-hand term by x. denominator of the N, the size of the population 1 * ator and (yi> y2' • • results in \ V „ P Equation xy —_L P P xiy, 3: as where Px is the . P = x Equation 6: P = expressed In order to = x|yi clarify the P P . .. P xjly! ... yi xi!yi P (x y ,x 112 ,...x.)> P xjlyi ...P VP P P L *k v*k x2,yk xjlyk all k • possible ' xj to ' calculate the ) that each disease yi, • • • . . . Xllyk' X2lyk' . xi,yi x2ly! P P for the ;|yi P Equa¬ tion under consideration and the incidence of each of the patients' symptoms in each of these diseases P (P etc.). These statistical data, required in disease \\ must be 7: may rewrite yi P P y, x|y VP P in Equation we expanded form as V2, yk exists in the presence of symptoms Xi, x2, Xj from statistical information concerning the incidence of each disease P in the popula- kyk XIYk . an Equation 7, of use yi:xi' v k the product of the probabilities of the individual symptoms which make up the set occurring in disease yi. This is occurring . probability (P . which is an expression of Bayes' rule for the prob¬ ability of causes. Now, in fact, any symptom com¬ plex (x) may be represented as a series of inde¬ Xj. Thus, the pendent symptoms \\, x2, conditional probability (P ) of symptom complex . = With this expression it is hVPykx|yk «Ilk and 5 results in P y,ix = , Equation 9: x all k x . xyi P * Combining equations 3, 4, a -^— probability 5: = possible. of symptoms x occur¬ in with of those diseases. If one any patient ring these diseases are considered mutually exclusive, it follows that Equation . X the non-uniform distribution of xb in diseases yi, yk and not due to a direct causal relation¬ ship between xa and xb. In the selection of symp¬ toms to be used in a particular field care must be taken to adhere to this criterion as closely as With = P can be true only if xb is uniformly throughout the population. This means P P 1 and that x„ is xblyi xbly2 xb,yk of no diagnostic value. For this reason Equation 8 must be an inequality. In spite of this, these symp¬ toms for present purposes are truly independent of each other as long as this inequality is due only distributed that Pv tion 6 in P P y,ix 1 |X ab However, this y2, By the same reasoning the probability of disease yi occurring in the presence of symptom complex x Equation 4: = X to yi may be written P . meaning of independence right-hand term of Equation 9, may be compiled and stored in a form (punched cards, punched paper tape, or magnetic tape) that will make them readily available for an electronic digi¬ tal computer to extract the pertinent numbers (de¬ pending upon the symptoms presented by the patient) for carrying out the calculation called for by the equation. Because Equation 9 uses only probabilities in¬ volving those symptoms actually present in the patient under consideration, the absence of a par¬ ticular symptom does not influence the diagnosis. Thus, in order to make use of the fact that the ab¬ sence of a symptom may have a bearing on the probability of a given disease being present, Equa¬ tion 9 is modified to give Downloaded From: http://jama.jamanetwork.com/ by a New York University User on 05/29/2015 Equation Congenital Heart Disease Because the accuracy of a diagnosis of congenital Application 10: p p yil(VY " V (1 X8lyi Xjlyi V (1 "P. ..)... P X8lyk yk Xllyk Xj'yk «11 k yl ~ x iy il where the bar above x8 in the left-hand term indicates that symptom x8 is not present in the patient under consideration. Because P X8lyi heart disease from clinical symptoms may be checked by cardiac catheterization and/or findings surgery, and because relatively objective clinical may be easily obtained, this field was chosen for a pilot study. The list of symptoms and at findings ing symbols, — Symptoms Code' BW {S BW BW BW BW BW BW BW I X4 age 1 mo. to 1 yr. age 1 to 20 yr. = I X« = Ut = BW BW BW BW BW BW BW B t diastolic thrill loudest in left 2nd diastolic X25 = murmur loudest in left 2nd interspace ¡continuóos murmur loudest in left 2nd 2nd interspace murmur loudest in ¡diastolic murmur loudest in 2nd interspace chest systolic murmur heard best over continuous murmur heard best over posterior chest ¡systolic right right BW BW BW BW BW ECG axis less than 0° ¡R wave greater than 1.2 mv in lead Vi X-A7- R' or qR pattern in lead Vi ¡R wave greater than 2.0 mv in lead Vb ¡T wave In lead Va inverted (no digitalis) diastolic murmur loudest at apex t' \X41 early late diastolic murmur loudest at apex holo-systolic murmur loudest in left 4th interspace I X48: mid-systolic murmur loudest in left 4th interspace holo-diastolic murmur loudest in left 4th interspace I X45 early diastolic murmur loudest in left 4th interspace 1X85= M X3«: •|X40: w X4 7= + < X4S= w XiO- BW X50= mid-systolic murmur with thrill loudest in 2nd interspace ¡holo-systolic murmur with thrill loudest in 2nd interspace ¡mid-systolic murmur without thrill loudest in left interspace ¡holo-systolic murmur without thrill loudest in left interspace ¡murmur louder than gr 3/6 left left 2nd 2nd tlndicates presence of mutually exclusive symptoms within brackets as special cases (see text). *B: symptom used on brown check-off sheet; W: symptom used on which must be handled white check-off sheet. Diseases yi =normal septal defect without pulmonary stenosis or pulmonary hypertension* defect with pulmonary stenosis ¡=atrial septal ys y4 ¡=atrial septal defect with pulmonary hypertension* endocardial cushion defect (A-V commune) —complete y5 yo =partial anomalous pulmonary venous connections (without a trial septal defect) y7 =total anomalous pulmonary venous connections (supradiaphragmatic) ys =¡tricuspid atresia without transposition =Ebstein's anomaly of tricuspid valve y9 yio=ventricular septal defect with valvular pulmonary stenosis 11=ventricular septal defect with infundibular stenosis y yi2=pulmonary stenosis, valvular (with or without probe-patent fora¬ men ovale) yis=pulinonary stenosis, infundibular (with or without probe-patent foramen ovale) yii=pulmonary atresia y i.-»—pulmonary artery stenosis (peripheral) yio=pulmonary hypertension,* isolated yi7=aortic-pulmonary window yis=patent ductus arteriosus without pulmonary hypertension* yi9=pulmonary arteriovenous fistula y20=mitral stenosis y2i=primary myocardial disease y22=anomalous origin of left coronary artery y23=aortic valvular stenosis y24=subaortic stenosis y25=coarctation of aorta yso—truncus arteriosus y27—transposed great vessels y28=corrected transposition ya =atrial y20=absent aortic arch yso=ventricular septal defect without pulmonary hypertension* y3i=ventrieular septal defect with pulmonary hypertension* y8s=patertt ductus arteriosus with pulmonary hypertension* y8a=tricuspid atresia with transposition ^Pulmonary hypertension is defined systemic arterial pressure. as pulmonary artery pressure ^ interspace posterior I X2' J X2: accentuated 2nd heart sound in left 2nd interspace 1X2! ^X2ft= diminished 2nd heart sound in left 2nd interspace xso= ¡right ventricular hyperactivity by palpation X3i: ¡forceful apical thrust :pulsatile liver :absent or diminished femoral pulsation I X341 ECG axis more than 110° \X2' BW BW w systolic systolic space ¡systolic murmur without interspace X24 = BW BW w syncope ,X23 = BW W X14: i BW BW W chest — BW W W W W easy Xii: X12: X13: systolic f Lxifl— B BW BW fatigue orthopnea pain repeated respiratory infections Xio: murmur loudest at apex murmur loudest at apex murmur loudest in left 4th interspace X17 < X18= diastolic murmur loudest in left 4th interspace continuous murmur loudest in left 4th interspace murmur with thrill loudest in left 2nd inter¬ B BW BW BW BW cyanosis, intermittent cyanosis, differential squatting dyspnea X15: XlÖ = B B >20 yrs. cyanosis, mild cyanosis, severe (with clubbing) = JX5 concerning the incidence of each Table 2.—List of Diseases Included in Differential Diagnosis S 1 Table 1.—List of Symptoms to Be Evaluated by Physician is tical information the probability of symptom x8 not occurring in a patient with disease yi. Thus, the absence of a symptom is treated as a discrete event when Equa¬ tion 10 is used, and the probability that a symptom is absent can be obtained directly from the proba¬ bility figure for the presence of the symptom. resent study, with their correspond¬ displayed in Tables 1 and 2. Statis¬ diseases used in this represents the probability of symptom x8 occurring complement (1 P ) must rep- in disease yi, its to of these symptoms in each of these diseases is presented in Table 3. The numbers in the first column represent the incidence of each disease (Py) in the subpopulation made up of patients referred to this laboratory and suspected of having congen¬ ital heart disease. The rest of Table 3 is a matrix with symptoms along the horizontal axis and dis¬ eases listed vertically. For instance, the number 0.02 at the intercept of symptom x4 and disease V2 the probability or incidence of represents P , mild cyanosis occurring in a patient with atrial septal defect without pulmonary hypertension. (In this study pulmonary hypertension is arbitrarily defined as pulmonary artery pressure equal to or greater than aortic pressure.) Several things about this symptom-disease matrix require explanation. Listed among the diseases is a category called normal. The incidence of normal Downloaded From: http://jama.jamanetwork.com/ by a New York University User on 05/29/2015 (P ) study in this referred is 0.10 since 10% of the Since the patient can belong in only one of the three age groups, these three "symptoms" cannot be considered independent of one another. Thus if the patient's age is between 1 and 20 years, P patients laboratory for heart catheterization by physiologic studies, which include dye-dilution curves. The figures in the incidence column and symptoms xt, x-2, and x:i (age) may vary from one population to the next, while the other data, which express the probability of each symp¬ are to this normal is used in probability used in this case. Furthermore it is important that care be taken not to include in the list of symptoms any two symptoms which invariably occur together, since this strongly suggests interdependence and a causal relationship between them. For instance, disease, should remain constant from one laboratory to the next. Each of these probabil¬ ities in the matrix was determined by us from (1) a careful review of published data of others, par¬ tom in each ticularly Keith and co-workers,3 (2) 10 but the complement of the for the other two age groups is not Equation review of data Table 3 — Diseases yi . Incidence xi xa x:i xi x-, xo X7 xs x¡> xio xu X12 0.100 01 10 40 60 0(1 01 (Kl 05 02 20 25 10 01 05 00 01 35 03 01 01 01 01 10 50 10 70 30 50 01 02 20 30 50 00 7(1 05 05 01 80 01 40 00 50 00 00 01 01 15 (Il (Kl 10 (Il 00 00 00 22 30 40 80 80 75 75 00 00 (Il 50 01 00 00 00 00 00 80 01 01 01 01 50 90 y2 y3 .081 .005 311 60 60 yt .001 10 20 y5 y« yys y» .027 20 10 50 yio yn yi2 yia yi4 yir, yi« yi7 yis yia y2o y21 y22 y2a y24 y25 y2o y27 y2s y2o yao .006 .001 .018 .001 .054 .063 .045 .013 .014 .001 .013 .001 .072 .002 .(108 .013 .001 .(136 .009 .054 .005 .063 .001 .001 .252 20 50 1(1 40 40 20 20 40 70 48 45 55 55 10 02 45 05 10 15 10 05 01 01 65 10 05 30 22 25 30 44 25 30 1(1 10 01 01 00 01 00 00 00 90 09 05 10 10 01 10 90 05 45 50 01 01 10 45 «0 45 10 40 40 50 00 01 (Il 00 01 01 (Il 01 01 01 45 01 (Il 01 01 30 01 05 01 45 01 01 00 (Il 10 10 (Il 01 (Il 01 01 00 01 01 01 01 00 01 20 10 (Il 30 20 30 01 00 6(1 60 05 01 01 01 01 01 (Il 50 01 20 30 20 20 2(1 70 7(1 10 10 10 50 90 30 60 yai ys2 .081 .008 15 30 30 yai .009 40 70 70 30 50 29 29 80 80 70 40 10 30 39 70 00 30 01 15 01 01 (Il (Il 60 10 40 55 30 01 30 01 05 50 01 01 05 10 01 (Il 00 (10 10 00 80 00 10 00 05 10 50 00 obtained from 1,035 patients referred to this labo¬ ratory for diagnostic catheterization, and (3) esti¬ mates based upon the pathologic physiology of the defect in the case of rare defects in which adequate statistics were not available. Notice that each patient is classified according to age into one of three categories, 1 month to 1 year, 1 year to 20 years, and over 20 years of age. The patient's age is treated as a symptom. For in¬ stance, the number 0.70 occurring at the intercept of x-2 and ylf! indicates that this symptom (age 1 to 20 years) will occur in 70 out of 100 patients with infundibular pulmonary stenosis coming to this laboratory. In this way, then, the fact is recognized that the age of the patient does influence the prob¬ ability that he has any given diagnosis. 01 01 leía 05 X14 xi-, xie X17 xis xid xao X21 xzz 03 05 01 70 80 10 10 (Il 02 05 57 05 01 02 02 30 05 15 90 10 10 40 01 40 20 05 05 10 02 20 60 05 05 05 80 10 30 75 20 15 90 90 (¡5 05 05 05 20 15 15 10 10 05 05 20 05 01 01 05 99 05 01 95 01 01 01 10 01 02 02 02 15 02 02 05 25 02 02 02 02 07 02 02 02 (12 02 10 05 05 05 00 02 02 05 05 02 20 40 10 10 20 20 (15 10 05 40 20 01 30 80 20 80 90 70 01 70 10 20 1(1 50 40 50 50 30 30 20 05 (15 01 05 (Il 15 05 20 15 60 30 30 30 30 70 01 10 20 (Il 30 01 05 01 01 10 50 20 (¡0 30 70 20 80 30 90 05 05 20 10 20 20 05 01 2(1 15 20 20 05 15* 20 clubbing arate 01 01 01 01 01 01 2(1 01 10 01 01 15 30 05 01 05 10 22 20 25 10 01 10 10 (Il 01 15 20 10 30 of the 20 02 15 05 10 65 95 20 20 10 10 25 02 02 02 05 02 02 05 01 05 20 05 65 65 05 01 02 25 (Il 01 05 05 01 05 10 05 05 05 01 00 50 02 80 05 20 15 05 10 10 10 05 01 02 10 10 15 02 20 02 30 20 20 02 02 01 35 20 35 05 10 20 02 10 02 01 10 10 10 (12 01 02 02 02 02 02 02 30 05 05 02 05 02 05 02 01 05 05 01 01 01 (Il 02 10 10 03 10 1(1 01 02 02 05 05 02 02 05 70 02 02 02 05 30 02 50 02 1)2 05 02 02 30 10 05 20 02 95 10 10 10 30 02 02 05 05 05 20 05 02 02 01 20 05 02 02 10 13 10 10 02 02 02 02 02 20 02 05 02 05 02 02 10 20 20 20 25 20 01 05 25 70 70 01 4(1 10 01 (Il 40 02 10 10 05 02 02 02 90 01 10 20 02 02 05 02 02 20 02 35 (Il 01 (Il 02 05 04 01 05 (Il 1(1 (Il 10 05 15 01 02 00 05 02 05 fingers symptom since it 01 02 10 was 7(1 50 50 10 10 10 70 05 not occurs 02 02 02 included in the 05 as a same 05 25 10 10 sep¬ patients heart disease who exhibit severe it is included as a part of the definition of severe cyanosis. Inclusion of redun¬ dant (interdependent) symptoms would result in an unreal increase in the probability of those diseases having a high incidence of these symptoms when these symptoms are present and a falsely low prob¬ ability when these symptoms are absent. There are other symptoms in the list which are mutually exclusive. For instance, the existence of x,-, excludes by definition x4, x6, and x7. Thus, it would be an error to consider the absence of X4, x«, and x7 as additional pieces of information once x-, is known to be present. On the other hand, the abwith congenital cyanosis. Instead, Downloaded From: http://jama.jamanetwork.com/ by a New York University User on 05/29/2015 of sence through x4 x7 in particular a case prepared from a check-off list of symptoms on which the physician, from his examination of the patient, has marked those symptoms presented by the patient. (X-ray data are not presented in this paper but are being evaluated for inclusion in the symptom list at the present time.) From this information, the computer then cal¬ culates, with use of Equation 9 or 10, the probabil¬ ity of each of the 33 congenital heart diseases being (no important fact and must be recog¬ cyanosis) nized by using in Equation 10 the complement of the sum of the probabilities of each of these symp¬ toms occurring in the distase in question, which is -P -P 1-Px -Px x x is an . 4 |y1 5 |y 6 1 |y1 7 |y1 Groups of mutually exclusive symptoms are indi¬ cated by brackets in Table 1, and a complete list of mutually exclusive symptoms, together with instructions about what data should be used in solving Equation 10 in any particular case, is given in Table 4. Symptom-Disease present in the patient under consideration. The dis¬ with probability greater than 1% are printed out at the end of the calculation together with their respective probabilities. Two symptom lists are eases Matrix Symptoms X23 X24 X2S X28 Xn X28 X39 X30 X31 X:i; X33 X;U X35 X36 X37 X.ts X39 X40 X41 Xli X43 X44 X4S 05 02 03 01 01 01 01 01 00 01 05 01 15 02 02 02 02 02 04 05 05 85 05 20 70 02 02 02 02 30 05 02 01 03 20 05 95 02 10 85 02 10 02 10 15 02 15 02 02 02 02 02 02 20 02 20 02 20 01 01 01 02 02 02 01 01 01 10 02 90 02 02 01 05 01 10 10 01 35 60 60 10 20 20 01 10 02 02 85 60 10 10 02 02 02 02 02 02 02 02 02 02 45 20 20 60 60 90 02 00 02 02 01 01 01 85 85 10 02 02 10 90 02 02 01 01 01 30 02 40 02 05 02 02 02 01 02 01 01 10 02 02 01 01 01 30 02 02 02 70 03 05 01 01 01 05 05 20 20 10 10 01 01 50 05 50 20 01 01 20 20 20 10 30 20 20 10 50 02 01 15 15 01 01 10 05 02 02 02 25 30 50 01 02 05 25 10 02 02 90 02 75 01 80 01 25 01 02 60 01 01 01 01 05 01 70 40 15 05 02 01 15 15 02 01 20 01 20 05 02 01 05 70 02 70 01 05 01 02 70 85 85 02 01 01 01 85 70 01 01 01 01 01 01 01 15 00 01 01 01 05 05 01 40 50 40 03 01 01 01 01 02 15 60 30 10 01 01 01 01 01 01 20 02 20 10 01 01 01 01 01 15 10 01 01 01 01 01 20 20 10 40 20 20 90 30 90 10 01 90 05 05 05 05 10 02 02 02 02 02 02 02 02 02 20 02 02 20 02 02 85 02 02 05 02 02 01 02 10 01 95 01 02 02 01 05 95 15 05 10 02 05 05 05 01 01 01 01 05 02 02 01 01 01 01 10 80 05 05 02 02 02 02 01 02' 05 01 02 02 02 02 02 05 02 01 01 01 02 01 01 02 02 70 01 01 02 01 85 10 10 10 10 01 10 95 30 Use of 02 10 80 01 01 01 10 01 01 20 20 01 01 01 01 01 01 01 40 40 01 05 01 01 30 10 02 02 05 05 02 01 10 01 02 01 95 02 01 95 02 01 01 01 01 01 01 01 95 95 95 10 95 02 01 02 15 02 02 60 05 10 02 10 10 02 02 50 05 02 05 05 50 05 05 06 02 02 10 05 02 02 40 05 05 02 40 20 02 02 90 90 10 02 20 02 02 20 02 01 01 10 10 10 01 05 15 15 05 10 02 02 02 70 02 02 02 01 02 01 02 02 01 01 01 40 30 05 01 01 01 01 01 20 02 50 05 40 10 01 10 10 30 20 10 02 02 10 40 05 10 05 01 01 30 02 30 05 02 10 30 01 05 20 01 05 05 30 02 90 02 02 10 02 01 01 05 05 99 01 02 01 10 05 05 30 40 02 02 02 02 10 10 85 85 05 02 02 02 20 40 30 20 10 10 70 05 10 05 05 90 80 05 01 30 01 70 01 70 01 02 Computer Because of the large number of calculations re¬ quired to make each diagnosis in the example (congenital heart disease) used in this paper, it is necessary to use a digital computer if Equation 10 is to be solved in a practical way. This equation can be solved by almost any general-purpose elec¬ tronic digital computer which has the capability of "floating decimal point" operation. The incidence of each symptom in each disease shown in the matrix is transferred to punch cards. These disease cards, together with cards which contain the pro¬ gram telling the computer what operations to per¬ form, are transferred into the computer memory by a card-reading machine. Another punched card is 75 75 02 10 05 10 05 05 15 15 02 70 15 15 40 20 20 10 10 15 10 04 05 05 10 90 05 10 10 05 05 05 02 02 02 02 20 01 02 10 01 02 02 02 02 02 01 02 02 02 02 05 40 30 30 30 92 30 10 30 15 45 05 05 10 10 10 X46 Xli 00 05 60 05 00 01 01 20 01 20 02 01 20 15 15 02 02 60 05 05 68 68 01 02 01 05 00 15 05 02 05 02 02 02 02 02 05 25 02 02 02 01 02 02 10 30 02 25 02 05 60 01 02 02 01 01 01 01 02 02 02 02 05 02 01 Xlt» Xr.o 80 90 05 01 10 60 38 01 01 20 70 30 20 05 02 02 20 05 25 05 90 25 25 25 02 25 05 10 20 10 05 01 01 02 02 90 40 20 60 80 70 50 50 80 02 75 10 85 10 30 10 10 70 05 05 01 10 10 90 90 02 02 01 01 20 20 10 10 20 40 30 30 30 10 02 02 02 02 05 10 02 02 02 05 05 02 01 01 02 02 02 02 10 03 05 10 10 03 05 10 01 01 10 01 02 01 02 02 02 05 02 05 05 10 10 01 20 30 30 10 30 80 05 10 01 02 40 80 20 60 20 05 02 01 05 10 02 XlH 01 05 05 10 10 10 01 01 05 10 10 30 30 30 30 10 05 20 30 65 40 50 60 20 85 50 20 50 checked off by the clinician after his examination of each patient. On one list (brown sheet) murmurs are described only as to timing and location, while on the other list (white sheet) the time-course of intensity of the murmurs is included (see Table 1). Equation 10 is solved with each of these sets of symptoms and the resulting differential diagnoses are compared. Although the calculation based on the white sheet often gave a higher probability to the correct diagnosis, this was not consistently the case, particularly in instances in which classifica¬ tion of the time-course of murmur intensity was difficult even with the help of a phonocardiogram. The point to be made here is that in applying this approach to diagnosis a compromise must be reached between two alternatives: (1) the desir- Downloaded From: http://jama.jamanetwork.com/ by a New York University User on 05/29/2015 Table 4.—Terms to Be Used in Equation 10 in Cases of Mutually Exclusive Symptoms Check Sheet Symptoms B/W XI—X3 B/W X4—XT Symptom Present X2 X3 PX3 XI Pxi X5 PX5 Pxe PX7 XG X7 none B/W X20—X27 X26 X27 neither B/W XS8—X21) X34—XSB X30—X37 PX29 X34 X30 X37 X17—X19 none B X20—X23 X22 X23 PX23 X20 X21 X20, X22 X21, X22 none W Xl», X42—X45 X19 X42 X43 X44 X45 X42, Xl4 X43, Xl4 X42, X45 X43, X45 none W X40—X41 W PXl9 PX42 PX42 PX44 PX43 PX44 PX42 PX45 PX43 PX45 (1—PxisHl—PX42—Px4s) (1—PX44—PX4B) X22 Xi« X48 X49 Xi«, X22 X47, X22 Xl8, X22 X49, X22 X23 none (1—PX40—PX41) (1—PXlO—PX47—PX46—PX4») (1-PX22) (1—PX22) (1—PX22) (1 —PX22) PX22 PXlO PX47 PX48 PX49 PX40 PX22 PX47 PX22 PX4S PX22 PX49 PX22 PX23 (1—PX22XI—PX23) (1—PXlO —PXlT—PX48—PX40) of using as much information (2) the limitations in accuracy ability as possible, with which the more detailed information can be observed in the patient and the necessary statistical data can be obtained. Example.—The case shown in Table 5 illustrates the effect of using both positive and negative in¬ formation in making a diagnosis. The list of symp¬ toms indicates that the patient was over 20 years of age and complained of easy fatigue and orthopnea; his pulmonary second sound was diminished; his electrocardiogram exhibited an axis greater than Disease Prob- No. ability yn yio yie 0.33 .28 x29 . yia .11 .14 X34 . yi3 .04 X30 . . Diagnosis With Equation 10 Disease No. y12 yia yio yn yio Prob- ability 0.62 Diagnosis With Equation 10 and Without Xn Disease No. .21 yi2 yi3 .07 yio Prob- ability 0.73 .24 .02 .04 .03 tion 10 is used, the probability of isolated pulmo¬ nary stenosis became 0.83 while the probability of tetralogy of Fallot was only 0.11. This patient was to have y]2 both by physiologic studies and at surgery. Also illustrated in this table is the way in which this approach may be used to evaluate the contri¬ bution made toward a diagnosis by any given symptom. Here the calculation has been carried out with and without the symptom of orthopnea ( xn ). Had this patient not complained of orthopnea the probability of isolated pulmonary stenosis (yl2 and yi3) would have been 0.97 as compared to 0.83 when orthopnea was considered present. This, of course, results from the fact that orthopnea rarely occurs in patients with yi2 or y13. Since the presence or absence of just one symptom may make a real difference in the differential diagnosis, as in this instance, it is apparent that each symptom on the list must be accurately evaluated in every case if the correct probabilities are to be calculated. For this reason, only the most objective symptoms should be included in the definition of the original list for any study, even if this must be done at some sacrifice of detail. In the case of the present study we are under the impression from our experience to date that symptoms 10 through 14 detract from the ac¬ curacy of diagnosis as often as they contribute, because of the difficulty involved in assessing their actual presence or absence in many cases plus the inaccuracy of the available statistical data regard¬ ing the incidence of these symptoms in each of these diseases. These 5 symptoms might well be eliminated from the list. later found (1—PX44—PX46) PX43 (1—PX44—PX4B) PX44 (1—PX42—PX13) P.X46 (1—PX42—PX13) PXll X47 and (1—Px2o— PXii)(l—PX22)(1—PX23) PX40 X22—X23 Xn PX21 PX22 X40 X4U—X49 xio. PX2Ü PX22 Xli neither Symptoms x 3. (1-Pxi7)(l—Px18)(l—PX19) PX17 PX18 PX20 (1—PX22) PX21 (1—PX22) PX22 (1—PX20—PX2l) X17, X18 greater than 1.2 mv in lead Vi; Diagnosis With Equation 9 PX36 PX37 PX19 wave Table 5.—Test Case Illustrating Effect of Including Negative Information (1—PX34—Pxsb) X19 R by phonocardiogram he had a midsystolic mur¬ without a thrill, which was of equal intensity in the pulmonary (2nd left interspace) and the pre¬ cordial (4th left interspace) area. Calculation of the probabilities for each disease with use only of the positive information (Equation 9) resulted in a higher probability for tetralogy of Fallot (yl0 and yn) than for isolated pulmonary stenosis (y12 and y 13). However, when both positive and negative in¬ formation were taken into account, as when Equa- PX34 PX3B Xl8 Xl7 an mur, (1—Px2s—PX29) (1—Pxs«—PX37) PX17 (1—PX18) PX18 (1—PX17) neither B (1—Px2e—PX27) X2B neither B/W (1—PX4—Pxs—Pxo—Px7) PX2S X35 Equation 10 PX28 PX27 X28 neither B/W Term Used in Pxi Px2 xi 110° and and Downloaded From: http://jama.jamanetwork.com/ by a New York University User on 05/29/2015 Experience to Date differential diagnosis obtained Evaluation of with Because the this approach represents an estimation of probabili¬ ties in which the statistical data of Table 3 are used, it is impossible from a limited number of cases to evaluate its accuracy. However, it is ap¬ parent from our experience to date with 36 cases that the most probable diagnosis estimated with Equation 10 agrees with the actual diagnosis made by physiologic studies and observation at surgery at least as often as does the most probable diagno¬ sis estimated by 3 experienced cardiologists from the same clinical information. Furthermore, the differential diagnosis resulting from solution of the frequently more complete and, in retro¬ spect, often appears more logical to the clinicians than the differential diagnosis listed by each of them prior to seeing the equation's prediction. It must be emphasized that Equation 10 was derived directly from the definition of conditional probability. Thus, any evaluation of the accuracy of the predictions made by this approach should be considered as testing the adequacy of the matrix equation of the diagnosis and what symptoms are unimpor¬ tant? A solution of Equation 10 for any given set of symptoms provides an objective, reproducible stand¬ ard against which students may check the accu¬ racy of their own deductions from these same symp¬ toms. The effect of modifying the symptom set in any desired fashion on the differential may be readily observed. approach to the teaching of diagnosis of congenital heart disease is in current use at this hospital and has met with enthusiastic acceptance by medical students. This Appendix is of statistical data and not of the equation. Given the correct original data matrix and accurate ob¬ servations of the patient, the calculated probabili¬ ties will be correct. Final refinement of the present data matrix must await the accumulation of suffi¬ cient data for calculation of new probabilities (P ). Since the presence or absence of each xilyk and follow-up information almost invariably yields the diagnosis with certainty, the data for satisfactory recalcula¬ tion of symptom incidence are routinely accumu¬ lating. The computer will be used to recalculate its own data matrix when the amount of data is sufficient. symptom is determined in each Aids to case, Teaching That an explicit expression of the logic used in medical diagnosis has potential usefulness as a tool for teaching diagnosis to medical students and phy¬ sicians seems apparent. The approach here pre¬ sented provides a framework within which any diagnostic problem may be formulated and criti¬ cally analyzed. Often the very act of attempting to formulate the problem in terms required for application of Equation 10 results in new insight by providing answers to such questions as: 1. What is the exact definition of each symptom and each disease? 2. Are certain symptoms interdependent and oth¬ ers mutually exclusive? 3. What symptoms are important determinants diagnosis use of Equation 10, consider the of a simple population consisting of just 2 dis¬ eases (yi and y2) and 3 independent symptoms (xi, x2, and X3). The relative incidence of these 2 dis¬ eases and the probability of each symptom in each disease are shown in the matrix below. To illustrate the case Incidence yi . y2 . 0.23 0.77 Xi xa x3 0.1 0.8 0.7 0.2 0.6 0.5 If the patient to be diagnosed presents with symptoms Xi and x3, Equation 10 would be solved with use of the following numbers to make the diagnosis: p y t I(x , x , 2* 1 1. - x = 3 _0.23(0.1)(l-0.7)(0.6)_ =0.016 0.23(0.1X1-0.7K0.6)+0.77(0.8X1-0.2K0.5) and y2,(VVX3) = _0.77(0.8X1-0.2X0-5)__ft q84 0.23(0.1X1-0.7X0.6)+0.77(0.8X1-0.2X0.5) 325 Eighth Ave., Salt Lake City 3 (Dr. Warner). The computer program may be obtained from Corporation, 6071 2nd Ave., Detroit 32. Burroughs This work was supported by the Burroughs Corporation and by a grant from the National Institutes of Health. References Feller, W., An Introduction to Probability Theory and Application, New York: John Wiley & Sons, Inc., 1960. 2. Ledley, R. S., and Lusted, L. B., Use of Electronic Computers to Aid in Medical Diagnosis, Proc Inst Radio Engineers 47:1970-1977 (Nov.) 1959. 3. Keith, J. D., and others: Heart Disease in Infancy and Childhood, New York: The Macmillan Co., 1958. 1. Its Downloaded From: http://jama.jamanetwork.com/ by a New York University User on 05/29/2015